yilsem/aiturk-hermes-ide
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases 6 Wiki Activity

Import AITURK IDE 1.0.0-beta.1 from Hermes 63279301; preserve MIT license

Browse Source
This commit is contained in:
yilsem
2026-09-05 13:26:46 +03:00
commit 03634b1ca3
11340 changed files with 3442369 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
website/docs/user-guide/skills/optional/mlops/mlops-accelerate.md
+366
View File
@@ -0,0 +1,366 @@
---
title: "Accelerate — Run PyTorch training across GPUs with minimal changes"
sidebar_label: "Accelerate"
description: "Run PyTorch training across GPUs with minimal changes"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Accelerate
Run PyTorch training across GPUs with minimal changes.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/accelerate` |
| Path | `optional-skills/mlops\accelerate` |
| Version | `1.0.1` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `accelerate`, `torch`, `transformers` |
| Platforms | linux, macos, windows |
| Tags | `Distributed Training`, `HuggingFace`, `Accelerate`, `DeepSpeed`, `FSDP`, `Mixed Precision`, `PyTorch`, `DDP`, `Unified API`, `Simple` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# HuggingFace Accelerate - Unified Distributed Training
## Quick start
Accelerate simplifies distributed training to 4 lines of code.
**Installation**:
```bash
pip install accelerate
```
**Convert PyTorch script** (4 lines):
```python
import torch
+ from accelerate import Accelerator
+ accelerator = Accelerator()
model = torch.nn.Transformer()
optimizer = torch.optim.Adam(model.parameters())
dataloader = torch.utils.data.DataLoader(dataset)
+ model, optimizer, dataloader = accelerator.prepare(model, optimizer, dataloader)
for batch in dataloader:
optimizer.zero_grad()
loss = model(batch)
- loss.backward()
+ accelerator.backward(loss)
optimizer.step()
```
**Run** (single command):
```bash
accelerate launch train.py
```
## Common workflows
### Workflow 1: From single GPU to multi-GPU
**Original script**:
```python
# train.py
import torch
model = torch.nn.Linear(10, 2).to('cuda')
optimizer = torch.optim.Adam(model.parameters())
dataloader = torch.utils.data.DataLoader(dataset, batch_size=32)
for epoch in range(10):
for batch in dataloader:
batch = batch.to('cuda')
optimizer.zero_grad()
loss = model(batch).mean()
loss.backward()
optimizer.step()
```
**With Accelerate** (4 lines added):
```python
# train.py
import torch
from accelerate import Accelerator # +1
accelerator = Accelerator() # +2
model = torch.nn.Linear(10, 2)
optimizer = torch.optim.Adam(model.parameters())
dataloader = torch.utils.data.DataLoader(dataset, batch_size=32)
model, optimizer, dataloader = accelerator.prepare(model, optimizer, dataloader) # +3
for epoch in range(10):
for batch in dataloader:
# No .to('cuda') needed - automatic!
optimizer.zero_grad()
loss = model(batch).mean()
accelerator.backward(loss) # +4
optimizer.step()
```
**Configure** (interactive):
```bash
accelerate config
```
**Questions**:
- Which machine? (single/multi GPU/TPU/CPU)
- How many machines? (1)
- Mixed precision? (no/fp16/bf16/fp8)
- DeepSpeed? (no/yes)
**Launch** (works on any setup):
```bash
# Single GPU
accelerate launch train.py
# Multi-GPU (8 GPUs)
accelerate launch --multi_gpu --num_processes 8 train.py
# Multi-node
accelerate launch --multi_gpu --num_processes 16 \
--num_machines 2 --machine_rank 0 \
--main_process_ip $MASTER_ADDR \
train.py
```
### Workflow 2: Mixed precision training
**Enable FP16/BF16**:
```python
from accelerate import Accelerator
# FP16 (with gradient scaling)
accelerator = Accelerator(mixed_precision='fp16')
# BF16 (no scaling, more stable)
accelerator = Accelerator(mixed_precision='bf16')
# FP8 (H100+)
accelerator = Accelerator(mixed_precision='fp8')
model, optimizer, dataloader = accelerator.prepare(model, optimizer, dataloader)
# Everything else is automatic!
for batch in dataloader:
with accelerator.autocast(): # Optional, done automatically
loss = model(batch)
accelerator.backward(loss)
```
### Workflow 3: DeepSpeed ZeRO integration
**Enable DeepSpeed ZeRO-2** (pass a `DeepSpeedPlugin`, not a raw dict):
```python
from accelerate import Accelerator, DeepSpeedPlugin
deepspeed_plugin = DeepSpeedPlugin(
zero_stage=2, # ZeRO-2
offload_optimizer_device="none", # or "cpu" to offload
gradient_accumulation_steps=4,
)
accelerator = Accelerator(
mixed_precision='bf16',
deepspeed_plugin=deepspeed_plugin, # DeepSpeedPlugin instance (or dict[str, DeepSpeedPlugin])
)
# Same code as before!
model, optimizer, dataloader = accelerator.prepare(model, optimizer, dataloader)
```
**Or point at a full DeepSpeed JSON config via the plugin**:
```python
from accelerate import Accelerator, DeepSpeedPlugin
# hf_ds_config accepts a path to a DeepSpeed config JSON (or a dict)
deepspeed_plugin = DeepSpeedPlugin(hf_ds_config="ds_config.json")
accelerator = Accelerator(mixed_precision='bf16', deepspeed_plugin=deepspeed_plugin)
```
**ds_config.json** (a raw DeepSpeed config — passed via the plugin, NOT via `--config_file`):
```json
{
"fp16": {"enabled": false},
"bf16": {"enabled": true},
"zero_optimization": {
"stage": 2,
"offload_optimizer": {"device": "cpu"},
"allgather_bucket_size": 5e8,
"reduce_bucket_size": 5e8
}
}
```
**Or via interactive config**:
```bash
accelerate config
# Select: DeepSpeed → ZeRO-2
# This writes an accelerate YAML config (default: ~/.cache/huggingface/accelerate/default_config.yaml)
```
**Launch** (`--config_file` expects an accelerate YAML, not a raw DeepSpeed JSON):
```bash
# Uses the default accelerate config written by `accelerate config`
accelerate launch train.py
# Or point at a specific accelerate YAML
accelerate launch --config_file accelerate_deepspeed.yaml train.py
```
### Workflow 4: FSDP (Fully Sharded Data Parallel)
**Enable FSDP**:
```python
from accelerate import Accelerator, FullyShardedDataParallelPlugin
fsdp_plugin = FullyShardedDataParallelPlugin(
sharding_strategy="FULL_SHARD", # ZeRO-3 equivalent
auto_wrap_policy="transformer_based_wrap", # valid: transformer_based_wrap | size_based_wrap | no_wrap
cpu_offload=False
)
accelerator = Accelerator(
mixed_precision='bf16',
fsdp_plugin=fsdp_plugin
)
model, optimizer, dataloader = accelerator.prepare(model, optimizer, dataloader)
```
**Or via config**:
```bash
accelerate config
# Select: FSDP → Full Shard → No CPU Offload
```
### Workflow 5: Gradient accumulation
**Accumulate gradients**:
```python
from accelerate import Accelerator
accelerator = Accelerator(gradient_accumulation_steps=4)
model, optimizer, dataloader = accelerator.prepare(model, optimizer, dataloader)
for batch in dataloader:
with accelerator.accumulate(model): # Handles accumulation
optimizer.zero_grad()
loss = model(batch)
accelerator.backward(loss)
optimizer.step()
```
**Effective batch size**: `batch_size * num_gpus * gradient_accumulation_steps`
## When to use vs alternatives
**Use Accelerate when**:
- Want simplest distributed training
- Need single script for any hardware
- Use HuggingFace ecosystem
- Want flexibility (DDP/DeepSpeed/FSDP/Megatron)
- Need quick prototyping
**Key advantages**:
- **4 lines**: Minimal code changes
- **Unified API**: Same code for DDP, DeepSpeed, FSDP, Megatron
- **Automatic**: Device placement, mixed precision, sharding
- **Interactive config**: No manual launcher setup
- **Single launch**: Works everywhere
**Use alternatives instead**:
- **PyTorch Lightning**: Need callbacks, high-level abstractions
- **Ray Train**: Multi-node orchestration, hyperparameter tuning
- **DeepSpeed**: Direct API control, advanced features
- **Raw DDP**: Maximum control, minimal abstraction
## Common issues
**Issue: Wrong device placement**
Don't manually move to device:
```python
# WRONG
batch = batch.to('cuda')
# CORRECT
# Accelerate handles it automatically after prepare()
```
**Issue: Gradient accumulation not working**
Use context manager:
```python
# CORRECT
with accelerator.accumulate(model):
optimizer.zero_grad()
accelerator.backward(loss)
optimizer.step()
```
**Issue: Checkpointing in distributed**
Use accelerator methods:
```python
# Save only on main process
if accelerator.is_main_process:
accelerator.save_state('checkpoint/')
# Load on all processes
accelerator.load_state('checkpoint/')
```
**Issue: Different results with FSDP**
Ensure same random seed:
```python
from accelerate.utils import set_seed
set_seed(42)
```
## Advanced topics
**Megatron integration**: See [references/megatron-integration.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\accelerate/references/megatron-integration.md) for tensor parallelism, pipeline parallelism, and sequence parallelism setup.
**Custom plugins**: See [references/custom-plugins.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\accelerate/references/custom-plugins.md) for creating custom distributed plugins and advanced configuration.
**Performance tuning**: See [references/performance.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\accelerate/references/performance.md) for profiling, memory optimization, and best practices.
## Hardware requirements
- **CPU**: Works (slow)
- **Single GPU**: Works
- **Multi-GPU**: DDP (default), DeepSpeed, or FSDP
- **Multi-node**: DDP, DeepSpeed, FSDP, Megatron
- **TPU**: Supported
- **Apple MPS**: Supported
**Launcher requirements**:
- **DDP**: `torch.distributed.run` (built-in)
- **DeepSpeed**: `deepspeed` (pip install deepspeed)
- **FSDP**: PyTorch 1.12+ (built-in)
- **Megatron**: Custom setup
## Resources
- Docs: https://huggingface.co/docs/accelerate
- GitHub: https://github.com/huggingface/accelerate
- Version: 1.11.0+
- Tutorial: "Accelerate your scripts"
- Examples: https://github.com/huggingface/accelerate/tree/main/examples
- Used by: HuggingFace Transformers, TRL, PEFT, all HF libraries
website/docs/user-guide/skills/optional/mlops/mlops-chroma.md
+425
View File
@@ -0,0 +1,425 @@
---
title: "Chroma — Embedding database for RAG and semantic search"
sidebar_label: "Chroma"
description: "Embedding database for RAG and semantic search"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Chroma
Embedding database for RAG and semantic search.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/chroma` |
| Path | `optional-skills/mlops\chroma` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `chromadb`, `sentence-transformers` |
| Platforms | linux, macos, windows |
| Tags | `RAG`, `Chroma`, `Vector Database`, `Embeddings`, `Semantic Search`, `Open Source`, `Self-Hosted`, `Document Retrieval`, `Metadata Filtering` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Chroma - Open-Source Embedding Database
The AI-native database for building LLM applications with memory.
## When to use Chroma
**Use Chroma when:**
- Building RAG (retrieval-augmented generation) applications
- Need local/self-hosted vector database
- Want open-source solution (Apache 2.0)
- Prototyping in notebooks
- Semantic search over documents
- Storing embeddings with metadata
**Metrics**:
- **24,300+ GitHub stars**
- **1,900+ forks**
- **v1.3.3** (stable, weekly releases)
- **Apache 2.0 license**
**Use alternatives instead**:
- **Pinecone**: Managed cloud, auto-scaling
- **FAISS**: Pure similarity search, no metadata
- **Weaviate**: Production ML-native database
- **Qdrant**: High performance, Rust-based
## Quick start
### Installation
```bash
# Python
pip install chromadb
# JavaScript/TypeScript
npm install chromadb @chroma-core/default-embed
```
### Basic usage (Python)
```python
import chromadb
# Create client
client = chromadb.Client()
# Create collection
collection = client.create_collection(name="my_collection")
# Add documents
collection.add(
documents=["This is document 1", "This is document 2"],
metadatas=[{"source": "doc1"}, {"source": "doc2"}],
ids=["id1", "id2"]
)
# Query
results = collection.query(
query_texts=["document about topic"],
n_results=2
)
print(results)
```
## Core operations
### 1. Create collection
```python
# Simple collection
collection = client.create_collection("my_docs")
# With custom embedding function
from chromadb.utils import embedding_functions
openai_ef = embedding_functions.OpenAIEmbeddingFunction(
api_key="your-key",
model_name="text-embedding-3-small"
)
collection = client.create_collection(
name="my_docs",
embedding_function=openai_ef
)
# Get existing collection
collection = client.get_collection("my_docs")
# Delete collection
client.delete_collection("my_docs")
```
### 2. Add documents
```python
# Add with auto-generated IDs
collection.add(
documents=["Doc 1", "Doc 2", "Doc 3"],
metadatas=[
{"source": "web", "category": "tutorial"},
{"source": "pdf", "page": 5},
{"source": "api", "timestamp": "2025-01-01"}
],
ids=["id1", "id2", "id3"]
)
# Add with custom embeddings
collection.add(
embeddings=[[0.1, 0.2, ...], [0.3, 0.4, ...]],
documents=["Doc 1", "Doc 2"],
ids=["id1", "id2"]
)
```
### 3. Query (similarity search)
```python
# Basic query
results = collection.query(
query_texts=["machine learning tutorial"],
n_results=5
)
# Query with filters
results = collection.query(
query_texts=["Python programming"],
n_results=3,
where={"source": "web"}
)
# Query with metadata filters
results = collection.query(
query_texts=["advanced topics"],
where={
"$and": [
{"category": "tutorial"},
{"difficulty": {"$gte": 3}}
]
}
)
# Access results
print(results["documents"]) # List of matching documents
print(results["metadatas"]) # Metadata for each doc
print(results["distances"]) # Similarity scores
print(results["ids"]) # Document IDs
```
### 4. Get documents
```python
# Get by IDs
docs = collection.get(
ids=["id1", "id2"]
)
# Get with filters
docs = collection.get(
where={"category": "tutorial"},
limit=10
)
# Get all documents
docs = collection.get()
```
### 5. Update documents
```python
# Update document content
collection.update(
ids=["id1"],
documents=["Updated content"],
metadatas=[{"source": "updated"}]
)
```
### 6. Delete documents
```python
# Delete by IDs
collection.delete(ids=["id1", "id2"])
# Delete with filter
collection.delete(
where={"source": "outdated"}
)
```
## Persistent storage
```python
# Persist to disk
client = chromadb.PersistentClient(path="./chroma_db")
collection = client.create_collection("my_docs")
collection.add(documents=["Doc 1"], ids=["id1"])
# Data persisted automatically
# Reload later with same path
client = chromadb.PersistentClient(path="./chroma_db")
collection = client.get_collection("my_docs")
```
## Embedding functions
### Default (Sentence Transformers)
```python
# Uses sentence-transformers by default
collection = client.create_collection("my_docs")
# Default model: all-MiniLM-L6-v2
```
### OpenAI
```python
from chromadb.utils import embedding_functions
openai_ef = embedding_functions.OpenAIEmbeddingFunction(
api_key="your-key",
model_name="text-embedding-3-small"
)
collection = client.create_collection(
name="openai_docs",
embedding_function=openai_ef
)
```
### HuggingFace
```python
huggingface_ef = embedding_functions.HuggingFaceEmbeddingFunction(
api_key="your-key",
model_name="sentence-transformers/all-mpnet-base-v2"
)
collection = client.create_collection(
name="hf_docs",
embedding_function=huggingface_ef
)
```
### Custom embedding function
```python
from chromadb import Documents, EmbeddingFunction, Embeddings
class MyEmbeddingFunction(EmbeddingFunction):
def __call__(self, input: Documents) -> Embeddings:
# Your embedding logic
return embeddings
my_ef = MyEmbeddingFunction()
collection = client.create_collection(
name="custom_docs",
embedding_function=my_ef
)
```
## Metadata filtering
```python
# Exact match
results = collection.query(
query_texts=["query"],
where={"category": "tutorial"}
)
# Comparison operators
results = collection.query(
query_texts=["query"],
where={"page": {"$gt": 10}} # $gt, $gte, $lt, $lte, $ne
)
# Logical operators
results = collection.query(
query_texts=["query"],
where={
"$and": [
{"category": "tutorial"},
{"difficulty": {"$lte": 3}}
]
} # Also: $or
)
# Contains
results = collection.query(
query_texts=["query"],
where={"tags": {"$in": ["python", "ml"]}}
)
```
## LangChain integration
```python
from langchain_chroma import Chroma
from langchain_openai import OpenAIEmbeddings
from langchain.text_splitter import RecursiveCharacterTextSplitter
# Split documents
text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000)
docs = text_splitter.split_documents(documents)
# Create Chroma vector store
vectorstore = Chroma.from_documents(
documents=docs,
embedding=OpenAIEmbeddings(),
persist_directory="./chroma_db"
)
# Query
results = vectorstore.similarity_search("machine learning", k=3)
# As retriever
retriever = vectorstore.as_retriever(search_kwargs={"k": 5})
```
## LlamaIndex integration
```python
from llama_index.vector_stores.chroma import ChromaVectorStore
from llama_index.core import VectorStoreIndex, StorageContext
import chromadb
# Initialize Chroma
db = chromadb.PersistentClient(path="./chroma_db")
collection = db.get_or_create_collection("my_collection")
# Create vector store
vector_store = ChromaVectorStore(chroma_collection=collection)
storage_context = StorageContext.from_defaults(vector_store=vector_store)
# Create index
index = VectorStoreIndex.from_documents(
documents,
storage_context=storage_context
)
# Query
query_engine = index.as_query_engine()
response = query_engine.query("What is machine learning?")
```
## Server mode
```python
# Run Chroma server
# Terminal: chroma run --path ./chroma_db --port 8000
# Connect to server
import chromadb
from chromadb.config import Settings
client = chromadb.HttpClient(
host="localhost",
port=8000,
settings=Settings(anonymized_telemetry=False)
)
# Use as normal
collection = client.get_or_create_collection("my_docs")
```
## Best practices
1. **Use persistent client** - Don't lose data on restart
2. **Add metadata** - Enables filtering and tracking
3. **Batch operations** - Add multiple docs at once
4. **Choose right embedding model** - Balance speed/quality
5. **Use filters** - Narrow search space
6. **Unique IDs** - Avoid collisions
7. **Regular backups** - Copy chroma_db directory
8. **Monitor collection size** - Scale up if needed
9. **Test embedding functions** - Ensure quality
10. **Use server mode for production** - Better for multi-user
## Performance
| Operation | Latency | Notes |
|-----------|---------|-------|
| Add 100 docs | ~1-3s | With embedding |
| Query (top 10) | ~50-200ms | Depends on collection size |
| Metadata filter | ~10-50ms | Fast with proper indexing |
## Resources
- **GitHub**: https://github.com/chroma-core/chroma ⭐ 24,300+
- **Docs**: https://docs.trychroma.com
- **Discord**: https://discord.gg/MMeYNTmh3x
- **Version**: 1.3.3+
- **License**: Apache 2.0
website/docs/user-guide/skills/optional/mlops/mlops-clip.md
+272
View File
@@ -0,0 +1,272 @@
---
title: "Clip — Zero-shot image classification and image-text search"
sidebar_label: "Clip"
description: "Zero-shot image classification and image-text search"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Clip
Zero-shot image classification and image-text search.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/clip` |
| Path | `optional-skills/mlops\clip` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `transformers`, `torch`, `pillow` |
| Platforms | linux, macos, windows |
| Tags | `Multimodal`, `CLIP`, `Vision-Language`, `Zero-Shot`, `Image Classification`, `OpenAI`, `Image Search`, `Cross-Modal Retrieval`, `Content Moderation` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# CLIP - Contrastive Language-Image Pre-Training
OpenAI's model that understands images from natural language.
## When to use CLIP
**Use when:**
- Zero-shot image classification (no training data needed)
- Image-text similarity/matching
- Semantic image search
- Content moderation (detect NSFW, violence)
- Visual question answering
- Cross-modal retrieval (image→text, text→image)
**Metrics**:
- **25,300+ GitHub stars**
- Trained on 400M image-text pairs
- Matches ResNet-50 on ImageNet (zero-shot)
- MIT License
**Use alternatives instead**:
- **BLIP-2**: Better captioning
- **LLaVA**: Vision-language chat
- **Segment Anything**: Image segmentation
## Quick start
### Installation
```bash
pip install git+https://github.com/openai/CLIP.git
pip install torch torchvision ftfy regex tqdm
```
### Zero-shot classification
```python
import torch
import clip
from PIL import Image
# Load model
device = "cuda" if torch.cuda.is_available() else "cpu"
model, preprocess = clip.load("ViT-B/32", device=device)
# Load image
image = preprocess(Image.open("photo.jpg")).unsqueeze(0).to(device)
# Define possible labels
text = clip.tokenize(["a dog", "a cat", "a bird", "a car"]).to(device)
# Compute similarity
with torch.no_grad():
image_features = model.encode_image(image)
text_features = model.encode_text(text)
# Cosine similarity
logits_per_image, logits_per_text = model(image, text)
probs = logits_per_image.softmax(dim=-1).cpu().numpy()
# Print results
labels = ["a dog", "a cat", "a bird", "a car"]
for label, prob in zip(labels, probs[0]):
print(f"{label}: {prob:.2%}")
```
## Available models
```python
# Models (sorted by size)
models = [
"RN50", # ResNet-50
"RN101", # ResNet-101
"ViT-B/32", # Vision Transformer (recommended)
"ViT-B/16", # Better quality, slower
"ViT-L/14", # Best quality, slowest
]
model, preprocess = clip.load("ViT-B/32")
```
| Model | Parameters | Speed | Quality |
|-------|------------|-------|---------|
| RN50 | 102M | Fast | Good |
| ViT-B/32 | 151M | Medium | Better |
| ViT-L/14 | 428M | Slow | Best |
## Image-text similarity
```python
# Compute embeddings
image_features = model.encode_image(image)
text_features = model.encode_text(text)
# Normalize
image_features /= image_features.norm(dim=-1, keepdim=True)
text_features /= text_features.norm(dim=-1, keepdim=True)
# Cosine similarity
similarity = (image_features @ text_features.T).item()
print(f"Similarity: {similarity:.4f}")
```
## Semantic image search
```python
# Index images
image_paths = ["img1.jpg", "img2.jpg", "img3.jpg"]
image_embeddings = []
for img_path in image_paths:
image = preprocess(Image.open(img_path)).unsqueeze(0).to(device)
with torch.no_grad():
embedding = model.encode_image(image)
embedding /= embedding.norm(dim=-1, keepdim=True)
image_embeddings.append(embedding)
image_embeddings = torch.cat(image_embeddings)
# Search with text query
query = "a sunset over the ocean"
text_input = clip.tokenize([query]).to(device)
with torch.no_grad():
text_embedding = model.encode_text(text_input)
text_embedding /= text_embedding.norm(dim=-1, keepdim=True)
# Find most similar images
similarities = (text_embedding @ image_embeddings.T).squeeze(0)
top_k = similarities.topk(3)
for idx, score in zip(top_k.indices, top_k.values):
print(f"{image_paths[idx]}: {score:.3f}")
```
## Content moderation
```python
# Define categories
categories = [
"safe for work",
"not safe for work",
"violent content",
"graphic content"
]
text = clip.tokenize(categories).to(device)
# Check image
with torch.no_grad():
logits_per_image, _ = model(image, text)
probs = logits_per_image.softmax(dim=-1)
# Get classification
max_idx = probs.argmax().item()
max_prob = probs[0, max_idx].item()
print(f"Category: {categories[max_idx]} ({max_prob:.2%})")
```
## Batch processing
```python
# Process multiple images
images = [preprocess(Image.open(f"img{i}.jpg")) for i in range(10)]
images = torch.stack(images).to(device)
with torch.no_grad():
image_features = model.encode_image(images)
image_features /= image_features.norm(dim=-1, keepdim=True)
# Batch text
texts = ["a dog", "a cat", "a bird"]
text_tokens = clip.tokenize(texts).to(device)
with torch.no_grad():
text_features = model.encode_text(text_tokens)
text_features /= text_features.norm(dim=-1, keepdim=True)
# Similarity matrix (10 images × 3 texts)
similarities = image_features @ text_features.T
print(similarities.shape) # (10, 3)
```
## Integration with vector databases
```python
# Store CLIP embeddings in Chroma/FAISS
import chromadb
client = chromadb.Client()
collection = client.create_collection("image_embeddings")
# Add image embeddings
for img_path, embedding in zip(image_paths, image_embeddings):
collection.add(
embeddings=[embedding.cpu().numpy().tolist()],
metadatas=[{"path": img_path}],
ids=[img_path]
)
# Query with text
query = "a sunset"
text_embedding = model.encode_text(clip.tokenize([query]))
results = collection.query(
query_embeddings=[text_embedding.cpu().numpy().tolist()],
n_results=5
)
```
## Best practices
1. **Use ViT-B/32 for most cases** - Good balance
2. **Normalize embeddings** - Required for cosine similarity
3. **Batch processing** - More efficient
4. **Cache embeddings** - Expensive to recompute
5. **Use descriptive labels** - Better zero-shot performance
6. **GPU recommended** - 10-50× faster
7. **Preprocess images** - Use provided preprocess function
## Performance
| Operation | CPU | GPU (V100) |
|-----------|-----|------------|
| Image encoding | ~200ms | ~20ms |
| Text encoding | ~50ms | ~5ms |
| Similarity compute | <1ms | <1ms |
## Limitations
1. **Not for fine-grained tasks** - Best for broad categories
2. **Requires descriptive text** - Vague labels perform poorly
3. **Biased on web data** - May have dataset biases
4. **No bounding boxes** - Whole image only
5. **Limited spatial understanding** - Position/counting weak
## Resources
- **GitHub**: https://github.com/openai/CLIP ⭐ 25,300+
- **Paper**: https://arxiv.org/abs/2103.00020
- **Colab**: https://colab.research.google.com/github/openai/clip/
- **License**: MIT
website/docs/user-guide/skills/optional/mlops/mlops-evaluation-evaluating-llms-harness.md
+512
View File
@@ -0,0 +1,512 @@
---
title: "Evaluating Llms Harness — lm-eval-harness: benchmark LLMs (MMLU, GSM8K, etc.)"
sidebar_label: "Evaluating Llms Harness"
description: "lm-eval-harness: benchmark LLMs (MMLU, GSM8K, etc.)"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Evaluating Llms Harness
lm-eval-harness: benchmark LLMs (MMLU, GSM8K, etc.).
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/evaluating-llms-harness` |
| Path | `optional-skills/mlops\evaluation\evaluating-llms-harness` |
| Version | `1.0.1` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `lm-eval`, `transformers`, `vllm` |
| Platforms | linux, macos |
| Tags | `Evaluation`, `LM Evaluation Harness`, `Benchmarking`, `MMLU`, `HumanEval`, `GSM8K`, `EleutherAI`, `Model Quality`, `Academic Benchmarks`, `Industry Standard` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# lm-evaluation-harness - LLM Benchmarking
## What's inside
Evaluates LLMs across 60+ academic benchmarks (MMLU, HumanEval, GSM8K, TruthfulQA, HellaSwag). Use when benchmarking model quality, comparing models, reporting academic results, or tracking training progress. Industry standard used by EleutherAI, HuggingFace, and major labs. Supports HuggingFace, vLLM, APIs.
## Quick start
lm-evaluation-harness evaluates LLMs across 60+ academic benchmarks using standardized prompts and metrics.
**Installation**:
```bash
pip install lm-eval
```
**Evaluate any HuggingFace model**:
```bash
lm_eval --model hf \
--model_args pretrained=meta-llama/Llama-2-7b-hf \
--tasks mmlu,gsm8k,hellaswag \
--device cuda:0 \
--batch_size 8
```
**View available tasks**:
```bash
lm-eval ls tasks
```
## Common workflows
### Workflow 1: Standard benchmark evaluation
Evaluate model on core benchmarks (MMLU, GSM8K, HumanEval).
Copy this checklist:
```
Benchmark Evaluation:
- [ ] Step 1: Choose benchmark suite
- [ ] Step 2: Configure model
- [ ] Step 3: Run evaluation
- [ ] Step 4: Analyze results
```
**Step 1: Choose benchmark suite**
**Core reasoning benchmarks**:
- **MMLU** (Massive Multitask Language Understanding) - 57 subjects, multiple choice
- **GSM8K** - Grade school math word problems
- **HellaSwag** - Common sense reasoning
- **TruthfulQA** - Truthfulness and factuality
- **ARC** (AI2 Reasoning Challenge) - Science questions
**Code benchmarks**:
- **HumanEval** - Python code generation (164 problems)
- **MBPP** (Mostly Basic Python Problems) - Python coding
**Standard suite** (recommended for model releases):
```bash
--tasks mmlu,gsm8k,hellaswag,truthfulqa,arc_challenge
```
**Step 2: Configure model**
**HuggingFace model**:
```bash
lm_eval --model hf \
--model_args pretrained=meta-llama/Llama-2-7b-hf,dtype=bfloat16 \
--tasks mmlu \
--device cuda:0 \
--batch_size auto # Auto-detect optimal batch size
```
**Quantized model (4-bit/8-bit)**:
```bash
lm_eval --model hf \
--model_args pretrained=meta-llama/Llama-2-7b-hf,load_in_4bit=True \
--tasks mmlu \
--device cuda:0
```
**Custom checkpoint**:
```bash
lm_eval --model hf \
--model_args pretrained=/path/to/my-model,tokenizer=/path/to/tokenizer \
--tasks mmlu \
--device cuda:0
```
**Step 3: Run evaluation**
```bash
# Full MMLU evaluation (57 subjects)
lm_eval --model hf \
--model_args pretrained=meta-llama/Llama-2-7b-hf \
--tasks mmlu \
--num_fewshot 5 \ # 5-shot evaluation (standard)
--batch_size 8 \
--output_path results/ \
--log_samples # Save individual predictions
# Multiple benchmarks at once
lm_eval --model hf \
--model_args pretrained=meta-llama/Llama-2-7b-hf \
--tasks mmlu,gsm8k,hellaswag,truthfulqa,arc_challenge \
--num_fewshot 5 \
--batch_size 8 \
--output_path results/llama2-7b-eval.json
```
**Step 4: Analyze results**
Results saved to `results/llama2-7b-eval.json`:
```json
{
"results": {
"mmlu": {
"acc": 0.459,
"acc_stderr": 0.004
},
"gsm8k": {
"exact_match": 0.142,
"exact_match_stderr": 0.006
},
"hellaswag": {
"acc_norm": 0.765,
"acc_norm_stderr": 0.004
}
},
"config": {
"model": "hf",
"model_args": "pretrained=meta-llama/Llama-2-7b-hf",
"num_fewshot": 5
}
}
```
### Workflow 2: Track training progress
Evaluate checkpoints during training.
```
Training Progress Tracking:
- [ ] Step 1: Set up periodic evaluation
- [ ] Step 2: Choose quick benchmarks
- [ ] Step 3: Automate evaluation
- [ ] Step 4: Plot learning curves
```
**Step 1: Set up periodic evaluation**
Evaluate every N training steps:
```bash
#!/bin/bash
# eval_checkpoint.sh
CHECKPOINT_DIR=$1
STEP=$2
lm_eval --model hf \
--model_args pretrained=$CHECKPOINT_DIR/checkpoint-$STEP \
--tasks gsm8k,hellaswag \
--num_fewshot 0 \ # 0-shot for speed
--batch_size 16 \
--output_path results/step-$STEP.json
```
**Step 2: Choose quick benchmarks**
Fast benchmarks for frequent evaluation:
- **HellaSwag**: ~10 minutes on 1 GPU
- **GSM8K**: ~5 minutes
- **PIQA**: ~2 minutes
Avoid for frequent eval (too slow):
- **MMLU**: ~2 hours (57 subjects)
- **HumanEval**: Requires code execution
**Step 3: Automate evaluation**
Integrate with training script:
```python
# In training loop
if step % eval_interval == 0:
model.save_pretrained(f"checkpoints/step-{step}")
# Run evaluation
os.system(f"./eval_checkpoint.sh checkpoints step-{step}")
```
Or use PyTorch Lightning callbacks:
```python
from pytorch_lightning import Callback
class EvalHarnessCallback(Callback):
def on_validation_epoch_end(self, trainer, pl_module):
step = trainer.global_step
checkpoint_path = f"checkpoints/step-{step}"
# Save checkpoint
trainer.save_checkpoint(checkpoint_path)
# Run lm-eval
os.system(f"lm_eval --model hf --model_args pretrained={checkpoint_path} ...")
```
**Step 4: Plot learning curves**
```python
import json
import matplotlib.pyplot as plt
# Load all results
steps = []
mmlu_scores = []
for file in sorted(glob.glob("results/step-*.json")):
with open(file) as f:
data = json.load(f)
step = int(file.split("-")[1].split(".")[0])
steps.append(step)
mmlu_scores.append(data["results"]["mmlu"]["acc"])
# Plot
plt.plot(steps, mmlu_scores)
plt.xlabel("Training Step")
plt.ylabel("MMLU Accuracy")
plt.title("Training Progress")
plt.savefig("training_curve.png")
```
### Workflow 3: Compare multiple models
Benchmark suite for model comparison.
```
Model Comparison:
- [ ] Step 1: Define model list
- [ ] Step 2: Run evaluations
- [ ] Step 3: Generate comparison table
```
**Step 1: Define model list**
```bash
# models.txt
meta-llama/Llama-2-7b-hf
meta-llama/Llama-2-13b-hf
mistralai/Mistral-7B-v0.1
microsoft/phi-2
```
**Step 2: Run evaluations**
```bash
#!/bin/bash
# eval_all_models.sh
TASKS="mmlu,gsm8k,hellaswag,truthfulqa"
while read model; do
echo "Evaluating $model"
# Extract model name for output file
model_name=$(echo $model | sed 's/\//-/g')
lm_eval --model hf \
--model_args pretrained=$model,dtype=bfloat16 \
--tasks $TASKS \
--num_fewshot 5 \
--batch_size auto \
--output_path results/$model_name.json
done < models.txt
```
**Step 3: Generate comparison table**
```python
import json
import pandas as pd
models = [
"meta-llama-Llama-2-7b-hf",
"meta-llama-Llama-2-13b-hf",
"mistralai-Mistral-7B-v0.1",
"microsoft-phi-2"
]
tasks = ["mmlu", "gsm8k", "hellaswag", "truthfulqa"]
results = []
for model in models:
with open(f"results/{model}.json") as f:
data = json.load(f)
row = {"Model": model.replace("-", "/")}
for task in tasks:
# Get primary metric for each task
metrics = data["results"][task]
if "acc" in metrics:
row[task.upper()] = f"{metrics['acc']:.3f}"
elif "exact_match" in metrics:
row[task.upper()] = f"{metrics['exact_match']:.3f}"
results.append(row)
df = pd.DataFrame(results)
print(df.to_markdown(index=False))
```
Output:
```
| Model | MMLU | GSM8K | HELLASWAG | TRUTHFULQA |
|------------------------|-------|-------|-----------|------------|
| meta-llama/Llama-2-7b | 0.459 | 0.142 | 0.765 | 0.391 |
| meta-llama/Llama-2-13b | 0.549 | 0.287 | 0.801 | 0.430 |
| mistralai/Mistral-7B | 0.626 | 0.395 | 0.812 | 0.428 |
| microsoft/phi-2 | 0.560 | 0.613 | 0.682 | 0.447 |
```
### Workflow 4: Evaluate with vLLM (faster inference)
Use vLLM backend for 5-10x faster evaluation.
```
vLLM Evaluation:
- [ ] Step 1: Install vLLM
- [ ] Step 2: Configure vLLM backend
- [ ] Step 3: Run evaluation
```
**Step 1: Install vLLM**
```bash
pip install vllm
```
**Step 2: Configure vLLM backend**
```bash
lm_eval --model vllm \
--model_args pretrained=meta-llama/Llama-2-7b-hf,tensor_parallel_size=1,dtype=auto,gpu_memory_utilization=0.8 \
--tasks mmlu \
--batch_size auto
```
**Step 3: Run evaluation**
vLLM is 5-10× faster than standard HuggingFace:
```bash
# Standard HF: ~2 hours for MMLU on 7B model
lm_eval --model hf \
--model_args pretrained=meta-llama/Llama-2-7b-hf \
--tasks mmlu \
--batch_size 8
# vLLM: ~15-20 minutes for MMLU on 7B model
lm_eval --model vllm \
--model_args pretrained=meta-llama/Llama-2-7b-hf,tensor_parallel_size=2 \
--tasks mmlu \
--batch_size auto
```
## When to use vs alternatives
**Use lm-evaluation-harness when:**
- Benchmarking models for academic papers
- Comparing model quality across standard tasks
- Tracking training progress
- Reporting standardized metrics (everyone uses same prompts)
- Need reproducible evaluation
**Use alternatives instead:**
- **HELM** (Stanford): Broader evaluation (fairness, efficiency, calibration)
- **AlpacaEval**: Instruction-following evaluation with LLM judges
- **MT-Bench**: Conversational multi-turn evaluation
- **Custom scripts**: Domain-specific evaluation
## Common issues
**Issue: Evaluation too slow**
Use vLLM backend:
```bash
lm_eval --model vllm \
--model_args pretrained=model-name,tensor_parallel_size=2
```
Or reduce fewshot examples:
```bash
--num_fewshot 0 # Instead of 5
```
Or evaluate subset of MMLU:
```bash
--tasks mmlu_stem # Only STEM subjects
```
**Issue: Out of memory**
Reduce batch size:
```bash
--batch_size 1 # Or --batch_size auto
```
Use quantization:
```bash
--model_args pretrained=model-name,load_in_8bit=True
```
Enable CPU offloading:
```bash
--model_args pretrained=model-name,device_map=auto,offload_folder=offload
```
**Issue: Different results than reported**
Check fewshot count:
```bash
--num_fewshot 5 # Most papers use 5-shot
```
Check exact task name:
```bash
--tasks mmlu # Not mmlu_direct or mmlu_fewshot
```
Verify model and tokenizer match:
```bash
--model_args pretrained=model-name,tokenizer=same-model-name
```
**Issue: HumanEval not executing code**
Code-executing tasks (HumanEval, MBPP, etc.) are gated behind an explicit
confirmation flag — you must pass `--confirm_run_unsafe_code` to run them:
```bash
lm_eval --model hf \
--model_args pretrained=model-name \
--tasks humaneval \
--confirm_run_unsafe_code # Required to run tasks that execute generated code
```
Without this flag lm-eval refuses to run the task rather than silently skipping
code execution.
## Advanced topics
**Benchmark descriptions**: See [references/benchmark-guide.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\evaluation\evaluating-llms-harness/references/benchmark-guide.md) for detailed description of all 60+ tasks, what they measure, and interpretation.
**Custom tasks**: See [references/custom-tasks.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\evaluation\evaluating-llms-harness/references/custom-tasks.md) for creating domain-specific evaluation tasks.
**API evaluation**: See [references/api-evaluation.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\evaluation\evaluating-llms-harness/references/api-evaluation.md) for evaluating OpenAI, Anthropic, and other API models.
**Multi-GPU strategies**: See [references/distributed-eval.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\evaluation\evaluating-llms-harness/references/distributed-eval.md) for data parallel and tensor parallel evaluation.
## Hardware requirements
- **GPU**: NVIDIA (CUDA 11.8+), works on CPU (very slow)
- **VRAM**:
- 7B model: 16GB (bf16) or 8GB (8-bit)
- 13B model: 28GB (bf16) or 14GB (8-bit)
- 70B model: Requires multi-GPU or quantization
- **Time** (7B model, single A100):
- HellaSwag: 10 minutes
- GSM8K: 5 minutes
- MMLU (full): 2 hours
- HumanEval: 20 minutes
## Resources
- GitHub: https://github.com/EleutherAI/lm-evaluation-harness
- Docs: https://github.com/EleutherAI/lm-evaluation-harness/tree/main/docs
- Task library: 60+ tasks including MMLU, GSM8K, HumanEval, TruthfulQA, HellaSwag, ARC, WinoGrande, etc.
- Leaderboard: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard (uses this harness)
website/docs/user-guide/skills/optional/mlops/mlops-evaluation-weights-and-biases.md
+613
View File
@@ -0,0 +1,613 @@
---
title: "Weights And Biases — W&B: log ML experiments, sweeps, model registry, dashboards"
sidebar_label: "Weights And Biases"
description: "W&B: log ML experiments, sweeps, model registry, dashboards"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Weights And Biases
W&B: log ML experiments, sweeps, model registry, dashboards.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/weights-and-biases` |
| Path | `optional-skills/mlops\evaluation\weights-and-biases` |
| Version | `1.0.1` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `wandb` |
| Platforms | linux, macos, windows |
| Tags | `MLOps`, `Weights And Biases`, `WandB`, `Experiment Tracking`, `Hyperparameter Tuning`, `Model Registry`, `Collaboration`, `Real-Time Visualization`, `PyTorch`, `TensorFlow`, `HuggingFace` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Weights & Biases: ML Experiment Tracking & MLOps
## When to Use This Skill
Use Weights & Biases (W&B) when you need to:
- **Track ML experiments** with automatic metric logging
- **Visualize training** in real-time dashboards
- **Compare runs** across hyperparameters and configurations
- **Optimize hyperparameters** with automated sweeps
- **Manage model registry** with versioning and lineage
- **Collaborate on ML projects** with team workspaces
- **Track artifacts** (datasets, models, code) with lineage
**Users**: 200,000+ ML practitioners | **GitHub Stars**: 10.5k+ | **Integrations**: 100+
## Installation
```bash
# Install W&B
pip install wandb
# Login (creates API key)
wandb login
# Or set API key programmatically
export WANDB_API_KEY=your_api_key_here
```
## Quick Start
### Basic Experiment Tracking
```python
import wandb
# Initialize a run
run = wandb.init(
project="my-project",
config={
"learning_rate": 0.001,
"epochs": 10,
"batch_size": 32,
"architecture": "ResNet50"
}
)
# Training loop
for epoch in range(run.config.epochs):
# Your training code
train_loss = train_epoch()
val_loss = validate()
# Log metrics
wandb.log({
"epoch": epoch,
"train/loss": train_loss,
"val/loss": val_loss,
"train/accuracy": train_acc,
"val/accuracy": val_acc
})
# Finish the run
wandb.finish()
```
### With PyTorch
```python
import torch
import wandb
# Initialize
wandb.init(project="pytorch-demo", config={
"lr": 0.001,
"epochs": 10
})
# Access config
config = wandb.config
# Training loop
for epoch in range(config.epochs):
for batch_idx, (data, target) in enumerate(train_loader):
# Forward pass
output = model(data)
loss = criterion(output, target)
# Backward pass
optimizer.zero_grad()
loss.backward()
optimizer.step()
# Log every 100 batches
if batch_idx % 100 == 0:
wandb.log({
"loss": loss.item(),
"epoch": epoch,
"batch": batch_idx
})
# Save model
torch.save(model.state_dict(), "model.pth")
wandb.save("model.pth") # Upload to W&B
wandb.finish()
```
## Core Concepts
### 1. Projects and Runs
**Project**: Collection of related experiments
**Run**: Single execution of your training script
```python
# Create/use project
run = wandb.init(
project="image-classification",
name="resnet50-experiment-1", # Optional run name
tags=["baseline", "resnet"], # Organize with tags
notes="First baseline run" # Add notes
)
# Each run has unique ID
print(f"Run ID: {run.id}")
print(f"Run URL: {run.url}")
```
### 2. Configuration Tracking
Track hyperparameters automatically:
```python
config = {
# Model architecture
"model": "ResNet50",
"pretrained": True,
# Training params
"learning_rate": 0.001,
"batch_size": 32,
"epochs": 50,
"optimizer": "Adam",
# Data params
"dataset": "ImageNet",
"augmentation": "standard"
}
wandb.init(project="my-project", config=config)
# Access config during training
lr = wandb.config.learning_rate
batch_size = wandb.config.batch_size
```
### 3. Metric Logging
```python
# Log scalars
wandb.log({"loss": 0.5, "accuracy": 0.92})
# Log multiple metrics
wandb.log({
"train/loss": train_loss,
"train/accuracy": train_acc,
"val/loss": val_loss,
"val/accuracy": val_acc,
"learning_rate": current_lr,
"epoch": epoch
})
# Log with custom x-axis
wandb.log({"loss": loss}, step=global_step)
# Log media (images, audio, video)
wandb.log({"examples": [wandb.Image(img) for img in images]})
# Log histograms
wandb.log({"gradients": wandb.Histogram(gradients)})
# Log tables
table = wandb.Table(columns=["id", "prediction", "ground_truth"])
wandb.log({"predictions": table})
```
### 4. Model Checkpointing
```python
import torch
import wandb
# Save model checkpoint
checkpoint = {
'epoch': epoch,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'loss': loss,
}
torch.save(checkpoint, 'checkpoint.pth')
# Upload to W&B
wandb.save('checkpoint.pth')
# Or use Artifacts (recommended)
artifact = wandb.Artifact('model', type='model')
artifact.add_file('checkpoint.pth')
wandb.log_artifact(artifact)
```
## Hyperparameter Sweeps
Automatically search for optimal hyperparameters.
### Define Sweep Configuration
```python
sweep_config = {
'method': 'bayes', # or 'grid', 'random'
'metric': {
'name': 'val/accuracy',
'goal': 'maximize'
},
'parameters': {
'learning_rate': {
'distribution': 'log_uniform_values',
'min': 1e-5,
'max': 1e-1
},
'batch_size': {
'values': [16, 32, 64, 128]
},
'optimizer': {
'values': ['adam', 'sgd', 'rmsprop']
},
'dropout': {
'distribution': 'uniform',
'min': 0.1,
'max': 0.5
}
}
}
# Initialize sweep
sweep_id = wandb.sweep(sweep_config, project="my-project")
```
### Define Training Function
```python
def train():
# Initialize run
run = wandb.init()
# Access sweep parameters
lr = wandb.config.learning_rate
batch_size = wandb.config.batch_size
optimizer_name = wandb.config.optimizer
# Build model with sweep config
model = build_model(wandb.config)
optimizer = get_optimizer(optimizer_name, lr)
# Training loop
for epoch in range(NUM_EPOCHS):
train_loss = train_epoch(model, optimizer, batch_size)
val_acc = validate(model)
# Log metrics
wandb.log({
"train/loss": train_loss,
"val/accuracy": val_acc
})
# Run sweep
wandb.agent(sweep_id, function=train, count=50) # Run 50 trials
```
### Sweep Strategies
```python
# Grid search - exhaustive
sweep_config = {
'method': 'grid',
'parameters': {
'lr': {'values': [0.001, 0.01, 0.1]},
'batch_size': {'values': [16, 32, 64]}
}
}
# Random search
sweep_config = {
'method': 'random',
'parameters': {
'lr': {'distribution': 'uniform', 'min': 0.0001, 'max': 0.1},
'dropout': {'distribution': 'uniform', 'min': 0.1, 'max': 0.5}
}
}
# Bayesian optimization (recommended)
sweep_config = {
'method': 'bayes',
'metric': {'name': 'val/loss', 'goal': 'minimize'},
'parameters': {
'lr': {'distribution': 'log_uniform_values', 'min': 1e-5, 'max': 1e-1}
}
}
```
## Artifacts
Track datasets, models, and other files with lineage.
### Log Artifacts
```python
# Create artifact
artifact = wandb.Artifact(
name='training-dataset',
type='dataset',
description='ImageNet training split',
metadata={'size': '1.2M images', 'split': 'train'}
)
# Add files
artifact.add_file('data/train.csv')
artifact.add_dir('data/images/')
# Log artifact
wandb.log_artifact(artifact)
```
### Use Artifacts
```python
# Download and use artifact
run = wandb.init(project="my-project")
# Download artifact
artifact = run.use_artifact('training-dataset:latest')
artifact_dir = artifact.download()
# Use the data
data = load_data(f"{artifact_dir}/train.csv")
```
### Model Registry
```python
# Log model as artifact
model_artifact = wandb.Artifact(
name='resnet50-model',
type='model',
metadata={'architecture': 'ResNet50', 'accuracy': 0.95}
)
model_artifact.add_file('model.pth')
wandb.log_artifact(model_artifact, aliases=['best', 'production'])
# Link to model registry
run.link_artifact(model_artifact, 'model-registry/production-models')
```
## Integration Examples
### HuggingFace Transformers
```python
from transformers import Trainer, TrainingArguments
import wandb
# Initialize W&B
wandb.init(project="hf-transformers")
# Training arguments with W&B
training_args = TrainingArguments(
output_dir="./results",
report_to="wandb", # Enable W&B logging
run_name="bert-finetuning",
logging_steps=100,
save_steps=500
)
# Trainer automatically logs to W&B
trainer = Trainer(
model=model,
args=training_args,
train_dataset=train_dataset,
eval_dataset=eval_dataset
)
trainer.train()
```
### PyTorch Lightning
```python
from pytorch_lightning import Trainer
from pytorch_lightning.loggers import WandbLogger
import wandb
# Create W&B logger
wandb_logger = WandbLogger(
project="lightning-demo",
log_model=True # Log model checkpoints
)
# Use with Trainer
trainer = Trainer(
logger=wandb_logger,
max_epochs=10
)
trainer.fit(model, datamodule=dm)
```
### Keras/TensorFlow
```python
import wandb
from wandb.integration.keras import WandbMetricsLogger, WandbModelCheckpoint
# Initialize
wandb.init(project="keras-demo")
# Add callbacks (the monolithic WandbCallback was removed;
# use the dedicated callbacks from wandb.integration.keras instead)
model.fit(
x_train, y_train,
validation_data=(x_val, y_val),
epochs=10,
callbacks=[
WandbMetricsLogger(), # Auto-logs metrics
WandbModelCheckpoint("models/model-{epoch}") # Saves checkpoints
]
)
```
## Visualization & Analysis
### Custom Charts
```python
# Log custom visualizations
import matplotlib.pyplot as plt
fig, ax = plt.subplots()
ax.plot(x, y)
wandb.log({"custom_plot": wandb.Image(fig)})
# Log confusion matrix
wandb.log({"conf_mat": wandb.plot.confusion_matrix(
probs=None,
y_true=ground_truth,
preds=predictions,
class_names=class_names
)})
```
### Reports
Create shareable reports in W&B UI:
- Combine runs, charts, and text
- Markdown support
- Embeddable visualizations
- Team collaboration
## Best Practices
### 1. Organize with Tags and Groups
```python
wandb.init(
project="my-project",
tags=["baseline", "resnet50", "imagenet"],
group="resnet-experiments", # Group related runs
job_type="train" # Type of job
)
```
### 2. Log Everything Relevant
```python
# Log system metrics
wandb.log({
"gpu/util": gpu_utilization,
"gpu/memory": gpu_memory_used,
"cpu/util": cpu_utilization
})
# Log code version
wandb.log({"git_commit": git_commit_hash})
# Log data splits
wandb.log({
"data/train_size": len(train_dataset),
"data/val_size": len(val_dataset)
})
```
### 3. Use Descriptive Names
```python
# ✅ Good: Descriptive run names
wandb.init(
project="nlp-classification",
name="bert-base-lr0.001-bs32-epoch10"
)
# ❌ Bad: Generic names
wandb.init(project="nlp", name="run1")
```
### 4. Save Important Artifacts
```python
# Save final model
artifact = wandb.Artifact('final-model', type='model')
artifact.add_file('model.pth')
wandb.log_artifact(artifact)
# Save predictions for analysis
predictions_table = wandb.Table(
columns=["id", "input", "prediction", "ground_truth"],
data=predictions_data
)
wandb.log({"predictions": predictions_table})
```
### 5. Use Offline Mode for Unstable Connections
```python
import os
# Enable offline mode
os.environ["WANDB_MODE"] = "offline"
wandb.init(project="my-project")
# ... your code ...
# Sync later
# wandb sync <run_directory>
```
## Team Collaboration
### Share Runs
```python
# Runs are automatically shareable via URL
run = wandb.init(project="team-project")
print(f"Share this URL: {run.url}")
```
### Team Projects
- Create team account at wandb.ai
- Add team members
- Set project visibility (private/public)
- Use team-level artifacts and model registry
## Pricing
- **Free**: Unlimited public projects, 100GB storage
- **Academic**: Free for students/researchers
- **Teams**: $50/seat/month, private projects, unlimited storage
- **Enterprise**: Custom pricing, on-prem options
## Resources
- **Documentation**: https://docs.wandb.ai
- **GitHub**: https://github.com/wandb/wandb (10.5k+ stars)
- **Examples**: https://github.com/wandb/examples
- **Community**: https://wandb.ai/community
- **Discord**: https://wandb.me/discord
## See Also
- `references/sweeps.md` - Comprehensive hyperparameter optimization guide
- `references/artifacts.md` - Data and model versioning patterns
- `references/integrations.md` - Framework-specific examples
website/docs/user-guide/skills/optional/mlops/mlops-faiss.md
+240
View File
@@ -0,0 +1,240 @@
---
title: "Faiss — Fast vector similarity search at billion scale"
sidebar_label: "Faiss"
description: "Fast vector similarity search at billion scale"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Faiss
Fast vector similarity search at billion scale.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/faiss` |
| Path | `optional-skills/mlops\faiss` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `faiss-cpu`, `faiss-gpu`, `numpy` |
| Platforms | linux, macos |
| Tags | `RAG`, `FAISS`, `Similarity Search`, `Vector Search`, `Facebook AI`, `GPU Acceleration`, `Billion-Scale`, `K-NN`, `HNSW`, `High Performance`, `Large Scale` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# FAISS - Efficient Similarity Search
Facebook AI's library for billion-scale vector similarity search.
## When to use FAISS
**Use FAISS when:**
- Need fast similarity search on large vector datasets (millions/billions)
- GPU acceleration required
- Pure vector similarity (no metadata filtering needed)
- High throughput, low latency critical
- Offline/batch processing of embeddings
**Metrics**:
- **31,700+ GitHub stars**
- Meta/Facebook AI Research
- **Handles billions of vectors**
- **C++** with Python bindings
**Use alternatives instead**:
- **Chroma/Pinecone**: Need metadata filtering
- **Weaviate**: Need full database features
- **Annoy**: Simpler, fewer features
## Quick start
### Installation
```bash
# CPU only
pip install faiss-cpu
# GPU support
pip install faiss-gpu
```
### Basic usage
```python
import faiss
import numpy as np
# Create sample data (1000 vectors, 128 dimensions)
d = 128
nb = 1000
vectors = np.random.random((nb, d)).astype('float32')
# Create index
index = faiss.IndexFlatL2(d) # L2 distance
index.add(vectors) # Add vectors
# Search
k = 5 # Find 5 nearest neighbors
query = np.random.random((1, d)).astype('float32')
distances, indices = index.search(query, k)
print(f"Nearest neighbors: {indices}")
print(f"Distances: {distances}")
```
## Index types
### 1. Flat (exact search)
```python
# L2 (Euclidean) distance
index = faiss.IndexFlatL2(d)
# Inner product (cosine similarity if normalized)
index = faiss.IndexFlatIP(d)
# Slowest, most accurate
```
### 2. IVF (inverted file) - Fast approximate
```python
# Create quantizer
quantizer = faiss.IndexFlatL2(d)
# IVF index with 100 clusters
nlist = 100
index = faiss.IndexIVFFlat(quantizer, d, nlist)
# Train on data
index.train(vectors)
# Add vectors
index.add(vectors)
# Search (nprobe = clusters to search)
index.nprobe = 10
distances, indices = index.search(query, k)
```
### 3. HNSW (Hierarchical NSW) - Best quality/speed
```python
# HNSW index
M = 32 # Number of connections per layer
index = faiss.IndexHNSWFlat(d, M)
# No training needed
index.add(vectors)
# Search
distances, indices = index.search(query, k)
```
### 4. Product Quantization - Memory efficient
```python
# PQ reduces memory by 16-32×
m = 8 # Number of subquantizers
nbits = 8
index = faiss.IndexPQ(d, m, nbits)
# Train and add
index.train(vectors)
index.add(vectors)
```
## Save and load
```python
# Save index
faiss.write_index(index, "large.index")
# Load index
index = faiss.read_index("large.index")
# Continue using
distances, indices = index.search(query, k)
```
## GPU acceleration
```python
# Single GPU
res = faiss.StandardGpuResources()
index_cpu = faiss.IndexFlatL2(d)
index_gpu = faiss.index_cpu_to_gpu(res, 0, index_cpu) # GPU 0
# Multi-GPU
index_gpu = faiss.index_cpu_to_all_gpus(index_cpu)
# 10-100× faster than CPU
```
## LangChain integration
```python
from langchain_community.vectorstores import FAISS
from langchain_openai import OpenAIEmbeddings
# Create FAISS vector store
vectorstore = FAISS.from_documents(docs, OpenAIEmbeddings())
# Save
vectorstore.save_local("faiss_index")
# Load
vectorstore = FAISS.load_local(
"faiss_index",
OpenAIEmbeddings(),
allow_dangerous_deserialization=True
)
# Search
results = vectorstore.similarity_search("query", k=5)
```
## LlamaIndex integration
```python
from llama_index.vector_stores.faiss import FaissVectorStore
import faiss
# Create FAISS index
d = 1536
faiss_index = faiss.IndexFlatL2(d)
vector_store = FaissVectorStore(faiss_index=faiss_index)
```
## Best practices
1. **Choose right index type** - Flat for &lt;10K, IVF for 10K-1M, HNSW for quality
2. **Normalize for cosine** - Use IndexFlatIP with normalized vectors
3. **Use GPU for large datasets** - 10-100× faster
4. **Save trained indices** - Training is expensive
5. **Tune nprobe/ef_search** - Balance speed/accuracy
6. **Monitor memory** - PQ for large datasets
7. **Batch queries** - Better GPU utilization
## Performance
| Index Type | Build Time | Search Time | Memory | Accuracy |
|------------|------------|-------------|--------|----------|
| Flat | Fast | Slow | High | 100% |
| IVF | Medium | Fast | Medium | 95-99% |
| HNSW | Slow | Fastest | High | 99% |
| PQ | Medium | Fast | Low | 90-95% |
## Resources
- **GitHub**: https://github.com/facebookresearch/faiss ⭐ 31,700+
- **Wiki**: https://github.com/facebookresearch/faiss/wiki
- **License**: MIT
website/docs/user-guide/skills/optional/mlops/mlops-flash-attention.md
+388
View File
@@ -0,0 +1,388 @@
---
title: "Flash Attention — Speed up long-sequence transformer training and inference"
sidebar_label: "Flash Attention"
description: "Speed up long-sequence transformer training and inference"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Flash Attention
Speed up long-sequence transformer training and inference.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/flash-attention` |
| Path | `optional-skills/mlops\flash-attention` |
| Version | `1.0.1` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `flash-attn`, `torch`, `transformers` |
| Platforms | linux, macos |
| Tags | `Optimization`, `Flash Attention`, `Attention Optimization`, `Memory Efficiency`, `Speed Optimization`, `Long Context`, `PyTorch`, `SDPA`, `H100`, `FP8`, `Transformers` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Flash Attention - Fast Memory-Efficient Attention
## Quick start
Flash Attention provides 2-4x speedup and 10-20x memory reduction for transformer attention through IO-aware tiling and recomputation.
**PyTorch native (easiest, PyTorch 2.2+)**:
```python
import torch
import torch.nn.functional as F
q = torch.randn(2, 8, 512, 64, device='cuda', dtype=torch.float16) # [batch, heads, seq, dim]
k = torch.randn(2, 8, 512, 64, device='cuda', dtype=torch.float16)
v = torch.randn(2, 8, 512, 64, device='cuda', dtype=torch.float16)
# Automatically uses Flash Attention if available
out = F.scaled_dot_product_attention(q, k, v)
```
**flash-attn library (more features)**:
```bash
pip install flash-attn --no-build-isolation
```
```python
from flash_attn import flash_attn_func
# q, k, v: [batch, seqlen, nheads, headdim]
out = flash_attn_func(q, k, v, dropout_p=0.0, causal=True)
```
## Common workflows
### Workflow 1: Enable in existing PyTorch model
Copy this checklist:
```
Flash Attention Integration:
- [ ] Step 1: Check PyTorch version (≥2.2)
- [ ] Step 2: Enable Flash Attention backend
- [ ] Step 3: Verify speedup with profiling
- [ ] Step 4: Test accuracy matches baseline
```
**Step 1: Check PyTorch version**
```bash
python -c "import torch; print(torch.__version__)"
# Should be ≥2.2.0
```
If &lt;2.2, upgrade:
```bash
pip install --upgrade torch
```
**Step 2: Enable Flash Attention backend**
Replace standard attention:
```python
# Before (standard attention)
attn_weights = torch.softmax(q @ k.transpose(-2, -1) / math.sqrt(d_k), dim=-1)
out = attn_weights @ v
# After (Flash Attention)
import torch.nn.functional as F
out = F.scaled_dot_product_attention(q, k, v, attn_mask=mask)
```
Force Flash Attention backend (`torch.backends.cuda.sdp_kernel` is deprecated; use
`torch.nn.attention.sdpa_kernel` with `SDPBackend`):
```python
from torch.nn.attention import SDPBackend, sdpa_kernel
with sdpa_kernel(SDPBackend.FLASH_ATTENTION):
out = F.scaled_dot_product_attention(q, k, v)
```
**Step 3: Verify speedup with profiling**
```python
import torch.utils.benchmark as benchmark
def test_attention(use_flash):
q, k, v = [torch.randn(2, 8, 2048, 64, device='cuda', dtype=torch.float16) for _ in range(3)]
if use_flash:
from torch.nn.attention import SDPBackend, sdpa_kernel
with sdpa_kernel(SDPBackend.FLASH_ATTENTION):
return F.scaled_dot_product_attention(q, k, v)
else:
attn = (q @ k.transpose(-2, -1) / 8.0).softmax(dim=-1)
return attn @ v
# Benchmark
t_flash = benchmark.Timer(stmt='test_attention(True)', globals=globals())
t_standard = benchmark.Timer(stmt='test_attention(False)', globals=globals())
print(f"Flash: {t_flash.timeit(100).mean:.3f}s")
print(f"Standard: {t_standard.timeit(100).mean:.3f}s")
```
Expected: 2-4x speedup for sequences >512 tokens.
**Step 4: Test accuracy matches baseline**
```python
# Compare outputs
q, k, v = [torch.randn(1, 8, 512, 64, device='cuda', dtype=torch.float16) for _ in range(3)]
# Flash Attention
out_flash = F.scaled_dot_product_attention(q, k, v)
# Standard attention
attn_weights = torch.softmax(q @ k.transpose(-2, -1) / 8.0, dim=-1)
out_standard = attn_weights @ v
# Check difference
diff = (out_flash - out_standard).abs().max()
print(f"Max difference: {diff:.6f}")
# Should be <1e-3 for float16
```
### Workflow 2: Use flash-attn library for advanced features
For multi-query attention, sliding window, or H100 FP8.
Copy this checklist:
```
flash-attn Library Setup:
- [ ] Step 1: Install flash-attn library
- [ ] Step 2: Modify attention code
- [ ] Step 3: Enable advanced features
- [ ] Step 4: Benchmark performance
```
**Step 1: Install flash-attn library**
```bash
# NVIDIA GPUs (CUDA 12.0+)
pip install flash-attn --no-build-isolation
# Verify installation
python -c "from flash_attn import flash_attn_func; print('Success')"
```
**Step 2: Modify attention code**
```python
from flash_attn import flash_attn_func
# Input: [batch_size, seq_len, num_heads, head_dim]
# Transpose from [batch, heads, seq, dim] if needed
q = q.transpose(1, 2) # [batch, seq, heads, dim]
k = k.transpose(1, 2)
v = v.transpose(1, 2)
out = flash_attn_func(
q, k, v,
dropout_p=0.1,
causal=True, # For autoregressive models
window_size=(-1, -1), # No sliding window
softmax_scale=None # Auto-scale
)
out = out.transpose(1, 2) # Back to [batch, heads, seq, dim]
```
**Step 3: Enable advanced features**
Multi-query attention (shared K/V across heads):
```python
from flash_attn import flash_attn_func
# q: [batch, seq, num_q_heads, dim]
# k, v: [batch, seq, num_kv_heads, dim] # Fewer KV heads
out = flash_attn_func(q, k, v) # Automatically handles MQA
```
Sliding window attention (local attention):
```python
# Only attend to window of 256 tokens before/after
out = flash_attn_func(
q, k, v,
window_size=(256, 256), # (left, right) window
causal=True
)
```
**Step 4: Benchmark performance**
```python
import torch
from flash_attn import flash_attn_func
import time
q, k, v = [torch.randn(4, 4096, 32, 64, device='cuda', dtype=torch.float16) for _ in range(3)]
# Warmup
for _ in range(10):
_ = flash_attn_func(q, k, v)
# Benchmark
torch.cuda.synchronize()
start = time.time()
for _ in range(100):
out = flash_attn_func(q, k, v)
torch.cuda.synchronize()
end = time.time()
print(f"Time per iteration: {(end-start)/100*1000:.2f}ms")
print(f"Memory allocated: {torch.cuda.max_memory_allocated()/1e9:.2f}GB")
```
### Workflow 3: H100 FP8 optimization (FlashAttention-3)
For maximum performance on Hopper GPUs (H100).
> **Important:** The pip package `flash-attn` (2.8.x) ships **FlashAttention-2 only** — it does
> **not** contain FA3 or FP8 H100 kernels, and `flash_attn_func` does **not** auto-use FP8.
> FlashAttention-3 is a separate **beta** build compiled from source from the repo's `hopper/`
> directory, exposed via the `flash_attn_interface` module. FA3 supports FP16/BF16 forward+backward
> and **FP8 forward only**.
```
FP8 Setup:
- [ ] Step 1: Verify Hopper (H100) GPU available
- [ ] Step 2: Build & install FlashAttention-3 from source (hopper/)
- [ ] Step 3: Use the FA3 interface (FP8 forward)
```
**Step 1: Verify H100 GPU**
```bash
nvidia-smi --query-gpu=name --format=csv
# Should show "H100" or "H800"
```
**Step 2: Build & install FlashAttention-3 from source**
FA3 is NOT included in `pip install flash-attn`. Build it from the `hopper/` subdirectory:
```bash
git clone https://github.com/Dao-AILab/flash-attention.git
cd flash-attention/hopper
python setup.py install
# (compilation is heavy and requires a CUDA toolchain + Hopper GPU)
```
**Step 3: Use the FA3 interface (FP8 forward)**
FA3 exposes its own module `flash_attn_interface` (distinct from the FA2 `flash_attn`).
FP8 is a **forward-only** path and expects `float8_e4m3fn` inputs:
```python
import torch
from flash_attn_interface import flash_attn_func # FA3 (hopper build), not `flash_attn`
# q, k, v: [batch, seqlen, nheads, headdim]
q = torch.randn(2, 4096, 32, 64, device='cuda', dtype=torch.float16)
k = torch.randn(2, 4096, 32, 64, device='cuda', dtype=torch.float16)
v = torch.randn(2, 4096, 32, 64, device='cuda', dtype=torch.float16)
# FP8 forward (inference / forward-only): cast to float8_e4m3fn
q_fp8 = q.to(torch.float8_e4m3fn)
k_fp8 = k.to(torch.float8_e4m3fn)
v_fp8 = v.to(torch.float8_e4m3fn)
out = flash_attn_func(q_fp8, k_fp8, v_fp8, causal=True)
# FP16/BF16 forward+backward is also supported by the FA3 interface.
```
## When to use vs alternatives
**Use Flash Attention when:**
- Training transformers with sequences >512 tokens
- Running inference with long context (>2K tokens)
- GPU memory constrained (OOM with standard attention)
- Need 2-4x speedup without accuracy loss
- Using PyTorch 2.2+ or can install flash-attn
**Use alternatives instead:**
- **Standard attention**: Sequences &lt;256 tokens (overhead not worth it)
- **xFormers**: Need more attention variants (not just speed)
- **Memory-efficient attention**: CPU inference (Flash Attention needs GPU)
## Common issues
**Issue: ImportError: cannot import flash_attn**
Install with no-build-isolation flag:
```bash
pip install flash-attn --no-build-isolation
```
Or install CUDA toolkit first:
```bash
conda install cuda -c nvidia
pip install flash-attn --no-build-isolation
```
**Issue: Slower than expected (no speedup)**
Flash Attention benefits increase with sequence length:
- &lt;512 tokens: Minimal speedup (10-20%)
- 512-2K tokens: 2-3x speedup
- >2K tokens: 3-4x speedup
Check sequence length is sufficient.
**Issue: RuntimeError: CUDA error**
Verify GPU supports Flash Attention:
```python
import torch
print(torch.cuda.get_device_capability())
# Should be ≥(7, 5) for Turing+
```
Flash Attention requires:
- Ampere (A100, A10): ✅ Full support
- Turing (T4): ✅ Supported
- Volta (V100): ❌ Not supported
**Issue: Accuracy degradation**
Check dtype is float16 or bfloat16 (not float32):
```python
q = q.to(torch.float16) # Or torch.bfloat16
```
Flash Attention uses float16/bfloat16 for speed. Float32 not supported.
## Advanced topics
**Integration with HuggingFace Transformers**: See [references/transformers-integration.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\flash-attention/references/transformers-integration.md) for enabling Flash Attention in BERT, GPT, Llama models.
**Performance benchmarks**: See [references/benchmarks.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\flash-attention/references/benchmarks.md) for detailed speed and memory comparisons across GPUs and sequence lengths.
## Hardware requirements
- **GPU**: NVIDIA Ampere+ (A100, A10, A30) or AMD MI200+
- **VRAM**: Same as standard attention (Flash Attention doesn't increase memory)
- **CUDA**: 12.0+ (11.8 minimum)
- **PyTorch**: 2.2+ for native support
**Not supported**: V100 (Volta), CPU inference
## Resources
- Paper: "FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness" (NeurIPS 2022)
- Paper: "FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning" (ICLR 2024)
- Blog: https://tridao.me/blog/2024/flash3/
- GitHub: https://github.com/Dao-AILab/flash-attention
- PyTorch docs: https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html
website/docs/user-guide/skills/optional/mlops/mlops-guidance.md
+595
View File
@@ -0,0 +1,595 @@
---
title: "Guidance — Constrain LLM output with grammars; guarantee valid JSON"
sidebar_label: "Guidance"
description: "Constrain LLM output with grammars; guarantee valid JSON"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Guidance
Constrain LLM output with grammars; guarantee valid JSON.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/guidance` |
| Path | `optional-skills/mlops\guidance` |
| Version | `1.0.1` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `guidance`, `transformers` |
| Platforms | linux, macos, windows |
| Tags | `Prompt Engineering`, `Guidance`, `Constrained Generation`, `Structured Output`, `JSON Validation`, `Grammar`, `Microsoft Research`, `Format Enforcement`, `Multi-Step Workflows` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Guidance: Constrained LLM Generation
## When to Use This Skill
Use Guidance when you need to:
- **Control LLM output syntax** with regex or grammars
- **Guarantee valid JSON/XML/code** generation
- **Reduce latency** vs traditional prompting approaches
- **Enforce structured formats** (dates, emails, IDs, etc.)
- **Build multi-step workflows** with Pythonic control flow
- **Prevent invalid outputs** through grammatical constraints
**GitHub Stars**: 18,000+ | **From**: Microsoft Research
## Installation
```bash
# Base installation
pip install guidance
# With specific backends
pip install guidance[transformers] # Hugging Face models
pip install guidance[llama_cpp] # llama.cpp models
```
## Quick Start
### Basic Example: Structured Generation
```python
from guidance import models, gen
# Load model (supports OpenAI, Transformers, llama.cpp)
lm = models.OpenAI("gpt-4")
# Generate with constraints
result = lm + "The capital of France is " + gen("capital", max_tokens=5)
print(result["capital"]) # "Paris"
```
### Chat format with a local model
> **Constraint support requires local logit access.** Regex, `select()`, and
> grammar-based constrained generation only work with local backends
> (`Transformers`, `LlamaCpp`). Remote API backends (`OpenAI`, and Azure
> variants) support unconstrained `gen()` / chat only — they cannot enforce
> token-level constraints. guidance 0.3.x has no `models.Anthropic` class.
```python
from guidance import models, gen, system, user, assistant
# Local model (supports constrained generation)
lm = models.Transformers("microsoft/Phi-4-mini-instruct")
# Use context managers for chat format
with system():
lm += "You are a helpful assistant."
with user():
lm += "What is the capital of France?"
with assistant():
lm += gen(max_tokens=20)
```
## Core Concepts
### 1. Context Managers
Guidance uses Pythonic context managers for chat-style interactions.
```python
from guidance import system, user, assistant, gen
lm = models.Transformers("microsoft/Phi-4-mini-instruct")
# System message
with system():
lm += "You are a JSON generation expert."
# User message
with user():
lm += "Generate a person object with name and age."
# Assistant response
with assistant():
lm += gen("response", max_tokens=100)
print(lm["response"])
```
**Benefits:**
- Natural chat flow
- Clear role separation
- Easy to read and maintain
### 2. Constrained Generation
Guidance ensures outputs match specified patterns using regex or grammars.
#### Regex Constraints
```python
from guidance import models, gen
lm = models.Transformers("microsoft/Phi-4-mini-instruct")
# Constrain to valid email format
lm += "Email: " + gen("email", regex=r"[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\.[a-zA-Z]{2,}")
# Constrain to date format (YYYY-MM-DD)
lm += "Date: " + gen("date", regex=r"\d{4}-\d{2}-\d{2}")
# Constrain to phone number
lm += "Phone: " + gen("phone", regex=r"\d{3}-\d{3}-\d{4}")
print(lm["email"]) # Guaranteed valid email
print(lm["date"]) # Guaranteed YYYY-MM-DD format
```
**How it works:**
- Regex converted to grammar at token level
- Invalid tokens filtered during generation
- Model can only produce matching outputs
#### Selection Constraints
```python
from guidance import models, gen, select
lm = models.Transformers("microsoft/Phi-4-mini-instruct")
# Constrain to specific choices
lm += "Sentiment: " + select(["positive", "negative", "neutral"], name="sentiment")
# Multiple-choice selection
lm += "Best answer: " + select(
["A) Paris", "B) London", "C) Berlin", "D) Madrid"],
name="answer"
)
print(lm["sentiment"]) # One of: positive, negative, neutral
print(lm["answer"]) # One of: A, B, C, or D
```
### 3. Token Healing
Guidance automatically "heals" token boundaries between prompt and generation.
**Problem:** Tokenization creates unnatural boundaries.
```python
# Without token healing
prompt = "The capital of France is "
# Last token: " is "
# First generated token might be " Par" (with leading space)
# Result: "The capital of France is Paris" (double space!)
```
**Solution:** Guidance backs up one token and regenerates.
```python
from guidance import models, gen
lm = models.Transformers("microsoft/Phi-4-mini-instruct")
# Token healing enabled by default
lm += "The capital of France is " + gen("capital", max_tokens=5)
# Result: "The capital of France is Paris" (correct spacing)
```
**Benefits:**
- Natural text boundaries
- No awkward spacing issues
- Better model performance (sees natural token sequences)
### 4. Grammar-Based Generation
Define complex structures by composing grammar functions. The template-string
`grammar=` form is not part of current guidance — build grammars from
composable functions, or use `guidance.json()` for JSON.
```python
from guidance import models, gen
from guidance import json as gen_json
from pydantic import BaseModel, Field
lm = models.Transformers("microsoft/Phi-4-mini-instruct")
# JSON via a Pydantic schema (guidance.json compiles the schema to a grammar)
class Person(BaseModel):
name: str = Field(pattern=r"[A-Za-z ]+")
age: int
email: str = Field(pattern=r"[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\.[a-zA-Z]{2,}")
lm += gen_json(name="person", schema=Person)
print(lm["person"]) # Guaranteed valid JSON matching the schema
# Or compose grammar functions directly:
grammar = "name=" + gen("name", regex=r"[A-Za-z ]+") + " age=" + gen("age", regex=r"[0-9]+")
lm += grammar
```
**Use cases:**
- Complex structured outputs
- Nested data structures
- Programming language syntax
- Domain-specific languages
### 5. Guidance Functions
Create reusable generation patterns with the `@guidance` decorator.
```python
from guidance import guidance, gen, models
@guidance
def generate_person(lm):
"""Generate a person with name and age."""
lm += "Name: " + gen("name", max_tokens=20, stop="\n")
lm += "\nAge: " + gen("age", regex=r"[0-9]+", max_tokens=3)
return lm
# Use the function
lm = models.Transformers("microsoft/Phi-4-mini-instruct")
lm = generate_person(lm)
print(lm["name"])
print(lm["age"])
```
**Stateful Functions:**
```python
@guidance(stateless=False)
def react_agent(lm, question, tools, max_rounds=5):
"""ReAct agent with tool use."""
lm += f"Question: {question}\n\n"
for i in range(max_rounds):
# Thought
lm += f"Thought {i+1}: " + gen("thought", stop="\n")
# Action
lm += "\nAction: " + select(list(tools.keys()), name="action")
# Execute tool
tool_result = tools[lm["action"]]()
lm += f"\nObservation: {tool_result}\n\n"
# Check if done
lm += "Done? " + select(["Yes", "No"], name="done")
if lm["done"] == "Yes":
break
# Final answer
lm += "\nFinal Answer: " + gen("answer", max_tokens=100)
return lm
```
## Backend Configuration
### OpenAI (remote — unconstrained only)
> Remote API backends cannot do constrained generation (regex/select/grammar);
> use them only for plain chat/`gen()`. For constraints, use a local backend.
```python
from guidance import models
lm = models.OpenAI(
model="gpt-4o-mini",
api_key="your-api-key" # Or set OPENAI_API_KEY env var
)
```
### Local Models (Transformers)
```python
from guidance.models import Transformers
lm = Transformers(
"microsoft/Phi-4-mini-instruct",
device="cuda" # Or "cpu"
)
```
### Local Models (llama.cpp)
```python
from guidance.models import LlamaCpp
lm = LlamaCpp(
model_path="/path/to/model.gguf",
n_ctx=4096,
n_gpu_layers=35
)
```
## Common Patterns
### Pattern 1: JSON Generation
```python
from guidance import models, gen, system, user, assistant
lm = models.Transformers("microsoft/Phi-4-mini-instruct")
with system():
lm += "You generate valid JSON."
with user():
lm += "Generate a user profile with name, age, and email."
with assistant():
lm += """{
"name": """ + gen("name", regex=r'"[A-Za-z ]+"', max_tokens=30) + """,
"age": """ + gen("age", regex=r"[0-9]+", max_tokens=3) + """,
"email": """ + gen("email", regex=r'"[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\.[a-zA-Z]{2,}"', max_tokens=50) + """
}"""
print(lm) # Valid JSON guaranteed
```
### Pattern 2: Classification
```python
from guidance import models, gen, select
lm = models.Transformers("microsoft/Phi-4-mini-instruct")
text = "This product is amazing! I love it."
lm += f"Text: {text}\n"
lm += "Sentiment: " + select(["positive", "negative", "neutral"], name="sentiment")
lm += "\nConfidence: " + gen("confidence", regex=r"[0-9]+", max_tokens=3) + "%"
print(f"Sentiment: {lm['sentiment']}")
print(f"Confidence: {lm['confidence']}%")
```
### Pattern 3: Multi-Step Reasoning
```python
from guidance import models, gen, guidance
@guidance
def chain_of_thought(lm, question):
"""Generate answer with step-by-step reasoning."""
lm += f"Question: {question}\n\n"
# Generate multiple reasoning steps
for i in range(3):
lm += f"Step {i+1}: " + gen(f"step_{i+1}", stop="\n", max_tokens=100) + "\n"
# Final answer
lm += "\nTherefore, the answer is: " + gen("answer", max_tokens=50)
return lm
lm = models.Transformers("microsoft/Phi-4-mini-instruct")
lm = chain_of_thought(lm, "What is 15% of 200?")
print(lm["answer"])
```
### Pattern 4: ReAct Agent
```python
from guidance import models, gen, select, guidance
@guidance(stateless=False)
def react_agent(lm, question):
"""ReAct agent with tool use."""
tools = {
"calculator": lambda expr: eval(expr),
"search": lambda query: f"Search results for: {query}",
}
lm += f"Question: {question}\n\n"
for round in range(5):
# Thought
lm += f"Thought: " + gen("thought", stop="\n") + "\n"
# Action selection
lm += "Action: " + select(["calculator", "search", "answer"], name="action")
if lm["action"] == "answer":
lm += "\nFinal Answer: " + gen("answer", max_tokens=100)
break
# Action input
lm += "\nAction Input: " + gen("action_input", stop="\n") + "\n"
# Execute tool
if lm["action"] in tools:
result = tools[lm["action"]](lm["action_input"])
lm += f"Observation: {result}\n\n"
return lm
lm = models.Transformers("microsoft/Phi-4-mini-instruct")
lm = react_agent(lm, "What is 25 * 4 + 10?")
print(lm["answer"])
```
### Pattern 5: Data Extraction
```python
from guidance import models, gen, guidance
@guidance
def extract_entities(lm, text):
"""Extract structured entities from text."""
lm += f"Text: {text}\n\n"
# Extract person
lm += "Person: " + gen("person", stop="\n", max_tokens=30) + "\n"
# Extract organization
lm += "Organization: " + gen("organization", stop="\n", max_tokens=30) + "\n"
# Extract date
lm += "Date: " + gen("date", regex=r"\d{4}-\d{2}-\d{2}", max_tokens=10) + "\n"
# Extract location
lm += "Location: " + gen("location", stop="\n", max_tokens=30) + "\n"
return lm
text = "Tim Cook announced at Apple Park on 2024-09-15 in Cupertino."
lm = models.Transformers("microsoft/Phi-4-mini-instruct")
lm = extract_entities(lm, text)
print(f"Person: {lm['person']}")
print(f"Organization: {lm['organization']}")
print(f"Date: {lm['date']}")
print(f"Location: {lm['location']}")
```
## Best Practices
### 1. Use Regex for Format Validation
```python
# ✅ Good: Regex ensures valid format
lm += "Email: " + gen("email", regex=r"[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\.[a-zA-Z]{2,}")
# ❌ Bad: Free generation may produce invalid emails
lm += "Email: " + gen("email", max_tokens=50)
```
### 2. Use select() for Fixed Categories
```python
# ✅ Good: Guaranteed valid category
lm += "Status: " + select(["pending", "approved", "rejected"], name="status")
# ❌ Bad: May generate typos or invalid values
lm += "Status: " + gen("status", max_tokens=20)
```
### 3. Leverage Token Healing
```python
# Token healing is enabled by default
# No special action needed - just concatenate naturally
lm += "The capital is " + gen("capital") # Automatic healing
```
### 4. Use stop Sequences
```python
# ✅ Good: Stop at newline for single-line outputs
lm += "Name: " + gen("name", stop="\n")
# ❌ Bad: May generate multiple lines
lm += "Name: " + gen("name", max_tokens=50)
```
### 5. Create Reusable Functions
```python
# ✅ Good: Reusable pattern
@guidance
def generate_person(lm):
lm += "Name: " + gen("name", stop="\n")
lm += "\nAge: " + gen("age", regex=r"[0-9]+")
return lm
# Use multiple times
lm = generate_person(lm)
lm += "\n\n"
lm = generate_person(lm)
```
### 6. Balance Constraints
```python
# ✅ Good: Reasonable constraints
lm += gen("name", regex=r"[A-Za-z ]+", max_tokens=30)
# ❌ Too strict: May fail or be very slow
lm += gen("name", regex=r"^(John|Jane)$", max_tokens=10)
```
## Comparison to Alternatives
| Feature | Guidance | Instructor | Outlines | LMQL |
|---------|----------|------------|----------|------|
| Regex Constraints | ✅ Yes | ❌ No | ✅ Yes | ✅ Yes |
| Grammar Support | ✅ CFG | ❌ No | ✅ CFG | ✅ CFG |
| Pydantic Validation | ❌ No | ✅ Yes | ✅ Yes | ❌ No |
| Token Healing | ✅ Yes | ❌ No | ✅ Yes | ❌ No |
| Local Models | ✅ Yes | ⚠️ Limited | ✅ Yes | ✅ Yes |
| API Models | ✅ Yes | ✅ Yes | ⚠️ Limited | ✅ Yes |
| Pythonic Syntax | ✅ Yes | ✅ Yes | ✅ Yes | ❌ SQL-like |
| Learning Curve | Low | Low | Medium | High |
**When to choose Guidance:**
- Need regex/grammar constraints
- Want token healing
- Building complex workflows with control flow
- Using local models (Transformers, llama.cpp)
- Prefer Pythonic syntax
**When to choose alternatives:**
- Instructor: Need Pydantic validation with automatic retrying
- Outlines: Need JSON schema validation
- LMQL: Prefer declarative query syntax
## Performance Characteristics
**Latency Reduction:**
- 30-50% faster than traditional prompting for constrained outputs
- Token healing reduces unnecessary regeneration
- Grammar constraints prevent invalid token generation
**Memory Usage:**
- Minimal overhead vs unconstrained generation
- Grammar compilation cached after first use
- Efficient token filtering at inference time
**Token Efficiency:**
- Prevents wasted tokens on invalid outputs
- No need for retry loops
- Direct path to valid outputs
## Resources
- **Documentation**: https://guidance.readthedocs.io
- **GitHub**: https://github.com/guidance-ai/guidance (18k+ stars)
- **Notebooks**: https://github.com/guidance-ai/guidance/tree/main/notebooks
- **Discord**: Community support available
## See Also
- `references/constraints.md` - Comprehensive regex and grammar patterns
- `references/backends.md` - Backend-specific configuration
- `references/examples.md` - Production-ready examples
website/docs/user-guide/skills/optional/mlops/mlops-huggingface-tokenizers.md
+535
View File
@@ -0,0 +1,535 @@
---
title: "Huggingface Tokenizers — Fast BPE/WordPiece tokenization and custom vocab training"
sidebar_label: "Huggingface Tokenizers"
description: "Fast BPE/WordPiece tokenization and custom vocab training"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Huggingface Tokenizers
Fast BPE/WordPiece tokenization and custom vocab training.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/huggingface-tokenizers` |
| Path | `optional-skills/mlops\huggingface-tokenizers` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `tokenizers`, `transformers`, `datasets` |
| Platforms | linux, macos, windows |
| Tags | `Tokenization`, `HuggingFace`, `BPE`, `WordPiece`, `Unigram`, `Fast Tokenization`, `Rust`, `Custom Tokenizer`, `Alignment Tracking`, `Production` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# HuggingFace Tokenizers - Fast Tokenization for NLP
Fast, production-ready tokenizers with Rust performance and Python ease-of-use.
## When to use HuggingFace Tokenizers
**Use HuggingFace Tokenizers when:**
- Need extremely fast tokenization (&lt;20s per GB of text)
- Training custom tokenizers from scratch
- Want alignment tracking (token → original text position)
- Building production NLP pipelines
- Need to tokenize large corpora efficiently
**Performance**:
- **Speed**: &lt;20 seconds to tokenize 1GB on CPU
- **Implementation**: Rust core with Python/Node.js bindings
- **Efficiency**: 10-100× faster than pure Python implementations
**Use alternatives instead**:
- **SentencePiece**: Language-independent, used by T5/ALBERT
- **tiktoken**: OpenAI's BPE tokenizer for GPT models
- **transformers AutoTokenizer**: Loading pretrained only (uses this library internally)
## Quick start
### Installation
```bash
# Install tokenizers
pip install tokenizers
# With transformers integration
pip install tokenizers transformers
```
### Load pretrained tokenizer
```python
from tokenizers import Tokenizer
# Load from HuggingFace Hub
tokenizer = Tokenizer.from_pretrained("bert-base-uncased")
# Encode text
output = tokenizer.encode("Hello, how are you?")
print(output.tokens) # ['hello', ',', 'how', 'are', 'you', '?']
print(output.ids) # [7592, 1010, 2129, 2024, 2017, 1029]
# Decode back
text = tokenizer.decode(output.ids)
print(text) # "hello, how are you?"
```
### Train custom BPE tokenizer
```python
from tokenizers import Tokenizer
from tokenizers.models import BPE
from tokenizers.trainers import BpeTrainer
from tokenizers.pre_tokenizers import Whitespace
# Initialize tokenizer with BPE model
tokenizer = Tokenizer(BPE(unk_token="[UNK]"))
tokenizer.pre_tokenizer = Whitespace()
# Configure trainer
trainer = BpeTrainer(
vocab_size=30000,
special_tokens=["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"],
min_frequency=2
)
# Train on files
files = ["train.txt", "validation.txt"]
tokenizer.train(files, trainer)
# Save
tokenizer.save("my-tokenizer.json")
```
**Training time**: ~1-2 minutes for 100MB corpus, ~10-20 minutes for 1GB
### Batch encoding with padding
```python
# Enable padding
tokenizer.enable_padding(pad_id=3, pad_token="[PAD]")
# Encode batch
texts = ["Hello world", "This is a longer sentence"]
encodings = tokenizer.encode_batch(texts)
for encoding in encodings:
print(encoding.ids)
# [101, 7592, 2088, 102, 3, 3, 3]
# [101, 2023, 2003, 1037, 2936, 6251, 102]
```
## Tokenization algorithms
### BPE (Byte-Pair Encoding)
**How it works**:
1. Start with character-level vocabulary
2. Find most frequent character pair
3. Merge into new token, add to vocabulary
4. Repeat until vocabulary size reached
**Used by**: GPT-2, GPT-3, RoBERTa, BART, DeBERTa
```python
from tokenizers import Tokenizer
from tokenizers.models import BPE
from tokenizers.trainers import BpeTrainer
from tokenizers.pre_tokenizers import ByteLevel
tokenizer = Tokenizer(BPE(unk_token="<|endoftext|>"))
tokenizer.pre_tokenizer = ByteLevel()
trainer = BpeTrainer(
vocab_size=50257,
special_tokens=["<|endoftext|>"],
min_frequency=2
)
tokenizer.train(files=["data.txt"], trainer=trainer)
```
**Advantages**:
- Handles OOV words well (breaks into subwords)
- Flexible vocabulary size
- Good for morphologically rich languages
**Trade-offs**:
- Tokenization depends on merge order
- May split common words unexpectedly
### WordPiece
**How it works**:
1. Start with character vocabulary
2. Score merge pairs: `frequency(pair) / (frequency(first) × frequency(second))`
3. Merge highest scoring pair
4. Repeat until vocabulary size reached
**Used by**: BERT, DistilBERT, MobileBERT
```python
from tokenizers import Tokenizer
from tokenizers.models import WordPiece
from tokenizers.trainers import WordPieceTrainer
from tokenizers.pre_tokenizers import Whitespace
from tokenizers.normalizers import BertNormalizer
tokenizer = Tokenizer(WordPiece(unk_token="[UNK]"))
tokenizer.normalizer = BertNormalizer(lowercase=True)
tokenizer.pre_tokenizer = Whitespace()
trainer = WordPieceTrainer(
vocab_size=30522,
special_tokens=["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"],
continuing_subword_prefix="##"
)
tokenizer.train(files=["corpus.txt"], trainer=trainer)
```
**Advantages**:
- Prioritizes meaningful merges (high score = semantically related)
- Used successfully in BERT (state-of-the-art results)
**Trade-offs**:
- Unknown words become `[UNK]` if no subword match
- Saves vocabulary, not merge rules (larger files)
### Unigram
**How it works**:
1. Start with large vocabulary (all substrings)
2. Compute loss for corpus with current vocabulary
3. Remove tokens with minimal impact on loss
4. Repeat until vocabulary size reached
**Used by**: ALBERT, T5, mBART, XLNet (via SentencePiece)
```python
from tokenizers import Tokenizer
from tokenizers.models import Unigram
from tokenizers.trainers import UnigramTrainer
tokenizer = Tokenizer(Unigram())
trainer = UnigramTrainer(
vocab_size=8000,
special_tokens=["<unk>", "<s>", "</s>"],
unk_token="<unk>"
)
tokenizer.train(files=["data.txt"], trainer=trainer)
```
**Advantages**:
- Probabilistic (finds most likely tokenization)
- Works well for languages without word boundaries
- Handles diverse linguistic contexts
**Trade-offs**:
- Computationally expensive to train
- More hyperparameters to tune
## Tokenization pipeline
Complete pipeline: **Normalization → Pre-tokenization → Model → Post-processing**
### Normalization
Clean and standardize text:
```python
from tokenizers.normalizers import NFD, StripAccents, Lowercase, Sequence
tokenizer.normalizer = Sequence([
NFD(), # Unicode normalization (decompose)
Lowercase(), # Convert to lowercase
StripAccents() # Remove accents
])
# Input: "Héllo WORLD"
# After normalization: "hello world"
```
**Common normalizers**:
- `NFD`, `NFC`, `NFKD`, `NFKC` - Unicode normalization forms
- `Lowercase()` - Convert to lowercase
- `StripAccents()` - Remove accents (é → e)
- `Strip()` - Remove whitespace
- `Replace(pattern, content)` - Regex replacement
### Pre-tokenization
Split text into word-like units:
```python
from tokenizers.pre_tokenizers import Whitespace, Punctuation, Sequence, ByteLevel
# Split on whitespace and punctuation
tokenizer.pre_tokenizer = Sequence([
Whitespace(),
Punctuation()
])
# Input: "Hello, world!"
# After pre-tokenization: ["Hello", ",", "world", "!"]
```
**Common pre-tokenizers**:
- `Whitespace()` - Split on spaces, tabs, newlines
- `ByteLevel()` - GPT-2 style byte-level splitting
- `Punctuation()` - Isolate punctuation
- `Digits(individual_digits=True)` - Split digits individually
- `Metaspace()` - Replace spaces with ▁ (SentencePiece style)
### Post-processing
Add special tokens for model input:
```python
from tokenizers.processors import TemplateProcessing
# BERT-style: [CLS] sentence [SEP]
tokenizer.post_processor = TemplateProcessing(
single="[CLS] $A [SEP]",
pair="[CLS] $A [SEP] $B [SEP]",
special_tokens=[
("[CLS]", 1),
("[SEP]", 2),
],
)
```
**Common patterns**:
```python
# GPT-2: sentence <|endoftext|>
TemplateProcessing(
single="$A <|endoftext|>",
special_tokens=[("<|endoftext|>", 50256)]
)
# RoBERTa: <s> sentence </s>
TemplateProcessing(
single="<s> $A </s>",
pair="<s> $A </s> </s> $B </s>",
special_tokens=[("<s>", 0), ("</s>", 2)]
)
```
## Alignment tracking
Track token positions in original text:
```python
output = tokenizer.encode("Hello, world!")
# Get token offsets
for token, offset in zip(output.tokens, output.offsets):
start, end = offset
print(f"{token:10} → [{start:2}, {end:2}): {text[start:end]!r}")
# Output:
# hello → [ 0, 5): 'Hello'
# , → [ 5, 6): ','
# world → [ 7, 12): 'world'
# ! → [12, 13): '!'
```
**Use cases**:
- Named entity recognition (map predictions back to text)
- Question answering (extract answer spans)
- Token classification (align labels to original positions)
## Integration with transformers
### Load with AutoTokenizer
```python
from transformers import AutoTokenizer
# AutoTokenizer automatically uses fast tokenizers
tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
# Check if using fast tokenizer
print(tokenizer.is_fast) # True
# Access underlying tokenizers.Tokenizer
fast_tokenizer = tokenizer.backend_tokenizer
print(type(fast_tokenizer)) # <class 'tokenizers.Tokenizer'>
```
### Convert custom tokenizer to transformers
```python
from tokenizers import Tokenizer
from transformers import PreTrainedTokenizerFast
# Train custom tokenizer
tokenizer = Tokenizer(BPE())
# ... train tokenizer ...
tokenizer.save("my-tokenizer.json")
# Wrap for transformers
transformers_tokenizer = PreTrainedTokenizerFast(
tokenizer_file="my-tokenizer.json",
unk_token="[UNK]",
pad_token="[PAD]",
cls_token="[CLS]",
sep_token="[SEP]",
mask_token="[MASK]"
)
# Use like any transformers tokenizer
outputs = transformers_tokenizer(
"Hello world",
padding=True,
truncation=True,
max_length=512,
return_tensors="pt"
)
```
## Common patterns
### Train from iterator (large datasets)
```python
from datasets import load_dataset
# Load dataset
dataset = load_dataset("wikitext", "wikitext-103-raw-v1", split="train")
# Create batch iterator
def batch_iterator(batch_size=1000):
for i in range(0, len(dataset), batch_size):
yield dataset[i:i + batch_size]["text"]
# Train tokenizer
tokenizer.train_from_iterator(
batch_iterator(),
trainer=trainer,
length=len(dataset) # For progress bar
)
```
**Performance**: Processes 1GB in ~10-20 minutes
### Enable truncation and padding
```python
# Enable truncation
tokenizer.enable_truncation(max_length=512)
# Enable padding
tokenizer.enable_padding(
pad_id=tokenizer.token_to_id("[PAD]"),
pad_token="[PAD]",
length=512 # Fixed length, or None for batch max
)
# Encode with both
output = tokenizer.encode("This is a long sentence that will be truncated...")
print(len(output.ids)) # 512
```
### Multi-processing
```python
from tokenizers import Tokenizer
from multiprocessing import Pool
# Load tokenizer
tokenizer = Tokenizer.from_file("tokenizer.json")
def encode_batch(texts):
return tokenizer.encode_batch(texts)
# Process large corpus in parallel
with Pool(8) as pool:
# Split corpus into chunks
chunk_size = 1000
chunks = [corpus[i:i+chunk_size] for i in range(0, len(corpus), chunk_size)]
# Encode in parallel
results = pool.map(encode_batch, chunks)
```
**Speedup**: 5-8× with 8 cores
## Performance benchmarks
### Training speed
| Corpus Size | BPE (30k vocab) | WordPiece (30k) | Unigram (8k) |
|-------------|-----------------|-----------------|--------------|
| 10 MB | 15 sec | 18 sec | 25 sec |
| 100 MB | 1.5 min | 2 min | 4 min |
| 1 GB | 15 min | 20 min | 40 min |
**Hardware**: 16-core CPU, tested on English Wikipedia
### Tokenization speed
| Implementation | 1 GB corpus | Throughput |
|----------------|-------------|---------------|
| Pure Python | ~20 minutes | ~50 MB/min |
| HF Tokenizers | ~15 seconds | ~4 GB/min |
| **Speedup** | **80×** | **80×** |
**Test**: English text, average sentence length 20 words
### Memory usage
| Task | Memory |
|-------------------------|---------|
| Load tokenizer | ~10 MB |
| Train BPE (30k vocab) | ~200 MB |
| Encode 1M sentences | ~500 MB |
## Supported models
Pre-trained tokenizers available via `from_pretrained()`:
**BERT family**:
- `bert-base-uncased`, `bert-large-cased`
- `distilbert-base-uncased`
- `roberta-base`, `roberta-large`
**GPT family**:
- `gpt2`, `gpt2-medium`, `gpt2-large`
- `distilgpt2`
**T5 family**:
- `t5-small`, `t5-base`, `t5-large`
- `google/flan-t5-xxl`
**Other**:
- `facebook/bart-base`, `facebook/mbart-large-cc25`
- `albert-base-v2`, `albert-xlarge-v2`
- `xlm-roberta-base`, `xlm-roberta-large`
Browse all: https://huggingface.co/models?library=tokenizers
## References
- **[Training Guide](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\huggingface-tokenizers/references/training.md)** - Train custom tokenizers, configure trainers, handle large datasets
- **[Algorithms Deep Dive](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\huggingface-tokenizers/references/algorithms.md)** - BPE, WordPiece, Unigram explained in detail
- **[Pipeline Components](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\huggingface-tokenizers/references/pipeline.md)** - Normalizers, pre-tokenizers, post-processors, decoders
- **[Transformers Integration](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\huggingface-tokenizers/references/integration.md)** - AutoTokenizer, PreTrainedTokenizerFast, special tokens
## Resources
- **Docs**: https://huggingface.co/docs/tokenizers
- **GitHub**: https://github.com/huggingface/tokenizers ⭐ 9,000+
- **Version**: 0.20.0+
- **Course**: https://huggingface.co/learn/nlp-course/chapter6/1
- **Paper**: BPE (Sennrich et al., 2016), WordPiece (Schuster & Nakajima, 2012)
website/docs/user-guide/skills/optional/mlops/mlops-inference-llama-cpp.md
+267
View File
@@ -0,0 +1,267 @@
---
title: "Llama Cpp — llama.cpp local GGUF inference + HF Hub model discovery"
sidebar_label: "Llama Cpp"
description: "llama.cpp local GGUF inference + HF Hub model discovery"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Llama Cpp
llama.cpp local GGUF inference + HF Hub model discovery.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/llama-cpp` |
| Path | `optional-skills/mlops\inference\llama-cpp` |
| Version | `2.1.2` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `llama-cpp-python>=0.2.0` |
| Platforms | linux, macos, windows |
| Tags | `llama.cpp`, `GGUF`, `Quantization`, `Hugging Face Hub`, `CPU Inference`, `Apple Silicon`, `Edge Deployment`, `AMD GPUs`, `Intel GPUs`, `NVIDIA`, `URL-first` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# llama.cpp + GGUF
Use this skill for local GGUF inference, quant selection, or Hugging Face repo discovery for llama.cpp.
## When to use
- Run local models on CPU, Apple Silicon, CUDA, ROCm, or Intel GPUs
- Find the right GGUF for a specific Hugging Face repo
- Build a `llama-server` or `llama-cli` command from the Hub
- Search the Hub for models that already support llama.cpp
- Enumerate available `.gguf` files and sizes for a repo
- Decide between Q4/Q5/Q6/IQ variants for the user's RAM or VRAM
## Model Discovery workflow
Prefer URL workflows before asking for `hf`, Python, or custom scripts.
1. Search for candidate repos on the Hub:
- Base: `https://huggingface.co/models?apps=llama.cpp&sort=trending`
- Add `search=<term>` for a model family
- Add `num_parameters=min:0,max:24B` or similar when the user has size constraints
2. Open the repo with the llama.cpp local-app view:
- `https://huggingface.co/<repo>?local-app=llama.cpp`
3. Treat the local-app snippet as the source of truth when it is visible:
- copy the exact `llama-server` or `llama-cli` command
- report the recommended quant exactly as HF shows it
4. Read the same `?local-app=llama.cpp` URL as page text or HTML and extract the section under `Hardware compatibility`:
- prefer its exact quant labels and sizes over generic tables
- keep repo-specific labels such as `UD-Q4_K_M` or `IQ4_NL_XL`
- if that section is not visible in the fetched page source, say so and fall back to the tree API plus generic quant guidance
5. Query the tree API to confirm what actually exists:
- `https://huggingface.co/api/models/<repo>/tree/main?recursive=true`
- keep entries where `type` is `file` and `path` ends with `.gguf`
- use `path` and `size` as the source of truth for filenames and byte sizes
- separate quantized checkpoints from `mmproj-*.gguf` projector files and `BF16/` shard files
- use `https://huggingface.co/<repo>/tree/main` only as a human fallback
6. If the local-app snippet is not text-visible, reconstruct the command from the repo plus the chosen quant:
- shorthand quant selection: `llama-server -hf <repo>:<QUANT>`
- exact-file fallback: `llama-server --hf-repo <repo> --hf-file <filename.gguf>`
7. Only suggest conversion from Transformers weights if the repo does not already expose GGUF files.
## Quick start
### Install llama.cpp
```bash
# macOS / Linux (simplest)
brew install llama.cpp
```
```bash
winget install llama.cpp
```
```bash
git clone https://github.com/ggml-org/llama.cpp
cd llama.cpp
cmake -B build
cmake --build build --config Release
```
### Run directly from the Hugging Face Hub
```bash
llama-cli -hf bartowski/Llama-3.2-3B-Instruct-GGUF:Q8_0
```
```bash
llama-server -hf bartowski/Llama-3.2-3B-Instruct-GGUF:Q8_0
```
### Run an exact GGUF file from the Hub
Use this when the tree API shows custom file naming or the exact HF snippet is missing.
```bash
llama-server \
--hf-repo microsoft/Phi-3-mini-4k-instruct-gguf \
--hf-file Phi-3-mini-4k-instruct-q4.gguf \
-c 4096
```
### OpenAI-compatible server check
```bash
curl http://localhost:8080/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"messages": [
{"role": "user", "content": "Write a limerick about Python exceptions"}
]
}'
```
## Python bindings (llama-cpp-python)
`pip install llama-cpp-python` (CUDA: `CMAKE_ARGS="-DGGML_CUDA=on" pip install llama-cpp-python --force-reinstall --no-cache-dir`; Metal: `CMAKE_ARGS="-DGGML_METAL=on" ...`).
### Basic generation
```python
from llama_cpp import Llama
llm = Llama(
model_path="./model-q4_k_m.gguf",
n_ctx=4096,
n_gpu_layers=35, # 0 for CPU, 99 to offload everything
n_threads=8,
)
out = llm("What is machine learning?", max_tokens=256, temperature=0.7)
print(out["choices"][0]["text"])
```
### Chat + streaming
```python
llm = Llama(
model_path="./model-q4_k_m.gguf",
n_ctx=4096,
n_gpu_layers=35,
chat_format="llama-3", # or "chatml", "mistral", etc.
)
resp = llm.create_chat_completion(
messages=[
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "What is Python?"},
],
max_tokens=256,
)
print(resp["choices"][0]["message"]["content"])
# Streaming
for chunk in llm("Explain quantum computing:", max_tokens=256, stream=True):
print(chunk["choices"][0]["text"], end="", flush=True)
```
### Embeddings
```python
llm = Llama(model_path="./model-q4_k_m.gguf", embedding=True, n_gpu_layers=35)
vec = llm.embed("This is a test sentence.")
print(f"Embedding dimension: {len(vec)}")
```
You can also load a GGUF straight from the Hub:
```python
llm = Llama.from_pretrained(
repo_id="bartowski/Llama-3.2-3B-Instruct-GGUF",
filename="*Q4_K_M.gguf",
n_gpu_layers=35,
)
```
## Choosing a quant
Use the Hub page first, generic heuristics second.
- Prefer the exact quant that HF marks as compatible for the user's hardware profile.
- For general chat, start with `Q4_K_M`.
- For code or technical work, prefer `Q5_K_M` or `Q6_K` if memory allows.
- For very tight RAM budgets, consider `Q3_K_M`, `IQ` variants, or `Q2` variants only if the user explicitly prioritizes fit over quality.
- For multimodal repos, mention `mmproj-*.gguf` separately. The projector is not the main model file.
- Do not normalize repo-native labels. If the page says `UD-Q4_K_M`, report `UD-Q4_K_M`.
## Extracting available GGUFs from a repo
When the user asks what GGUFs exist, return:
- filename
- file size
- quant label
- whether it is a main model or an auxiliary projector
Ignore unless requested:
- README
- BF16 shard files
- imatrix blobs or calibration artifacts
Use the tree API for this step:
- `https://huggingface.co/api/models/<repo>/tree/main?recursive=true`
For a repo like `unsloth/Qwen3.6-35B-A3B-GGUF`, the local-app page can show quant chips such as `UD-Q4_K_M`, `UD-Q5_K_M`, `UD-Q6_K`, and `Q8_0`, while the tree API exposes exact file paths such as `Qwen3.6-35B-A3B-UD-Q4_K_M.gguf` and `Qwen3.6-35B-A3B-Q8_0.gguf` with byte sizes. Use the tree API to turn a quant label into an exact filename.
## Search patterns
Use these URL shapes directly:
```text
https://huggingface.co/models?apps=llama.cpp&sort=trending
https://huggingface.co/models?search=<term>&apps=llama.cpp&sort=trending
https://huggingface.co/models?search=<term>&apps=llama.cpp&num_parameters=min:0,max:24B&sort=trending
https://huggingface.co/<repo>?local-app=llama.cpp
https://huggingface.co/api/models/<repo>/tree/main?recursive=true
https://huggingface.co/<repo>/tree/main
```
## Output format
When answering discovery requests, prefer a compact structured result like:
```text
Repo: <repo>
Recommended quant from HF: <label> (<size>)
llama-server: <command>
Other GGUFs:
- <filename> - <size>
- <filename> - <size>
Source URLs:
- <local-app URL>
- <tree API URL>
```
## References
- **[hub-discovery.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\inference\llama-cpp/references/hub-discovery.md)** - URL-only Hugging Face workflows, search patterns, GGUF extraction, and command reconstruction
- **[advanced-usage.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\inference\llama-cpp/references/advanced-usage.md)** — speculative decoding, batched inference, grammar-constrained generation, LoRA, multi-GPU, custom builds, benchmark scripts
- **[quantization.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\inference\llama-cpp/references/quantization.md)** — quant quality tradeoffs, when to use Q4/Q5/Q6/IQ, model size scaling, imatrix
- **[server.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\inference\llama-cpp/references/server.md)** — direct-from-Hub server launch, OpenAI API endpoints, Docker deployment, NGINX load balancing, monitoring
- **[optimization.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\inference\llama-cpp/references/optimization.md)** — CPU threading, BLAS, GPU offload heuristics, batch tuning, benchmarks
- **[troubleshooting.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\inference\llama-cpp/references/troubleshooting.md)** — install/convert/quantize/inference/server issues, Apple Silicon, debugging
## Resources
- **GitHub**: https://github.com/ggml-org/llama.cpp
- **Hugging Face GGUF + llama.cpp docs**: https://huggingface.co/docs/hub/gguf-llamacpp
- **Hugging Face Local Apps docs**: https://huggingface.co/docs/hub/main/local-apps
- **Hugging Face Local Agents docs**: https://huggingface.co/docs/hub/agents-local
- **Example local-app page**: https://huggingface.co/unsloth/Qwen3.6-35B-A3B-GGUF?local-app=llama.cpp
- **Example tree API**: https://huggingface.co/api/models/unsloth/Qwen3.6-35B-A3B-GGUF/tree/main?recursive=true
- **Example llama.cpp search**: https://huggingface.co/models?num_parameters=min:0,max:24B&apps=llama.cpp&sort=trending
- **License**: MIT
website/docs/user-guide/skills/optional/mlops/mlops-inference-outlines.md
+679
View File
@@ -0,0 +1,679 @@
---
title: "Outlines — Outlines: structured JSON/regex/Pydantic LLM generation"
sidebar_label: "Outlines"
description: "Outlines: structured JSON/regex/Pydantic LLM generation"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Outlines
Outlines: structured JSON/regex/Pydantic LLM generation.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/outlines` |
| Path | `optional-skills/mlops\inference\outlines` |
| Version | `1.0.1` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `outlines`, `transformers`, `vllm`, `pydantic` |
| Platforms | linux, macos, windows |
| Tags | `Prompt Engineering`, `Outlines`, `Structured Generation`, `JSON Schema`, `Pydantic`, `Local Models`, `Grammar-Based Generation`, `vLLM`, `Transformers`, `Type Safety` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Outlines: Structured Text Generation
## When to Use This Skill
Use Outlines when you need to:
- **Guarantee valid JSON/XML/code** structure during generation
- **Use Pydantic models** for type-safe outputs
- **Support local models** (Transformers, llama.cpp, vLLM)
- **Maximize inference speed** with zero-overhead structured generation
- **Generate against JSON schemas** automatically
- **Control token sampling** at the grammar level
**GitHub Stars**: 12,000+ | **From**: dottxt.ai (formerly .txt)
> **API note (Outlines 1.x):** This skill targets the current v1 API.
> The pre-1.0 helpers (`outlines.models.transformers(...)`,
> `outlines.generate.json/choice/regex/...`) have been **removed**. In v1 you
> create a model with `outlines.from_transformers(...)` (or `from_vllm`,
> `from_llamacpp`, `from_openai`) and then **call the model directly** with an
> output type: `model(prompt, output_type)`. JSON/Pydantic outputs are returned
> as a **JSON string** — validate with `YourModel.model_validate_json(result)`.
## Installation
```bash
# Base installation
pip install outlines
# With specific backends
pip install outlines transformers # Hugging Face models
pip install outlines llama-cpp-python # llama.cpp
pip install outlines vllm # vLLM for high-throughput
```
## Quick Start
### Basic Example: Classification
```python
import outlines
from typing import Literal
from transformers import AutoModelForCausalLM, AutoTokenizer
MODEL_NAME = "microsoft/Phi-3-mini-4k-instruct"
# v1: wrap a Transformers model + tokenizer
model = outlines.from_transformers(
AutoModelForCausalLM.from_pretrained(MODEL_NAME, device_map="auto"),
AutoTokenizer.from_pretrained(MODEL_NAME),
)
# Call the model directly with an output type
prompt = "Sentiment of 'This product is amazing!': "
sentiment = model(prompt, Literal["positive", "negative", "neutral"])
print(sentiment) # "positive" (guaranteed one of these)
```
### With Pydantic Models
```python
from pydantic import BaseModel
import outlines
from transformers import AutoModelForCausalLM, AutoTokenizer
class User(BaseModel):
name: str
age: int
email: str
MODEL_NAME = "microsoft/Phi-3-mini-4k-instruct"
model = outlines.from_transformers(
AutoModelForCausalLM.from_pretrained(MODEL_NAME, device_map="auto"),
AutoTokenizer.from_pretrained(MODEL_NAME),
)
# Generate structured output (returns a JSON string)
prompt = "Extract user: John Doe, 30 years old, john@example.com"
result = model(prompt, User, max_new_tokens=200)
user = User.model_validate_json(result) # parse into the Pydantic model
print(user.name) # "John Doe"
print(user.age) # 30
print(user.email) # "john@example.com"
```
## Core Concepts
### 1. Constrained Token Sampling
Outlines constrains token generation at the logit level using a compiled
automaton derived from your output type.
**How it works:**
1. Convert the output type (JSON/Pydantic/regex/`Literal`) to a schema/grammar
2. Compile the grammar into a token-level automaton
3. Filter invalid tokens at each step during generation
4. Fast-forward when only one valid token exists
**Benefits:**
- **Zero overhead**: Filtering happens at token level
- **Speed improvement**: Fast-forward through deterministic paths
- **Guaranteed validity**: Invalid outputs impossible
```python
import outlines
from pydantic import BaseModel
from transformers import AutoModelForCausalLM, AutoTokenizer
class Person(BaseModel):
name: str
age: int
model = outlines.from_transformers(
AutoModelForCausalLM.from_pretrained("microsoft/Phi-3-mini-4k-instruct", device_map="auto"),
AutoTokenizer.from_pretrained("microsoft/Phi-3-mini-4k-instruct"),
)
result = model("Generate person: Alice, 25", Person)
person = Person.model_validate_json(result)
```
### 2. Output Types
In v1 you pass the desired **output type** directly as the second argument.
#### Multiple choice (`Literal`)
```python
from typing import Literal
sentiment = model("Review: This is great!", Literal["positive", "negative", "neutral"])
# Result: one of the three choices
```
#### JSON via Pydantic
```python
from pydantic import BaseModel
class Product(BaseModel):
name: str
price: float
in_stock: bool
result = model("Extract: iPhone 15, $999, available", Product)
product = Product.model_validate_json(result) # valid Product instance
```
#### Regex (pass a regex string)
```python
# Generate text matching a regex pattern
phone = model("Generate phone number:", r"[0-9]{3}-[0-9]{3}-[0-9]{4}")
# Result: "555-123-4567" (guaranteed to match the pattern)
```
#### Numeric types
```python
# Pass the Python type directly
age = model("Person's age:", int) # guaranteed integer
price = model("Product price:", float) # guaranteed float
```
### 3. Model Backends
Outlines supports multiple local and API-based backends via `from_*` factories.
#### Transformers (Hugging Face)
```python
import outlines
from transformers import AutoModelForCausalLM, AutoTokenizer
model = outlines.from_transformers(
AutoModelForCausalLM.from_pretrained("microsoft/Phi-3-mini-4k-instruct", device_map="auto"),
AutoTokenizer.from_pretrained("microsoft/Phi-3-mini-4k-instruct"),
)
result = model(prompt, YourModel)
```
#### llama.cpp
```python
import outlines
from llama_cpp import Llama
llm = Llama("./models/llama-3.1-8b-instruct.Q4_K_M.gguf", n_gpu_layers=35, n_ctx=4096)
model = outlines.from_llamacpp(llm)
result = model(prompt, YourModel)
```
#### vLLM (High Throughput)
```python
import outlines
from vllm import LLM
llm = LLM("meta-llama/Llama-3.1-8B-Instruct", tensor_parallel_size=2)
model = outlines.from_vllm(llm)
result = model(prompt, YourModel)
```
#### OpenAI (server-side constrained JSON)
```python
import outlines
from openai import OpenAI
client = OpenAI()
model = outlines.from_openai(client, "gpt-4o-mini")
# API backends support JSON-schema style structured output
result = model(prompt, YourModel)
```
### 4. Pydantic Integration
Outlines has first-class Pydantic support with automatic schema translation.
Generation returns a JSON string; call `model_validate_json` to get an instance.
#### Basic Models
```python
from pydantic import BaseModel, Field
class Article(BaseModel):
title: str = Field(description="Article title")
author: str = Field(description="Author name")
word_count: int = Field(description="Number of words", gt=0)
tags: list[str] = Field(description="List of tags")
result = model("Generate article about AI", Article, max_new_tokens=300)
article = Article.model_validate_json(result)
print(article.title)
print(article.word_count) # Guaranteed > 0
```
#### Nested Models
```python
class Address(BaseModel):
street: str
city: str
country: str
class Person(BaseModel):
name: str
age: int
address: Address # Nested model
result = model("Generate person in New York", Person)
person = Person.model_validate_json(result)
print(person.address.city) # "New York"
```
#### Enums and Literals
```python
from enum import Enum
from typing import Literal
class Status(str, Enum):
PENDING = "pending"
APPROVED = "approved"
REJECTED = "rejected"
class Application(BaseModel):
applicant: str
status: Status # Must be one of enum values
priority: Literal["low", "medium", "high"] # Must be one of literals
result = model("Generate application", Application)
app = Application.model_validate_json(result)
print(app.status) # Status.PENDING (or APPROVED/REJECTED)
```
## Common Patterns
### Pattern 1: Data Extraction
```python
from pydantic import BaseModel
import outlines
from transformers import AutoModelForCausalLM, AutoTokenizer
class CompanyInfo(BaseModel):
name: str
founded_year: int
industry: str
employees: int
model = outlines.from_transformers(
AutoModelForCausalLM.from_pretrained("microsoft/Phi-3-mini-4k-instruct", device_map="auto"),
AutoTokenizer.from_pretrained("microsoft/Phi-3-mini-4k-instruct"),
)
text = """
Apple Inc. was founded in 1976 in the technology industry.
The company employs approximately 164,000 people worldwide.
"""
prompt = f"Extract company information:\n{text}\n\nCompany:"
company = CompanyInfo.model_validate_json(model(prompt, CompanyInfo, max_new_tokens=200))
print(f"Name: {company.name}")
print(f"Founded: {company.founded_year}")
print(f"Industry: {company.industry}")
print(f"Employees: {company.employees}")
```
### Pattern 2: Classification
```python
from typing import Literal
from pydantic import BaseModel
# Binary classification
result = model("Email: Buy now! 50% off!", Literal["spam", "not_spam"])
# Multi-class classification
category = model(
"Article: Apple announces new iPhone...",
Literal["technology", "business", "sports", "entertainment"],
)
# With confidence
class Classification(BaseModel):
label: Literal["positive", "negative", "neutral"]
confidence: float
out = model("Review: This product is okay, nothing special", Classification)
result = Classification.model_validate_json(out)
```
### Pattern 3: Structured Forms
```python
class UserProfile(BaseModel):
full_name: str
age: int
email: str
phone: str
country: str
interests: list[str]
prompt = """
Extract user profile from:
Name: Alice Johnson
Age: 28
Email: alice@example.com
Phone: 555-0123
Country: USA
Interests: hiking, photography, cooking
"""
profile = UserProfile.model_validate_json(model(prompt, UserProfile, max_new_tokens=250))
print(profile.full_name)
print(profile.interests) # ["hiking", "photography", "cooking"]
```
### Pattern 4: Multi-Entity Extraction
```python
from typing import Literal
class Entity(BaseModel):
name: str
type: Literal["PERSON", "ORGANIZATION", "LOCATION"]
class DocumentEntities(BaseModel):
entities: list[Entity]
text = "Tim Cook met with Satya Nadella at Microsoft headquarters in Redmond."
prompt = f"Extract entities from: {text}"
result = DocumentEntities.model_validate_json(model(prompt, DocumentEntities, max_new_tokens=300))
for entity in result.entities:
print(f"{entity.name} ({entity.type})")
```
### Pattern 5: Code Generation
```python
class PythonFunction(BaseModel):
function_name: str
parameters: list[str]
docstring: str
body: str
prompt = "Generate a Python function to calculate factorial"
func = PythonFunction.model_validate_json(model(prompt, PythonFunction, max_new_tokens=300))
print(f"def {func.function_name}({', '.join(func.parameters)}):")
print(f' """{func.docstring}"""')
print(f" {func.body}")
```
### Pattern 6: Batch Processing
```python
import outlines
from transformers import AutoModelForCausalLM, AutoTokenizer
from pydantic import BaseModel
class Person(BaseModel):
name: str
age: int
model = outlines.from_transformers(
AutoModelForCausalLM.from_pretrained("microsoft/Phi-3-mini-4k-instruct", device_map="auto"),
AutoTokenizer.from_pretrained("microsoft/Phi-3-mini-4k-instruct"),
)
texts = [
"John is 30 years old",
"Alice is 25 years old",
"Bob is 40 years old",
]
# v1 accepts a list of prompts for batched generation
prompts = [f"Extract from: {t}" for t in texts]
outputs = model(prompts, Person, max_new_tokens=100)
people = [Person.model_validate_json(o) for o in outputs]
for person in people:
print(f"{person.name}: {person.age}")
```
## Backend Configuration
### Transformers
```python
import outlines
from transformers import AutoModelForCausalLM, AutoTokenizer
MODEL_NAME = "microsoft/Phi-3-mini-4k-instruct"
# Basic usage
model = outlines.from_transformers(
AutoModelForCausalLM.from_pretrained(MODEL_NAME, device_map="auto"),
AutoTokenizer.from_pretrained(MODEL_NAME),
)
# GPU + dtype configuration is set on the HF model itself
import torch
model = outlines.from_transformers(
AutoModelForCausalLM.from_pretrained(MODEL_NAME, device_map="cuda", torch_dtype=torch.float16),
AutoTokenizer.from_pretrained(MODEL_NAME),
)
# Popular models
for name in [
"meta-llama/Llama-3.1-8B-Instruct",
"mistralai/Mistral-7B-Instruct-v0.3",
"Qwen/Qwen2.5-7B-Instruct",
]:
model = outlines.from_transformers(
AutoModelForCausalLM.from_pretrained(name, device_map="auto"),
AutoTokenizer.from_pretrained(name),
)
```
### llama.cpp
```python
import outlines
from llama_cpp import Llama
# Load GGUF model
llm = Llama(
"./models/llama-3.1-8b.Q4_K_M.gguf",
n_ctx=4096, # Context window
n_gpu_layers=35, # GPU layers
n_threads=8, # CPU threads
)
model = outlines.from_llamacpp(llm)
# Full GPU offload: set n_gpu_layers=-1 on the Llama object
```
### vLLM (Production)
```python
import outlines
from vllm import LLM
# Single GPU
model = outlines.from_vllm(LLM("meta-llama/Llama-3.1-8B-Instruct"))
# Multi-GPU
model = outlines.from_vllm(LLM("meta-llama/Llama-3.1-70B-Instruct", tensor_parallel_size=4))
# With quantization
model = outlines.from_vllm(LLM("meta-llama/Llama-3.1-8B-Instruct", quantization="awq"))
```
## Best Practices
### 1. Use Specific Types
```python
# ✅ Good: Specific types
class Product(BaseModel):
name: str
price: float # Not str
quantity: int # Not str
in_stock: bool # Not str
# ❌ Bad: Everything as string
class Product(BaseModel):
name: str
price: str # Should be float
quantity: str # Should be int
```
### 2. Add Constraints
```python
from pydantic import Field
# ✅ Good: With constraints
class User(BaseModel):
name: str = Field(min_length=1, max_length=100)
age: int = Field(ge=0, le=120)
email: str = Field(pattern=r"^[\w\.-]+@[\w\.-]+\.\w+$")
# ❌ Bad: No constraints
class User(BaseModel):
name: str
age: int
email: str
```
### 3. Use Enums for Categories
```python
# ✅ Good: Enum for fixed set
class Priority(str, Enum):
LOW = "low"
MEDIUM = "medium"
HIGH = "high"
class Task(BaseModel):
title: str
priority: Priority
# ❌ Bad: Free-form string
class Task(BaseModel):
title: str
priority: str # Can be anything
```
### 4. Provide Context in Prompts
```python
# ✅ Good: Clear context
prompt = """
Extract product information from the following text.
Text: iPhone 15 Pro costs $999 and is currently in stock.
Product:
"""
# ❌ Bad: Minimal context
prompt = "iPhone 15 Pro costs $999 and is currently in stock."
```
### 5. Handle Optional Fields
```python
from typing import Optional
# ✅ Good: Optional fields for incomplete data
class Article(BaseModel):
title: str # Required
author: Optional[str] = None # Optional
date: Optional[str] = None # Optional
tags: list[str] = [] # Default empty list
# Can succeed even if author/date missing
```
### 6. Always Validate JSON Output
```python
# v1 returns a JSON string for Pydantic/JSON output types.
result = model(prompt, Article) # str
article = Article.model_validate_json(result) # Article instance
```
## Comparison to Alternatives
| Feature | Outlines | Instructor | Guidance | LMQL |
|---------|----------|------------|----------|------|
| Pydantic Support | ✅ Native | ✅ Native | ✅ Yes | ❌ No |
| JSON Schema | ✅ Yes | ✅ Yes | ✅ Yes | ✅ Yes |
| Regex Constraints | ✅ Yes | ❌ No | ✅ Yes | ✅ Yes |
| Local Models | ✅ Full | ⚠️ Limited | ✅ Full | ✅ Full |
| API Models | ✅ Yes | ✅ Full | ✅ Yes | ✅ Full |
| Zero Overhead | ✅ Yes | ❌ No | ⚠️ Partial | ✅ Yes |
| Automatic Retrying | ❌ No | ✅ Yes | ❌ No | ❌ No |
| Learning Curve | Low | Low | Low | High |
**When to choose Outlines:**
- Using local models (Transformers, llama.cpp, vLLM)
- Need maximum inference speed
- Want Pydantic model support
- Require zero-overhead structured generation
- Control token sampling process
**When to choose alternatives:**
- Instructor: Need API models with automatic retrying
- Guidance: Need token healing and complex workflows
- LMQL: Prefer declarative query syntax
## Performance Characteristics
**Speed:**
- **Zero overhead**: Structured generation as fast as unconstrained
- **Fast-forward optimization**: Skips deterministic tokens
- **1.2-2x faster** than post-generation validation approaches
**Memory:**
- Automaton compiled once per output type (cached)
- Minimal runtime overhead
- Efficient with vLLM for high throughput
**Accuracy:**
- **100% valid outputs** (guaranteed by the constrained automaton)
- No retry loops needed
- Deterministic token filtering
## Resources
- **Documentation**: https://dottxt-ai.github.io/outlines/
- **GitHub**: https://github.com/dottxt-ai/outlines (12k+ stars)
- **Discord**: https://discord.gg/R9DSu34mGd
- **Blog**: https://blog.dottxt.co
## See Also
- `references/json_generation.md` - Comprehensive JSON and Pydantic patterns
- `references/backends.md` - Backend-specific configuration
- `references/examples.md` - Production-ready examples
website/docs/user-guide/skills/optional/mlops/mlops-inference-serving-llms-vllm.md
+387
View File
@@ -0,0 +1,387 @@
---
title: "Serving Llms Vllm — vLLM: high-throughput LLM serving, OpenAI API, quantization"
sidebar_label: "Serving Llms Vllm"
description: "vLLM: high-throughput LLM serving, OpenAI API, quantization"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Serving Llms Vllm
vLLM: high-throughput LLM serving, OpenAI API, quantization.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/serving-llms-vllm` |
| Path | `optional-skills/mlops\inference\serving-llms-vllm` |
| Version | `1.0.1` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `vllm`, `torch`, `transformers` |
| Platforms | linux, macos |
| Tags | `vLLM`, `Inference Serving`, `PagedAttention`, `Continuous Batching`, `High Throughput`, `Production`, `OpenAI API`, `Quantization`, `Tensor Parallelism` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# vLLM - High-Performance LLM Serving
## When to use
Use when deploying production LLM APIs, optimizing inference latency/throughput, or serving models with limited GPU memory. Supports OpenAI-compatible endpoints, quantization (GPTQ/AWQ/FP8), and tensor parallelism.
## Quick start
vLLM achieves 24x higher throughput than standard transformers through PagedAttention (block-based KV cache) and continuous batching (mixing prefill/decode requests).
**Installation**:
```bash
pip install vllm
```
**Basic offline inference**:
```python
from vllm import LLM, SamplingParams
llm = LLM(model="meta-llama/Meta-Llama-3-8B-Instruct")
sampling = SamplingParams(temperature=0.7, max_tokens=256)
outputs = llm.generate(["Explain quantum computing"], sampling)
print(outputs[0].outputs[0].text)
```
**OpenAI-compatible server**:
```bash
vllm serve meta-llama/Meta-Llama-3-8B-Instruct
# Query with OpenAI SDK
python -c "
from openai import OpenAI
client = OpenAI(base_url='http://localhost:8000/v1', api_key='EMPTY')
print(client.chat.completions.create(
model='meta-llama/Meta-Llama-3-8B-Instruct',
messages=[{'role': 'user', 'content': 'Hello!'}]
).choices[0].message.content)
"
```
## Common workflows
### Workflow 1: Production API deployment
Copy this checklist and track progress:
```
Deployment Progress:
- [ ] Step 1: Configure server settings
- [ ] Step 2: Test with limited traffic
- [ ] Step 3: Enable monitoring
- [ ] Step 4: Deploy to production
- [ ] Step 5: Verify performance metrics
```
**Step 1: Configure server settings**
Choose configuration based on your model size:
```bash
# For 7B-13B models on single GPU
vllm serve meta-llama/Meta-Llama-3-8B-Instruct \
--gpu-memory-utilization 0.9 \
--max-model-len 8192 \
--port 8000
# For 30B-70B models with tensor parallelism
vllm serve meta-llama/Meta-Llama-3-70B-Instruct \
--tensor-parallel-size 4 \
--gpu-memory-utilization 0.9 \
--quantization awq \
--port 8000
# For production with caching (Prometheus metrics are exposed
# automatically at /metrics on the API port)
vllm serve meta-llama/Meta-Llama-3-8B-Instruct \
--gpu-memory-utilization 0.9 \
--enable-prefix-caching \
--port 8000 \
--host 0.0.0.0
```
**Step 2: Test with limited traffic**
Run load test before production:
```bash
# Install load testing tool
pip install locust
# Create test_load.py with sample requests
# Run: locust -f test_load.py --host http://localhost:8000
```
Verify TTFT (time to first token) &lt; 500ms and throughput > 100 req/sec.
**Step 3: Enable monitoring**
vLLM exposes Prometheus metrics at `/metrics` on the API port (default 8000):
```bash
curl http://localhost:8000/metrics | grep vllm
```
Key metrics to monitor:
- `vllm:time_to_first_token_seconds` - Latency
- `vllm:num_requests_running` - Active requests
- `vllm:gpu_cache_usage_perc` - KV cache utilization
**Step 4: Deploy to production**
Use Docker for consistent deployment:
```bash
# Run vLLM in Docker
docker run --gpus all -p 8000:8000 \
vllm/vllm-openai:latest \
--model meta-llama/Meta-Llama-3-8B-Instruct \
--gpu-memory-utilization 0.9 \
--enable-prefix-caching
```
**Step 5: Verify performance metrics**
Check that deployment meets targets:
- TTFT &lt; 500ms (for short prompts)
- Throughput > target req/sec
- GPU utilization > 80%
- No OOM errors in logs
### Workflow 2: Offline batch inference
For processing large datasets without server overhead.
Copy this checklist:
```
Batch Processing:
- [ ] Step 1: Prepare input data
- [ ] Step 2: Configure LLM engine
- [ ] Step 3: Run batch inference
- [ ] Step 4: Process results
```
**Step 1: Prepare input data**
```python
# Load prompts from file
prompts = []
with open("prompts.txt") as f:
prompts = [line.strip() for line in f]
print(f"Loaded {len(prompts)} prompts")
```
**Step 2: Configure LLM engine**
```python
from vllm import LLM, SamplingParams
llm = LLM(
model="meta-llama/Meta-Llama-3-8B-Instruct",
tensor_parallel_size=2, # Use 2 GPUs
gpu_memory_utilization=0.9,
max_model_len=4096
)
sampling = SamplingParams(
temperature=0.7,
top_p=0.95,
max_tokens=512,
stop=["</s>", "\n\n"]
)
```
**Step 3: Run batch inference**
vLLM automatically batches requests for efficiency:
```python
# Process all prompts in one call
outputs = llm.generate(prompts, sampling)
# vLLM handles batching internally
# No need to manually chunk prompts
```
**Step 4: Process results**
```python
# Extract generated text
results = []
for output in outputs:
prompt = output.prompt
generated = output.outputs[0].text
results.append({
"prompt": prompt,
"generated": generated,
"tokens": len(output.outputs[0].token_ids)
})
# Save to file
import json
with open("results.jsonl", "w") as f:
for result in results:
f.write(json.dumps(result) + "\n")
print(f"Processed {len(results)} prompts")
```
### Workflow 3: Quantized model serving
Fit large models in limited GPU memory.
```
Quantization Setup:
- [ ] Step 1: Choose quantization method
- [ ] Step 2: Find or create quantized model
- [ ] Step 3: Launch with quantization flag
- [ ] Step 4: Verify accuracy
```
**Step 1: Choose quantization method**
- **AWQ**: Best for 70B models, minimal accuracy loss
- **GPTQ**: Wide model support, good compression
- **FP8**: Fastest on H100 GPUs
**Step 2: Find or create quantized model**
Use pre-quantized models from HuggingFace:
```bash
# Search for AWQ models
# Example: TheBloke/Llama-2-70B-AWQ
```
**Step 3: Launch with quantization flag**
```bash
# Using pre-quantized model
vllm serve TheBloke/Llama-2-70B-AWQ \
--quantization awq \
--tensor-parallel-size 1 \
--gpu-memory-utilization 0.95
# Results: 70B model in ~40GB VRAM
```
**Step 4: Verify accuracy**
Test outputs match expected quality:
```python
# Compare quantized vs non-quantized responses
# Verify task-specific performance unchanged
```
## When to use vs alternatives
**Use vLLM when:**
- Deploying production LLM APIs (100+ req/sec)
- Serving OpenAI-compatible endpoints
- Limited GPU memory but need large models
- Multi-user applications (chatbots, assistants)
- Need low latency with high throughput
**Use alternatives instead:**
- **llama.cpp**: CPU/edge inference, single-user
- **HuggingFace transformers**: Research, prototyping, one-off generation
- **TensorRT-LLM**: NVIDIA-only, need absolute maximum performance
- **Text-Generation-Inference**: Already in HuggingFace ecosystem
## Common issues
**Issue: Out of memory during model loading**
Reduce memory usage:
```bash
vllm serve MODEL \
--gpu-memory-utilization 0.7 \
--max-model-len 4096
```
Or use quantization:
```bash
vllm serve MODEL --quantization awq
```
**Issue: Slow first token (TTFT > 1 second)**
Enable prefix caching for repeated prompts:
```bash
vllm serve MODEL --enable-prefix-caching
```
For long prompts, enable chunked prefill:
```bash
vllm serve MODEL --enable-chunked-prefill
```
**Issue: Model not found error**
Use `--trust-remote-code` for custom models:
```bash
vllm serve MODEL --trust-remote-code
```
**Issue: Low throughput (&lt;50 req/sec)**
Increase concurrent sequences:
```bash
vllm serve MODEL --max-num-seqs 512
```
Check GPU utilization with `nvidia-smi` - should be >80%.
**Issue: Inference slower than expected**
Verify tensor parallelism uses power of 2 GPUs:
```bash
vllm serve MODEL --tensor-parallel-size 4 # Not 3
```
Enable speculative decoding for faster generation (pass config as JSON;
`--speculative-model` was removed in favor of `--speculative-config`):
```bash
vllm serve MODEL \
--speculative-config '{"model": "DRAFT_MODEL", "num_speculative_tokens": 5, "method": "draft_model"}'
```
## Advanced topics
**Server deployment patterns**: See [references/server-deployment.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\inference\serving-llms-vllm/references/server-deployment.md) for Docker, Kubernetes, and load balancing configurations.
**Performance optimization**: See [references/optimization.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\inference\serving-llms-vllm/references/optimization.md) for PagedAttention tuning, continuous batching details, and benchmark results.
**Quantization guide**: See [references/quantization.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\inference\serving-llms-vllm/references/quantization.md) for AWQ/GPTQ/FP8 setup, model preparation, and accuracy comparisons.
**Troubleshooting**: See [references/troubleshooting.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\inference\serving-llms-vllm/references/troubleshooting.md) for detailed error messages, debugging steps, and performance diagnostics.
## Hardware requirements
- **Small models (7B-13B)**: 1x A10 (24GB) or A100 (40GB)
- **Medium models (30B-40B)**: 2x A100 (40GB) with tensor parallelism
- **Large models (70B+)**: 4x A100 (40GB) or 2x A100 (80GB), use AWQ/GPTQ
Supported platforms: NVIDIA (primary), AMD ROCm, Intel GPUs, TPUs
## Resources
- Official docs: https://docs.vllm.ai
- GitHub: https://github.com/vllm-project/vllm
- Paper: "Efficient Memory Management for Large Language Model Serving with PagedAttention" (SOSP 2023)
- Community: https://discuss.vllm.ai
website/docs/user-guide/skills/optional/mlops/mlops-instructor.md
+759
View File
@@ -0,0 +1,759 @@
---
title: "Instructor — Structured LLM outputs validated with Pydantic"
sidebar_label: "Instructor"
description: "Structured LLM outputs validated with Pydantic"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Instructor
Structured LLM outputs validated with Pydantic.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/instructor` |
| Path | `optional-skills/mlops\instructor` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `instructor`, `pydantic`, `openai`, `anthropic` |
| Platforms | linux, macos, windows |
| Tags | `Prompt Engineering`, `Instructor`, `Structured Output`, `Pydantic`, `Data Extraction`, `JSON Parsing`, `Type Safety`, `Validation`, `Streaming`, `OpenAI`, `Anthropic` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Instructor: Structured LLM Outputs
## When to Use This Skill
Use Instructor when you need to:
- **Extract structured data** from LLM responses reliably
- **Validate outputs** against Pydantic schemas automatically
- **Retry failed extractions** with automatic error handling
- **Parse complex JSON** with type safety and validation
- **Stream partial results** for real-time processing
- **Support multiple LLM providers** with consistent API
**GitHub Stars**: 15,000+ | **Battle-tested**: 100,000+ developers
## Installation
```bash
# Base installation
pip install instructor
# With specific providers
pip install "instructor[anthropic]" # Anthropic Claude
pip install "instructor[openai]" # OpenAI
pip install "instructor[all]" # All providers
```
## Quick Start
### Basic Example: Extract User Data
```python
import instructor
from pydantic import BaseModel
from anthropic import Anthropic
# Define output structure
class User(BaseModel):
name: str
age: int
email: str
# Create instructor client
client = instructor.from_anthropic(Anthropic())
# Extract structured data
user = client.messages.create(
model="claude-sonnet-4-5-20250929",
max_tokens=1024,
messages=[{
"role": "user",
"content": "John Doe is 30 years old. His email is john@example.com"
}],
response_model=User
)
print(user.name) # "John Doe"
print(user.age) # 30
print(user.email) # "john@example.com"
```
### With OpenAI
```python
from openai import OpenAI
client = instructor.from_openai(OpenAI())
user = client.chat.completions.create(
model="gpt-4o-mini",
response_model=User,
messages=[{"role": "user", "content": "Extract: Alice, 25, alice@email.com"}]
)
```
## Core Concepts
### 1. Response Models (Pydantic)
Response models define the structure and validation rules for LLM outputs.
#### Basic Model
```python
from pydantic import BaseModel, Field
class Article(BaseModel):
title: str = Field(description="Article title")
author: str = Field(description="Author name")
word_count: int = Field(description="Number of words", gt=0)
tags: list[str] = Field(description="List of relevant tags")
article = client.messages.create(
model="claude-sonnet-4-5-20250929",
max_tokens=1024,
messages=[{
"role": "user",
"content": "Analyze this article: [article text]"
}],
response_model=Article
)
```
**Benefits:**
- Type safety with Python type hints
- Automatic validation (word_count > 0)
- Self-documenting with Field descriptions
- IDE autocomplete support
#### Nested Models
```python
class Address(BaseModel):
street: str
city: str
country: str
class Person(BaseModel):
name: str
age: int
address: Address # Nested model
person = client.messages.create(
model="claude-sonnet-4-5-20250929",
max_tokens=1024,
messages=[{
"role": "user",
"content": "John lives at 123 Main St, Boston, USA"
}],
response_model=Person
)
print(person.address.city) # "Boston"
```
#### Optional Fields
```python
from typing import Optional
class Product(BaseModel):
name: str
price: float
discount: Optional[float] = None # Optional
description: str = Field(default="No description") # Default value
# LLM doesn't need to provide discount or description
```
#### Enums for Constraints
```python
from enum import Enum
class Sentiment(str, Enum):
POSITIVE = "positive"
NEGATIVE = "negative"
NEUTRAL = "neutral"
class Review(BaseModel):
text: str
sentiment: Sentiment # Only these 3 values allowed
review = client.messages.create(
model="claude-sonnet-4-5-20250929",
max_tokens=1024,
messages=[{
"role": "user",
"content": "This product is amazing!"
}],
response_model=Review
)
print(review.sentiment) # Sentiment.POSITIVE
```
### 2. Validation
Pydantic validates LLM outputs automatically. If validation fails, Instructor retries.
#### Built-in Validators
```python
from pydantic import Field, EmailStr, HttpUrl
class Contact(BaseModel):
name: str = Field(min_length=2, max_length=100)
age: int = Field(ge=0, le=120) # 0 <= age <= 120
email: EmailStr # Validates email format
website: HttpUrl # Validates URL format
# If LLM provides invalid data, Instructor retries automatically
```
#### Custom Validators
```python
from pydantic import field_validator
class Event(BaseModel):
name: str
date: str
attendees: int
@field_validator('date')
def validate_date(cls, v):
"""Ensure date is in YYYY-MM-DD format."""
import re
if not re.match(r'\d{4}-\d{2}-\d{2}', v):
raise ValueError('Date must be YYYY-MM-DD format')
return v
@field_validator('attendees')
def validate_attendees(cls, v):
"""Ensure positive attendees."""
if v < 1:
raise ValueError('Must have at least 1 attendee')
return v
```
#### Model-Level Validation
```python
from pydantic import model_validator
class DateRange(BaseModel):
start_date: str
end_date: str
@model_validator(mode='after')
def check_dates(self):
"""Ensure end_date is after start_date."""
from datetime import datetime
start = datetime.strptime(self.start_date, '%Y-%m-%d')
end = datetime.strptime(self.end_date, '%Y-%m-%d')
if end < start:
raise ValueError('end_date must be after start_date')
return self
```
### 3. Automatic Retrying
Instructor retries automatically when validation fails, providing error feedback to the LLM.
```python
# Retries up to 3 times if validation fails
user = client.messages.create(
model="claude-sonnet-4-5-20250929",
max_tokens=1024,
messages=[{
"role": "user",
"content": "Extract user from: John, age unknown"
}],
response_model=User,
max_retries=3 # Default is 3
)
# If age can't be extracted, Instructor tells the LLM:
# "Validation error: age - field required"
# LLM tries again with better extraction
```
**How it works:**
1. LLM generates output
2. Pydantic validates
3. If invalid: Error message sent back to LLM
4. LLM tries again with error feedback
5. Repeats up to max_retries
### 4. Streaming
Stream partial results for real-time processing.
#### Streaming Partial Objects
```python
from instructor import Partial
class Story(BaseModel):
title: str
content: str
tags: list[str]
# Stream partial updates as LLM generates
for partial_story in client.messages.create_partial(
model="claude-sonnet-4-5-20250929",
max_tokens=1024,
messages=[{
"role": "user",
"content": "Write a short sci-fi story"
}],
response_model=Story
):
print(f"Title: {partial_story.title}")
print(f"Content so far: {partial_story.content[:100]}...")
# Update UI in real-time
```
#### Streaming Iterables
```python
class Task(BaseModel):
title: str
priority: str
# Stream list items as they're generated
tasks = client.messages.create_iterable(
model="claude-sonnet-4-5-20250929",
max_tokens=1024,
messages=[{
"role": "user",
"content": "Generate 10 project tasks"
}],
response_model=Task
)
for task in tasks:
print(f"- {task.title} ({task.priority})")
# Process each task as it arrives
```
## Provider Configuration
### Anthropic Claude
```python
import instructor
from anthropic import Anthropic
client = instructor.from_anthropic(
Anthropic(api_key="your-api-key")
)
# Use with Claude models
response = client.messages.create(
model="claude-sonnet-4-5-20250929",
max_tokens=1024,
messages=[...],
response_model=YourModel
)
```
### OpenAI
```python
from openai import OpenAI
client = instructor.from_openai(
OpenAI(api_key="your-api-key")
)
response = client.chat.completions.create(
model="gpt-4o-mini",
response_model=YourModel,
messages=[...]
)
```
### Local Models (Ollama)
```python
from openai import OpenAI
# Point to local Ollama server
client = instructor.from_openai(
OpenAI(
base_url="http://localhost:11434/v1",
api_key="ollama" # Required but ignored
),
mode=instructor.Mode.JSON
)
response = client.chat.completions.create(
model="llama3.1",
response_model=YourModel,
messages=[...]
)
```
## Common Patterns
### Pattern 1: Data Extraction from Text
```python
class CompanyInfo(BaseModel):
name: str
founded_year: int
industry: str
employees: int
headquarters: str
text = """
Tesla, Inc. was founded in 2003. It operates in the automotive and energy
industry with approximately 140,000 employees. The company is headquartered
in Austin, Texas.
"""
company = client.messages.create(
model="claude-sonnet-4-5-20250929",
max_tokens=1024,
messages=[{
"role": "user",
"content": f"Extract company information from: {text}"
}],
response_model=CompanyInfo
)
```
### Pattern 2: Classification
```python
class Category(str, Enum):
TECHNOLOGY = "technology"
FINANCE = "finance"
HEALTHCARE = "healthcare"
EDUCATION = "education"
OTHER = "other"
class ArticleClassification(BaseModel):
category: Category
confidence: float = Field(ge=0.0, le=1.0)
keywords: list[str]
classification = client.messages.create(
model="claude-sonnet-4-5-20250929",
max_tokens=1024,
messages=[{
"role": "user",
"content": "Classify this article: [article text]"
}],
response_model=ArticleClassification
)
```
### Pattern 3: Multi-Entity Extraction
```python
class Person(BaseModel):
name: str
role: str
class Organization(BaseModel):
name: str
industry: str
class Entities(BaseModel):
people: list[Person]
organizations: list[Organization]
locations: list[str]
text = "Tim Cook, CEO of Apple, announced at the event in Cupertino..."
entities = client.messages.create(
model="claude-sonnet-4-5-20250929",
max_tokens=1024,
messages=[{
"role": "user",
"content": f"Extract all entities from: {text}"
}],
response_model=Entities
)
for person in entities.people:
print(f"{person.name} - {person.role}")
```
### Pattern 4: Structured Analysis
```python
class SentimentAnalysis(BaseModel):
overall_sentiment: Sentiment
positive_aspects: list[str]
negative_aspects: list[str]
suggestions: list[str]
score: float = Field(ge=-1.0, le=1.0)
review = "The product works well but setup was confusing..."
analysis = client.messages.create(
model="claude-sonnet-4-5-20250929",
max_tokens=1024,
messages=[{
"role": "user",
"content": f"Analyze this review: {review}"
}],
response_model=SentimentAnalysis
)
```
### Pattern 5: Batch Processing
```python
def extract_person(text: str) -> Person:
return client.messages.create(
model="claude-sonnet-4-5-20250929",
max_tokens=1024,
messages=[{
"role": "user",
"content": f"Extract person from: {text}"
}],
response_model=Person
)
texts = [
"John Doe is a 30-year-old engineer",
"Jane Smith, 25, works in marketing",
"Bob Johnson, age 40, software developer"
]
people = [extract_person(text) for text in texts]
```
## Advanced Features
### Union Types
```python
from typing import Union
class TextContent(BaseModel):
type: str = "text"
content: str
class ImageContent(BaseModel):
type: str = "image"
url: HttpUrl
caption: str
class Post(BaseModel):
title: str
content: Union[TextContent, ImageContent] # Either type
# LLM chooses appropriate type based on content
```
### Dynamic Models
```python
from pydantic import create_model
# Create model at runtime
DynamicUser = create_model(
'User',
name=(str, ...),
age=(int, Field(ge=0)),
email=(EmailStr, ...)
)
user = client.messages.create(
model="claude-sonnet-4-5-20250929",
max_tokens=1024,
messages=[...],
response_model=DynamicUser
)
```
### Custom Modes
```python
# For providers without native structured outputs
client = instructor.from_anthropic(
Anthropic(),
mode=instructor.Mode.JSON # JSON mode
)
# Available modes:
# - Mode.ANTHROPIC_TOOLS (recommended for Claude)
# - Mode.JSON (fallback)
# - Mode.TOOLS (OpenAI tools)
```
### Context Management
```python
# Single-use client
with instructor.from_anthropic(Anthropic()) as client:
result = client.messages.create(
model="claude-sonnet-4-5-20250929",
max_tokens=1024,
messages=[...],
response_model=YourModel
)
# Client closed automatically
```
## Error Handling
### Handling Validation Errors
```python
from pydantic import ValidationError
try:
user = client.messages.create(
model="claude-sonnet-4-5-20250929",
max_tokens=1024,
messages=[...],
response_model=User,
max_retries=3
)
except ValidationError as e:
print(f"Failed after retries: {e}")
# Handle gracefully
except Exception as e:
print(f"API error: {e}")
```
### Custom Error Messages
```python
class ValidatedUser(BaseModel):
name: str = Field(description="Full name, 2-100 characters")
age: int = Field(description="Age between 0 and 120", ge=0, le=120)
email: EmailStr = Field(description="Valid email address")
class Config:
# Custom error messages
json_schema_extra = {
"examples": [
{
"name": "John Doe",
"age": 30,
"email": "john@example.com"
}
]
}
```
## Best Practices
### 1. Clear Field Descriptions
```python
# ❌ Bad: Vague
class Product(BaseModel):
name: str
price: float
# ✅ Good: Descriptive
class Product(BaseModel):
name: str = Field(description="Product name from the text")
price: float = Field(description="Price in USD, without currency symbol")
```
### 2. Use Appropriate Validation
```python
# ✅ Good: Constrain values
class Rating(BaseModel):
score: int = Field(ge=1, le=5, description="Rating from 1 to 5 stars")
review: str = Field(min_length=10, description="Review text, at least 10 chars")
```
### 3. Provide Examples in Prompts
```python
messages = [{
"role": "user",
"content": """Extract person info from: "John, 30, engineer"
Example format:
{
"name": "John Doe",
"age": 30,
"occupation": "engineer"
}"""
}]
```
### 4. Use Enums for Fixed Categories
```python
# ✅ Good: Enum ensures valid values
class Status(str, Enum):
PENDING = "pending"
APPROVED = "approved"
REJECTED = "rejected"
class Application(BaseModel):
status: Status # LLM must choose from enum
```
### 5. Handle Missing Data Gracefully
```python
class PartialData(BaseModel):
required_field: str
optional_field: Optional[str] = None
default_field: str = "default_value"
# LLM only needs to provide required_field
```
## Comparison to Alternatives
| Feature | Instructor | Manual JSON | LangChain | DSPy |
|---------|------------|-------------|-----------|------|
| Type Safety | ✅ Yes | ❌ No | ⚠️ Partial | ✅ Yes |
| Auto Validation | ✅ Yes | ❌ No | ❌ No | ⚠️ Limited |
| Auto Retry | ✅ Yes | ❌ No | ❌ No | ✅ Yes |
| Streaming | ✅ Yes | ❌ No | ✅ Yes | ❌ No |
| Multi-Provider | ✅ Yes | ⚠️ Manual | ✅ Yes | ✅ Yes |
| Learning Curve | Low | Low | Medium | High |
**When to choose Instructor:**
- Need structured, validated outputs
- Want type safety and IDE support
- Require automatic retries
- Building data extraction systems
**When to choose alternatives:**
- DSPy: Need prompt optimization
- LangChain: Building complex chains
- Manual: Simple, one-off extractions
## Resources
- **Documentation**: https://python.useinstructor.com
- **GitHub**: https://github.com/jxnl/instructor (15k+ stars)
- **Cookbook**: https://python.useinstructor.com/examples
- **Discord**: Community support available
## See Also
- `references/validation.md` - Advanced validation patterns
- `references/providers.md` - Provider-specific configuration
- `references/examples.md` - Real-world use cases
website/docs/user-guide/skills/optional/mlops/mlops-lambda-labs.md
+568
View File
@@ -0,0 +1,568 @@
---
title: "Lambda Labs — On-demand GPU cloud instances for ML training"
sidebar_label: "Lambda Labs"
description: "On-demand GPU cloud instances for ML training"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Lambda Labs
On-demand GPU cloud instances for ML training.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/lambda-labs` |
| Path | `optional-skills/mlops\lambda-labs` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `lambda-cloud-client>=1.0.0` |
| Platforms | linux, macos, windows |
| Tags | `Infrastructure`, `GPU Cloud`, `Training`, `Inference`, `Lambda Labs` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Lambda Labs GPU Cloud
Guide to running ML workloads on Lambda Labs GPU cloud with on-demand instances and 1-Click Clusters.
## When to use Lambda Labs
**Use Lambda Labs when:**
- Need dedicated GPU instances with full SSH access
- Running long training jobs (hours to days)
- Want simple pricing with no egress fees
- Need persistent storage across sessions
- Require high-performance multi-node clusters (16-512 GPUs)
- Want pre-installed ML stack (Lambda Stack with PyTorch, CUDA, NCCL)
**Key features:**
- **GPU variety**: B200, H100, GH200, A100, A10, A6000, V100
- **Lambda Stack**: Pre-installed PyTorch, TensorFlow, CUDA, cuDNN, NCCL
- **Persistent filesystems**: Keep data across instance restarts
- **1-Click Clusters**: 16-512 GPU Slurm clusters with InfiniBand
- **Simple pricing**: Pay-per-minute, no egress fees
- **Global regions**: 12+ regions worldwide
**Use alternatives instead:**
- **Modal**: For serverless, auto-scaling workloads
- **SkyPilot**: For multi-cloud orchestration and cost optimization
- **RunPod**: For cheaper spot instances and serverless endpoints
- **Vast.ai**: For GPU marketplace with lowest prices
## Quick start
### Account setup
1. Create account at https://lambda.ai
2. Add payment method
3. Generate API key from dashboard
4. Add SSH key (required before launching instances)
### Launch via console
1. Go to https://cloud.lambda.ai/instances
2. Click "Launch instance"
3. Select GPU type and region
4. Choose SSH key
5. Optionally attach filesystem
6. Launch and wait 3-15 minutes
### Connect via SSH
```bash
# Get instance IP from console
ssh ubuntu@<INSTANCE-IP>
# Or with specific key
ssh -i ~/.ssh/lambda_key ubuntu@<INSTANCE-IP>
```
## GPU instances
### Available GPUs
| GPU | VRAM | Price/GPU/hr | Best For |
|-----|------|--------------|----------|
| B200 SXM6 | 180 GB | $4.99 | Largest models, fastest training |
| H100 SXM | 80 GB | $2.99-3.29 | Large model training |
| H100 PCIe | 80 GB | $2.49 | Cost-effective H100 |
| GH200 | 96 GB | $1.49 | Single-GPU large models |
| A100 80GB | 80 GB | $1.79 | Production training |
| A100 40GB | 40 GB | $1.29 | Standard training |
| A10 | 24 GB | $0.75 | Inference, fine-tuning |
| A6000 | 48 GB | $0.80 | Good VRAM/price ratio |
| V100 | 16 GB | $0.55 | Budget training |
### Instance configurations
```
8x GPU: Best for distributed training (DDP, FSDP)
4x GPU: Large models, multi-GPU training
2x GPU: Medium workloads
1x GPU: Fine-tuning, inference, development
```
### Launch times
- Single-GPU: 3-5 minutes
- Multi-GPU: 10-15 minutes
## Lambda Stack
All instances come with Lambda Stack pre-installed:
```bash
# Included software
- Ubuntu 22.04 LTS
- NVIDIA drivers (latest)
- CUDA 12.x
- cuDNN 8.x
- NCCL (for multi-GPU)
- PyTorch (latest)
- TensorFlow (latest)
- JAX
- JupyterLab
```
### Verify installation
```bash
# Check GPU
nvidia-smi
# Check PyTorch
python -c "import torch; print(torch.cuda.is_available())"
# Check CUDA version
nvcc --version
```
## Python API
### Installation
```bash
pip install lambda-cloud-client
```
### Authentication
```python
import os
import lambda_cloud_client
# Configure with API key
configuration = lambda_cloud_client.Configuration(
host="https://cloud.lambdalabs.com/api/v1",
access_token=os.environ["LAMBDA_API_KEY"]
)
```
### List available instances
```python
with lambda_cloud_client.ApiClient(configuration) as api_client:
api = lambda_cloud_client.DefaultApi(api_client)
# Get available instance types
types = api.instance_types()
for name, info in types.data.items():
print(f"{name}: {info.instance_type.description}")
```
### Launch instance
```python
from lambda_cloud_client.models import LaunchInstanceRequest
request = LaunchInstanceRequest(
region_name="us-west-1",
instance_type_name="gpu_1x_h100_sxm5",
ssh_key_names=["my-ssh-key"],
file_system_names=["my-filesystem"], # Optional
name="training-job"
)
response = api.launch_instance(request)
instance_id = response.data.instance_ids[0]
print(f"Launched: {instance_id}")
```
### List running instances
```python
instances = api.list_instances()
for instance in instances.data:
print(f"{instance.name}: {instance.ip} ({instance.status})")
```
### Terminate instance
```python
from lambda_cloud_client.models import TerminateInstanceRequest
request = TerminateInstanceRequest(
instance_ids=[instance_id]
)
api.terminate_instance(request)
```
### SSH key management
```python
from lambda_cloud_client.models import AddSshKeyRequest
# Add SSH key
request = AddSshKeyRequest(
name="my-key",
public_key="ssh-rsa AAAA..."
)
api.add_ssh_key(request)
# List keys
keys = api.list_ssh_keys()
# Delete key
api.delete_ssh_key(key_id)
```
## CLI with curl
### List instance types
```bash
curl -u $LAMBDA_API_KEY: \
https://cloud.lambdalabs.com/api/v1/instance-types | jq
```
### Launch instance
```bash
curl -u $LAMBDA_API_KEY: \
-X POST https://cloud.lambdalabs.com/api/v1/instance-operations/launch \
-H "Content-Type: application/json" \
-d '{
"region_name": "us-west-1",
"instance_type_name": "gpu_1x_h100_sxm5",
"ssh_key_names": ["my-key"]
}' | jq
```
### Terminate instance
```bash
curl -u $LAMBDA_API_KEY: \
-X POST https://cloud.lambdalabs.com/api/v1/instance-operations/terminate \
-H "Content-Type: application/json" \
-d '{"instance_ids": ["<INSTANCE-ID>"]}' | jq
```
## Persistent storage
### Filesystems
Filesystems persist data across instance restarts:
```bash
# Mount location
/lambda/nfs/<FILESYSTEM_NAME>
# Example: save checkpoints
python train.py --checkpoint-dir /lambda/nfs/my-storage/checkpoints
```
### Create filesystem
1. Go to Storage in Lambda console
2. Click "Create filesystem"
3. Select region (must match instance region)
4. Name and create
### Attach to instance
Filesystems must be attached at instance launch time:
- Via console: Select filesystem when launching
- Via API: Include `file_system_names` in launch request
### Best practices
<!-- ascii-guard-ignore -->
```bash
# Store on filesystem (persists)
/lambda/nfs/storage/
├── datasets/
├── checkpoints/
├── models/
└── outputs/
# Local SSD (faster, ephemeral)
~/ (instance home)
└── working/ # Temporary files
```
<!-- ascii-guard-ignore-end -->
## SSH configuration
### Add SSH key
```bash
# Generate key locally
ssh-keygen -t ed25519 -f ~/.ssh/lambda_key
# Add public key to Lambda console
# Or via API
```
### Multiple keys
```bash
# On instance, add more keys
echo 'ssh-rsa AAAA...' >> ~/.ssh/authorized_keys
```
### Import from GitHub
```bash
# On instance
ssh-import-id gh:username
```
### SSH tunneling
```bash
# Forward Jupyter
ssh -L 8888:localhost:8888 ubuntu@<IP>
# Forward TensorBoard
ssh -L 6006:localhost:6006 ubuntu@<IP>
# Multiple ports
ssh -L 8888:localhost:8888 -L 6006:localhost:6006 ubuntu@<IP>
```
## JupyterLab
### Launch from console
1. Go to Instances page
2. Click "Launch" in Cloud IDE column
3. JupyterLab opens in browser
### Manual access
```bash
# On instance
jupyter lab --ip=0.0.0.0 --port=8888
# From local machine with tunnel
ssh -L 8888:localhost:8888 ubuntu@<IP>
# Open http://localhost:8888
```
## Training workflows
### Single-GPU training
```bash
# SSH to instance
ssh ubuntu@<IP>
# Clone repo
git clone https://github.com/user/project
cd project
# Install dependencies
pip install -r requirements.txt
# Train
python train.py --epochs 100 --checkpoint-dir /lambda/nfs/storage/checkpoints
```
### Multi-GPU training (single node)
```python
# train_ddp.py
import torch
import torch.distributed as dist
from torch.nn.parallel import DistributedDataParallel as DDP
def main():
dist.init_process_group("nccl")
rank = dist.get_rank()
device = rank % torch.cuda.device_count()
model = MyModel().to(device)
model = DDP(model, device_ids=[device])
# Training loop...
if __name__ == "__main__":
main()
```
```bash
# Launch with torchrun (8 GPUs)
torchrun --nproc_per_node=8 train_ddp.py
```
### Checkpoint to filesystem
```python
import os
checkpoint_dir = "/lambda/nfs/my-storage/checkpoints"
os.makedirs(checkpoint_dir, exist_ok=True)
# Save checkpoint
torch.save({
'epoch': epoch,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'loss': loss,
}, f"{checkpoint_dir}/checkpoint_{epoch}.pt")
```
## 1-Click Clusters
### Overview
High-performance Slurm clusters with:
- 16-512 NVIDIA H100 or B200 GPUs
- NVIDIA Quantum-2 400 Gb/s InfiniBand
- GPUDirect RDMA at 3200 Gb/s
- Pre-installed distributed ML stack
### Included software
- Ubuntu 22.04 LTS + Lambda Stack
- NCCL, Open MPI
- PyTorch with DDP and FSDP
- TensorFlow
- OFED drivers
### Storage
- 24 TB NVMe per compute node (ephemeral)
- Lambda filesystems for persistent data
### Multi-node training
```bash
# On Slurm cluster
srun --nodes=4 --ntasks-per-node=8 --gpus-per-node=8 \
torchrun --nnodes=4 --nproc_per_node=8 \
--rdzv_backend=c10d --rdzv_endpoint=$MASTER_ADDR:29500 \
train.py
```
## Networking
### Bandwidth
- Inter-instance (same region): up to 200 Gbps
- Internet outbound: 20 Gbps max
### Firewall
- Default: Only port 22 (SSH) open
- Configure additional ports in Lambda console
- ICMP traffic allowed by default
### Private IPs
```bash
# Find private IP
ip addr show | grep 'inet '
```
## Common workflows
### Workflow 1: Fine-tuning LLM
```bash
# 1. Launch 8x H100 instance with filesystem
# 2. SSH and setup
ssh ubuntu@<IP>
pip install transformers accelerate peft
# 3. Download model to filesystem
python -c "
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained('meta-llama/Llama-2-7b-hf')
model.save_pretrained('/lambda/nfs/storage/models/llama-2-7b')
"
# 4. Fine-tune with checkpoints on filesystem
accelerate launch --num_processes 8 train.py \
--model_path /lambda/nfs/storage/models/llama-2-7b \
--output_dir /lambda/nfs/storage/outputs \
--checkpoint_dir /lambda/nfs/storage/checkpoints
```
### Workflow 2: Batch inference
```bash
# 1. Launch A10 instance (cost-effective for inference)
# 2. Run inference
python inference.py \
--model /lambda/nfs/storage/models/fine-tuned \
--input /lambda/nfs/storage/data/inputs.jsonl \
--output /lambda/nfs/storage/data/outputs.jsonl
```
## Cost optimization
### Choose right GPU
| Task | Recommended GPU |
|------|-----------------|
| LLM fine-tuning (7B) | A100 40GB |
| LLM fine-tuning (70B) | 8x H100 |
| Inference | A10, A6000 |
| Development | V100, A10 |
| Maximum performance | B200 |
### Reduce costs
1. **Use filesystems**: Avoid re-downloading data
2. **Checkpoint frequently**: Resume interrupted training
3. **Right-size**: Don't over-provision GPUs
4. **Terminate idle**: No auto-stop, manually terminate
### Monitor usage
- Dashboard shows real-time GPU utilization
- API for programmatic monitoring
## Common issues
| Issue | Solution |
|-------|----------|
| Instance won't launch | Check region availability, try different GPU |
| SSH connection refused | Wait for instance to initialize (3-15 min) |
| Data lost after terminate | Use persistent filesystems |
| Slow data transfer | Use filesystem in same region |
| GPU not detected | Reboot instance, check drivers |
## References
- **[Advanced Usage](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\lambda-labs/references/advanced-usage.md)** - Multi-node training, API automation
- **[Troubleshooting](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\lambda-labs/references/troubleshooting.md)** - Common issues and solutions
## Resources
- **Documentation**: https://docs.lambda.ai
- **Console**: https://cloud.lambda.ai
- **Pricing**: https://lambda.ai/instances
- **Support**: https://support.lambdalabs.com
- **Blog**: https://lambda.ai/blog
website/docs/user-guide/skills/optional/mlops/mlops-llava.md
+323
View File
@@ -0,0 +1,323 @@
---
title: "Llava — Vision-language chat: VQA, captioning, image dialogue"
sidebar_label: "Llava"
description: "Vision-language chat: VQA, captioning, image dialogue"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Llava
Vision-language chat: VQA, captioning, image dialogue.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/llava` |
| Path | `optional-skills/mlops\llava` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `transformers`, `torch`, `pillow` |
| Platforms | linux, macos, windows |
| Tags | `LLaVA`, `Vision-Language`, `Multimodal`, `Visual Question Answering`, `Image Chat`, `CLIP`, `Vicuna`, `Conversational AI`, `Instruction Tuning`, `VQA` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# LLaVA - Large Language and Vision Assistant
Open-source vision-language model for conversational image understanding.
## When to use LLaVA
**Use when:**
- Building vision-language chatbots
- Visual question answering (VQA)
- Image description and captioning
- Multi-turn image conversations
- Visual instruction following
- Document understanding with images
**Metrics**:
- **23,000+ GitHub stars**
- GPT-4V level capabilities (targeted)
- Apache 2.0 License
- Multiple model sizes (7B-34B params)
**Use alternatives instead**:
- **GPT-4V**: Highest quality, API-based
- **CLIP**: Simple zero-shot classification
- **BLIP-2**: Better for captioning only
- **Flamingo**: Research, not open-source
## Quick start
### Installation
```bash
# Clone repository
git clone https://github.com/haotian-liu/LLaVA
cd LLaVA
# Install
pip install -e .
```
### Basic usage
```python
from llava.model.builder import load_pretrained_model
from llava.mm_utils import get_model_name_from_path, process_images, tokenizer_image_token
from llava.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN
from llava.conversation import conv_templates
from PIL import Image
import torch
# Load model
model_path = "liuhaotian/llava-v1.5-7b"
tokenizer, model, image_processor, context_len = load_pretrained_model(
model_path=model_path,
model_base=None,
model_name=get_model_name_from_path(model_path)
)
# Load image
image = Image.open("image.jpg")
image_tensor = process_images([image], image_processor, model.config)
image_tensor = image_tensor.to(model.device, dtype=torch.float16)
# Create conversation
conv = conv_templates["llava_v1"].copy()
conv.append_message(conv.roles[0], DEFAULT_IMAGE_TOKEN + "\nWhat is in this image?")
conv.append_message(conv.roles[1], None)
prompt = conv.get_prompt()
# Generate response
input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).to(model.device)
with torch.inference_mode():
output_ids = model.generate(
input_ids,
images=image_tensor,
do_sample=True,
temperature=0.2,
max_new_tokens=512
)
response = tokenizer.decode(output_ids[0], skip_special_tokens=True).strip()
print(response)
```
## Available models
| Model | Parameters | VRAM | Quality |
|-------|------------|------|---------|
| LLaVA-v1.5-7B | 7B | ~14 GB | Good |
| LLaVA-v1.5-13B | 13B | ~28 GB | Better |
| LLaVA-v1.6-34B | 34B | ~70 GB | Best |
```python
# Load different models
model_7b = "liuhaotian/llava-v1.5-7b"
model_13b = "liuhaotian/llava-v1.5-13b"
model_34b = "liuhaotian/llava-v1.6-34b"
# 4-bit quantization for lower VRAM
load_4bit = True # Reduces VRAM by ~4×
```
## CLI usage
```bash
# Single image query
python -m llava.serve.cli \
--model-path liuhaotian/llava-v1.5-7b \
--image-file image.jpg \
--query "What is in this image?"
# Multi-turn conversation
python -m llava.serve.cli \
--model-path liuhaotian/llava-v1.5-7b \
--image-file image.jpg
# Then type questions interactively
```
## Web UI (Gradio)
```bash
# Launch Gradio interface
python -m llava.serve.gradio_web_server \
--model-path liuhaotian/llava-v1.5-7b \
--load-4bit # Optional: reduce VRAM
# Access at http://localhost:7860
```
## Multi-turn conversations
```python
# Initialize conversation
conv = conv_templates["llava_v1"].copy()
# Turn 1
conv.append_message(conv.roles[0], DEFAULT_IMAGE_TOKEN + "\nWhat is in this image?")
conv.append_message(conv.roles[1], None)
response1 = generate(conv, model, image) # "A dog playing in a park"
# Turn 2
conv.messages[-1][1] = response1 # Add previous response
conv.append_message(conv.roles[0], "What breed is the dog?")
conv.append_message(conv.roles[1], None)
response2 = generate(conv, model, image) # "Golden Retriever"
# Turn 3
conv.messages[-1][1] = response2
conv.append_message(conv.roles[0], "What time of day is it?")
conv.append_message(conv.roles[1], None)
response3 = generate(conv, model, image)
```
## Common tasks
### Image captioning
```python
question = "Describe this image in detail."
response = ask(model, image, question)
```
### Visual question answering
```python
question = "How many people are in the image?"
response = ask(model, image, question)
```
### Object detection (textual)
```python
question = "List all the objects you can see in this image."
response = ask(model, image, question)
```
### Scene understanding
```python
question = "What is happening in this scene?"
response = ask(model, image, question)
```
### Document understanding
```python
question = "What is the main topic of this document?"
response = ask(model, document_image, question)
```
## Training custom model
```bash
# Stage 1: Feature alignment (558K image-caption pairs)
bash scripts/v1_5/pretrain.sh
# Stage 2: Visual instruction tuning (150K instruction data)
bash scripts/v1_5/finetune.sh
```
## Quantization (reduce VRAM)
```python
# 4-bit quantization
tokenizer, model, image_processor, context_len = load_pretrained_model(
model_path="liuhaotian/llava-v1.5-13b",
model_base=None,
model_name=get_model_name_from_path("liuhaotian/llava-v1.5-13b"),
load_4bit=True # Reduces VRAM ~4×
)
# 8-bit quantization
load_8bit=True # Reduces VRAM ~2×
```
## Best practices
1. **Start with 7B model** - Good quality, manageable VRAM
2. **Use 4-bit quantization** - Reduces VRAM significantly
3. **GPU required** - CPU inference extremely slow
4. **Clear prompts** - Specific questions get better answers
5. **Multi-turn conversations** - Maintain conversation context
6. **Temperature 0.2-0.7** - Balance creativity/consistency
7. **max_new_tokens 512-1024** - For detailed responses
8. **Batch processing** - Process multiple images sequentially
## Performance
| Model | VRAM (FP16) | VRAM (4-bit) | Speed (tokens/s) |
|-------|-------------|--------------|------------------|
| 7B | ~14 GB | ~4 GB | ~20 |
| 13B | ~28 GB | ~8 GB | ~12 |
| 34B | ~70 GB | ~18 GB | ~5 |
*On A100 GPU*
## Benchmarks
LLaVA achieves competitive scores on:
- **VQAv2**: 78.5%
- **GQA**: 62.0%
- **MM-Vet**: 35.4%
- **MMBench**: 64.3%
## Limitations
1. **Hallucinations** - May describe things not in image
2. **Spatial reasoning** - Struggles with precise locations
3. **Small text** - Difficulty reading fine print
4. **Object counting** - Imprecise for many objects
5. **VRAM requirements** - Need powerful GPU
6. **Inference speed** - Slower than CLIP
## Integration with frameworks
### LangChain
```python
from langchain.llms.base import LLM
class LLaVALLM(LLM):
def _call(self, prompt, stop=None):
# Custom LLaVA inference
return response
llm = LLaVALLM()
```
### Gradio App
```python
import gradio as gr
def chat(image, text, history):
response = ask_llava(model, image, text)
return response
demo = gr.ChatInterface(
chat,
additional_inputs=[gr.Image(type="pil")],
title="LLaVA Chat"
)
demo.launch()
```
## Resources
- **GitHub**: https://github.com/haotian-liu/LLaVA ⭐ 23,000+
- **Paper**: https://arxiv.org/abs/2304.08485
- **Demo**: https://llava.hliu.cc
- **Models**: https://huggingface.co/liuhaotian
- **License**: Apache 2.0
website/docs/user-guide/skills/optional/mlops/mlops-modal.md
+368
View File
@@ -0,0 +1,368 @@
---
title: "Modal — Serverless GPU cloud for ML jobs and model APIs"
sidebar_label: "Modal"
description: "Serverless GPU cloud for ML jobs and model APIs"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Modal
Serverless GPU cloud for ML jobs and model APIs.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/modal` |
| Path | `optional-skills/mlops\modal` |
| Version | `1.0.1` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `modal>=1.0` |
| Platforms | linux, macos, windows |
| Tags | `Infrastructure`, `Serverless`, `GPU`, `Cloud`, `Deployment`, `Modal` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Modal Serverless GPU
Guide to running ML workloads on Modal's serverless GPU cloud platform.
## When to use Modal
**Use Modal when:**
- Running GPU-intensive ML workloads without managing infrastructure
- Deploying ML models as auto-scaling APIs
- Running batch processing jobs (training, inference, data processing)
- Need pay-per-second GPU pricing without idle costs
- Prototyping ML applications quickly
- Running scheduled jobs (cron-like workloads)
**Key features:**
- **Serverless GPUs**: T4, L4, A10G, L40S, A100, H100, H200, B200 on-demand
- **Python-native**: Define infrastructure in Python code, no YAML
- **Auto-scaling**: Scale to zero, scale to 100+ GPUs instantly
- **Sub-second cold starts**: Rust-based infrastructure for fast container launches
- **Container caching**: Image layers cached for rapid iteration
- **Web endpoints**: Deploy functions as REST APIs with zero-downtime updates
**Use alternatives instead:**
- **RunPod**: For longer-running pods with persistent state
- **Lambda Labs**: For reserved GPU instances
- **SkyPilot**: For multi-cloud orchestration and cost optimization
- **Kubernetes**: For complex multi-service architectures
## Quick start
### Installation
```bash
pip install modal
modal setup # Opens browser for authentication
```
### Hello World with GPU
```python
import modal
app = modal.App("hello-gpu")
@app.function(gpu="T4")
def gpu_info():
import subprocess
return subprocess.run(["nvidia-smi"], capture_output=True, text=True).stdout
@app.local_entrypoint()
def main():
print(gpu_info.remote())
```
Run: `modal run hello_gpu.py`
### Basic inference endpoint
```python
import modal
app = modal.App("text-generation")
image = modal.Image.debian_slim().pip_install("transformers", "torch", "accelerate")
@app.cls(gpu="A10G", image=image)
class TextGenerator:
@modal.enter()
def load_model(self):
from transformers import pipeline
self.pipe = pipeline("text-generation", model="gpt2", device=0)
@modal.method()
def generate(self, prompt: str) -> str:
return self.pipe(prompt, max_length=100)[0]["generated_text"]
@app.local_entrypoint()
def main():
print(TextGenerator().generate.remote("Hello, world"))
```
## Core concepts
### Key components
| Component | Purpose |
|-----------|---------|
| `App` | Container for functions and resources |
| `Function` | Serverless function with compute specs |
| `Cls` | Class-based functions with lifecycle hooks |
| `Image` | Container image definition |
| `Volume` | Persistent storage for models/data |
| `Secret` | Secure credential storage |
### Execution modes
| Command | Description |
|---------|-------------|
| `modal run script.py` | Execute and exit |
| `modal serve script.py` | Development with live reload |
| `modal deploy script.py` | Persistent cloud deployment |
## GPU configuration
### Available GPUs
| GPU | VRAM | Best For |
|-----|------|----------|
| `T4` | 16GB | Budget inference, small models |
| `L4` | 24GB | Inference, Ada Lovelace arch |
| `A10G` | 24GB | Training/inference, 3.3x faster than T4 |
| `L40S` | 48GB | Recommended for inference (best cost/perf) |
| `A100-40GB` | 40GB | Large model training |
| `A100-80GB` | 80GB | Very large models |
| `H100` | 80GB | Fastest, FP8 + Transformer Engine |
| `H200` | 141GB | Auto-upgrade from H100, 4.8TB/s bandwidth |
| `B200` | Latest | Blackwell architecture |
### GPU specification patterns
```python
# Single GPU
@app.function(gpu="A100")
# Specific memory variant
@app.function(gpu="A100-80GB")
# Multiple GPUs (up to 8)
@app.function(gpu="H100:4")
# GPU with fallbacks
@app.function(gpu=["H100", "A100", "L40S"])
# Any available GPU
@app.function(gpu="any")
```
## Container images
```python
# Basic image with pip
image = modal.Image.debian_slim(python_version="3.11").pip_install(
"torch==2.1.0", "transformers==4.36.0", "accelerate"
)
# From CUDA base
image = modal.Image.from_registry(
"nvidia/cuda:12.1.0-cudnn8-devel-ubuntu22.04",
add_python="3.11"
).pip_install("torch", "transformers")
# With system packages
image = modal.Image.debian_slim().apt_install("git", "ffmpeg").pip_install("whisper")
```
## Persistent storage
```python
volume = modal.Volume.from_name("model-cache", create_if_missing=True)
@app.function(gpu="A10G", volumes={"/models": volume})
def load_model():
import os
model_path = "/models/llama-7b"
if not os.path.exists(model_path):
model = download_model()
model.save_pretrained(model_path)
volume.commit() # Persist changes
return load_from_path(model_path)
```
## Web endpoints
### FastAPI endpoint decorator
```python
@app.function()
@modal.fastapi_endpoint(method="POST")
def predict(text: str) -> dict:
return {"result": model.predict(text)}
```
### Full ASGI app
```python
from fastapi import FastAPI
web_app = FastAPI()
@web_app.post("/predict")
async def predict(text: str):
return {"result": await model.predict.remote.aio(text)}
@app.function()
@modal.asgi_app()
def fastapi_app():
return web_app
```
### Web endpoint types
| Decorator | Use Case |
|-----------|----------|
| `@modal.fastapi_endpoint()` | Simple function → API |
| `@modal.asgi_app()` | Full FastAPI/Starlette apps |
| `@modal.wsgi_app()` | Django/Flask apps |
| `@modal.web_server(port)` | Arbitrary HTTP servers |
## Dynamic batching
```python
@app.function()
@modal.batched(max_batch_size=32, wait_ms=100)
async def batch_predict(inputs: list[str]) -> list[dict]:
# Inputs automatically batched
return model.batch_predict(inputs)
```
## Secrets management
```bash
# Create secret
modal secret create huggingface HF_TOKEN=hf_xxx
```
```python
@app.function(secrets=[modal.Secret.from_name("huggingface")])
def download_model():
import os
token = os.environ["HF_TOKEN"]
```
## Scheduling
```python
@app.function(schedule=modal.Cron("0 0 * * *")) # Daily midnight
def daily_job():
pass
@app.function(schedule=modal.Period(hours=1))
def hourly_job():
pass
```
## Performance optimization
### Cold start mitigation
```python
# Modal 1.0 autoscaler params: scaledown_window (was container_idle_timeout).
# Input concurrency moved to the @modal.concurrent decorator.
@app.function(scaledown_window=300) # Keep warm 5 min
@modal.concurrent(max_inputs=10) # Handle concurrent requests per container
def inference():
pass
```
### Model loading best practices
```python
@app.cls(gpu="A100")
class Model:
@modal.enter() # Run once at container start
def load(self):
self.model = load_model() # Load during warm-up
@modal.method()
def predict(self, x):
return self.model(x)
```
## Parallel processing
```python
@app.function()
def process_item(item):
return expensive_computation(item)
@app.function()
def run_parallel():
items = list(range(1000))
# Fan out to parallel containers
results = list(process_item.map(items))
return results
```
## Common configuration
```python
@app.function(
gpu="A100",
memory=32768, # 32GB RAM
cpu=4, # 4 CPU cores
timeout=3600, # 1 hour max
scaledown_window=120, # Keep warm 2 min (was container_idle_timeout)
retries=3, # Retry on failure
max_containers=10, # Max concurrent containers (was concurrency_limit)
min_containers=1, # Keep N containers warm (was keep_warm)
)
def my_function():
pass
```
> **Modal 1.0 autoscaler renames** (see the [migration guide](https://modal.com/docs/guide/modal-1-0-migration)):
> - `container_idle_timeout` → `scaledown_window`
> - `concurrency_limit` → `max_containers`
> - `keep_warm` → `min_containers`
> - `allow_concurrent_inputs=N` → the `@modal.concurrent(max_inputs=N)` decorator
## Debugging
```python
# Test locally
if __name__ == "__main__":
result = my_function.local()
# View logs
# modal app logs my-app
```
## Common issues
| Issue | Solution |
|-------|----------|
| Cold start latency | Increase `scaledown_window`, use `@modal.enter()` |
| GPU OOM | Use larger GPU (`A100-80GB`), enable gradient checkpointing |
| Image build fails | Pin dependency versions, check CUDA compatibility |
| Timeout errors | Increase `timeout`, add checkpointing |
## References
- **[Advanced Usage](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\modal/references/advanced-usage.md)** - Multi-GPU, distributed training, cost optimization
- **[Troubleshooting](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\modal/references/troubleshooting.md)** - Common issues and solutions
## Resources
- **Documentation**: https://modal.com/docs
- **Examples**: https://github.com/modal-labs/modal-examples
- **Pricing**: https://modal.com/pricing
- **Discord**: https://discord.gg/modal
website/docs/user-guide/skills/optional/mlops/mlops-models-huggingface-hub.md
+100
View File
@@ -0,0 +1,100 @@
---
title: "Huggingface Hub — HuggingFace hf CLI: search/download/upload models, datasets"
sidebar_label: "Huggingface Hub"
description: "HuggingFace hf CLI: search/download/upload models, datasets"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Huggingface Hub
HuggingFace hf CLI: search/download/upload models, datasets.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/huggingface-hub` |
| Path | `optional-skills/mlops\models\huggingface-hub` |
| Version | `1.0.1` |
| Author | Hugging Face |
| License | MIT |
| Platforms | linux, macos, windows |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Hugging Face CLI (`hf`) Reference Guide
The `hf` command is the modern command-line interface for interacting with the Hugging Face Hub, providing tools to manage repositories, models, datasets, and Spaces.
> **IMPORTANT:** The `hf` command replaces the now deprecated `huggingface-cli` command.
## Quick Start
* **Installation:** `curl -LsSf https://hf.co/cli/install.sh | bash -s`
* **Help:** Use `hf --help` to view all available functions and real-world examples.
* **Authentication:** Recommended via `HF_TOKEN` environment variable or the `--token` flag.
---
## Core Commands
### General Operations
* `hf download REPO_ID`: Download files from the Hub.
* `hf upload REPO_ID`: Upload files/folders (recommended for single-commit; also handles resumable uploads of large directories).
* `hf upload-large-folder REPO_ID LOCAL_PATH`: **[Deprecated]** — use `hf upload` instead.
* `hf sync`: Sync files between a local directory and a bucket.
* `hf env` / `hf version`: View environment and version details.
### Authentication (`hf auth`)
* `login` / `logout`: Manage sessions using tokens from [huggingface.co/settings/tokens](https://huggingface.co/settings/tokens).
* `list` / `switch`: Manage and toggle between multiple stored access tokens.
* `whoami`: Identify the currently logged-in account.
### Repository Management (`hf repos`)
* `create` / `delete`: Create or permanently remove repositories.
* `duplicate`: Clone a model, dataset, or Space to a new ID.
* `move`: Transfer a repository between namespaces.
* `branch` / `tag`: Manage Git-like references.
* `delete-files`: Remove specific files using patterns.
---
## Specialized Hub Interactions
### Datasets & Models
* **Datasets:** `hf datasets list`, `info`, and `parquet` (list parquet URLs).
* **SQL Queries:** `hf datasets sql SQL` — Execute raw SQL via DuckDB against dataset parquet URLs.
* **Models:** `hf models list` and `info`.
* **Papers:** `hf papers ls` — View daily papers.
### Discussions & Pull Requests (`hf discussions`)
* Manage the lifecycle of Hub contributions: `list`, `create`, `info`, `comment`, `close`, `reopen`, and `rename`.
* `diff`: View changes in a PR.
* `merge`: Finalize pull requests.
### Infrastructure & Compute
* **Endpoints:** Deploy and manage Inference Endpoints (`deploy`, `pause`, `resume`, `scale-to-zero`, `catalog`).
* **Jobs:** Run compute tasks on HF infrastructure. Includes `hf jobs uv` for running Python scripts with inline dependencies and `stats` for resource monitoring.
* **Spaces:** Manage interactive apps. Includes `dev-mode` and `hot-reload` for Python files without full restarts.
### Storage & Automation
* **Buckets:** Full S3-like bucket management (`create`, `cp`, `mv`, `rm`, `sync`).
* **Cache:** Manage local storage with `list`, `prune` (remove detached revisions), and `verify` (checksum checks).
* **Webhooks:** Automate workflows by managing Hub webhooks (`create`, `watch`, `enable`/`disable`).
* **Collections:** Organize Hub items into collections (`add-item`, `update`, `list`).
---
## Advanced Usage & Tips
### Global Flags
* `--format json`: Produces machine-readable output for automation.
* `-q` / `--quiet`: Limits output to IDs only.
### Extensions & Skills
* **Extensions:** Extend CLI functionality via GitHub repositories using `hf extensions install REPO_ID`.
* **Skills:** Manage AI assistant skills with `hf skills add`.
website/docs/user-guide/skills/optional/mlops/mlops-models-segment-anything-model.md
+525
View File
@@ -0,0 +1,525 @@
---
title: "Segment Anything Model — SAM: zero-shot image segmentation via points, boxes, masks"
sidebar_label: "Segment Anything Model"
description: "SAM: zero-shot image segmentation via points, boxes, masks"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Segment Anything Model
SAM: zero-shot image segmentation via points, boxes, masks.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/segment-anything-model` |
| Path | `optional-skills/mlops\models\segment-anything-model` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `segment-anything`, `transformers>=4.30.0`, `torch>=1.7.0` |
| Platforms | linux, macos, windows |
| Tags | `Multimodal`, `Image Segmentation`, `Computer Vision`, `SAM`, `Zero-Shot` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Segment Anything Model (SAM)
Guide to using Meta AI's Segment Anything Model for zero-shot image segmentation.
## When to use SAM
**Use SAM when:**
- Need to segment any object in images without task-specific training
- Building interactive annotation tools with point/box prompts
- Generating training data for other vision models
- Need zero-shot transfer to new image domains
- Building object detection/segmentation pipelines
- Processing medical, satellite, or domain-specific images
**Key features:**
- **Zero-shot segmentation**: Works on any image domain without fine-tuning
- **Flexible prompts**: Points, bounding boxes, or previous masks
- **Automatic segmentation**: Generate all object masks automatically
- **High quality**: Trained on 1.1 billion masks from 11 million images
- **Multiple model sizes**: ViT-B (fastest), ViT-L, ViT-H (most accurate)
- **ONNX export**: Deploy in browsers and edge devices
**Use alternatives instead:**
- **YOLO/Detectron2**: For real-time object detection with classes
- **Mask2Former**: For semantic/panoptic segmentation with categories
- **GroundingDINO + SAM**: For text-prompted segmentation
- **SAM 2**: For video segmentation tasks
## Quick start
### Installation
```bash
# From GitHub
pip install git+https://github.com/facebookresearch/segment-anything.git
# Optional dependencies
pip install opencv-python pycocotools matplotlib
# Or use HuggingFace transformers
pip install transformers
```
### Download checkpoints
```bash
# ViT-H (largest, most accurate) - 2.4GB
wget https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth
# ViT-L (medium) - 1.2GB
wget https://dl.fbaipublicfiles.com/segment_anything/sam_vit_l_0b3195.pth
# ViT-B (smallest, fastest) - 375MB
wget https://dl.fbaipublicfiles.com/segment_anything/sam_vit_b_01ec64.pth
```
### Basic usage with SamPredictor
```python
import numpy as np
from segment_anything import sam_model_registry, SamPredictor
# Load model
sam = sam_model_registry["vit_h"](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\models\segment-anything-model/checkpoint="sam_vit_h_4b8939.pth")
sam.to(device="cuda")
# Create predictor
predictor = SamPredictor(sam)
# Set image (computes embeddings once)
image = cv2.imread("image.jpg")
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
predictor.set_image(image)
# Predict with point prompts
input_point = np.array([[500, 375]]) # (x, y) coordinates
input_label = np.array([1]) # 1 = foreground, 0 = background
masks, scores, logits = predictor.predict(
point_coords=input_point,
point_labels=input_label,
multimask_output=True # Returns 3 mask options
)
# Select best mask
best_mask = masks[np.argmax(scores)]
```
### HuggingFace Transformers
```python
import torch
from PIL import Image
from transformers import SamModel, SamProcessor
# Load model and processor
model = SamModel.from_pretrained("facebook/sam-vit-huge")
processor = SamProcessor.from_pretrained("facebook/sam-vit-huge")
model.to("cuda")
# Process image with point prompt
image = Image.open("image.jpg")
input_points = [[[450, 600]]] # Batch of points
inputs = processor(image, input_points=input_points, return_tensors="pt")
inputs = {k: v.to("cuda") for k, v in inputs.items()}
# Generate masks
with torch.no_grad():
outputs = model(**inputs)
# Post-process masks to original size
masks = processor.image_processor.post_process_masks(
outputs.pred_masks.cpu(),
inputs["original_sizes"].cpu(),
inputs["reshaped_input_sizes"].cpu()
)
```
## Core concepts
### Model architecture
<!-- ascii-guard-ignore -->
<!-- ascii-guard-ignore -->
```
SAM Architecture:
┌─────────────────┐ ┌─────────────────┐ ┌─────────────────┐
│ Image Encoder │────▶│ Prompt Encoder │────▶│ Mask Decoder │
│ (ViT) │ │ (Points/Boxes) │ │ (Transformer) │
└─────────────────┘ └─────────────────┘ └─────────────────┘
│ │ │
Image Embeddings Prompt Embeddings Masks + IoU
(computed once) (per prompt) predictions
```
<!-- ascii-guard-ignore-end -->
<!-- ascii-guard-ignore-end -->
### Model variants
| Model | Checkpoint | Size | Speed | Accuracy |
|-------|------------|------|-------|----------|
| ViT-H | `vit_h` | 2.4 GB | Slowest | Best |
| ViT-L | `vit_l` | 1.2 GB | Medium | Good |
| ViT-B | `vit_b` | 375 MB | Fastest | Good |
### Prompt types
| Prompt | Description | Use Case |
|--------|-------------|----------|
| Point (foreground) | Click on object | Single object selection |
| Point (background) | Click outside object | Exclude regions |
| Bounding box | Rectangle around object | Larger objects |
| Previous mask | Low-res mask input | Iterative refinement |
## Interactive segmentation
### Point prompts
```python
# Single foreground point
input_point = np.array([[500, 375]])
input_label = np.array([1])
masks, scores, logits = predictor.predict(
point_coords=input_point,
point_labels=input_label,
multimask_output=True
)
# Multiple points (foreground + background)
input_points = np.array([[500, 375], [600, 400], [450, 300]])
input_labels = np.array([1, 1, 0]) # 2 foreground, 1 background
masks, scores, logits = predictor.predict(
point_coords=input_points,
point_labels=input_labels,
multimask_output=False # Single mask when prompts are clear
)
```
### Box prompts
```python
# Bounding box [x1, y1, x2, y2]
input_box = np.array([425, 600, 700, 875])
masks, scores, logits = predictor.predict(
box=input_box,
multimask_output=False
)
```
### Combined prompts
```python
# Box + points for precise control
masks, scores, logits = predictor.predict(
point_coords=np.array([[500, 375]]),
point_labels=np.array([1]),
box=np.array([400, 300, 700, 600]),
multimask_output=False
)
```
### Iterative refinement
```python
# Initial prediction
masks, scores, logits = predictor.predict(
point_coords=np.array([[500, 375]]),
point_labels=np.array([1]),
multimask_output=True
)
# Refine with additional point using previous mask
masks, scores, logits = predictor.predict(
point_coords=np.array([[500, 375], [550, 400]]),
point_labels=np.array([1, 0]), # Add background point
mask_input=logits[np.argmax(scores)][None, :, :], # Use best mask
multimask_output=False
)
```
## Automatic mask generation
### Basic automatic segmentation
```python
from segment_anything import SamAutomaticMaskGenerator
# Create generator
mask_generator = SamAutomaticMaskGenerator(sam)
# Generate all masks
masks = mask_generator.generate(image)
# Each mask contains:
# - segmentation: binary mask
# - bbox: [x, y, w, h]
# - area: pixel count
# - predicted_iou: quality score
# - stability_score: robustness score
# - point_coords: generating point
```
### Customized generation
```python
mask_generator = SamAutomaticMaskGenerator(
model=sam,
points_per_side=32, # Grid density (more = more masks)
pred_iou_thresh=0.88, # Quality threshold
stability_score_thresh=0.95, # Stability threshold
crop_n_layers=1, # Multi-scale crops
crop_n_points_downscale_factor=2,
min_mask_region_area=100, # Remove tiny masks
)
masks = mask_generator.generate(image)
```
### Filtering masks
```python
# Sort by area (largest first)
masks = sorted(masks, key=lambda x: x['area'], reverse=True)
# Filter by predicted IoU
high_quality = [m for m in masks if m['predicted_iou'] > 0.9]
# Filter by stability score
stable_masks = [m for m in masks if m['stability_score'] > 0.95]
```
## Batched inference
### Multiple images
```python
# Process multiple images efficiently
images = [cv2.imread(f"image_{i}.jpg") for i in range(10)]
all_masks = []
for image in images:
predictor.set_image(image)
masks, _, _ = predictor.predict(
point_coords=np.array([[500, 375]]),
point_labels=np.array([1]),
multimask_output=True
)
all_masks.append(masks)
```
### Multiple prompts per image
```python
# Process multiple prompts efficiently (one image encoding)
predictor.set_image(image)
# Batch of point prompts
points = [
np.array([[100, 100]]),
np.array([[200, 200]]),
np.array([[300, 300]])
]
all_masks = []
for point in points:
masks, scores, _ = predictor.predict(
point_coords=point,
point_labels=np.array([1]),
multimask_output=True
)
all_masks.append(masks[np.argmax(scores)])
```
## ONNX deployment
### Export model
```bash
python scripts/export_onnx_model.py \
--checkpoint sam_vit_h_4b8939.pth \
--model-type vit_h \
--output sam_onnx.onnx \
--return-single-mask
```
### Use ONNX model
```python
import onnxruntime
# Load ONNX model
ort_session = onnxruntime.InferenceSession("sam_onnx.onnx")
# Run inference (image embeddings computed separately)
masks = ort_session.run(
None,
{
"image_embeddings": image_embeddings,
"point_coords": point_coords,
"point_labels": point_labels,
"mask_input": np.zeros((1, 1, 256, 256), dtype=np.float32),
"has_mask_input": np.array([0], dtype=np.float32),
"orig_im_size": np.array([h, w], dtype=np.float32)
}
)
```
## Common workflows
### Workflow 1: Annotation tool
```python
import cv2
# Load model
predictor = SamPredictor(sam)
predictor.set_image(image)
def on_click(event, x, y, flags, param):
if event == cv2.EVENT_LBUTTONDOWN:
# Foreground point
masks, scores, _ = predictor.predict(
point_coords=np.array([[x, y]]),
point_labels=np.array([1]),
multimask_output=True
)
# Display best mask
display_mask(masks[np.argmax(scores)])
```
### Workflow 2: Object extraction
```python
def extract_object(image, point):
"""Extract object at point with transparent background."""
predictor.set_image(image)
masks, scores, _ = predictor.predict(
point_coords=np.array([point]),
point_labels=np.array([1]),
multimask_output=True
)
best_mask = masks[np.argmax(scores)]
# Create RGBA output
rgba = np.zeros((image.shape[0], image.shape[1], 4), dtype=np.uint8)
rgba[:, :, :3] = image
rgba[:, :, 3] = best_mask * 255
return rgba
```
### Workflow 3: Medical image segmentation
```python
# Process medical images (grayscale to RGB)
medical_image = cv2.imread("scan.png", cv2.IMREAD_GRAYSCALE)
rgb_image = cv2.cvtColor(medical_image, cv2.COLOR_GRAY2RGB)
predictor.set_image(rgb_image)
# Segment region of interest
masks, scores, _ = predictor.predict(
box=np.array([x1, y1, x2, y2]), # ROI bounding box
multimask_output=True
)
```
## Output format
### Mask data structure
```python
# SamAutomaticMaskGenerator output
{
"segmentation": np.ndarray, # H×W binary mask
"bbox": [x, y, w, h], # Bounding box
"area": int, # Pixel count
"predicted_iou": float, # 0-1 quality score
"stability_score": float, # 0-1 robustness score
"crop_box": [x, y, w, h], # Generation crop region
"point_coords": [[x, y]], # Input point
}
```
### COCO RLE format
```python
from pycocotools import mask as mask_utils
# Encode mask to RLE
rle = mask_utils.encode(np.asfortranarray(mask.astype(np.uint8)))
rle["counts"] = rle["counts"].decode("utf-8")
# Decode RLE to mask
decoded_mask = mask_utils.decode(rle)
```
## Performance optimization
### GPU memory
```python
# Use smaller model for limited VRAM
sam = sam_model_registry["vit_b"](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\models\segment-anything-model/checkpoint="sam_vit_b_01ec64.pth")
# Process images in batches
# Clear CUDA cache between large batches
torch.cuda.empty_cache()
```
### Speed optimization
```python
# Use half precision
sam = sam.half()
# Reduce points for automatic generation
mask_generator = SamAutomaticMaskGenerator(
model=sam,
points_per_side=16, # Default is 32
)
# Use ONNX for deployment
# Export with --return-single-mask for faster inference
```
## Common issues
| Issue | Solution |
|-------|----------|
| Out of memory | Use ViT-B model, reduce image size |
| Slow inference | Use ViT-B, reduce points_per_side |
| Poor mask quality | Try different prompts, use box + points |
| Edge artifacts | Use stability_score filtering |
| Small objects missed | Increase points_per_side |
## References
- **[Advanced Usage](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\models\segment-anything-model/references/advanced-usage.md)** - Batching, fine-tuning, integration
- **[Troubleshooting](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\models\segment-anything-model/references/troubleshooting.md)** - Common issues and solutions
## Resources
- **GitHub**: https://github.com/facebookresearch/segment-anything
- **Paper**: https://arxiv.org/abs/2304.02643
- **Demo**: https://segment-anything.com
- **SAM 2 (Video)**: https://github.com/facebookresearch/segment-anything-2
- **HuggingFace**: https://huggingface.co/facebook/sam-vit-huge
website/docs/user-guide/skills/optional/mlops/mlops-nemo-curator.md
+415
View File
@@ -0,0 +1,415 @@
---
title: "Nemo Curator — Curate LLM training data: dedupe, filter, PII redaction"
sidebar_label: "Nemo Curator"
description: "Curate LLM training data: dedupe, filter, PII redaction"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Nemo Curator
Curate LLM training data: dedupe, filter, PII redaction.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/nemo-curator` |
| Path | `optional-skills/mlops\nemo-curator` |
| Version | `1.0.1` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `nemo-curator`, `cudf`, `dask`, `rapids` |
| Platforms | linux, macos |
| Tags | `Data Processing`, `NeMo Curator`, `Data Curation`, `GPU Acceleration`, `Deduplication`, `Quality Filtering`, `NVIDIA`, `RAPIDS`, `PII Redaction`, `Multimodal`, `LLM Training Data` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# NeMo Curator - GPU-Accelerated Data Curation
NVIDIA's toolkit for preparing high-quality training data for LLMs.
## When to use NeMo Curator
**Use NeMo Curator when:**
- Preparing LLM training data from web scrapes (Common Crawl)
- Need fast deduplication (16× faster than CPU)
- Curating multi-modal datasets (text, images, video, audio)
- Filtering low-quality or toxic content
- Scaling data processing across GPU cluster
**Performance**:
- **16× faster** fuzzy deduplication (8TB RedPajama v2)
- **40% lower TCO** vs CPU alternatives
- **Near-linear scaling** across GPU nodes
**Use alternatives instead**:
- **datatrove**: CPU-based, open-source data processing
- **dolma**: Allen AI's data toolkit
- **Ray Data**: General ML data processing (no curation focus)
## Quick start
### Installation
```bash
# NeMo Curator 1.x installs with uv. Extras use hyphens (PyPI-normalized):
# text-cuda12 / text-cpu (and image/video/audio/math variants), or `all`.
# Text curation (CUDA 12)
uv pip install "nemo-curator[text-cuda12]"
# All modalities
uv pip install "nemo-curator[all]"
# CPU-only text (slower)
uv pip install "nemo-curator[text-cpu]"
```
### Basic text curation pipeline
> **Major version rewrite (1.x):** NeMo Curator was rewritten around a **Ray-based
> pipeline/stage architecture**. The old `DocumentDataset` + `nemo_curator.modules.*` /
> `ScoreFilter` / `Modify` call-the-object-on-a-dataset API from 0.x is gone. In 1.x you
> compose `ProcessingStage`s into a `Pipeline` and run it with an executor. The exact
> stage/import surface differs per modality — treat the examples in this skill below as
> **conceptual** (0.x-style) and follow the current
> [quickstart](https://github.com/NVIDIA-NeMo/Curator/blob/main/tutorials/quickstart.py)
> and [text guide](https://docs.nvidia.com/nemo/curator/latest/get-started/text) for the
> exact 1.x APIs rather than copying imports verbatim.
Shape of a 1.x pipeline (from the upstream quickstart):
```python
from nemo_curator.pipeline import Pipeline
from nemo_curator.stages.base import ProcessingStage
from nemo_curator.stages.resources import Resources
from nemo_curator.backends.xenna import XennaExecutor
from nemo_curator.core.client import RayClient
# 1. Define/compose stages (load -> filter -> dedupe -> classify -> write).
# Each stage declares its own Resources (CPU cores, GPU memory, replicas).
pipeline = Pipeline(name="curation", stages=[...])
# 2. Run it with an executor (Ray-backed).
client = RayClient()
client.start()
pipeline.run(XennaExecutor())
client.stop()
```
The 0.x-style snippets in the sections that follow illustrate the *concepts* (quality
filtering, exact/fuzzy/semantic dedup, PII redaction, classifier filtering). For runnable
1.x code, map each concept onto the corresponding stage from the modality guide.
## Data curation pipeline
### Stage 1: Quality filtering
```python
from nemo_curator.filters import (
WordCountFilter,
RepeatedLinesFilter,
UrlRatioFilter,
NonAlphaNumericFilter
)
# Apply 30+ heuristic filters
from nemo_curator import ScoreFilter
# Word count filter
dataset = dataset.filter(WordCountFilter(min_words=50, max_words=100000))
# Remove repetitive content
dataset = dataset.filter(RepeatedLinesFilter(max_repeated_line_fraction=0.3))
# URL ratio filter
dataset = dataset.filter(UrlRatioFilter(max_url_ratio=0.2))
```
### Stage 2: Deduplication
**Exact deduplication**:
```python
from nemo_curator.modules import ExactDuplicates
# Remove exact duplicates
deduped = ExactDuplicates(id_field="id", text_field="text")(dataset)
```
**Fuzzy deduplication** (16× faster on GPU):
```python
from nemo_curator.modules import FuzzyDuplicates
# MinHash + LSH deduplication
fuzzy_dedup = FuzzyDuplicates(
id_field="id",
text_field="text",
num_hashes=260, # MinHash parameters
num_buckets=20,
hash_method="md5"
)
deduped = fuzzy_dedup(dataset)
```
**Semantic deduplication**:
```python
from nemo_curator.modules import SemanticDuplicates
# Embedding-based deduplication
semantic_dedup = SemanticDuplicates(
id_field="id",
text_field="text",
embedding_model="sentence-transformers/all-MiniLM-L6-v2",
threshold=0.8 # Cosine similarity threshold
)
deduped = semantic_dedup(dataset)
```
### Stage 3: PII redaction
```python
from nemo_curator.modules import Modify
from nemo_curator.modifiers import PIIRedactor
# Redact personally identifiable information
pii_redactor = PIIRedactor(
supported_entities=["EMAIL_ADDRESS", "PHONE_NUMBER", "PERSON", "LOCATION"],
anonymize_action="replace" # or "redact"
)
redacted = Modify(pii_redactor)(dataset)
```
### Stage 4: Classifier filtering
```python
from nemo_curator.classifiers import QualityClassifier
# Quality classification
quality_clf = QualityClassifier(
model_path="nvidia/quality-classifier-deberta",
batch_size=256,
device="cuda"
)
# Filter low-quality documents
high_quality = dataset.filter(lambda doc: quality_clf(doc["text"]) > 0.5)
```
## GPU acceleration
### GPU vs CPU performance
| Operation | CPU (16 cores) | GPU (A100) | Speedup |
|-----------|----------------|------------|---------|
| Fuzzy dedup (8TB) | 120 hours | 7.5 hours | 16× |
| Exact dedup (1TB) | 8 hours | 0.5 hours | 16× |
| Quality filtering | 2 hours | 0.2 hours | 10× |
### Multi-GPU scaling
```python
from nemo_curator import get_client
import dask_cuda
# Initialize GPU cluster
client = get_client(cluster_type="gpu", n_workers=8)
# Process with 8 GPUs
deduped = FuzzyDuplicates(...)(dataset)
```
## Multi-modal curation
### Image curation
```python
from nemo_curator.image import (
AestheticFilter,
NSFWFilter,
CLIPEmbedder
)
# Aesthetic scoring
aesthetic_filter = AestheticFilter(threshold=5.0)
filtered_images = aesthetic_filter(image_dataset)
# NSFW detection
nsfw_filter = NSFWFilter(threshold=0.9)
safe_images = nsfw_filter(filtered_images)
# Generate CLIP embeddings
clip_embedder = CLIPEmbedder(model="openai/clip-vit-base-patch32")
image_embeddings = clip_embedder(safe_images)
```
### Video curation
```python
from nemo_curator.video import (
SceneDetector,
ClipExtractor,
InternVideo2Embedder
)
# Detect scenes
scene_detector = SceneDetector(threshold=27.0)
scenes = scene_detector(video_dataset)
# Extract clips
clip_extractor = ClipExtractor(min_duration=2.0, max_duration=10.0)
clips = clip_extractor(scenes)
# Generate embeddings
video_embedder = InternVideo2Embedder()
video_embeddings = video_embedder(clips)
```
### Audio curation
```python
from nemo_curator.audio import (
ASRInference,
WERFilter,
DurationFilter
)
# ASR transcription
asr = ASRInference(model="nvidia/stt_en_fastconformer_hybrid_large_pc")
transcribed = asr(audio_dataset)
# Filter by WER (word error rate)
wer_filter = WERFilter(max_wer=0.3)
high_quality_audio = wer_filter(transcribed)
# Duration filtering
duration_filter = DurationFilter(min_duration=1.0, max_duration=30.0)
filtered_audio = duration_filter(high_quality_audio)
```
## Common patterns
### Web scrape curation (Common Crawl)
```python
from nemo_curator import ScoreFilter, Modify
from nemo_curator.filters import *
from nemo_curator.modules import *
from nemo_curator.datasets import DocumentDataset
# Load Common Crawl data
dataset = DocumentDataset.read_parquet("common_crawl/*.parquet")
# Pipeline
pipeline = [
# 1. Quality filtering
WordCountFilter(min_words=100, max_words=50000),
RepeatedLinesFilter(max_repeated_line_fraction=0.2),
SymbolToWordRatioFilter(max_symbol_to_word_ratio=0.3),
UrlRatioFilter(max_url_ratio=0.3),
# 2. Language filtering
LanguageIdentificationFilter(target_languages=["en"]),
# 3. Deduplication
ExactDuplicates(id_field="id", text_field="text"),
FuzzyDuplicates(id_field="id", text_field="text", num_hashes=260),
# 4. PII redaction
PIIRedactor(),
# 5. NSFW filtering
NSFWClassifier(threshold=0.8)
]
# Execute
for stage in pipeline:
dataset = stage(dataset)
# Save
dataset.to_parquet("curated_common_crawl/")
```
### Distributed processing
```python
from nemo_curator import get_client
from dask_cuda import LocalCUDACluster
# Multi-GPU cluster
cluster = LocalCUDACluster(n_workers=8)
client = get_client(cluster=cluster)
# Process large dataset
dataset = DocumentDataset.read_parquet("s3://large_dataset/*.parquet")
deduped = FuzzyDuplicates(...)(dataset)
# Cleanup
client.close()
cluster.close()
```
## Performance benchmarks
### Fuzzy deduplication (8TB RedPajama v2)
- **CPU (256 cores)**: 120 hours
- **GPU (8× A100)**: 7.5 hours
- **Speedup**: 16×
### Exact deduplication (1TB)
- **CPU (64 cores)**: 8 hours
- **GPU (4× A100)**: 0.5 hours
- **Speedup**: 16×
### Quality filtering (100GB)
- **CPU (32 cores)**: 2 hours
- **GPU (2× A100)**: 0.2 hours
- **Speedup**: 10×
## Cost comparison
**CPU-based curation** (AWS c5.18xlarge × 10):
- Cost: $3.60/hour × 10 = $36/hour
- Time for 8TB: 120 hours
- **Total**: $4,320
**GPU-based curation** (AWS p4d.24xlarge × 2):
- Cost: $32.77/hour × 2 = $65.54/hour
- Time for 8TB: 7.5 hours
- **Total**: $491.55
**Savings**: 89% reduction ($3,828 saved)
## Supported data formats
- **Input**: Parquet, JSONL, CSV
- **Output**: Parquet (recommended), JSONL
- **WebDataset**: TAR archives for multi-modal
## Use cases
**Production deployments**:
- NVIDIA used NeMo Curator to prepare Nemotron-4 training data
- Open-source datasets curated: RedPajama v2, The Pile
## References
- **[Filtering Guide](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\nemo-curator/references/filtering.md)** - 30+ quality filters, heuristics
- **[Deduplication Guide](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\nemo-curator/references/deduplication.md)** - Exact, fuzzy, semantic methods
## Resources
- **GitHub**: https://github.com/NVIDIA-NeMo/Curator
- **Docs**: https://docs.nvidia.com/nemo/curator/latest/
- **Version**: 1.2.0 (1.x is a Ray-based pipeline rewrite — see the quickstart before copying 0.x snippets)
- **License**: Apache 2.0
website/docs/user-guide/skills/optional/mlops/mlops-obliteratus.md
+360
View File
@@ -0,0 +1,360 @@
---
title: "Obliteratus — OBLITERATUS: abliterate LLM refusals (diff-in-means)"
sidebar_label: "Obliteratus"
description: "OBLITERATUS: abliterate LLM refusals (diff-in-means)"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Obliteratus
OBLITERATUS: abliterate LLM refusals (diff-in-means).
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/obliteratus` |
| Path | `optional-skills/mlops\obliteratus` |
| Version | `2.0.0` |
| Author | Hermes Agent |
| License | MIT |
| Dependencies | `obliteratus`, `torch`, `transformers`, `bitsandbytes`, `accelerate`, `safetensors` |
| Platforms | linux, macos |
| Tags | `Abliteration`, `Uncensoring`, `Refusal-Removal`, `LLM`, `Weight-Projection`, `SVD`, `Mechanistic-Interpretability`, `HuggingFace`, `Model-Surgery` |
| Related skills | [`serving-llms-vllm`](/docs/user-guide/skills/optional/mlops/mlops-inference-serving-llms-vllm), [`llama-cpp`](/docs/user-guide/skills/optional/mlops/mlops-inference-llama-cpp), [`huggingface-tokenizers`](/docs/user-guide/skills/optional/mlops/mlops-huggingface-tokenizers) |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# OBLITERATUS Skill
## What's inside
9 CLI methods, 28 analysis modules, 116 model presets across 5 compute tiers, tournament evaluation, and telemetry-driven recommendations.
Remove refusal behaviors (guardrails) from open-weight LLMs without retraining or fine-tuning. Uses mechanistic interpretability techniques — including diff-in-means, SVD, whitened SVD, LEACE concept erasure, SAE decomposition, Bayesian kernel projection, and more — to identify and surgically excise refusal directions from model weights while preserving reasoning capabilities.
**License warning:** OBLITERATUS is AGPL-3.0. NEVER import it as a Python library. Always invoke via CLI (`obliteratus` command) or subprocess. This keeps Hermes Agent's MIT license clean.
## Video Guide
Walkthrough of OBLITERATUS used by a Hermes agent to abliterate Gemma:
https://www.youtube.com/watch?v=8fG9BrNTeHs ("OBLITERATUS: An AI Agent Removed Gemma 4's Safety Guardrails")
Useful when the user wants a visual overview of the end-to-end workflow before running it themselves.
## When to Use This Skill
Trigger when the user:
- Wants to "uncensor" or "abliterate" an LLM
- Asks about removing refusal/guardrails from a model
- Wants to create an uncensored version of Llama, Qwen, Mistral, etc.
- Mentions "refusal removal", "abliteration", "weight projection"
- Wants to analyze how a model's refusal mechanism works
- References OBLITERATUS, abliterator, or refusal directions
## Step 1: Installation
Check if already installed:
```bash
obliteratus --version 2>/dev/null && echo "INSTALLED" || echo "NOT INSTALLED"
```
If not installed, clone and install from GitHub:
```bash
git clone https://github.com/elder-plinius/OBLITERATUS.git
cd OBLITERATUS
pip install -e .
# For Gradio web UI support:
# pip install -e ".[spaces]"
```
**IMPORTANT:** Confirm with user before installing. This pulls in ~5-10GB of dependencies (PyTorch, Transformers, bitsandbytes, etc.).
## Step 2: Check Hardware
Before anything, check what GPU is available:
```bash
python3 -c "
import torch
if torch.cuda.is_available():
gpu = torch.cuda.get_device_name(0)
vram = torch.cuda.get_device_properties(0).total_memory / 1024**3
print(f'GPU: {gpu}')
print(f'VRAM: {vram:.1f} GB')
if vram < 4: print('TIER: tiny (models under 1B)')
elif vram < 8: print('TIER: small (models 1-4B)')
elif vram < 16: print('TIER: medium (models 4-9B with 4bit quant)')
elif vram < 32: print('TIER: large (models 8-32B with 4bit quant)')
else: print('TIER: frontier (models 32B+)')
else:
print('NO GPU - only tiny models (under 1B) on CPU')
"
```
### VRAM Requirements (with 4-bit quantization)
| VRAM | Max Model Size | Example Models |
|:---------|:----------------|:--------------------------------------------|
| CPU only | ~1B params | GPT-2, TinyLlama, SmolLM |
| 4-8 GB | ~4B params | Qwen2.5-1.5B, Phi-3.5 mini, Llama 3.2 3B |
| 8-16 GB | ~9B params | Llama 3.1 8B, Mistral 7B, Gemma 2 9B |
| 24 GB | ~32B params | Qwen3-32B, Llama 3.1 70B (tight), Command-R |
| 48 GB+ | ~72B+ params | Qwen2.5-72B, DeepSeek-R1 |
| Multi-GPU| 200B+ params | Llama 3.1 405B, DeepSeek-V3 (685B MoE) |
## Step 3: Browse Available Models & Get Recommendations
```bash
# Browse models by compute tier
obliteratus models --tier medium
# Get architecture info for a specific model
obliteratus info <model_name>
# Get telemetry-driven recommendation for best method & params
obliteratus recommend <model_name>
obliteratus recommend <model_name> --insights # global cross-architecture rankings
```
## Step 4: Choose a Method
### Method Selection Guide
**Default / recommended for most cases: `advanced`.** It uses multi-direction SVD with norm-preserving projection and is well-tested.
| Situation | Recommended Method | Why |
|:----------------------------------|:-------------------|:-----------------------------------------|
| Default / most models | `advanced` | Multi-direction SVD, norm-preserving, reliable |
| Quick test / prototyping | `basic` | Fast, simple, good enough to evaluate |
| Dense model (Llama, Mistral) | `advanced` | Multi-direction, norm-preserving |
| MoE model (DeepSeek, Mixtral) | `nuclear` | Expert-granular, handles MoE complexity |
| Reasoning model (R1 distills) | `surgical` | CoT-aware, preserves chain-of-thought |
| Stubborn refusals persist | `aggressive` | Whitened SVD + head surgery + jailbreak |
| Want reversible changes | Use steering vectors (see Analysis section) |
| Maximum quality, time no object | `optimized` | Bayesian search for best parameters |
| Experimental auto-detection | `informed` | Auto-detects alignment type — experimental, may not always outperform advanced |
### 9 CLI Methods
- **basic** — Single refusal direction via diff-in-means. Fast (~5-10 min for 8B).
- **advanced** (DEFAULT, RECOMMENDED) — Multiple SVD directions, norm-preserving projection, 2 refinement passes. Medium speed (~10-20 min).
- **aggressive** — Whitened SVD + jailbreak-contrastive + attention head surgery. Higher risk of coherence damage.
- **spectral_cascade** — DCT frequency-domain decomposition. Research/novel approach.
- **informed** — Runs analysis DURING abliteration to auto-configure. Experimental — slower and less predictable than advanced.
- **surgical** — SAE features + neuron masking + head surgery + per-expert. Very slow (~1-2 hrs). Best for reasoning models.
- **optimized** — Bayesian hyperparameter search (Optuna TPE). Longest runtime but finds optimal parameters.
- **inverted** — Flips the refusal direction. Model becomes actively willing.
- **nuclear** — Maximum force combo for stubborn MoE models. Expert-granular.
### Direction Extraction Methods (--direction-method flag)
- **diff_means** (default) — Simple difference-in-means between refused/complied activations. Robust.
- **svd** — Multi-direction SVD extraction. Better for complex alignment.
- **leace** — LEACE (Linear Erasure via Closed-form Estimation). Optimal linear erasure.
### 4 Python-API-Only Methods
(NOT available via CLI — require Python import, which violates AGPL boundary. Mention to user only if they explicitly want to use OBLITERATUS as a library in their own AGPL project.)
- failspy, gabliteration, heretic, rdo
## Step 5: Run Abliteration
### Standard usage
```bash
# Default method (advanced) — recommended for most models
obliteratus obliterate <model_name> --method advanced --output-dir ./abliterated-models
# With 4-bit quantization (saves VRAM)
obliteratus obliterate <model_name> --method advanced --quantization 4bit --output-dir ./abliterated-models
# Large models (70B+) — conservative defaults
obliteratus obliterate <model_name> --method advanced --quantization 4bit --large-model --output-dir ./abliterated-models
```
### Fine-tuning parameters
```bash
obliteratus obliterate <model_name> \
--method advanced \
--direction-method diff_means \
--n-directions 4 \
--refinement-passes 2 \
--regularization 0.1 \
--quantization 4bit \
--output-dir ./abliterated-models \
--contribute # opt-in telemetry for community research
```
### Key flags
| Flag | Description | Default |
|:-----|:------------|:--------|
| `--method` | Abliteration method | advanced |
| `--direction-method` | Direction extraction | diff_means |
| `--n-directions` | Number of refusal directions (1-32) | method-dependent |
| `--refinement-passes` | Iterative passes (1-5) | 2 |
| `--regularization` | Regularization strength (0.0-1.0) | 0.1 |
| `--quantization` | Load in 4bit or 8bit | none (full precision) |
| `--large-model` | Conservative defaults for 120B+ | false |
| `--output-dir` | Where to save the abliterated model | ./obliterated_model |
| `--contribute` | Share anonymized results for research | false |
| `--verify-sample-size` | Number of test prompts for refusal check | 20 |
| `--dtype` | Model dtype (float16, bfloat16) | auto |
### Other execution modes
```bash
# Interactive guided mode (hardware → model → preset)
obliteratus interactive
# Web UI (Gradio)
obliteratus ui --port 7860
# Run a full ablation study from YAML config
obliteratus run config.yaml --preset quick
# Tournament: pit all methods against each other
obliteratus tourney <model_name>
```
## Step 6: Verify Results
After abliteration, check the output metrics:
| Metric | Good Value | Warning |
|:-------|:-----------|:--------|
| Refusal rate | &lt; 5% (ideally ~0%) | > 10% means refusals persist |
| Perplexity change | &lt; 10% increase | > 15% means coherence damage |
| KL divergence | &lt; 0.1 | > 0.5 means significant distribution shift |
| Coherence | High / passes qualitative check | Degraded responses, repetition |
### If refusals persist (> 10%)
1. Try `aggressive` method
2. Increase `--n-directions` (e.g., 8 or 16)
3. Add `--refinement-passes 3`
4. Try `--direction-method svd` instead of diff_means
### If coherence is damaged (perplexity > 15% increase)
1. Reduce `--n-directions` (try 2)
2. Increase `--regularization` (try 0.3)
3. Reduce `--refinement-passes` to 1
4. Try `basic` method (gentler)
## Step 7: Use the Abliterated Model
The output is a standard HuggingFace model directory.
```bash
# Test locally with transformers
python3 -c "
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained('./abliterated-models/<model>')
tokenizer = AutoTokenizer.from_pretrained('./abliterated-models/<model>')
inputs = tokenizer('How do I pick a lock?', return_tensors='pt')
outputs = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
"
# Upload to HuggingFace Hub
huggingface-cli upload <username>/<model-name>-abliterated ./abliterated-models/<model>
# Serve with vLLM
vllm serve ./abliterated-models/<model>
```
## CLI Command Reference
| Command | Description |
|:--------|:------------|
| `obliteratus obliterate` | Main abliteration command |
| `obliteratus info <model>` | Print model architecture details |
| `obliteratus models --tier <tier>` | Browse curated models by compute tier |
| `obliteratus recommend <model>` | Telemetry-driven method/param suggestion |
| `obliteratus interactive` | Guided setup wizard |
| `obliteratus tourney <model>` | Tournament: all methods head-to-head |
| `obliteratus run <config.yaml>` | Execute ablation study from YAML |
| `obliteratus strategies` | List all registered ablation strategies |
| `obliteratus report <results.json>` | Regenerate visual reports |
| `obliteratus ui` | Launch Gradio web interface |
| `obliteratus aggregate` | Summarize community telemetry data |
## Analysis Modules
OBLITERATUS includes 28 analysis modules for mechanistic interpretability.
See `skill_view(name="obliteratus", file_path="references/analysis-modules.md")` for the full reference.
### Quick analysis commands
```bash
# Run specific analysis modules
obliteratus run analysis-config.yaml --preset quick
# Key modules to run first:
# - alignment_imprint: Fingerprint DPO/RLHF/CAI/SFT alignment method
# - concept_geometry: Single direction vs polyhedral cone
# - logit_lens: Which layer decides to refuse
# - anti_ouroboros: Self-repair risk score
# - causal_tracing: Causally necessary components
```
### Steering Vectors (Reversible Alternative)
Instead of permanent weight modification, use inference-time steering:
```python
# Python API only — for user's own projects
from obliteratus.analysis.steering_vectors import SteeringVectorFactory, SteeringHookManager
```
## Ablation Strategies
Beyond direction-based abliteration, OBLITERATUS includes structural ablation strategies:
- **Embedding Ablation** — Target embedding layer components
- **FFN Ablation** — Feed-forward network block removal
- **Head Pruning** — Attention head pruning
- **Layer Removal** — Full layer removal
List all available: `obliteratus strategies`
## Evaluation
OBLITERATUS includes built-in evaluation tools:
- Refusal rate benchmarking
- Perplexity comparison (before/after)
- LM Eval Harness integration for academic benchmarks
- Head-to-head competitor comparison
- Baseline performance tracking
## Platform Support
- **CUDA** — Full support (NVIDIA GPUs)
- **Apple Silicon (MLX)** — Supported via MLX backend
- **CPU** — Supported for tiny models (&lt; 1B params)
## YAML Config Templates
Load templates for reproducible runs via `skill_view`:
- `templates/abliteration-config.yaml` — Standard single-model config
- `templates/analysis-study.yaml` — Pre-abliteration analysis study
- `templates/batch-abliteration.yaml` — Multi-model batch processing
## Telemetry
OBLITERATUS can optionally contribute anonymized run data to a global research dataset.
Enable with `--contribute` flag. No personal data is collected — only model name, method, metrics.
## Common Pitfalls
1. **Don't use `informed` as default** — it's experimental and slower. Use `advanced` for reliable results.
2. **Models under ~1B respond poorly to abliteration** — their refusal behaviors are shallow and fragmented, making clean direction extraction difficult. Expect partial results (20-40% remaining refusal). Models 3B+ have cleaner refusal directions and respond much better (often 0% refusal with `advanced`).
3. **`aggressive` can make things worse** — on small models it can damage coherence and actually increase refusal rate. Only use it if `advanced` leaves > 10% refusals on a 3B+ model.
4. **Always check perplexity** — if it spikes > 15%, the model is damaged. Reduce aggressiveness.
5. **MoE models need special handling** — use `nuclear` method for Mixtral, DeepSeek-MoE, etc.
6. **Quantized models can't be re-quantized** — abliterate the full-precision model, then quantize the output.
7. **VRAM estimation is approximate** — 4-bit quant helps but peak usage can spike during extraction.
8. **Reasoning models are sensitive** — use `surgical` for R1 distills to preserve chain-of-thought.
9. **Check `obliteratus recommend`** — telemetry data may have better parameters than defaults.
10. **AGPL license** — never `import obliteratus` in MIT/Apache projects. CLI invocation only.
11. **Large models (70B+)** — always use `--large-model` flag for conservative defaults.
12. **Spectral certification RED is common** — the spectral check often flags "incomplete" even when practical refusal rate is 0%. Check actual refusal rate rather than relying on spectral certification alone.
## Complementary Skills
- **vllm** — Serve abliterated models with high throughput
- **gguf** — Convert abliterated models to GGUF for llama.cpp
- **huggingface-tokenizers** — Work with model tokenizers
website/docs/user-guide/skills/optional/mlops/mlops-peft.md
+452
View File
@@ -0,0 +1,452 @@
---
title: "Peft — Fine-tune large LLMs with LoRA on limited GPU memory"
sidebar_label: "Peft"
description: "Fine-tune large LLMs with LoRA on limited GPU memory"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Peft
Fine-tune large LLMs with LoRA on limited GPU memory.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/peft` |
| Path | `optional-skills/mlops\peft` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `peft>=0.13.0`, `transformers>=4.45.0`, `torch>=2.0.0`, `bitsandbytes>=0.43.0` |
| Platforms | linux, macos, windows |
| Tags | `Fine-Tuning`, `PEFT`, `LoRA`, `QLoRA`, `Parameter-Efficient`, `Adapters`, `Low-Rank`, `Memory Optimization`, `Multi-Adapter` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# PEFT (Parameter-Efficient Fine-Tuning)
Fine-tune LLMs by training &lt;1% of parameters using LoRA, QLoRA, and 25+ adapter methods.
## When to use PEFT
**Use PEFT/LoRA when:**
- Fine-tuning 7B-70B models on consumer GPUs (RTX 4090, A100)
- Need to train &lt;1% parameters (6MB adapters vs 14GB full model)
- Want fast iteration with multiple task-specific adapters
- Deploying multiple fine-tuned variants from one base model
**Use QLoRA (PEFT + quantization) when:**
- Fine-tuning 70B models on single 24GB GPU
- Memory is the primary constraint
- Can accept ~5% quality trade-off vs full fine-tuning
**Use full fine-tuning instead when:**
- Training small models (&lt;1B parameters)
- Need maximum quality and have compute budget
- Significant domain shift requires updating all weights
## Quick start
### Installation
```bash
# Basic installation
pip install peft
# With quantization support (recommended)
pip install peft bitsandbytes
# Full stack
pip install peft transformers accelerate bitsandbytes datasets
```
### LoRA fine-tuning (standard)
```python
from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments, Trainer
from peft import get_peft_model, LoraConfig, TaskType
from datasets import load_dataset
# Load base model
model_name = "meta-llama/Llama-3.1-8B"
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="auto", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_name)
tokenizer.pad_token = tokenizer.eos_token
# LoRA configuration
lora_config = LoraConfig(
task_type=TaskType.CAUSAL_LM,
r=16, # Rank (8-64, higher = more capacity)
lora_alpha=32, # Scaling factor (typically 2*r)
lora_dropout=0.05, # Dropout for regularization
target_modules=["q_proj", "v_proj", "k_proj", "o_proj"], # Attention layers
bias="none" # Don't train biases
)
# Apply LoRA
model = get_peft_model(model, lora_config)
model.print_trainable_parameters()
# Output: trainable params: 13,631,488 || all params: 8,043,307,008 || trainable%: 0.17%
# Prepare dataset
dataset = load_dataset("databricks/databricks-dolly-15k", split="train")
def tokenize(example):
text = f"### Instruction:\n{example['instruction']}\n\n### Response:\n{example['response']}"
return tokenizer(text, truncation=True, max_length=512, padding="max_length")
tokenized = dataset.map(tokenize, remove_columns=dataset.column_names)
# Training
training_args = TrainingArguments(
output_dir="./lora-llama",
num_train_epochs=3,
per_device_train_batch_size=4,
gradient_accumulation_steps=4,
learning_rate=2e-4,
fp16=True,
logging_steps=10,
save_strategy="epoch"
)
trainer = Trainer(
model=model,
args=training_args,
train_dataset=tokenized,
data_collator=lambda data: {"input_ids": torch.stack([f["input_ids"] for f in data]),
"attention_mask": torch.stack([f["attention_mask"] for f in data]),
"labels": torch.stack([f["input_ids"] for f in data])}
)
trainer.train()
# Save adapter only (6MB vs 16GB)
model.save_pretrained("./lora-llama-adapter")
```
### QLoRA fine-tuning (memory-efficient)
```python
from transformers import AutoModelForCausalLM, BitsAndBytesConfig
from peft import get_peft_model, LoraConfig, prepare_model_for_kbit_training
# 4-bit quantization config
bnb_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_quant_type="nf4", # NormalFloat4 (best for LLMs)
bnb_4bit_compute_dtype="bfloat16", # Compute in bf16
bnb_4bit_use_double_quant=True # Nested quantization
)
# Load quantized model
model = AutoModelForCausalLM.from_pretrained(
"meta-llama/Llama-3.1-70B",
quantization_config=bnb_config,
device_map="auto"
)
# Prepare for training (enables gradient checkpointing)
model = prepare_model_for_kbit_training(model)
# LoRA config for QLoRA
lora_config = LoraConfig(
r=64, # Higher rank for 70B
lora_alpha=128,
lora_dropout=0.1,
target_modules=["q_proj", "v_proj", "k_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
bias="none",
task_type="CAUSAL_LM"
)
model = get_peft_model(model, lora_config)
# 70B model now fits on single 24GB GPU!
```
## LoRA parameter selection
### Rank (r) - capacity vs efficiency
| Rank | Trainable Params | Memory | Quality | Use Case |
|------|-----------------|--------|---------|----------|
| 4 | ~3M | Minimal | Lower | Simple tasks, prototyping |
| **8** | ~7M | Low | Good | **Recommended starting point** |
| **16** | ~14M | Medium | Better | **General fine-tuning** |
| 32 | ~27M | Higher | High | Complex tasks |
| 64 | ~54M | High | Highest | Domain adaptation, 70B models |
### Alpha (lora_alpha) - scaling factor
```python
# Rule of thumb: alpha = 2 * rank
LoraConfig(r=16, lora_alpha=32) # Standard
LoraConfig(r=16, lora_alpha=16) # Conservative (lower learning rate effect)
LoraConfig(r=16, lora_alpha=64) # Aggressive (higher learning rate effect)
```
### Target modules by architecture
```python
# Llama / Mistral / Qwen
target_modules = ["q_proj", "v_proj", "k_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]
# GPT-2 / GPT-Neo
target_modules = ["c_attn", "c_proj", "c_fc"]
# Falcon
target_modules = ["query_key_value", "dense", "dense_h_to_4h", "dense_4h_to_h"]
# BLOOM
target_modules = ["query_key_value", "dense", "dense_h_to_4h", "dense_4h_to_h"]
# Auto-detect all linear layers
target_modules = "all-linear" # PEFT 0.6.0+
```
## Loading and merging adapters
### Load trained adapter
```python
from peft import PeftModel, AutoPeftModelForCausalLM
from transformers import AutoModelForCausalLM
# Option 1: Load with PeftModel
base_model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.1-8B")
model = PeftModel.from_pretrained(base_model, "./lora-llama-adapter")
# Option 2: Load directly (recommended)
model = AutoPeftModelForCausalLM.from_pretrained(
"./lora-llama-adapter",
device_map="auto"
)
```
### Merge adapter into base model
```python
# Merge for deployment (no adapter overhead)
merged_model = model.merge_and_unload()
# Save merged model
merged_model.save_pretrained("./llama-merged")
tokenizer.save_pretrained("./llama-merged")
# Push to Hub
merged_model.push_to_hub("username/llama-finetuned")
```
### Multi-adapter serving
```python
from peft import PeftModel
# Load base with first adapter
model = AutoPeftModelForCausalLM.from_pretrained("./adapter-task1")
# Load additional adapters
model.load_adapter("./adapter-task2", adapter_name="task2")
model.load_adapter("./adapter-task3", adapter_name="task3")
# Switch between adapters at runtime
model.set_adapter("task1") # Use task1 adapter
output1 = model.generate(**inputs)
model.set_adapter("task2") # Switch to task2
output2 = model.generate(**inputs)
# Disable adapters (use base model)
with model.disable_adapter():
base_output = model.generate(**inputs)
```
## PEFT methods comparison
| Method | Trainable % | Memory | Speed | Best For |
|--------|------------|--------|-------|----------|
| **LoRA** | 0.1-1% | Low | Fast | General fine-tuning |
| **QLoRA** | 0.1-1% | Very Low | Medium | Memory-constrained |
| AdaLoRA | 0.1-1% | Low | Medium | Automatic rank selection |
| IA3 | 0.01% | Minimal | Fastest | Few-shot adaptation |
| Prefix Tuning | 0.1% | Low | Medium | Generation control |
| Prompt Tuning | 0.001% | Minimal | Fast | Simple task adaptation |
| P-Tuning v2 | 0.1% | Low | Medium | NLU tasks |
### IA3 (minimal parameters)
```python
from peft import IA3Config
ia3_config = IA3Config(
target_modules=["q_proj", "v_proj", "k_proj", "down_proj"],
feedforward_modules=["down_proj"]
)
model = get_peft_model(model, ia3_config)
# Trains only 0.01% of parameters!
```
### Prefix Tuning
```python
from peft import PrefixTuningConfig
prefix_config = PrefixTuningConfig(
task_type="CAUSAL_LM",
num_virtual_tokens=20, # Prepended tokens
prefix_projection=True # Use MLP projection
)
model = get_peft_model(model, prefix_config)
```
## Integration patterns
### With TRL (SFTTrainer)
```python
from trl import SFTTrainer, SFTConfig
from peft import LoraConfig
lora_config = LoraConfig(r=16, lora_alpha=32, target_modules="all-linear")
trainer = SFTTrainer(
model=model,
args=SFTConfig(output_dir="./output", max_seq_length=512),
train_dataset=dataset,
peft_config=lora_config, # Pass LoRA config directly
)
trainer.train()
```
### With Axolotl (YAML config)
```yaml
# axolotl config.yaml
adapter: lora
lora_r: 16
lora_alpha: 32
lora_dropout: 0.05
lora_target_modules:
- q_proj
- v_proj
- k_proj
- o_proj
lora_target_linear: true # Target all linear layers
```
### With vLLM (inference)
```python
from vllm import LLM
from vllm.lora.request import LoRARequest
# Load base model with LoRA support
llm = LLM(model="meta-llama/Llama-3.1-8B", enable_lora=True)
# Serve with adapter
outputs = llm.generate(
prompts,
lora_request=LoRARequest("adapter1", 1, "./lora-adapter")
)
```
## Performance benchmarks
### Memory usage (Llama 3.1 8B)
| Method | GPU Memory | Trainable Params |
|--------|-----------|------------------|
| Full fine-tuning | 60+ GB | 8B (100%) |
| LoRA r=16 | 18 GB | 14M (0.17%) |
| QLoRA r=16 | 6 GB | 14M (0.17%) |
| IA3 | 16 GB | 800K (0.01%) |
### Training speed (A100 80GB)
| Method | Tokens/sec | vs Full FT |
|--------|-----------|------------|
| Full FT | 2,500 | 1x |
| LoRA | 3,200 | 1.3x |
| QLoRA | 2,100 | 0.84x |
### Quality (MMLU benchmark)
| Model | Full FT | LoRA | QLoRA |
|-------|---------|------|-------|
| Llama 2-7B | 45.3 | 44.8 | 44.1 |
| Llama 2-13B | 54.8 | 54.2 | 53.5 |
## Common issues
### CUDA OOM during training
```python
# Solution 1: Enable gradient checkpointing
model.gradient_checkpointing_enable()
# Solution 2: Reduce batch size + increase accumulation
TrainingArguments(
per_device_train_batch_size=1,
gradient_accumulation_steps=16
)
# Solution 3: Use QLoRA
from transformers import BitsAndBytesConfig
bnb_config = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4")
```
### Adapter not applying
```python
# Verify adapter is active
print(model.active_adapters) # Should show adapter name
# Check trainable parameters
model.print_trainable_parameters()
# Ensure model in training mode
model.train()
```
### Quality degradation
```python
# Increase rank
LoraConfig(r=32, lora_alpha=64)
# Target more modules
target_modules = "all-linear"
# Use more training data and epochs
TrainingArguments(num_train_epochs=5)
# Lower learning rate
TrainingArguments(learning_rate=1e-4)
```
## Best practices
1. **Start with r=8-16**, increase if quality insufficient
2. **Use alpha = 2 * rank** as starting point
3. **Target attention + MLP layers** for best quality/efficiency
4. **Enable gradient checkpointing** for memory savings
5. **Save adapters frequently** (small files, easy rollback)
6. **Evaluate on held-out data** before merging
7. **Use QLoRA for 70B+ models** on consumer hardware
## References
- **[Advanced Usage](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\peft/references/advanced-usage.md)** - DoRA, LoftQ, rank stabilization, custom modules
- **[Troubleshooting](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\peft/references/troubleshooting.md)** - Common errors, debugging, optimization
## Resources
- **GitHub**: https://github.com/huggingface/peft
- **Docs**: https://huggingface.co/docs/peft
- **LoRA Paper**: arXiv:2106.09685
- **QLoRA Paper**: arXiv:2305.14314
- **Models**: https://huggingface.co/models?library=peft
website/docs/user-guide/skills/optional/mlops/mlops-pinecone.md
+396
View File
@@ -0,0 +1,396 @@
---
title: "Pinecone — Managed vector DB for production RAG and search"
sidebar_label: "Pinecone"
description: "Managed vector DB for production RAG and search"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Pinecone
Managed vector DB for production RAG and search.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/pinecone` |
| Path | `optional-skills/mlops\pinecone` |
| Version | `1.0.1` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `pinecone` |
| Platforms | linux, macos, windows |
| Tags | `RAG`, `Pinecone`, `Vector Database`, `Managed Service`, `Serverless`, `Hybrid Search`, `Production`, `Auto-Scaling`, `Low Latency`, `Recommendations` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Pinecone - Managed Vector Database
The vector database for production AI applications.
## When to use Pinecone
**Use when:**
- Need managed, serverless vector database
- Production RAG applications
- Auto-scaling required
- Low latency critical (&lt;100ms)
- Don't want to manage infrastructure
- Need hybrid search (dense + sparse vectors)
**Metrics**:
- Fully managed SaaS
- Auto-scales to billions of vectors
- **p95 latency &lt;100ms**
- 99.9% uptime SLA
**Use alternatives instead**:
- **Chroma**: Self-hosted, open-source
- **FAISS**: Offline, pure similarity search
- **Weaviate**: Self-hosted with more features
## Quick start
### Installation
```bash
pip install pinecone
```
> Note: the old `pinecone-client` package is deprecated. Install `pinecone` (v5+; current 9.x). The import stays `from pinecone import Pinecone`.
### Basic usage
```python
from pinecone import Pinecone, ServerlessSpec
# Initialize
pc = Pinecone(api_key="your-api-key")
# Create index
pc.create_index(
name="my-index",
dimension=1536, # Must match embedding dimension
metric="cosine", # or "euclidean", "dotproduct"
spec=ServerlessSpec(cloud="aws", region="us-east-1")
)
# Connect to index
index = pc.Index("my-index")
# Upsert vectors
index.upsert(vectors=[
{"id": "vec1", "values": [0.1, 0.2, ...], "metadata": {"category": "A"}},
{"id": "vec2", "values": [0.3, 0.4, ...], "metadata": {"category": "B"}}
])
# Query
results = index.query(
vector=[0.1, 0.2, ...],
top_k=5,
include_metadata=True
)
print(results["matches"])
```
## Core operations
### Create index
```python
# Serverless (recommended)
pc.create_index(
name="my-index",
dimension=1536,
metric="cosine",
spec=ServerlessSpec(
cloud="aws", # or "gcp", "azure"
region="us-east-1"
)
)
# Pod-based (for consistent performance)
from pinecone import PodSpec
pc.create_index(
name="my-index",
dimension=1536,
metric="cosine",
spec=PodSpec(
environment="us-east1-gcp",
pod_type="p1.x1"
)
)
```
### Upsert vectors
```python
# Single upsert
index.upsert(vectors=[
{
"id": "doc1",
"values": [0.1, 0.2, ...], # 1536 dimensions
"metadata": {
"text": "Document content",
"category": "tutorial",
"timestamp": "2025-01-01"
}
}
])
# Batch upsert (recommended)
vectors = [
{"id": f"vec{i}", "values": embedding, "metadata": metadata}
for i, (embedding, metadata) in enumerate(zip(embeddings, metadatas))
]
index.upsert(vectors=vectors, batch_size=100)
```
### Query vectors
```python
# Basic query
results = index.query(
vector=[0.1, 0.2, ...],
top_k=10,
include_metadata=True,
include_values=False
)
# With metadata filtering
results = index.query(
vector=[0.1, 0.2, ...],
top_k=5,
filter={"category": {"$eq": "tutorial"}}
)
# Namespace query
results = index.query(
vector=[0.1, 0.2, ...],
top_k=5,
namespace="production"
)
# Access results
for match in results["matches"]:
print(f"ID: {match['id']}")
print(f"Score: {match['score']}")
print(f"Metadata: {match['metadata']}")
```
### Metadata filtering
```python
# Exact match
filter = {"category": "tutorial"}
# Comparison
filter = {"price": {"$gte": 100}} # $gt, $gte, $lt, $lte, $ne
# Logical operators
filter = {
"$and": [
{"category": "tutorial"},
{"difficulty": {"$lte": 3}}
]
} # Also: $or
# In operator
filter = {"tags": {"$in": ["python", "ml"]}}
```
## Namespaces
```python
# Partition data by namespace
index.upsert(
vectors=[{"id": "vec1", "values": [...]}],
namespace="user-123"
)
# Query specific namespace
results = index.query(
vector=[...],
namespace="user-123",
top_k=5
)
# List namespaces
stats = index.describe_index_stats()
print(stats['namespaces'])
```
## Hybrid search (dense + sparse)
```python
# Upsert with sparse vectors
index.upsert(vectors=[
{
"id": "doc1",
"values": [0.1, 0.2, ...], # Dense vector
"sparse_values": {
"indices": [10, 45, 123], # Token IDs
"values": [0.5, 0.3, 0.8] # TF-IDF scores
},
"metadata": {"text": "..."}
}
])
# Hybrid query
# NOTE: index.query() does NOT accept an `alpha` kwarg. Pinecone stores a
# single sparse-dense vector, so weighting must be applied by pre-scaling the
# query vectors before sending them. Use the hybrid_score_norm helper below
# (alpha * dense + (1 - alpha) * sparse; alpha=1 → pure dense, 0 → pure sparse).
def hybrid_score_norm(dense, sparse, alpha: float):
"""Scale dense/sparse query vectors for weighted hybrid search."""
if not 0 <= alpha <= 1:
raise ValueError("alpha must be between 0 and 1")
scaled_sparse = {
"indices": sparse["indices"],
"values": [v * (1 - alpha) for v in sparse["values"]],
}
return [v * alpha for v in dense], scaled_sparse
hdense, hsparse = hybrid_score_norm(
dense=[0.1, 0.2, ...],
sparse={"indices": [10, 45], "values": [0.5, 0.3]},
alpha=0.5, # 0=sparse, 1=dense, 0.5=balanced
)
results = index.query(
vector=hdense,
sparse_vector=hsparse,
top_k=5,
)
```
## LangChain integration
```python
from langchain_pinecone import PineconeVectorStore
from langchain_openai import OpenAIEmbeddings
# Create vector store
vectorstore = PineconeVectorStore.from_documents(
documents=docs,
embedding=OpenAIEmbeddings(),
index_name="my-index"
)
# Query
results = vectorstore.similarity_search("query", k=5)
# With metadata filter
results = vectorstore.similarity_search(
"query",
k=5,
filter={"category": "tutorial"}
)
# As retriever
retriever = vectorstore.as_retriever(search_kwargs={"k": 10})
```
## LlamaIndex integration
```python
from llama_index.vector_stores.pinecone import PineconeVectorStore
# Connect to Pinecone
pc = Pinecone(api_key="your-key")
pinecone_index = pc.Index("my-index")
# Create vector store
vector_store = PineconeVectorStore(pinecone_index=pinecone_index)
# Use in LlamaIndex
from llama_index.core import StorageContext, VectorStoreIndex
storage_context = StorageContext.from_defaults(vector_store=vector_store)
index = VectorStoreIndex.from_documents(documents, storage_context=storage_context)
```
## Index management
```python
# List indices
indexes = pc.list_indexes()
# Describe index
index_info = pc.describe_index("my-index")
print(index_info)
# Get index stats
stats = index.describe_index_stats()
print(f"Total vectors: {stats['total_vector_count']}")
print(f"Namespaces: {stats['namespaces']}")
# Delete index
pc.delete_index("my-index")
```
## Delete vectors
```python
# Delete by ID
index.delete(ids=["vec1", "vec2"])
# Delete by filter
index.delete(filter={"category": "old"})
# Delete all in namespace
index.delete(delete_all=True, namespace="test")
# Delete entire index
index.delete(delete_all=True)
```
## Best practices
1. **Use serverless** - Auto-scaling, cost-effective
2. **Batch upserts** - More efficient (100-200 per batch)
3. **Add metadata** - Enable filtering
4. **Use namespaces** - Isolate data by user/tenant
5. **Monitor usage** - Check Pinecone dashboard
6. **Optimize filters** - Index frequently filtered fields
7. **Test with free tier** - 1 index, 100K vectors free
8. **Use hybrid search** - Better quality
9. **Set appropriate dimensions** - Match embedding model
10. **Regular backups** - Export important data
## Performance
| Operation | Latency | Notes |
|-----------|---------|-------|
| Upsert | ~50-100ms | Per batch |
| Query (p50) | ~50ms | Depends on index size |
| Query (p95) | ~100ms | SLA target |
| Metadata filter | ~+10-20ms | Additional overhead |
## Pricing (as of 2025)
**Serverless**:
- $0.096 per million read units
- $0.06 per million write units
- $0.06 per GB storage/month
**Free tier**:
- 1 serverless index
- 100K vectors (1536 dimensions)
- Great for prototyping
## Resources
- **Website**: https://www.pinecone.io
- **Docs**: https://docs.pinecone.io
- **Console**: https://app.pinecone.io
- **Pricing**: https://www.pinecone.io/pricing
website/docs/user-guide/skills/optional/mlops/mlops-pytorch-fsdp.md
+97
View File
@@ -0,0 +1,97 @@
---
title: "Pytorch Fsdp — Fully sharded data-parallel training for large models"
sidebar_label: "Pytorch Fsdp"
description: "Fully sharded data-parallel training for large models"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Pytorch Fsdp
Fully sharded data-parallel training for large models.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/pytorch-fsdp` |
| Path | `optional-skills/mlops\pytorch-fsdp` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `torch>=2.0`, `transformers` |
| Platforms | linux, macos |
| Tags | `Distributed Training`, `PyTorch`, `FSDP`, `Data Parallel`, `Sharding`, `Mixed Precision`, `CPU Offloading`, `FSDP2`, `Large-Scale Training` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Pytorch-Fsdp Skill
Assistance with pytorch-fsdp development, generated from official documentation.
## When to Use This Skill
This skill should be triggered when:
- Working with pytorch-fsdp
- Asking about pytorch-fsdp features or APIs
- Implementing pytorch-fsdp solutions
- Debugging pytorch-fsdp code
- Learning pytorch-fsdp best practices
## Quick Reference
The full common-patterns catalog (~157k chars of runnable FSDP snippets) lives in
`references/common-patterns.md` — load it with `read_file` when you need wrapping,
sharding-strategy, checkpoint, or mixed-precision examples. Start there rather than
reconstructing FSDP incantations from memory.
## Reference Files
This skill includes comprehensive documentation in `references/`:
- **other.md** - Other documentation
Use `view` to read specific reference files when detailed information is needed.
## Working with This Skill
### For Beginners
Start with the getting_started or tutorials reference files for foundational concepts.
### For Specific Features
Use the appropriate category reference file (api, guides, etc.) for detailed information.
### For Code Examples
The quick reference section above contains common patterns extracted from the official docs.
## Resources
### references/
Organized documentation extracted from official sources. These files contain:
- Detailed explanations
- Code examples with language annotations
- Links to original documentation
- Table of contents for quick navigation
### scripts/
Add helper scripts here for common automation tasks.
### assets/
Add templates, boilerplate, or example projects here.
## Notes
- This skill was automatically generated from official documentation
- Reference files preserve the structure and examples from source docs
- Code examples include language detection for better syntax highlighting
- Quick reference patterns are extracted from common usage examples in the docs
## Updating
To refresh this skill with updated documentation:
1. Re-run the scraper with the same configuration
2. The skill will be rebuilt with the latest information
website/docs/user-guide/skills/optional/mlops/mlops-pytorch-lightning.md
+365
View File
@@ -0,0 +1,365 @@
---
title: "Pytorch Lightning — Clean training loops with built-in distributed support"
sidebar_label: "Pytorch Lightning"
description: "Clean training loops with built-in distributed support"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Pytorch Lightning
Clean training loops with built-in distributed support.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/pytorch-lightning` |
| Path | `optional-skills/mlops\pytorch-lightning` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `lightning`, `torch`, `transformers` |
| Platforms | linux, macos, windows |
| Tags | `PyTorch Lightning`, `Training Framework`, `Distributed Training`, `DDP`, `FSDP`, `DeepSpeed`, `High-Level API`, `Callbacks`, `Best Practices`, `Scalable` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# PyTorch Lightning - High-Level Training Framework
## Quick start
PyTorch Lightning organizes PyTorch code to eliminate boilerplate while maintaining flexibility.
**Installation**:
```bash
pip install lightning
```
**Convert PyTorch to Lightning** (3 steps):
```python
import lightning as L
import torch
from torch import nn
from torch.utils.data import DataLoader, Dataset
# Step 1: Define LightningModule (organize your PyTorch code)
class LitModel(L.LightningModule):
def __init__(self, hidden_size=128):
super().__init__()
self.model = nn.Sequential(
nn.Linear(28 * 28, hidden_size),
nn.ReLU(),
nn.Linear(hidden_size, 10)
)
def training_step(self, batch, batch_idx):
x, y = batch
y_hat = self.model(x)
loss = nn.functional.cross_entropy(y_hat, y)
self.log('train_loss', loss) # Auto-logged to TensorBoard
return loss
def configure_optimizers(self):
return torch.optim.Adam(self.parameters(), lr=1e-3)
# Step 2: Create data
train_loader = DataLoader(train_dataset, batch_size=32)
# Step 3: Train with Trainer (handles everything else!)
trainer = L.Trainer(max_epochs=10, accelerator='gpu', devices=2)
model = LitModel()
trainer.fit(model, train_loader)
```
**That's it!** Trainer handles:
- GPU/TPU/CPU switching
- Distributed training (DDP, FSDP, DeepSpeed)
- Mixed precision (FP16, BF16)
- Gradient accumulation
- Checkpointing
- Logging
- Progress bars
## Common workflows
### Workflow 1: From PyTorch to Lightning
**Original PyTorch code**:
```python
model = MyModel()
optimizer = torch.optim.Adam(model.parameters())
model.to('cuda')
for epoch in range(max_epochs):
for batch in train_loader:
batch = batch.to('cuda')
optimizer.zero_grad()
loss = model(batch)
loss.backward()
optimizer.step()
```
**Lightning version**:
```python
class LitModel(L.LightningModule):
def __init__(self):
super().__init__()
self.model = MyModel()
def training_step(self, batch, batch_idx):
loss = self.model(batch) # No .to('cuda') needed!
return loss
def configure_optimizers(self):
return torch.optim.Adam(self.parameters())
# Train
trainer = L.Trainer(max_epochs=10, accelerator='gpu')
trainer.fit(LitModel(), train_loader)
```
**Benefits**: 40+ lines → 15 lines, no device management, automatic distributed
### Workflow 2: Validation and testing
```python
class LitModel(L.LightningModule):
def __init__(self):
super().__init__()
self.model = MyModel()
def training_step(self, batch, batch_idx):
x, y = batch
y_hat = self.model(x)
loss = nn.functional.cross_entropy(y_hat, y)
self.log('train_loss', loss)
return loss
def validation_step(self, batch, batch_idx):
x, y = batch
y_hat = self.model(x)
val_loss = nn.functional.cross_entropy(y_hat, y)
acc = (y_hat.argmax(dim=1) == y).float().mean()
self.log('val_loss', val_loss)
self.log('val_acc', acc)
def test_step(self, batch, batch_idx):
x, y = batch
y_hat = self.model(x)
test_loss = nn.functional.cross_entropy(y_hat, y)
self.log('test_loss', test_loss)
def configure_optimizers(self):
return torch.optim.Adam(self.parameters(), lr=1e-3)
# Train with validation
trainer = L.Trainer(max_epochs=10)
trainer.fit(model, train_loader, val_loader)
# Test
trainer.test(model, test_loader)
```
**Automatic features**:
- Validation runs every epoch by default
- Metrics logged to TensorBoard
- Best model checkpointing based on val_loss
### Workflow 3: Distributed training (DDP)
```python
# Same code as single GPU!
model = LitModel()
# 8 GPUs with DDP (automatic!)
trainer = L.Trainer(
accelerator='gpu',
devices=8,
strategy='ddp' # Or 'fsdp', 'deepspeed'
)
trainer.fit(model, train_loader)
```
**Launch**:
```bash
# Single command, Lightning handles the rest
python train.py
```
**No changes needed**:
- Automatic data distribution
- Gradient synchronization
- Multi-node support (just set `num_nodes=2`)
### Workflow 4: Callbacks for monitoring
```python
from lightning.pytorch.callbacks import ModelCheckpoint, EarlyStopping, LearningRateMonitor
# Create callbacks
checkpoint = ModelCheckpoint(
monitor='val_loss',
mode='min',
save_top_k=3,
filename='model-{epoch:02d}-{val_loss:.2f}'
)
early_stop = EarlyStopping(
monitor='val_loss',
patience=5,
mode='min'
)
lr_monitor = LearningRateMonitor(logging_interval='epoch')
# Add to Trainer
trainer = L.Trainer(
max_epochs=100,
callbacks=[checkpoint, early_stop, lr_monitor]
)
trainer.fit(model, train_loader, val_loader)
```
**Result**:
- Auto-saves best 3 models
- Stops early if no improvement for 5 epochs
- Logs learning rate to TensorBoard
### Workflow 5: Learning rate scheduling
```python
class LitModel(L.LightningModule):
# ... (training_step, etc.)
def configure_optimizers(self):
optimizer = torch.optim.Adam(self.parameters(), lr=1e-3)
# Cosine annealing
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizer,
T_max=100,
eta_min=1e-5
)
return {
'optimizer': optimizer,
'lr_scheduler': {
'scheduler': scheduler,
'interval': 'epoch', # Update per epoch
'frequency': 1
}
}
# Learning rate auto-logged!
trainer = L.Trainer(max_epochs=100)
trainer.fit(model, train_loader)
```
## When to use vs alternatives
**Use PyTorch Lightning when**:
- Want clean, organized code
- Need production-ready training loops
- Switching between single GPU, multi-GPU, TPU
- Want built-in callbacks and logging
- Team collaboration (standardized structure)
**Key advantages**:
- **Organized**: Separates research code from engineering
- **Automatic**: DDP, FSDP, DeepSpeed with 1 line
- **Callbacks**: Modular training extensions
- **Reproducible**: Less boilerplate = fewer bugs
- **Tested**: 1M+ downloads/month, battle-tested
**Use alternatives instead**:
- **Accelerate**: Minimal changes to existing code, more flexibility
- **Ray Train**: Multi-node orchestration, hyperparameter tuning
- **Raw PyTorch**: Maximum control, learning purposes
- **Keras**: TensorFlow ecosystem
## Common issues
**Issue: Loss not decreasing**
Check data and model setup:
```python
# Add to training_step
def training_step(self, batch, batch_idx):
if batch_idx == 0:
print(f"Batch shape: {batch[0].shape}")
print(f"Labels: {batch[1]}")
loss = ...
return loss
```
**Issue: Out of memory**
Reduce batch size or use gradient accumulation:
```python
trainer = L.Trainer(
accumulate_grad_batches=4, # Effective batch = batch_size × 4
precision='bf16' # Or 'fp16', reduces memory 50%
)
```
**Issue: Validation not running**
Ensure you pass val_loader:
```python
# WRONG
trainer.fit(model, train_loader)
# CORRECT
trainer.fit(model, train_loader, val_loader)
```
**Issue: DDP spawns multiple processes unexpectedly**
Lightning auto-detects GPUs. Explicitly set devices:
```python
# Test on CPU first
trainer = L.Trainer(accelerator='cpu', devices=1)
# Then GPU
trainer = L.Trainer(accelerator='gpu', devices=1)
```
## Advanced topics
**Callbacks**: See [references/callbacks.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\pytorch-lightning/references/callbacks.md) for EarlyStopping, ModelCheckpoint, custom callbacks, and callback hooks.
**Distributed strategies**: See [references/distributed.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\pytorch-lightning/references/distributed.md) for DDP, FSDP, DeepSpeed ZeRO integration, multi-node setup.
**Hyperparameter tuning**: See [references/hyperparameter-tuning.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\pytorch-lightning/references/hyperparameter-tuning.md) for integration with Optuna, Ray Tune, and WandB sweeps.
## Hardware requirements
- **CPU**: Works (good for debugging)
- **Single GPU**: Works
- **Multi-GPU**: DDP (default), FSDP, or DeepSpeed
- **Multi-node**: DDP, FSDP, DeepSpeed
- **TPU**: Supported (8 cores)
- **Apple MPS**: Supported
**Precision options**:
- FP32 (default)
- FP16 (V100, older GPUs)
- BF16 (A100/H100, recommended)
- FP8 (H100)
## Resources
- Docs: https://lightning.ai/docs/pytorch/stable/
- GitHub: https://github.com/Lightning-AI/pytorch-lightning ⭐ 29,000+
- Version: 2.5.5+
- Examples: https://github.com/Lightning-AI/pytorch-lightning/tree/master/examples
- Discord: https://discord.gg/lightning-ai
- Used by: Kaggle winners, research labs, production teams
website/docs/user-guide/skills/optional/mlops/mlops-qdrant.md
+522
View File
@@ -0,0 +1,522 @@
---
title: "Qdrant — Vector search engine for production RAG systems"
sidebar_label: "Qdrant"
description: "Vector search engine for production RAG systems"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Qdrant
Vector search engine for production RAG systems.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/qdrant` |
| Path | `optional-skills/mlops\qdrant` |
| Version | `1.0.1` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `qdrant-client>=1.14.0` |
| Platforms | linux, macos, windows |
| Tags | `RAG`, `Vector Search`, `Qdrant`, `Semantic Search`, `Embeddings`, `Similarity Search`, `HNSW`, `Production`, `Distributed` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Qdrant - Vector Similarity Search Engine
High-performance vector database written in Rust for production RAG and semantic search.
## When to use Qdrant
**Use Qdrant when:**
- Building production RAG systems requiring low latency
- Need hybrid search (vectors + metadata filtering)
- Require horizontal scaling with sharding/replication
- Want on-premise deployment with full data control
- Need multi-vector storage per record (dense + sparse)
- Building real-time recommendation systems
**Key features:**
- **Rust-powered**: Memory-safe, high performance
- **Rich filtering**: Filter by any payload field during search
- **Multiple vectors**: Dense, sparse, multi-dense per point
- **Quantization**: Scalar, product, binary for memory efficiency
- **Distributed**: Raft consensus, sharding, replication
- **REST + gRPC**: Both APIs with full feature parity
**Use alternatives instead:**
- **Chroma**: Simpler setup, embedded use cases
- **FAISS**: Maximum raw speed, research/batch processing
- **Pinecone**: Fully managed, zero ops preferred
- **Weaviate**: GraphQL preference, built-in vectorizers
## Quick start
### Installation
```bash
# Python client
pip install qdrant-client
# Docker (recommended for development)
docker run -p 6333:6333 -p 6334:6334 qdrant/qdrant
# Docker with persistent storage
docker run -p 6333:6333 -p 6334:6334 \
-v $(pwd)/qdrant_storage:/qdrant/storage \
qdrant/qdrant
```
### Basic usage
```python
from qdrant_client import QdrantClient
from qdrant_client.models import Distance, VectorParams, PointStruct
# Connect to Qdrant
client = QdrantClient(host="localhost", port=6333)
# Create collection
client.create_collection(
collection_name="documents",
vectors_config=VectorParams(size=384, distance=Distance.COSINE)
)
# Insert vectors with payload
client.upsert(
collection_name="documents",
points=[
PointStruct(
id=1,
vector=[0.1, 0.2, ...], # 384-dim vector
payload={"title": "Doc 1", "category": "tech"}
),
PointStruct(
id=2,
vector=[0.3, 0.4, ...],
payload={"title": "Doc 2", "category": "science"}
)
]
)
# Search with filtering (query_points is the current API; client.search is removed in qdrant-client 1.14+)
response = client.query_points(
collection_name="documents",
query=[0.15, 0.25, ...],
query_filter={
"must": [{"key": "category", "match": {"value": "tech"}}]
},
limit=10
)
for point in response.points:
print(f"ID: {point.id}, Score: {point.score}, Payload: {point.payload}")
```
## Core concepts
### Points - Basic data unit
```python
from qdrant_client.models import PointStruct
# Point = ID + Vector(s) + Payload
point = PointStruct(
id=123, # Integer or UUID string
vector=[0.1, 0.2, 0.3, ...], # Dense vector
payload={ # Arbitrary JSON metadata
"title": "Document title",
"category": "tech",
"timestamp": 1699900000,
"tags": ["python", "ml"]
}
)
# Batch upsert (recommended)
client.upsert(
collection_name="documents",
points=[point1, point2, point3],
wait=True # Wait for indexing
)
```
### Collections - Vector containers
```python
from qdrant_client.models import VectorParams, Distance, HnswConfigDiff
# Create with HNSW configuration
client.create_collection(
collection_name="documents",
vectors_config=VectorParams(
size=384, # Vector dimensions
distance=Distance.COSINE # COSINE, EUCLID, DOT, MANHATTAN
),
hnsw_config=HnswConfigDiff(
m=16, # Connections per node (default 16)
ef_construct=100, # Build-time accuracy (default 100)
full_scan_threshold=10000 # Switch to brute force below this
),
on_disk_payload=True # Store payload on disk
)
# Collection info
info = client.get_collection("documents")
print(f"Points: {info.points_count}, Vectors: {info.vectors_count}")
```
### Distance metrics
| Metric | Use Case | Range |
|--------|----------|-------|
| `COSINE` | Text embeddings, normalized vectors | 0 to 2 |
| `EUCLID` | Spatial data, image features | 0 to ∞ |
| `DOT` | Recommendations, unnormalized | -∞ to ∞ |
| `MANHATTAN` | Sparse features, discrete data | 0 to ∞ |
## Search operations
### Basic search
```python
# Simple nearest neighbor search (returns a QueryResponse; use .points)
response = client.query_points(
collection_name="documents",
query=[0.1, 0.2, ...],
limit=10,
with_payload=True,
with_vectors=False # Don't return vectors (faster)
)
results = response.points
```
### Filtered search
```python
from qdrant_client.models import Filter, FieldCondition, MatchValue, Range
# Complex filtering
response = client.query_points(
collection_name="documents",
query=query_embedding,
query_filter=Filter(
must=[
FieldCondition(key="category", match=MatchValue(value="tech")),
FieldCondition(key="timestamp", range=Range(gte=1699000000))
],
must_not=[
FieldCondition(key="status", match=MatchValue(value="archived"))
]
),
limit=10
).points
# Shorthand filter syntax
response = client.query_points(
collection_name="documents",
query=query_embedding,
query_filter={
"must": [
{"key": "category", "match": {"value": "tech"}},
{"key": "price", "range": {"gte": 10, "lte": 100}}
]
},
limit=10
).points
```
### Batch search
```python
from qdrant_client.models import QueryRequest
# Multiple queries in one request (search_batch is replaced by query_batch_points)
responses = client.query_batch_points(
collection_name="documents",
requests=[
QueryRequest(query=[0.1, ...], limit=5),
QueryRequest(query=[0.2, ...], limit=5, filter={"must": [...]}),
QueryRequest(query=[0.3, ...], limit=10)
]
)
# Each element is a QueryResponse; use .points
for resp in responses:
for point in resp.points:
print(point.id, point.score)
```
## RAG integration
### With sentence-transformers
```python
from sentence_transformers import SentenceTransformer
from qdrant_client import QdrantClient
from qdrant_client.models import VectorParams, Distance, PointStruct
# Initialize
encoder = SentenceTransformer("all-MiniLM-L6-v2")
client = QdrantClient(host="localhost", port=6333)
# Create collection
client.create_collection(
collection_name="knowledge_base",
vectors_config=VectorParams(size=384, distance=Distance.COSINE)
)
# Index documents
documents = [
{"id": 1, "text": "Python is a programming language", "source": "wiki"},
{"id": 2, "text": "Machine learning uses algorithms", "source": "textbook"},
]
points = [
PointStruct(
id=doc["id"],
vector=encoder.encode(doc["text"]).tolist(),
payload={"text": doc["text"], "source": doc["source"]}
)
for doc in documents
]
client.upsert(collection_name="knowledge_base", points=points)
# RAG retrieval
def retrieve(query: str, top_k: int = 5) -> list[dict]:
query_vector = encoder.encode(query).tolist()
response = client.query_points(
collection_name="knowledge_base",
query=query_vector,
limit=top_k
)
return [{"text": r.payload["text"], "score": r.score} for r in response.points]
# Use in RAG pipeline
context = retrieve("What is Python?")
prompt = f"Context: {context}\n\nQuestion: What is Python?"
```
### With LangChain
```python
from langchain_community.vectorstores import Qdrant
from langchain_community.embeddings import HuggingFaceEmbeddings
embeddings = HuggingFaceEmbeddings(model_name="all-MiniLM-L6-v2")
vectorstore = Qdrant.from_documents(documents, embeddings, url="http://localhost:6333", collection_name="docs")
retriever = vectorstore.as_retriever(search_kwargs={"k": 5})
```
### With LlamaIndex
```python
from llama_index.vector_stores.qdrant import QdrantVectorStore
from llama_index.core import VectorStoreIndex, StorageContext
vector_store = QdrantVectorStore(client=client, collection_name="llama_docs")
storage_context = StorageContext.from_defaults(vector_store=vector_store)
index = VectorStoreIndex.from_documents(documents, storage_context=storage_context)
query_engine = index.as_query_engine()
```
## Multi-vector support
### Named vectors (different embedding models)
```python
from qdrant_client.models import VectorParams, Distance
# Collection with multiple vector types
client.create_collection(
collection_name="hybrid_search",
vectors_config={
"dense": VectorParams(size=384, distance=Distance.COSINE),
"sparse": VectorParams(size=30000, distance=Distance.DOT)
}
)
# Insert with named vectors
client.upsert(
collection_name="hybrid_search",
points=[
PointStruct(
id=1,
vector={
"dense": dense_embedding,
"sparse": sparse_embedding
},
payload={"text": "document text"}
)
]
)
# Search specific named vector (pass the vector name via `using`)
response = client.query_points(
collection_name="hybrid_search",
query=query_dense,
using="dense", # Specify which named vector to search
limit=10
)
results = response.points
```
### Sparse vectors (BM25, SPLADE)
```python
from qdrant_client.models import SparseVectorParams, SparseIndexParams, SparseVector
# Collection with sparse vectors
client.create_collection(
collection_name="sparse_search",
vectors_config={},
sparse_vectors_config={"text": SparseVectorParams(index=SparseIndexParams(on_disk=False))}
)
# Insert sparse vector
client.upsert(
collection_name="sparse_search",
points=[PointStruct(id=1, vector={"text": SparseVector(indices=[1, 5, 100], values=[0.5, 0.8, 0.2])}, payload={"text": "document"})]
)
```
## Quantization (memory optimization)
```python
from qdrant_client.models import ScalarQuantization, ScalarQuantizationConfig, ScalarType
# Scalar quantization (4x memory reduction)
client.create_collection(
collection_name="quantized",
vectors_config=VectorParams(size=384, distance=Distance.COSINE),
quantization_config=ScalarQuantization(
scalar=ScalarQuantizationConfig(
type=ScalarType.INT8,
quantile=0.99, # Clip outliers
always_ram=True # Keep quantized in RAM
)
)
)
# Search with rescoring
response = client.query_points(
collection_name="quantized",
query=query,
search_params={"quantization": {"rescore": True}}, # Rescore top results
limit=10
)
results = response.points
```
## Payload indexing
```python
from qdrant_client.models import PayloadSchemaType
# Create payload index for faster filtering
client.create_payload_index(
collection_name="documents",
field_name="category",
field_schema=PayloadSchemaType.KEYWORD
)
client.create_payload_index(
collection_name="documents",
field_name="timestamp",
field_schema=PayloadSchemaType.INTEGER
)
# Index types: KEYWORD, INTEGER, FLOAT, GEO, TEXT (full-text), BOOL
```
## Production deployment
### Qdrant Cloud
```python
from qdrant_client import QdrantClient
# Connect to Qdrant Cloud
client = QdrantClient(
url="https://your-cluster.cloud.qdrant.io",
api_key="your-api-key"
)
```
### Performance tuning
```python
# Optimize for search speed (higher recall)
client.update_collection(
collection_name="documents",
hnsw_config=HnswConfigDiff(ef_construct=200, m=32)
)
# Optimize for indexing speed (bulk loads)
client.update_collection(
collection_name="documents",
optimizer_config={"indexing_threshold": 20000}
)
```
## Best practices
1. **Batch operations** - Use batch upsert/search for efficiency
2. **Payload indexing** - Index fields used in filters
3. **Quantization** - Enable for large collections (>1M vectors)
4. **Sharding** - Use for collections >10M vectors
5. **On-disk storage** - Enable `on_disk_payload` for large payloads
6. **Connection pooling** - Reuse client instances
## Common issues
**Slow search with filters:**
```python
# Create payload index for filtered fields
client.create_payload_index(
collection_name="docs",
field_name="category",
field_schema=PayloadSchemaType.KEYWORD
)
```
**Out of memory:**
```python
# Enable quantization and on-disk storage
client.create_collection(
collection_name="large_collection",
vectors_config=VectorParams(size=384, distance=Distance.COSINE),
quantization_config=ScalarQuantization(...),
on_disk_payload=True
)
```
**Connection issues:**
```python
# Use timeout and retry
client = QdrantClient(
host="localhost",
port=6333,
timeout=30,
prefer_grpc=True # gRPC for better performance
)
```
## References
- **[Advanced Usage](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\qdrant/references/advanced-usage.md)** - Distributed mode, hybrid search, recommendations
- **[Troubleshooting](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\qdrant/references/troubleshooting.md)** - Common issues, debugging, performance tuning
## Resources
- **GitHub**: https://github.com/qdrant/qdrant (22k+ stars)
- **Docs**: https://qdrant.tech/documentation/
- **Python Client**: https://github.com/qdrant/qdrant-client
- **Cloud**: https://cloud.qdrant.io
- **Version**: 1.14.0+
- **License**: Apache 2.0
website/docs/user-guide/skills/optional/mlops/mlops-research-dspy.md
+609
View File
@@ -0,0 +1,609 @@
---
title: "Dspy — DSPy: declarative LM programs, auto-optimize prompts, RAG"
sidebar_label: "Dspy"
description: "DSPy: declarative LM programs, auto-optimize prompts, RAG"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Dspy
DSPy: declarative LM programs, auto-optimize prompts, RAG.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/dspy` |
| Path | `optional-skills/mlops\research\dspy` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `dspy`, `openai`, `anthropic` |
| Platforms | linux, macos, windows |
| Tags | `Prompt Engineering`, `DSPy`, `Declarative Programming`, `RAG`, `Agents`, `Prompt Optimization`, `LM Programming`, `Stanford NLP`, `Automatic Optimization`, `Modular AI` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# DSPy: Declarative Language Model Programming
## When to Use This Skill
Use DSPy when you need to:
- **Build complex AI systems** with multiple components and workflows
- **Program LMs declaratively** instead of manual prompt engineering
- **Optimize prompts automatically** using data-driven methods
- **Create modular AI pipelines** that are maintainable and portable
- **Improve model outputs systematically** with optimizers
- **Build RAG systems, agents, or classifiers** with better reliability
**GitHub Stars**: 22,000+ | **Created By**: Stanford NLP
## Installation
```bash
# Stable release
pip install dspy
# Latest development version
pip install git+https://github.com/stanfordnlp/dspy.git
# With specific LM providers
pip install dspy[openai] # OpenAI
pip install dspy[anthropic] # Anthropic Claude
pip install dspy[all] # All providers
```
## Quick Start
### Basic Example: Question Answering
```python
import dspy
# Configure your language model
lm = dspy.Claude(model="claude-sonnet-4-5-20250929")
dspy.settings.configure(lm=lm)
# Define a signature (input → output)
class QA(dspy.Signature):
"""Answer questions with short factual answers."""
question = dspy.InputField()
answer = dspy.OutputField(desc="often between 1 and 5 words")
# Create a module
qa = dspy.Predict(QA)
# Use it
response = qa(question="What is the capital of France?")
print(response.answer) # "Paris"
```
### Chain of Thought Reasoning
```python
import dspy
lm = dspy.Claude(model="claude-sonnet-4-5-20250929")
dspy.settings.configure(lm=lm)
# Use ChainOfThought for better reasoning
class MathProblem(dspy.Signature):
"""Solve math word problems."""
problem = dspy.InputField()
answer = dspy.OutputField(desc="numerical answer")
# ChainOfThought generates reasoning steps automatically
cot = dspy.ChainOfThought(MathProblem)
response = cot(problem="If John has 5 apples and gives 2 to Mary, how many does he have?")
print(response.rationale) # Shows reasoning steps
print(response.answer) # "3"
```
## Core Concepts
### 1. Signatures
Signatures define the structure of your AI task (inputs → outputs):
```python
# Inline signature (simple)
qa = dspy.Predict("question -> answer")
# Class signature (detailed)
class Summarize(dspy.Signature):
"""Summarize text into key points."""
text = dspy.InputField()
summary = dspy.OutputField(desc="bullet points, 3-5 items")
summarizer = dspy.ChainOfThought(Summarize)
```
**When to use each:**
- **Inline**: Quick prototyping, simple tasks
- **Class**: Complex tasks, type hints, better documentation
### 2. Modules
Modules are reusable components that transform inputs to outputs:
#### dspy.Predict
Basic prediction module:
```python
predictor = dspy.Predict("context, question -> answer")
result = predictor(context="Paris is the capital of France",
question="What is the capital?")
```
#### dspy.ChainOfThought
Generates reasoning steps before answering:
```python
cot = dspy.ChainOfThought("question -> answer")
result = cot(question="Why is the sky blue?")
print(result.rationale) # Reasoning steps
print(result.answer) # Final answer
```
#### dspy.ReAct
Agent-like reasoning with tools:
```python
from dspy.predict import ReAct
class SearchQA(dspy.Signature):
"""Answer questions using search."""
question = dspy.InputField()
answer = dspy.OutputField()
def search_tool(query: str) -> str:
"""Search Wikipedia."""
# Your search implementation
return results
react = ReAct(SearchQA, tools=[search_tool])
result = react(question="When was Python created?")
```
#### dspy.ProgramOfThought
Generates and executes code for reasoning:
```python
pot = dspy.ProgramOfThought("question -> answer")
result = pot(question="What is 15% of 240?")
# Generates: answer = 240 * 0.15
```
### 3. Optimizers
Optimizers improve your modules automatically using training data:
#### BootstrapFewShot
Learns from examples:
```python
from dspy.teleprompt import BootstrapFewShot
# Training data
trainset = [
dspy.Example(question="What is 2+2?", answer="4").with_inputs("question"),
dspy.Example(question="What is 3+5?", answer="8").with_inputs("question"),
]
# Define metric
def validate_answer(example, pred, trace=None):
return example.answer == pred.answer
# Optimize
optimizer = BootstrapFewShot(metric=validate_answer, max_bootstrapped_demos=3)
optimized_qa = optimizer.compile(qa, trainset=trainset)
# Now optimized_qa performs better!
```
#### MIPRO (Most Important Prompt Optimization)
Iteratively improves prompts:
```python
from dspy.teleprompt import MIPRO
optimizer = MIPRO(
metric=validate_answer,
num_candidates=10,
init_temperature=1.0
)
optimized_cot = optimizer.compile(
cot,
trainset=trainset,
num_trials=100
)
```
#### BootstrapFinetune
Creates datasets for model fine-tuning:
```python
from dspy.teleprompt import BootstrapFinetune
optimizer = BootstrapFinetune(metric=validate_answer)
optimized_module = optimizer.compile(qa, trainset=trainset)
# Exports training data for fine-tuning
```
### 4. Building Complex Systems
#### Multi-Stage Pipeline
```python
import dspy
class MultiHopQA(dspy.Module):
def __init__(self):
super().__init__()
self.retrieve = dspy.Retrieve(k=3)
self.generate_query = dspy.ChainOfThought("question -> search_query")
self.generate_answer = dspy.ChainOfThought("context, question -> answer")
def forward(self, question):
# Stage 1: Generate search query
search_query = self.generate_query(question=question).search_query
# Stage 2: Retrieve context
passages = self.retrieve(search_query).passages
context = "\n".join(passages)
# Stage 3: Generate answer
answer = self.generate_answer(context=context, question=question).answer
return dspy.Prediction(answer=answer, context=context)
# Use the pipeline
qa_system = MultiHopQA()
result = qa_system(question="Who wrote the book that inspired the movie Blade Runner?")
```
#### RAG System with Optimization
```python
import dspy
from dspy.retrieve.chromadb_rm import ChromadbRM
# Configure retriever
retriever = ChromadbRM(
collection_name="documents",
persist_directory="./chroma_db"
)
class RAG(dspy.Module):
def __init__(self, num_passages=3):
super().__init__()
self.retrieve = dspy.Retrieve(k=num_passages)
self.generate = dspy.ChainOfThought("context, question -> answer")
def forward(self, question):
context = self.retrieve(question).passages
return self.generate(context=context, question=question)
# Create and optimize
rag = RAG()
# Optimize with training data
from dspy.teleprompt import BootstrapFewShot
optimizer = BootstrapFewShot(metric=validate_answer)
optimized_rag = optimizer.compile(rag, trainset=trainset)
```
## LM Provider Configuration
### Anthropic Claude
```python
import dspy
lm = dspy.Claude(
model="claude-sonnet-4-5-20250929",
api_key="your-api-key", # Or set ANTHROPIC_API_KEY env var
max_tokens=1000,
temperature=0.7
)
dspy.settings.configure(lm=lm)
```
### OpenAI
```python
lm = dspy.OpenAI(
model="gpt-4",
api_key="your-api-key",
max_tokens=1000
)
dspy.settings.configure(lm=lm)
```
### Local Models (Ollama)
```python
lm = dspy.OllamaLocal(
model="llama3.1",
base_url="http://localhost:11434"
)
dspy.settings.configure(lm=lm)
```
### Multiple Models
```python
# Different models for different tasks
cheap_lm = dspy.OpenAI(model="gpt-3.5-turbo")
strong_lm = dspy.Claude(model="claude-sonnet-4-5-20250929")
# Use cheap model for retrieval, strong model for reasoning
with dspy.settings.context(lm=cheap_lm):
context = retriever(question)
with dspy.settings.context(lm=strong_lm):
answer = generator(context=context, question=question)
```
## Common Patterns
### Pattern 1: Structured Output
```python
from pydantic import BaseModel, Field
class PersonInfo(BaseModel):
name: str = Field(description="Full name")
age: int = Field(description="Age in years")
occupation: str = Field(description="Current job")
class ExtractPerson(dspy.Signature):
"""Extract person information from text."""
text = dspy.InputField()
person: PersonInfo = dspy.OutputField()
extractor = dspy.TypedPredictor(ExtractPerson)
result = extractor(text="John Doe is a 35-year-old software engineer.")
print(result.person.name) # "John Doe"
print(result.person.age) # 35
```
### Pattern 2: Assertion-Driven Optimization
```python
import dspy
from dspy.primitives.assertions import assert_transform_module, backtrack_handler
class MathQA(dspy.Module):
def __init__(self):
super().__init__()
self.solve = dspy.ChainOfThought("problem -> solution: float")
def forward(self, problem):
solution = self.solve(problem=problem).solution
# Assert solution is numeric
dspy.Assert(
isinstance(float(solution), float),
"Solution must be a number",
backtrack=backtrack_handler
)
return dspy.Prediction(solution=solution)
```
### Pattern 3: Self-Consistency
```python
import dspy
from collections import Counter
class ConsistentQA(dspy.Module):
def __init__(self, num_samples=5):
super().__init__()
self.qa = dspy.ChainOfThought("question -> answer")
self.num_samples = num_samples
def forward(self, question):
# Generate multiple answers
answers = []
for _ in range(self.num_samples):
result = self.qa(question=question)
answers.append(result.answer)
# Return most common answer
most_common = Counter(answers).most_common(1)[0][0]
return dspy.Prediction(answer=most_common)
```
### Pattern 4: Retrieval with Reranking
```python
class RerankedRAG(dspy.Module):
def __init__(self):
super().__init__()
self.retrieve = dspy.Retrieve(k=10)
self.rerank = dspy.Predict("question, passage -> relevance_score: float")
self.answer = dspy.ChainOfThought("context, question -> answer")
def forward(self, question):
# Retrieve candidates
passages = self.retrieve(question).passages
# Rerank passages
scored = []
for passage in passages:
score = float(self.rerank(question=question, passage=passage).relevance_score)
scored.append((score, passage))
# Take top 3
top_passages = [p for _, p in sorted(scored, reverse=True)[:3]]
context = "\n\n".join(top_passages)
# Generate answer
return self.answer(context=context, question=question)
```
## Evaluation and Metrics
### Custom Metrics
```python
def exact_match(example, pred, trace=None):
"""Exact match metric."""
return example.answer.lower() == pred.answer.lower()
def f1_score(example, pred, trace=None):
"""F1 score for text overlap."""
pred_tokens = set(pred.answer.lower().split())
gold_tokens = set(example.answer.lower().split())
if not pred_tokens:
return 0.0
precision = len(pred_tokens & gold_tokens) / len(pred_tokens)
recall = len(pred_tokens & gold_tokens) / len(gold_tokens)
if precision + recall == 0:
return 0.0
return 2 * (precision * recall) / (precision + recall)
```
### Evaluation
```python
from dspy.evaluate import Evaluate
# Create evaluator
evaluator = Evaluate(
devset=testset,
metric=exact_match,
num_threads=4,
display_progress=True
)
# Evaluate model
score = evaluator(qa_system)
print(f"Accuracy: {score}")
# Compare optimized vs unoptimized
score_before = evaluator(qa)
score_after = evaluator(optimized_qa)
print(f"Improvement: {score_after - score_before:.2%}")
```
## Best Practices
### 1. Start Simple, Iterate
```python
# Start with Predict
qa = dspy.Predict("question -> answer")
# Add reasoning if needed
qa = dspy.ChainOfThought("question -> answer")
# Add optimization when you have data
optimized_qa = optimizer.compile(qa, trainset=data)
```
### 2. Use Descriptive Signatures
```python
# ❌ Bad: Vague
class Task(dspy.Signature):
input = dspy.InputField()
output = dspy.OutputField()
# ✅ Good: Descriptive
class SummarizeArticle(dspy.Signature):
"""Summarize news articles into 3-5 key points."""
article = dspy.InputField(desc="full article text")
summary = dspy.OutputField(desc="bullet points, 3-5 items")
```
### 3. Optimize with Representative Data
```python
# Create diverse training examples
trainset = [
dspy.Example(question="factual", answer="...).with_inputs("question"),
dspy.Example(question="reasoning", answer="...").with_inputs("question"),
dspy.Example(question="calculation", answer="...").with_inputs("question"),
]
# Use validation set for metric
def metric(example, pred, trace=None):
return example.answer in pred.answer
```
### 4. Save and Load Optimized Models
```python
# Save
optimized_qa.save("models/qa_v1.json")
# Load
loaded_qa = dspy.ChainOfThought("question -> answer")
loaded_qa.load("models/qa_v1.json")
```
### 5. Monitor and Debug
```python
# Enable tracing
dspy.settings.configure(lm=lm, trace=[])
# Run prediction
result = qa(question="...")
# Inspect trace
for call in dspy.settings.trace:
print(f"Prompt: {call['prompt']}")
print(f"Response: {call['response']}")
```
## Comparison to Other Approaches
| Feature | Manual Prompting | LangChain | DSPy |
|---------|-----------------|-----------|------|
| Prompt Engineering | Manual | Manual | Automatic |
| Optimization | Trial & error | None | Data-driven |
| Modularity | Low | Medium | High |
| Type Safety | No | Limited | Yes (Signatures) |
| Portability | Low | Medium | High |
| Learning Curve | Low | Medium | Medium-High |
**When to choose DSPy:**
- You have training data or can generate it
- You need systematic prompt improvement
- You're building complex multi-stage systems
- You want to optimize across different LMs
**When to choose alternatives:**
- Quick prototypes (manual prompting)
- Simple chains with existing tools (LangChain)
- Custom optimization logic needed
## Resources
- **Documentation**: https://dspy.ai
- **GitHub**: https://github.com/stanfordnlp/dspy (22k+ stars)
- **Discord**: https://discord.gg/XCGy2WDCQB
- **Twitter**: @DSPyOSS
- **Paper**: "DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines"
## See Also
- `references/modules.md` - Detailed module guide (Predict, ChainOfThought, ReAct, ProgramOfThought)
- `references/optimizers.md` - Optimization algorithms (BootstrapFewShot, MIPRO, BootstrapFinetune)
- `references/examples.md` - Real-world examples (RAG, agents, classifiers)
website/docs/user-guide/skills/optional/mlops/mlops-saelens.md
+439
View File
@@ -0,0 +1,439 @@
---
title: "Saelens — Train sparse autoencoders to interpret model features"
sidebar_label: "Saelens"
description: "Train sparse autoencoders to interpret model features"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Saelens
Train sparse autoencoders to interpret model features.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/saelens` |
| Path | `optional-skills/mlops\saelens` |
| Version | `1.0.1` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `sae-lens>=6.0.0`, `transformer-lens>=2.0.0`, `torch>=2.0.0` |
| Platforms | linux, macos, windows |
| Tags | `Sparse Autoencoders`, `SAE`, `Mechanistic Interpretability`, `Feature Discovery`, `Superposition` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# SAELens: Sparse Autoencoders for Mechanistic Interpretability
SAELens is the primary library for training and analyzing Sparse Autoencoders (SAEs) - a technique for decomposing polysemantic neural network activations into sparse, interpretable features. Based on Anthropic's groundbreaking research on monosemanticity.
**GitHub**: [jbloomAus/SAELens](https://github.com/jbloomAus/SAELens) (1,100+ stars)
## The Problem: Polysemanticity & Superposition
Individual neurons in neural networks are **polysemantic** - they activate in multiple, semantically distinct contexts. This happens because models use **superposition** to represent more features than they have neurons, making interpretability difficult.
**SAEs solve this** by decomposing dense activations into sparse, monosemantic features - typically only a small number of features activate for any given input, and each feature corresponds to an interpretable concept.
## When to Use SAELens
**Use SAELens when you need to:**
- Discover interpretable features in model activations
- Understand what concepts a model has learned
- Study superposition and feature geometry
- Perform feature-based steering or ablation
- Analyze safety-relevant features (deception, bias, harmful content)
**Consider alternatives when:**
- You need basic activation analysis → Use **TransformerLens** directly
- You want causal intervention experiments → Use **pyvene** or **TransformerLens**
- You need production steering → Consider direct activation engineering
## Installation
```bash
pip install sae-lens
```
Requirements: Python 3.10+, transformer-lens>=2.0.0
## Core Concepts
### What SAEs Learn
SAEs are trained to reconstruct model activations through a sparse bottleneck:
```
Input Activation → Encoder → Sparse Features → Decoder → Reconstructed Activation
(d_model) ↓ (d_sae >> d_model) ↓ (d_model)
sparsity reconstruction
penalty loss
```
**Loss Function**: `MSE(original, reconstructed) + L1_coefficient × L1(features)`
### Key Validation (Anthropic Research)
In "Towards Monosemanticity", human evaluators found **70% of SAE features genuinely interpretable**. Features discovered include:
- DNA sequences, legal language, HTTP requests
- Hebrew text, nutrition statements, code syntax
- Sentiment, named entities, grammatical structures
## Workflow 1: Loading and Analyzing Pre-trained SAEs
### Step-by-Step
```python
from transformer_lens import HookedTransformer
from sae_lens import SAE
# 1. Load model and pre-trained SAE
model = HookedTransformer.from_pretrained("gpt2-small", device="cuda")
# In sae-lens v6, SAE.from_pretrained() returns JUST the SAE (not a tuple).
sae = SAE.from_pretrained(
release="gpt2-small-res-jb",
sae_id="blocks.8.hook_resid_pre",
device="cuda"
)
# If you also need the cfg dict and feature sparsity, use:
# sae, cfg_dict, sparsity = SAE.from_pretrained_with_cfg_and_sparsity(...)
# 2. Get model activations
tokens = model.to_tokens("The capital of France is Paris")
_, cache = model.run_with_cache(tokens)
activations = cache["resid_pre", 8] # [batch, pos, d_model]
# 3. Encode to SAE features
sae_features = sae.encode(activations) # [batch, pos, d_sae]
print(f"Active features: {(sae_features > 0).sum()}")
# 4. Find top features for each position
for pos in range(tokens.shape[1]):
top_features = sae_features[0, pos].topk(5)
token = model.to_str_tokens(tokens[0, pos:pos+1])[0]
print(f"Token '{token}': features {top_features.indices.tolist()}")
# 5. Reconstruct activations
reconstructed = sae.decode(sae_features)
reconstruction_error = (activations - reconstructed).norm()
```
### Available Pre-trained SAEs
| Release | Model | Layers |
|---------|-------|--------|
| `gpt2-small-res-jb` | GPT-2 Small | Multiple residual streams |
| `gemma-2b-res` | Gemma 2B | Residual streams |
| Various on HuggingFace | Search tag `saelens` | Various |
### Checklist
- [ ] Load model with TransformerLens
- [ ] Load matching SAE for target layer
- [ ] Encode activations to sparse features
- [ ] Identify top-activating features per token
- [ ] Validate reconstruction quality
## Workflow 2: Training a Custom SAE
### Step-by-Step
```python
from sae_lens import (
LanguageModelSAETrainingRunner,
LanguageModelSAERunnerConfig,
StandardTrainingSAEConfig,
LoggingConfig,
)
# 1. Configure training (v6 uses a NESTED config: SAE-specific options live in a
# `sae=` sub-config, and logging options live in a `logger=` sub-config).
# Note: `architecture`, `d_sae`, `l1_coefficient` etc. are now on the SAE sub-config,
# and legacy flat options like `hook_layer`, `activation_fn`, `log_to_wandb` were removed.
cfg = LanguageModelSAERunnerConfig(
# SAE architecture + sparsity (nested)
sae=StandardTrainingSAEConfig(
d_in=768, # Model dimension
d_sae=768 * 8, # Expansion factor of 8
l1_coefficient=8e-5, # Sparsity penalty
apply_b_dec_to_input=True,
normalize_activations="expected_average_only_in",
),
# Data-generating function (model + hook point)
model_name="gpt2-small",
hook_name="blocks.8.hook_resid_pre", # layer is inferred from hook_name (no hook_layer)
# Training
lr=4e-4,
l1_warm_up_steps=1000,
train_batch_size_tokens=4096,
training_tokens=100_000_000,
# Data
dataset_path="monology/pile-uncopyrighted",
context_size=128,
# Logging (nested)
logger=LoggingConfig(
log_to_wandb=True,
wandb_project="sae-training",
),
# Checkpointing
checkpoint_path="checkpoints",
n_checkpoints=5,
)
# 2. Train
trainer = LanguageModelSAETrainingRunner(cfg) # SAETrainingRunner still works as an alias
sae = trainer.run()
# 3. Evaluate
print(f"L0 (avg active features): {trainer.metrics['l0']}")
print(f"CE Loss Recovered: {trainer.metrics['ce_loss_score']}")
```
> **v6 migration note:** For other SAE types swap the `sae=` sub-config —
> `GatedTrainingSAEConfig`, `TopKTrainingSAEConfig` (set `k` directly), or
> `JumpReLUTrainingSAEConfig` (uses `l0_coefficient`). Legacy flat options
> (`architecture`, `expansion_factor`, `hook_layer`, `activation_fn`/`activation_fn_kwargs`,
> `use_ghost_grads`, ghost grads, b_dec/decoder init options) were removed in v6.
### Key Hyperparameters
| Parameter | Typical Value | Effect |
|-----------|---------------|--------|
| `d_sae` | 4-16× d_model | More features, higher capacity |
| `l1_coefficient` | 5e-5 to 1e-4 | Higher = sparser, less accurate |
| `lr` | 1e-4 to 1e-3 | Standard optimizer LR |
| `l1_warm_up_steps` | 500-2000 | Prevents early feature death |
### Evaluation Metrics
| Metric | Target | Meaning |
|--------|--------|---------|
| **L0** | 50-200 | Average active features per token |
| **CE Loss Score** | 80-95% | Cross-entropy recovered vs original |
| **Dead Features** | &lt;5% | Features that never activate |
| **Explained Variance** | >90% | Reconstruction quality |
### Checklist
- [ ] Choose target layer and hook point
- [ ] Set expansion factor (d_sae = 4-16× d_model)
- [ ] Tune L1 coefficient for desired sparsity
- [ ] Enable L1 warm-up to prevent dead features
- [ ] Monitor metrics during training (W&B)
- [ ] Validate L0 and CE loss recovery
- [ ] Check dead feature ratio
## Workflow 3: Feature Analysis and Steering
### Analyzing Individual Features
```python
from transformer_lens import HookedTransformer
from sae_lens import SAE
import torch
model = HookedTransformer.from_pretrained("gpt2-small", device="cuda")
sae = SAE.from_pretrained( # v6 returns just the SAE
release="gpt2-small-res-jb",
sae_id="blocks.8.hook_resid_pre",
device="cuda"
)
# Find what activates a specific feature
feature_idx = 1234
test_texts = [
"The scientist conducted an experiment",
"I love chocolate cake",
"The code compiles successfully",
"Paris is beautiful in spring",
]
for text in test_texts:
tokens = model.to_tokens(text)
_, cache = model.run_with_cache(tokens)
features = sae.encode(cache["resid_pre", 8])
activation = features[0, :, feature_idx].max().item()
print(f"{activation:.3f}: {text}")
```
### Feature Steering
```python
def steer_with_feature(model, sae, prompt, feature_idx, strength=5.0):
"""Add SAE feature direction to residual stream."""
tokens = model.to_tokens(prompt)
# Get feature direction from decoder
feature_direction = sae.W_dec[feature_idx] # [d_model]
def steering_hook(activation, hook):
# Add scaled feature direction at all positions
activation += strength * feature_direction
return activation
# Generate with steering
output = model.generate(
tokens,
max_new_tokens=50,
fwd_hooks=[("blocks.8.hook_resid_pre", steering_hook)]
)
return model.to_string(output[0])
```
### Feature Attribution
```python
# Which features most affect a specific output?
tokens = model.to_tokens("The capital of France is")
_, cache = model.run_with_cache(tokens)
# Get features at final position
features = sae.encode(cache["resid_pre", 8])[0, -1] # [d_sae]
# Get logit attribution per feature
# Feature contribution = feature_activation × decoder_weight × unembedding
W_dec = sae.W_dec # [d_sae, d_model]
W_U = model.W_U # [d_model, vocab]
# Contribution to "Paris" logit
paris_token = model.to_single_token(" Paris")
feature_contributions = features * (W_dec @ W_U[:, paris_token])
top_features = feature_contributions.topk(10)
print("Top features for 'Paris' prediction:")
for idx, val in zip(top_features.indices, top_features.values):
print(f" Feature {idx.item()}: {val.item():.3f}")
```
## Common Issues & Solutions
> All examples below use the v6 nested config: SAE-specific options go in the `sae=`
> sub-config (`StandardTrainingSAEConfig` / `TopKTrainingSAEConfig` / etc.), training
> knobs stay on the top-level `LanguageModelSAERunnerConfig`.
### Issue: High dead feature ratio
```python
from sae_lens import LanguageModelSAERunnerConfig, StandardTrainingSAEConfig
# WRONG: no warm-up, features die early
cfg = LanguageModelSAERunnerConfig(
sae=StandardTrainingSAEConfig(d_in=768, d_sae=768*8, l1_coefficient=1e-4),
l1_warm_up_steps=0, # Bad!
)
# RIGHT: warm up the L1 penalty (v6 removed ghost grads; warm-up is the lever now)
cfg = LanguageModelSAERunnerConfig(
sae=StandardTrainingSAEConfig(d_in=768, d_sae=768*8, l1_coefficient=8e-5),
l1_warm_up_steps=1000, # Gradually increase
)
```
### Issue: Poor reconstruction (low CE recovery)
```python
# Reduce sparsity penalty and/or add capacity (both on the SAE sub-config)
cfg = LanguageModelSAERunnerConfig(
sae=StandardTrainingSAEConfig(
d_in=768,
d_sae=768 * 16, # More capacity
l1_coefficient=5e-5, # Lower = better reconstruction
),
)
```
### Issue: Features not interpretable
```python
from sae_lens import LanguageModelSAERunnerConfig, StandardTrainingSAEConfig, TopKTrainingSAEConfig
# Increase sparsity (higher L1)
cfg = LanguageModelSAERunnerConfig(
sae=StandardTrainingSAEConfig(d_in=768, d_sae=768*8, l1_coefficient=1e-4),
)
# Or use a TopK SAE (k is set directly in v6, not via activation_fn_kwargs)
cfg = LanguageModelSAERunnerConfig(
sae=TopKTrainingSAEConfig(d_in=768, d_sae=768*8, k=50), # Exactly 50 active features
)
```
### Issue: Memory errors during training
```python
cfg = LanguageModelSAERunnerConfig(
sae=StandardTrainingSAEConfig(d_in=768, d_sae=768*8, l1_coefficient=8e-5),
train_batch_size_tokens=2048, # Reduce batch size
store_batch_size_prompts=4, # Fewer prompts in buffer
n_batches_in_buffer=8, # Smaller activation buffer
)
```
## Integration with Neuronpedia
Browse pre-trained SAE features at [neuronpedia.org](https://neuronpedia.org):
```python
# Features are indexed by SAE ID
# Example: gpt2-small layer 8 feature 1234
# → neuronpedia.org/gpt2-small/8-res-jb/1234
```
## Key Classes Reference
| Class | Purpose |
|-------|---------|
| `SAE` | Sparse Autoencoder model |
| `LanguageModelSAERunnerConfig` | Top-level training configuration (nests `sae=` and `logger=`) |
| `StandardTrainingSAEConfig` / `TopKTrainingSAEConfig` / `GatedTrainingSAEConfig` / `JumpReLUTrainingSAEConfig` | SAE-type-specific sub-configs (v6) |
| `LoggingConfig` | Logging/W&B sub-config (v6) |
| `LanguageModelSAETrainingRunner` | Training loop manager (alias: `SAETrainingRunner`) |
| `ActivationsStore` | Activation collection and batching |
| `HookedSAETransformer` | TransformerLens + SAE integration |
## Reference Documentation
For detailed API documentation, tutorials, and advanced usage, see the `references/` folder:
| File | Contents |
|------|----------|
| [references/README.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\saelens/references/README.md) | Overview and quick start guide |
| [references/api.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\saelens/references/api.md) | Complete API reference for SAE, TrainingSAE, configurations |
| [references/tutorials.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\saelens/references/tutorials.md) | Step-by-step tutorials for training, analysis, steering |
## External Resources
### Tutorials
- [Basic Loading & Analysis](https://github.com/jbloomAus/SAELens/blob/main/tutorials/basic_loading_and_analysing.ipynb)
- [Training a Sparse Autoencoder](https://github.com/jbloomAus/SAELens/blob/main/tutorials/training_a_sparse_autoencoder.ipynb)
- [ARENA SAE Curriculum](https://www.lesswrong.com/posts/LnHowHgmrMbWtpkxx/intro-to-superposition-and-sparse-autoencoders-colab)
### Papers
- [Towards Monosemanticity](https://transformer-circuits.pub/2023/monosemantic-features) - Anthropic (2023)
- [Scaling Monosemanticity](https://transformer-circuits.pub/2024/scaling-monosemanticity/) - Anthropic (2024)
- [Sparse Autoencoders Find Highly Interpretable Features](https://arxiv.org/abs/2309.08600) - Cunningham et al. (ICLR 2024)
### Official Documentation
- [SAELens Docs](https://jbloomaus.github.io/SAELens/)
- [Neuronpedia](https://neuronpedia.org) - Feature browser
## SAE Architectures
| Architecture | Description | Use Case |
|--------------|-------------|----------|
| **Standard** | ReLU + L1 penalty | General purpose |
| **Gated** | Learned gating mechanism | Better sparsity control |
| **TopK** | Exactly K active features | Consistent sparsity |
```python
from sae_lens import LanguageModelSAERunnerConfig, TopKTrainingSAEConfig
# TopK SAE (exactly 50 features active) — `k` is set on the SAE sub-config in v6
cfg = LanguageModelSAERunnerConfig(
sae=TopKTrainingSAEConfig(d_in=768, d_sae=768*8, k=50),
)
```
website/docs/user-guide/skills/optional/mlops/mlops-simpo.md
+237
View File
@@ -0,0 +1,237 @@
---
title: "Simpo — Reference-free preference alignment, simpler than DPO"
sidebar_label: "Simpo"
description: "Reference-free preference alignment, simpler than DPO"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Simpo
Reference-free preference alignment, simpler than DPO.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/simpo` |
| Path | `optional-skills/mlops\simpo` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `torch`, `transformers`, `datasets`, `trl`, `accelerate` |
| Platforms | linux, macos, windows |
| Tags | `Post-Training`, `SimPO`, `Preference Optimization`, `Alignment`, `DPO Alternative`, `Reference-Free`, `LLM Alignment`, `Efficient Training` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# SimPO - Simple Preference Optimization
## Quick start
SimPO is a reference-free preference optimization method that outperforms DPO without needing a reference model.
**Installation**:
```bash
# Create environment
conda create -n simpo python=3.10 && conda activate simpo
# Install PyTorch 2.2.2
# Visit: https://pytorch.org/get-started/locally/
# Install alignment-handbook
git clone https://github.com/huggingface/alignment-handbook.git
cd alignment-handbook
python -m pip install .
# Install Flash Attention 2
python -m pip install flash-attn --no-build-isolation
```
**Training** (Mistral 7B):
```bash
ACCELERATE_LOG_LEVEL=info accelerate launch \
--config_file accelerate_configs/deepspeed_zero3.yaml \
scripts/run_simpo.py \
training_configs/mistral-7b-base-simpo.yaml
```
## Common workflows
### Workflow 1: Train from base model (Mistral 7B)
**Config** (`mistral-7b-base-simpo.yaml`):
```yaml
# Model
model_name_or_path: mistralai/Mistral-7B-v0.1
torch_dtype: bfloat16
# Dataset
dataset_mixer:
HuggingFaceH4/ultrafeedback_binarized: 1.0
dataset_splits:
- train_prefs
- test_prefs
# SimPO hyperparameters
beta: 2.0 # Reward scaling (2.0-10.0)
gamma_beta_ratio: 0.5 # Target margin (0-1)
loss_type: sigmoid # sigmoid or hinge
sft_weight: 0.0 # Optional SFT regularization
# Training
learning_rate: 5e-7 # Critical: 3e-7 to 1e-6
num_train_epochs: 1
per_device_train_batch_size: 1
gradient_accumulation_steps: 8
# Output
output_dir: ./outputs/mistral-7b-simpo
```
**Launch training**:
```bash
accelerate launch --config_file accelerate_configs/deepspeed_zero3.yaml \
scripts/run_simpo.py training_configs/mistral-7b-base-simpo.yaml
```
### Workflow 2: Fine-tune instruct model (Llama 3 8B)
**Config** (`llama3-8b-instruct-simpo.yaml`):
```yaml
model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
dataset_mixer:
argilla/ultrafeedback-binarized-preferences-cleaned: 1.0
beta: 2.5
gamma_beta_ratio: 0.5
learning_rate: 5e-7
sft_weight: 0.1 # Add SFT loss to preserve capabilities
num_train_epochs: 1
per_device_train_batch_size: 2
gradient_accumulation_steps: 4
output_dir: ./outputs/llama3-8b-simpo
```
**Launch**:
```bash
accelerate launch --config_file accelerate_configs/deepspeed_zero3.yaml \
scripts/run_simpo.py training_configs/llama3-8b-instruct-simpo.yaml
```
### Workflow 3: Reasoning-intensive tasks (lower LR)
**For math/code tasks**:
```yaml
model_name_or_path: deepseek-ai/deepseek-math-7b-base
dataset_mixer:
argilla/distilabel-math-preference-dpo: 1.0
beta: 5.0 # Higher for stronger signal
gamma_beta_ratio: 0.7 # Larger margin
learning_rate: 3e-7 # Lower LR for reasoning
sft_weight: 0.0
num_train_epochs: 1
per_device_train_batch_size: 1
gradient_accumulation_steps: 16
```
## When to use vs alternatives
**Use SimPO when**:
- Want simpler training than DPO (no reference model)
- Have preference data (chosen/rejected pairs)
- Need better performance than DPO
- Limited compute resources
- Single-node training sufficient
**Algorithm selection**:
- **SimPO**: Simplest, best performance, no reference model
- **DPO**: Need reference model baseline, more conservative
- **PPO**: Maximum control, need reward model, complex setup
- **GRPO**: Memory-efficient RL, no critic
**Use alternatives instead**:
- **OpenRLHF**: Multi-node distributed training, PPO/GRPO
- **TRL**: Need multiple methods in one framework
- **DPO**: Established baseline comparison
## Common issues
**Issue: Loss divergence**
Reduce learning rate:
```yaml
learning_rate: 3e-7 # Reduce from 5e-7
```
Reduce beta:
```yaml
beta: 1.0 # Reduce from 2.0
```
**Issue: Model forgets capabilities**
Add SFT regularization:
```yaml
sft_weight: 0.1 # Add SFT loss component
```
**Issue: Poor preference separation**
Increase beta and margin:
```yaml
beta: 5.0 # Increase from 2.0
gamma_beta_ratio: 0.8 # Increase from 0.5
```
**Issue: OOM during training**
Reduce batch size:
```yaml
per_device_train_batch_size: 1
gradient_accumulation_steps: 16 # Maintain effective batch
```
Enable gradient checkpointing:
```yaml
gradient_checkpointing: true
```
## Advanced topics
**Loss functions**: See [references/loss-functions.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\simpo/references/loss-functions.md) for sigmoid vs hinge loss, mathematical formulations, and when to use each.
**Hyperparameter tuning**: See [references/hyperparameters.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\simpo/references/hyperparameters.md) for beta, gamma, learning rate selection guide, and model-size-specific recommendations.
**Dataset preparation**: See [references/datasets.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\simpo/references/datasets.md) for preference data formats, quality filtering, and custom dataset creation.
## Hardware requirements
- **GPU**: NVIDIA A100/H100 recommended
- **VRAM**:
- 7B model: 1× A100 40GB (DeepSpeed ZeRO-3)
- 8B model: 2× A100 40GB
- 70B model: 8× A100 80GB
- **Single-node**: DeepSpeed ZeRO-3 sufficient
- **Mixed precision**: BF16 recommended
**Memory optimization**:
- DeepSpeed ZeRO-3 (default config)
- Gradient checkpointing
- Flash Attention 2
## Resources
- Paper: https://arxiv.org/abs/2405.14734 (NeurIPS 2024)
- GitHub: https://github.com/princeton-nlp/SimPO
- Models: https://huggingface.co/princeton-nlp
- Alignment Handbook: https://github.com/huggingface/alignment-handbook
website/docs/user-guide/skills/optional/mlops/mlops-slime.md
+486
View File
@@ -0,0 +1,486 @@
---
title: "Slime — RL post-training for LLMs with Megatron and SGLang"
sidebar_label: "Slime"
description: "RL post-training for LLMs with Megatron and SGLang"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Slime
RL post-training for LLMs with Megatron and SGLang.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/slime` |
| Path | `optional-skills/mlops\slime` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `sglang-router>=0.2.3`, `ray`, `torch>=2.0.0`, `transformers>=4.40.0` |
| Platforms | linux, macos |
| Tags | `Reinforcement Learning`, `Megatron-LM`, `SGLang`, `GRPO`, `Post-Training`, `GLM` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# slime: LLM Post-Training Framework for RL Scaling
slime is an LLM post-training framework from Tsinghua's THUDM team, powering GLM-4.5, GLM-4.6, and GLM-4.7. It connects Megatron-LM for training with SGLang for high-throughput rollout generation.
## When to Use slime
**Choose slime when you need:**
- Megatron-LM native training with SGLang inference
- Custom data generation workflows with flexible data buffers
- Training GLM, Qwen3, DeepSeek V3, or Llama 3 models
- Research-grade framework with production backing (Z.ai)
**Consider alternatives when:**
- You need enterprise-grade stability features → use **miles**
- You want flexible backend swapping → use **verl**
- You need PyTorch-native abstractions → use **torchforge**
## Key Features
- **Training**: Megatron-LM with full parallelism support (TP, PP, DP, SP)
- **Rollout**: SGLang-based high-throughput generation with router
- **Data Buffer**: Flexible prompt management and sample storage
- **Models**: GLM-4.x, Qwen3, DeepSeek V3/R1, Llama 3
## Architecture Overview
<!-- ascii-guard-ignore -->
```
┌─────────────────────────────────────────────────────────┐
│ Data Buffer │
│ - Prompt initialization and management │
│ - Custom data generation and filtering │
│ - Rollout sample storage │
└─────────────┬───────────────────────────┬───────────────┘
│ │
┌─────────────▼───────────┐ ┌─────────────▼───────────────┐
│ Training (Megatron-LM) │ │ Rollout (SGLang + Router) │
│ - Actor model training │ │ - Response generation │
│ - Critic (optional) │ │ - Reward/verifier output │
│ - Weight sync to rollout│ │ - Multi-turn support │
└─────────────────────────┘ └─────────────────────────────┘
```
<!-- ascii-guard-ignore-end -->
## Installation
```bash
# Recommended: Docker
docker pull slimerl/slime:latest
docker run --rm --gpus all --ipc=host --shm-size=16g \
-it slimerl/slime:latest /bin/bash
# Inside container
cd /root/slime && pip install -e . --no-deps
```
### From Source
```bash
git clone https://github.com/THUDM/slime.git
cd slime
pip install -r requirements.txt
pip install -e .
```
## Quick Start: GRPO Training
```bash
# Source model configuration
source scripts/models/qwen3-4B.sh
# Launch training
python train.py \
--actor-num-nodes 1 \
--actor-num-gpus-per-node 4 \
--rollout-num-gpus 4 \
--advantage-estimator grpo \
--use-kl-loss --kl-loss-coef 0.001 \
--rollout-batch-size 32 \
--n-samples-per-prompt 8 \
--global-batch-size 256 \
--num-rollout 3000 \
--prompt-data /path/to/data.jsonl \
${MODEL_ARGS[@]} ${CKPT_ARGS[@]}
```
---
## Workflow 1: Standard GRPO Training
Use this workflow for training reasoning models with group-relative advantages.
### Prerequisites Checklist
- [ ] Docker environment or Megatron-LM + SGLang installed
- [ ] Model checkpoint (HuggingFace or Megatron format)
- [ ] Training data in JSONL format
### Step 1: Prepare Data
```python
# data.jsonl format
{"prompt": "What is 2 + 2?", "label": "4"}
{"prompt": "Solve: 3x = 12", "label": "x = 4"}
```
Or with chat format:
```python
{
"prompt": [
{"role": "system", "content": "You are a math tutor."},
{"role": "user", "content": "What is 15 + 27?"}
],
"label": "42"
}
```
### Step 2: Configure Model
Choose a pre-configured model script:
```bash
# List available models
ls scripts/models/
# glm4-9B.sh, qwen3-4B.sh, qwen3-30B-A3B.sh, deepseek-v3.sh, llama3-8B.sh, ...
# Source your model
source scripts/models/qwen3-4B.sh
```
### Step 3: Launch Training
```bash
python train.py \
--actor-num-nodes 1 \
--actor-num-gpus-per-node 8 \
--rollout-num-gpus 8 \
--advantage-estimator grpo \
--use-kl-loss \
--kl-loss-coef 0.001 \
--prompt-data /path/to/train.jsonl \
--input-key prompt \
--label-key label \
--apply-chat-template \
--rollout-batch-size 32 \
--n-samples-per-prompt 8 \
--global-batch-size 256 \
--num-rollout 3000 \
--save-interval 100 \
--eval-interval 50 \
${MODEL_ARGS[@]}
```
### Step 4: Monitor Training
- [ ] Check TensorBoard: `tensorboard --logdir outputs/`
- [ ] Verify reward curves are increasing
- [ ] Monitor GPU utilization across nodes
---
## Workflow 2: Asynchronous Training
Use async mode for higher throughput by overlapping rollout and training.
### When to Use Async
- Large models with long generation times
- High GPU idle time in synchronous mode
- Sufficient memory for buffering
### Launch Async Training
```bash
python train_async.py \
--actor-num-nodes 1 \
--actor-num-gpus-per-node 8 \
--rollout-num-gpus 8 \
--advantage-estimator grpo \
--async-buffer-size 4 \
--prompt-data /path/to/train.jsonl \
${MODEL_ARGS[@]}
```
### Async-Specific Parameters
```bash
--async-buffer-size 4 # Number of rollouts to buffer
--update-weights-interval 2 # Sync weights every N rollouts
```
---
## Workflow 3: Multi-Turn Agentic Training
Use this workflow for training agents with tool use or multi-step reasoning.
### Prerequisites
- [ ] Custom generate function for multi-turn logic
- [ ] Tool/environment interface
### Step 1: Define Custom Generate Function
```python
# custom_generate.py
async def custom_generate(args, samples, evaluation=False):
"""Multi-turn generation with tool calling."""
for sample in samples:
conversation = sample.prompt
for turn in range(args.max_turns):
# Generate response
response = await generate_single(conversation)
# Check for tool call
tool_call = extract_tool_call(response)
if tool_call:
tool_result = execute_tool(tool_call)
conversation.append({"role": "assistant", "content": response})
conversation.append({"role": "tool", "content": tool_result})
else:
break
sample.response = response
sample.reward = compute_reward(sample)
return samples
```
### Step 2: Launch with Custom Function
```bash
python train.py \
--custom-generate-function-path custom_generate.py \
--max-turns 5 \
--prompt-data /path/to/agent_data.jsonl \
${MODEL_ARGS[@]}
```
See `examples/search-r1/` for a complete multi-turn search example.
---
## Configuration Reference
### Three Argument Categories
slime uses three types of arguments:
**1. Megatron Arguments** (passed directly):
```bash
--tensor-model-parallel-size 2
--pipeline-model-parallel-size 1
--num-layers 32
--hidden-size 4096
```
**2. SGLang Arguments** (prefixed with `--sglang-`):
```bash
--sglang-mem-fraction-static 0.8
--sglang-context-length 8192
--sglang-log-level INFO
```
**3. slime Arguments**:
```bash
# Resource allocation
--actor-num-nodes 1
--actor-num-gpus-per-node 8
--rollout-num-gpus 8
--colocate # Share GPUs between training/inference
# Data
--prompt-data /path/to/data.jsonl
--input-key prompt
--label-key label
# Training loop
--num-rollout 3000
--rollout-batch-size 32
--n-samples-per-prompt 8
--global-batch-size 256
# Algorithm
--advantage-estimator grpo # or: gspo, ppo, reinforce_plus_plus
--use-kl-loss
--kl-loss-coef 0.001
```
### Key Constraints
```
rollout_batch_size × n_samples_per_prompt = global_batch_size × num_steps_per_rollout
```
Example: 32 × 8 = 256 × 1
---
## Data Buffer System
slime's data buffer enables flexible data management:
### Basic Data Source
```python
class RolloutDataSource:
def get_samples(self, num_samples):
"""Fetch prompts from dataset."""
return self.dataset.sample(num_samples)
def add_samples(self, samples):
"""Called after generation (no-op by default)."""
pass
```
### Buffered Data Source (Off-Policy)
```python
class RolloutDataSourceWithBuffer(RolloutDataSource):
def __init__(self):
self.buffer = []
def add_samples(self, samples):
"""Store generated samples for reuse."""
self.buffer.extend(samples)
def buffer_filter(self, args, buffer, num_samples):
"""Custom selection logic (prioritized, stratified, etc.)."""
return select_best(buffer, num_samples)
```
---
## Common Issues and Solutions
### Issue: SGLang Engine Crash
**Symptoms**: Inference engine dies mid-training
**Solutions**:
```bash
# Enable fault tolerance
--use-fault-tolerance
# Increase memory allocation
--sglang-mem-fraction-static 0.85
# Reduce batch size
--rollout-batch-size 16
```
### Issue: Weight Sync Timeout
**Symptoms**: Training hangs after rollout
**Solutions**:
```bash
# Increase sync interval
--update-weights-interval 5
# Use colocated mode (no network transfer)
--colocate
```
### Issue: OOM During Training
**Symptoms**: CUDA OOM in backward pass
**Solutions**:
```bash
# Enable gradient checkpointing
--recompute-activations
# Reduce micro-batch size
--micro-batch-size 1
# Enable sequence parallelism
--sequence-parallel
```
### Issue: Slow Data Loading
**Symptoms**: GPU idle during data fetch
**Solutions**:
```bash
# Increase data workers
--num-data-workers 4
# Use streaming dataset
--streaming-data
```
---
## Supported Models
| Model Family | Configurations |
|--------------|----------------|
| GLM | GLM-4.5, GLM-4.6, GLM-4.7, GLM-Z1-9B |
| Qwen | Qwen3 (4B, 8B, 30B-A3B), Qwen3-MoE, Qwen2.5 |
| DeepSeek | V3, V3.1, R1 |
| Llama | Llama 3 (8B, 70B) |
| Others | Kimi K2, Moonlight-16B |
Each model has pre-configured scripts in `scripts/models/`.
---
## Advanced Topics
### Co-location Mode
Share GPUs between training and inference to reduce memory:
```bash
python train.py \
--colocate \
--actor-num-gpus-per-node 8 \
--sglang-mem-fraction-static 0.4 \
${MODEL_ARGS[@]}
```
### Custom Reward Model
```python
# custom_rm.py
class CustomRewardModel:
def __init__(self, model_path):
self.model = load_model(model_path)
def compute_reward(self, prompts, responses):
inputs = self.tokenize(prompts, responses)
scores = self.model(inputs)
return scores.tolist()
```
```bash
--custom-rm-path custom_rm.py
```
### Evaluation Multi-Task
```bash
--eval-prompt-data aime /path/to/aime.jsonl \
--eval-prompt-data gsm8k /path/to/gsm8k.jsonl \
--n-samples-per-eval-prompt 16
```
---
## Resources
- **Documentation**: https://thudm.github.io/slime/
- **GitHub**: https://github.com/THUDM/slime
- **Blog**: https://lmsys.org/blog/2025-07-09-slime/
- **Examples**: See `examples/` directory for 14+ worked examples
website/docs/user-guide/skills/optional/mlops/mlops-stable-diffusion.md
+542
View File
@@ -0,0 +1,542 @@
---
title: "Stable Diffusion — Text-to-image generation, inpainting, and img2img"
sidebar_label: "Stable Diffusion"
description: "Text-to-image generation, inpainting, and img2img"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Stable Diffusion
Text-to-image generation, inpainting, and img2img.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/stable-diffusion` |
| Path | `optional-skills/mlops\stable-diffusion` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `diffusers>=0.30.0`, `transformers>=4.41.0`, `accelerate>=0.31.0`, `torch>=2.0.0` |
| Platforms | linux, macos, windows |
| Tags | `Image Generation`, `Stable Diffusion`, `Diffusers`, `Text-to-Image`, `Multimodal`, `Computer Vision` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Stable Diffusion Image Generation
Guide to generating images with Stable Diffusion using the HuggingFace Diffusers library.
## When to use Stable Diffusion
**Use Stable Diffusion when:**
- Generating images from text descriptions
- Performing image-to-image translation (style transfer, enhancement)
- Inpainting (filling in masked regions)
- Outpainting (extending images beyond boundaries)
- Creating variations of existing images
- Building custom image generation workflows
**Key features:**
- **Text-to-Image**: Generate images from natural language prompts
- **Image-to-Image**: Transform existing images with text guidance
- **Inpainting**: Fill masked regions with context-aware content
- **ControlNet**: Add spatial conditioning (edges, poses, depth)
- **LoRA Support**: Efficient fine-tuning and style adaptation
- **Multiple Models**: SD 1.5, SDXL, SD 3.0, Flux support
**Use alternatives instead:**
- **DALL-E 3**: For API-based generation without GPU
- **Midjourney**: For artistic, stylized outputs
- **Imagen**: For Google Cloud integration
- **Leonardo.ai**: For web-based creative workflows
## Quick start
### Installation
```bash
pip install diffusers transformers accelerate torch
pip install xformers # Optional: memory-efficient attention
```
### Basic text-to-image
```python
from diffusers import DiffusionPipeline
import torch
# Load pipeline (auto-detects model type)
pipe = DiffusionPipeline.from_pretrained(
"stable-diffusion-v1-5/stable-diffusion-v1-5",
torch_dtype=torch.float16
)
pipe.to("cuda")
# Generate image
image = pipe(
"A serene mountain landscape at sunset, highly detailed",
num_inference_steps=50,
guidance_scale=7.5
).images[0]
image.save("output.png")
```
### Using SDXL (higher quality)
```python
from diffusers import AutoPipelineForText2Image
import torch
pipe = AutoPipelineForText2Image.from_pretrained(
"stabilityai/stable-diffusion-xl-base-1.0",
torch_dtype=torch.float16,
variant="fp16"
)
pipe.to("cuda")
# Enable memory optimization
pipe.enable_model_cpu_offload()
image = pipe(
prompt="A futuristic city with flying cars, cinematic lighting",
height=1024,
width=1024,
num_inference_steps=30
).images[0]
```
## Architecture overview
### Three-pillar design
Diffusers is built around three core components:
<!-- ascii-guard-ignore -->
```
Pipeline (orchestration)
├── Model (neural networks)
│ ├── UNet / Transformer (noise prediction)
│ ├── VAE (latent encoding/decoding)
│ └── Text Encoder (CLIP/T5)
└── Scheduler (denoising algorithm)
```
<!-- ascii-guard-ignore-end -->
### Pipeline inference flow
```
Text Prompt → Text Encoder → Text Embeddings
↓
Random Noise → [Denoising Loop] ← Scheduler
↓
Predicted Noise
↓
VAE Decoder → Final Image
```
## Core concepts
### Pipelines
Pipelines orchestrate complete workflows:
| Pipeline | Purpose |
|----------|---------|
| `StableDiffusionPipeline` | Text-to-image (SD 1.x/2.x) |
| `StableDiffusionXLPipeline` | Text-to-image (SDXL) |
| `StableDiffusion3Pipeline` | Text-to-image (SD 3.0) |
| `FluxPipeline` | Text-to-image (Flux models) |
| `StableDiffusionImg2ImgPipeline` | Image-to-image |
| `StableDiffusionInpaintPipeline` | Inpainting |
### Schedulers
Schedulers control the denoising process:
| Scheduler | Steps | Quality | Use Case |
|-----------|-------|---------|----------|
| `EulerDiscreteScheduler` | 20-50 | Good | Default choice |
| `EulerAncestralDiscreteScheduler` | 20-50 | Good | More variation |
| `DPMSolverMultistepScheduler` | 15-25 | Excellent | Fast, high quality |
| `DDIMScheduler` | 50-100 | Good | Deterministic |
| `LCMScheduler` | 4-8 | Good | Very fast |
| `UniPCMultistepScheduler` | 15-25 | Excellent | Fast convergence |
### Swapping schedulers
```python
from diffusers import DPMSolverMultistepScheduler
# Swap for faster generation
pipe.scheduler = DPMSolverMultistepScheduler.from_config(
pipe.scheduler.config
)
# Now generate with fewer steps
image = pipe(prompt, num_inference_steps=20).images[0]
```
## Generation parameters
### Key parameters
| Parameter | Default | Description |
|-----------|---------|-------------|
| `prompt` | Required | Text description of desired image |
| `negative_prompt` | None | What to avoid in the image |
| `num_inference_steps` | 50 | Denoising steps (more = better quality) |
| `guidance_scale` | 7.5 | Prompt adherence (7-12 typical) |
| `height`, `width` | 512/1024 | Output dimensions (multiples of 8) |
| `generator` | None | Torch generator for reproducibility |
| `num_images_per_prompt` | 1 | Batch size |
### Reproducible generation
```python
import torch
generator = torch.Generator(device="cuda").manual_seed(42)
image = pipe(
prompt="A cat wearing a top hat",
generator=generator,
num_inference_steps=50
).images[0]
```
### Negative prompts
```python
image = pipe(
prompt="Professional photo of a dog in a garden",
negative_prompt="blurry, low quality, distorted, ugly, bad anatomy",
guidance_scale=7.5
).images[0]
```
## Image-to-image
Transform existing images with text guidance:
```python
from diffusers import AutoPipelineForImage2Image
from PIL import Image
pipe = AutoPipelineForImage2Image.from_pretrained(
"stable-diffusion-v1-5/stable-diffusion-v1-5",
torch_dtype=torch.float16
).to("cuda")
init_image = Image.open("input.jpg").resize((512, 512))
image = pipe(
prompt="A watercolor painting of the scene",
image=init_image,
strength=0.75, # How much to transform (0-1)
num_inference_steps=50
).images[0]
```
## Inpainting
Fill masked regions:
```python
from diffusers import AutoPipelineForInpainting
from PIL import Image
pipe = AutoPipelineForInpainting.from_pretrained(
"runwayml/stable-diffusion-inpainting",
torch_dtype=torch.float16
).to("cuda")
image = Image.open("photo.jpg")
mask = Image.open("mask.png") # White = inpaint region
result = pipe(
prompt="A red car parked on the street",
image=image,
mask_image=mask,
num_inference_steps=50
).images[0]
```
## ControlNet
Add spatial conditioning for precise control:
```python
from diffusers import StableDiffusionControlNetPipeline, ControlNetModel
import torch
# Load ControlNet for edge conditioning
controlnet = ControlNetModel.from_pretrained(
"lllyasviel/control_v11p_sd15_canny",
torch_dtype=torch.float16
)
pipe = StableDiffusionControlNetPipeline.from_pretrained(
"stable-diffusion-v1-5/stable-diffusion-v1-5",
controlnet=controlnet,
torch_dtype=torch.float16
).to("cuda")
# Use Canny edge image as control
control_image = get_canny_image(input_image)
image = pipe(
prompt="A beautiful house in the style of Van Gogh",
image=control_image,
num_inference_steps=30
).images[0]
```
### Available ControlNets
| ControlNet | Input Type | Use Case |
|------------|------------|----------|
| `canny` | Edge maps | Preserve structure |
| `openpose` | Pose skeletons | Human poses |
| `depth` | Depth maps | 3D-aware generation |
| `normal` | Normal maps | Surface details |
| `mlsd` | Line segments | Architectural lines |
| `scribble` | Rough sketches | Sketch-to-image |
## LoRA adapters
Load fine-tuned style adapters:
```python
from diffusers import DiffusionPipeline
pipe = DiffusionPipeline.from_pretrained(
"stable-diffusion-v1-5/stable-diffusion-v1-5",
torch_dtype=torch.float16
).to("cuda")
# Load LoRA weights
pipe.load_lora_weights("path/to/lora", weight_name="style.safetensors")
# Generate with LoRA style
image = pipe("A portrait in the trained style").images[0]
# Adjust LoRA strength
pipe.fuse_lora(lora_scale=0.8)
# Unload LoRA
pipe.unload_lora_weights()
```
### Multiple LoRAs
```python
# Load multiple LoRAs
pipe.load_lora_weights("lora1", adapter_name="style")
pipe.load_lora_weights("lora2", adapter_name="character")
# Set weights for each
pipe.set_adapters(["style", "character"], adapter_weights=[0.7, 0.5])
image = pipe("A portrait").images[0]
```
## Memory optimization
### Enable CPU offloading
```python
# Model CPU offload - moves models to CPU when not in use
pipe.enable_model_cpu_offload()
# Sequential CPU offload - more aggressive, slower
pipe.enable_sequential_cpu_offload()
```
### Attention slicing
```python
# Reduce memory by computing attention in chunks
pipe.enable_attention_slicing()
# Or specific chunk size
pipe.enable_attention_slicing("max")
```
### xFormers memory-efficient attention
```python
# Requires xformers package
pipe.enable_xformers_memory_efficient_attention()
```
### VAE slicing for large images
```python
# Decode latents in tiles for large images
pipe.enable_vae_slicing()
pipe.enable_vae_tiling()
```
## Model variants
### Loading different precisions
```python
# FP16 (recommended for GPU)
pipe = DiffusionPipeline.from_pretrained(
"model-id",
torch_dtype=torch.float16,
variant="fp16"
)
# BF16 (better precision, requires Ampere+ GPU)
pipe = DiffusionPipeline.from_pretrained(
"model-id",
torch_dtype=torch.bfloat16
)
```
### Loading specific components
```python
from diffusers import UNet2DConditionModel, AutoencoderKL
# Load custom VAE
vae = AutoencoderKL.from_pretrained("stabilityai/sd-vae-ft-mse")
# Use with pipeline
pipe = DiffusionPipeline.from_pretrained(
"stable-diffusion-v1-5/stable-diffusion-v1-5",
vae=vae,
torch_dtype=torch.float16
)
```
## Batch generation
Generate multiple images efficiently:
```python
# Multiple prompts
prompts = [
"A cat playing piano",
"A dog reading a book",
"A bird painting a picture"
]
images = pipe(prompts, num_inference_steps=30).images
# Multiple images per prompt
images = pipe(
"A beautiful sunset",
num_images_per_prompt=4,
num_inference_steps=30
).images
```
## Common workflows
### Workflow 1: High-quality generation
```python
from diffusers import StableDiffusionXLPipeline, DPMSolverMultistepScheduler
import torch
# 1. Load SDXL with optimizations
pipe = StableDiffusionXLPipeline.from_pretrained(
"stabilityai/stable-diffusion-xl-base-1.0",
torch_dtype=torch.float16,
variant="fp16"
)
pipe.to("cuda")
pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
pipe.enable_model_cpu_offload()
# 2. Generate with quality settings
image = pipe(
prompt="A majestic lion in the savanna, golden hour lighting, 8k, detailed fur",
negative_prompt="blurry, low quality, cartoon, anime, sketch",
num_inference_steps=30,
guidance_scale=7.5,
height=1024,
width=1024
).images[0]
```
### Workflow 2: Fast prototyping
```python
from diffusers import AutoPipelineForText2Image, LCMScheduler
import torch
# Use LCM for 4-8 step generation
pipe = AutoPipelineForText2Image.from_pretrained(
"stabilityai/stable-diffusion-xl-base-1.0",
torch_dtype=torch.float16
).to("cuda")
# Load LCM LoRA for fast generation
pipe.load_lora_weights("latent-consistency/lcm-lora-sdxl")
pipe.scheduler = LCMScheduler.from_config(pipe.scheduler.config)
pipe.fuse_lora()
# Generate in ~1 second
image = pipe(
"A beautiful landscape",
num_inference_steps=4,
guidance_scale=1.0
).images[0]
```
## Common issues
**CUDA out of memory:**
```python
# Enable memory optimizations
pipe.enable_model_cpu_offload()
pipe.enable_attention_slicing()
pipe.enable_vae_slicing()
# Or use lower precision
pipe = DiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16)
```
**Black/noise images:**
```python
# Check VAE configuration
# Use safety checker bypass if needed
pipe.safety_checker = None
# Ensure proper dtype consistency
pipe = pipe.to(dtype=torch.float16)
```
**Slow generation:**
```python
# Use faster scheduler
from diffusers import DPMSolverMultistepScheduler
pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
# Reduce steps
image = pipe(prompt, num_inference_steps=20).images[0]
```
## References
- **[Advanced Usage](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\stable-diffusion/references/advanced-usage.md)** - Custom pipelines, fine-tuning, deployment
- **[Troubleshooting](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\stable-diffusion/references/troubleshooting.md)** - Common issues and solutions
## Resources
- **Documentation**: https://huggingface.co/docs/diffusers
- **Repository**: https://github.com/huggingface/diffusers
- **Model Hub**: https://huggingface.co/models?library=diffusers
- **Discord**: https://discord.gg/diffusers
website/docs/user-guide/skills/optional/mlops/mlops-tensorrt-llm.md
+208
View File
@@ -0,0 +1,208 @@
---
title: "Tensorrt Llm — High-throughput LLM inference on NVIDIA GPUs"
sidebar_label: "Tensorrt Llm"
description: "High-throughput LLM inference on NVIDIA GPUs"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Tensorrt Llm
High-throughput LLM inference on NVIDIA GPUs.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/tensorrt-llm` |
| Path | `optional-skills/mlops\tensorrt-llm` |
| Version | `1.0.1` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `tensorrt-llm`, `torch` |
| Platforms | linux, macos |
| Tags | `Inference Serving`, `TensorRT-LLM`, `NVIDIA`, `Inference Optimization`, `High Throughput`, `Low Latency`, `Production`, `FP8`, `INT4`, `In-Flight Batching`, `Multi-GPU` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# TensorRT-LLM
NVIDIA's open-source library for optimizing LLM inference with high performance on NVIDIA GPUs.
## When to use TensorRT-LLM
**Use TensorRT-LLM when:**
- Deploying on NVIDIA GPUs (A100, H100, GB200)
- Need maximum throughput (24,000+ tokens/sec on Llama 3)
- Require low latency for real-time applications
- Working with quantized models (FP8, INT4, FP4)
- Scaling across multiple GPUs or nodes
**Use vLLM instead when:**
- Need simpler setup and Python-first API
- Want PagedAttention without TensorRT compilation
- Working with AMD GPUs or non-NVIDIA hardware
**Use llama.cpp instead when:**
- Deploying on CPU or Apple Silicon
- Need edge deployment without NVIDIA GPUs
- Want simpler GGUF quantization format
## Quick start
### Installation
```bash
# Docker (recommended) — images are on NGC (nvcr.io), not Docker Hub.
# Replace x.y.z with the desired version (e.g. 1.2.1). Browse tags on NGC:
# https://catalog.ngc.nvidia.com/orgs/nvidia/teams/tensorrt-llm/containers/release/tags
docker pull nvcr.io/nvidia/tensorrt-llm/release:x.y.z
# pip install (current stable GA)
pip install tensorrt_llm
# Requires CUDA 13.2.1, TensorRT 10.x, Python 3.10-3.12
```
### Basic inference
```python
from tensorrt_llm import LLM, SamplingParams
# Initialize model
llm = LLM(model="meta-llama/Meta-Llama-3-8B")
# Configure sampling
sampling_params = SamplingParams(
max_tokens=100,
temperature=0.7,
top_p=0.9
)
# Generate
prompts = ["Explain quantum computing"]
outputs = llm.generate(prompts, sampling_params)
for output in outputs:
print(output.text)
```
### Serving with trtllm-serve
```bash
# Start server (automatic model download and compilation)
trtllm-serve meta-llama/Meta-Llama-3-8B \
--tp_size 4 \ # Tensor parallelism (4 GPUs)
--max_batch_size 256 \
--max_num_tokens 4096
# Client request
curl -X POST http://localhost:8000/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "meta-llama/Meta-Llama-3-8B",
"messages": [{"role": "user", "content": "Hello!"}],
"temperature": 0.7,
"max_tokens": 100
}'
```
## Key features
### Performance optimizations
- **In-flight batching**: Dynamic batching during generation
- **Paged KV cache**: Efficient memory management
- **Flash Attention**: Optimized attention kernels
- **Quantization**: FP8, INT4, FP4 for 2-4× faster inference
- **CUDA graphs**: Reduced kernel launch overhead
### Parallelism
- **Tensor parallelism (TP)**: Split model across GPUs
- **Pipeline parallelism (PP)**: Layer-wise distribution
- **Expert parallelism**: For Mixture-of-Experts models
- **Multi-node**: Scale beyond single machine
### Advanced features
- **Speculative decoding**: Faster generation with draft models
- **LoRA serving**: Efficient multi-adapter deployment
- **Disaggregated serving**: Separate prefill and generation
## Common patterns
### Quantized model (FP8)
```python
from tensorrt_llm import LLM
# Load FP8 quantized model (2× faster, 50% memory)
llm = LLM(
model="meta-llama/Meta-Llama-3-70B",
dtype="fp8",
max_num_tokens=8192
)
# Inference same as before
outputs = llm.generate(["Summarize this article..."])
```
### Multi-GPU deployment
```python
# Tensor parallelism across 8 GPUs
llm = LLM(
model="meta-llama/Meta-Llama-3-405B",
tensor_parallel_size=8,
dtype="fp8"
)
```
### Batch inference
```python
# Process 100 prompts efficiently
prompts = [f"Question {i}: ..." for i in range(100)]
outputs = llm.generate(
prompts,
sampling_params=SamplingParams(max_tokens=200)
)
# Automatic in-flight batching for maximum throughput
```
## Performance benchmarks
**Meta Llama 3-8B** (H100 GPU):
- Throughput: 24,000 tokens/sec
- Latency: ~10ms per token
- vs PyTorch: **100× faster**
**Llama 3-70B** (8× A100 80GB):
- FP8 quantization: 2× faster than FP16
- Memory: 50% reduction with FP8
## Supported models
- **LLaMA family**: Llama 2, Llama 3, CodeLlama
- **GPT family**: GPT-2, GPT-J, GPT-NeoX
- **Qwen**: Qwen, Qwen2, QwQ
- **DeepSeek**: DeepSeek-V2, DeepSeek-V3
- **Mixtral**: Mixtral-8x7B, Mixtral-8x22B
- **Vision**: LLaVA, Phi-3-vision
- **100+ models** on HuggingFace
## References
- **[Optimization Guide](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\tensorrt-llm/references/optimization.md)** - Quantization, batching, KV cache tuning
- **[Multi-GPU Setup](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\tensorrt-llm/references/multi-gpu.md)** - Tensor/pipeline parallelism, multi-node
- **[Serving Guide](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\tensorrt-llm/references/serving.md)** - Production deployment, monitoring, autoscaling
## Resources
- **Docs**: https://nvidia.github.io/TensorRT-LLM/
- **GitHub**: https://github.com/NVIDIA/TensorRT-LLM
- **Models**: https://huggingface.co/models?library=tensorrt_llm
website/docs/user-guide/skills/optional/mlops/mlops-torchtitan.md
+404
View File
@@ -0,0 +1,404 @@
---
title: "Torchtitan — Pretrain LLMs at scale with PyTorch 4D parallelism"
sidebar_label: "Torchtitan"
description: "Pretrain LLMs at scale with PyTorch 4D parallelism"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Torchtitan
Pretrain LLMs at scale with PyTorch 4D parallelism.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/torchtitan` |
| Path | `optional-skills/mlops\torchtitan` |
| Version | `1.0.1` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `torch>=2.6.0`, `torchtitan>=0.2.0`, `torchao>=0.5.0` |
| Platforms | linux, macos |
| Tags | `Model Architecture`, `Distributed Training`, `TorchTitan`, `FSDP2`, `Tensor Parallel`, `Pipeline Parallel`, `Context Parallel`, `Float8`, `Llama`, `Pretraining` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# TorchTitan - PyTorch Native Distributed LLM Pretraining
## Quick start
TorchTitan is PyTorch's official platform for large-scale LLM pretraining with composable 4D parallelism (FSDP2, TP, PP, CP), achieving 65%+ speedups over baselines on H100 GPUs.
**Installation**:
```bash
# From PyPI (stable)
pip install torchtitan
# From source (latest features, requires PyTorch nightly)
git clone https://github.com/pytorch/torchtitan
cd torchtitan
pip install -r requirements.txt
```
**Download tokenizer**:
```bash
# Get HF token from https://huggingface.co/settings/tokens
python scripts/download_hf_assets.py --repo_id meta-llama/Llama-3.1-8B --assets tokenizer --hf_token=...
```
**Start training on 8 GPUs**:
```bash
# Configs are selected by name from the Python config registry
# (torchtitan/models/llama3/config_registry.py), not by TOML path
MODULE=llama3 CONFIG=llama3_8b ./run_train.sh
```
## Common workflows
### Workflow 1: Pretrain Llama 3.1 8B on single node
Copy this checklist:
```
Single Node Pretraining:
- [ ] Step 1: Download tokenizer
- [ ] Step 2: Configure training
- [ ] Step 3: Launch training
- [ ] Step 4: Monitor and checkpoint
```
**Step 1: Download tokenizer**
```bash
python scripts/download_hf_assets.py \
--repo_id meta-llama/Llama-3.1-8B \
--assets tokenizer \
--hf_token=YOUR_HF_TOKEN
```
**Step 2: Configure training**
In torchtitan's current layout, run configs are defined in a Python **config registry**
(`torchtitan/models/llama3/config_registry.py`) and selected by name via `CONFIG=<name>`
(or `--config <name>`). To customize, register your own config in the registry, or override
individual fields on the command line (e.g. `--optimizer.lr 3e-4 --training.steps 1000`).
The equivalent settings for an 8B run look like this (shown as fields; set them in the
registry entry or as `--section.key value` overrides):
```toml
# fields for a llama3 8B run (register in config_registry.py or pass as --overrides)
[job]
dump_folder = "./outputs"
description = "Llama 3.1 8B training"
[model]
name = "llama3"
flavor = "8B"
hf_assets_path = "./assets/hf/Llama-3.1-8B"
[optimizer]
name = "AdamW"
lr = 3e-4
[lr_scheduler]
warmup_steps = 200
[training]
local_batch_size = 2
seq_len = 8192
max_norm = 1.0
steps = 1000
dataset = "c4"
[parallelism]
data_parallel_shard_degree = -1 # Use all GPUs for FSDP
[activation_checkpoint]
mode = "selective"
selective_ac_option = "op"
[checkpoint]
enable = true
folder = "checkpoint"
interval = 500
```
**Step 3: Launch training**
```bash
# 8 GPUs on single node (config selected by name from the registry)
MODULE=llama3 CONFIG=llama3_8b ./run_train.sh
# Override individual fields on the command line
MODULE=llama3 CONFIG=llama3_8b ./run_train.sh --optimizer.lr 3e-4 --training.steps 1000
# Or explicitly with torchrun (run_train.sh wraps this)
torchrun --nproc_per_node=8 \
-m torchtitan.train \
--module llama3 --config llama3_8b
```
**Step 4: Monitor and checkpoint**
TensorBoard logs are saved to `./outputs/tb/`:
```bash
tensorboard --logdir ./outputs/tb
```
### Workflow 2: Multi-node training with SLURM
```
Multi-Node Training:
- [ ] Step 1: Configure parallelism for scale
- [ ] Step 2: Set up SLURM script
- [ ] Step 3: Submit job
- [ ] Step 4: Resume from checkpoint
```
**Step 1: Configure parallelism for scale**
For 70B model on 256 GPUs (32 nodes):
```toml
[parallelism]
data_parallel_shard_degree = 32 # FSDP across 32 ranks
tensor_parallel_degree = 8 # TP within node
pipeline_parallel_degree = 1 # No PP for 70B
context_parallel_degree = 1 # Increase for long sequences
```
**Step 2: Set up SLURM script**
```bash
#!/bin/bash
#SBATCH --job-name=llama70b
#SBATCH --nodes=32
#SBATCH --ntasks-per-node=8
#SBATCH --gpus-per-node=8
srun torchrun \
--nnodes=32 \
--nproc_per_node=8 \
--rdzv_backend=c10d \
--rdzv_endpoint=$MASTER_ADDR:$MASTER_PORT \
-m torchtitan.train \
--module llama3 --config llama3_70b
```
**Step 3: Submit job**
```bash
sbatch multinode_trainer.slurm
```
**Step 4: Resume from checkpoint**
Training auto-resumes if checkpoint exists in configured folder.
### Workflow 3: Enable Float8 training for H100s
Float8 provides 30-50% speedup on H100 GPUs.
```
Float8 Training:
- [ ] Step 1: Install torchao
- [ ] Step 2: Configure Float8
- [ ] Step 3: Launch with compile
```
**Step 1: Install torchao**
```bash
USE_CPP=0 pip install git+https://github.com/pytorch/ao.git
```
**Step 2: Configure Float8**
In the current torchtitan, Float8 is applied at config time via the `quantization`
parameter in your `model_registry()` call inside the config registry (not via a
`[quantize.linear.float8]` TOML section). Add a `Float8LinearConverter.Config`:
```python
# in torchtitan/models/llama3/config_registry.py (your model_registry(...) call)
from torchtitan.components.quantization import Float8LinearConverter
model_spec = model_registry(
"8B",
quantization=[
Float8LinearConverter.Config(
recipe_name="rowwise", # or "rowwise_with_gw_hp"
filter_fqns=["output"], # skip layers too small to benefit
model_compile_enabled=True, # requires torch.compile for competitive perf
),
],
)
```
Enable `torch.compile` in your run config too:
```toml
[compile]
enable = true
components = ["model", "loss"]
```
**Step 3: Launch with compile**
```bash
# Float8 config is baked into the registered config; just select it and enable compile
MODULE=llama3 CONFIG=llama3_8b ./run_train.sh --compile.enable
```
### Workflow 4: 4D parallelism for 405B models
```
4D Parallelism (FSDP + TP + PP + CP):
- [ ] Step 1: Create seed checkpoint
- [ ] Step 2: Configure 4D parallelism
- [ ] Step 3: Launch on 512 GPUs
```
**Step 1: Create seed checkpoint**
Required for consistent initialization across PP stages:
```bash
NGPU=1 MODULE=llama3 CONFIG=llama3_405b ./run_train.sh \
--checkpoint.enable \
--checkpoint.create_seed_checkpoint \
--parallelism.data_parallel_shard_degree 1 \
--parallelism.tensor_parallel_degree 1 \
--parallelism.pipeline_parallel_degree 1
```
**Step 2: Configure 4D parallelism**
```toml
[parallelism]
data_parallel_shard_degree = 8 # FSDP
tensor_parallel_degree = 8 # TP within node
pipeline_parallel_degree = 8 # PP across nodes
context_parallel_degree = 1 # CP for long sequences
[training]
local_batch_size = 32
seq_len = 8192
```
**Step 3: Launch on 512 GPUs**
```bash
# 64 nodes x 8 GPUs = 512 GPUs
srun torchrun --nnodes=64 --nproc_per_node=8 \
-m torchtitan.train \
--module llama3 --config llama3_405b
```
## When to use vs alternatives
**Use TorchTitan when:**
- Pretraining LLMs from scratch (8B to 405B+)
- Need PyTorch-native solution without third-party dependencies
- Require composable 4D parallelism (FSDP2, TP, PP, CP)
- Training on H100s with Float8 support
- Want interoperable checkpoints with torchtune/HuggingFace
**Use alternatives instead:**
- **Megatron-LM**: Maximum performance for NVIDIA-only deployments
- **DeepSpeed**: Broader ZeRO optimization ecosystem, inference support
- **Axolotl/TRL**: Fine-tuning rather than pretraining
- **LitGPT**: Educational, smaller-scale training
## Common issues
**Issue: Out of memory on large models**
Enable activation checkpointing and reduce batch size:
```toml
[activation_checkpoint]
mode = "full" # Instead of "selective"
[training]
local_batch_size = 1
```
Or use gradient accumulation:
```toml
[training]
local_batch_size = 1
global_batch_size = 32 # Accumulates gradients
```
**Issue: TP causes high memory with async collectives**
Set environment variable:
```bash
export TORCH_NCCL_AVOID_RECORD_STREAMS=1
```
**Issue: Float8 training not faster**
Float8 only benefits large GEMMs. Filter small layers via the converter's `filter_fqns`:
```python
from torchtitan.components.quantization import Float8LinearConverter
Float8LinearConverter.Config(
# add "auto_filter_small_kn" to auto-skip layers too small to benefit
filter_fqns=["attention.wk", "attention.wv", "output", "auto_filter_small_kn"],
model_compile_enabled=True,
)
```
**Issue: Checkpoint loading fails after parallelism change**
Use DCP's resharding capability:
```bash
# Convert sharded checkpoint to single file
python -m torch.distributed.checkpoint.format_utils \
dcp_to_torch checkpoint/step-1000 checkpoint.pt
```
**Issue: Pipeline parallelism initialization**
Create seed checkpoint first (see Workflow 4, Step 1).
## Supported models
| Model | Sizes | Status |
|-------|-------|--------|
| Llama 3.1 | 8B, 70B, 405B | Production |
| Llama 4 | Various | Experimental |
| DeepSeek V3 | 16B, 236B, 671B (MoE) | Experimental |
| GPT-OSS | 20B, 120B (MoE) | Experimental |
| Qwen 3 | Various | Experimental |
| Flux | Diffusion | Experimental |
## Performance benchmarks (H100)
| Model | GPUs | Parallelism | TPS/GPU | Techniques |
|-------|------|-------------|---------|------------|
| Llama 8B | 8 | FSDP | 5,762 | Baseline |
| Llama 8B | 8 | FSDP+compile+FP8 | 8,532 | +48% |
| Llama 70B | 256 | FSDP+TP+AsyncTP | 876 | 2D parallel |
| Llama 405B | 512 | FSDP+TP+PP | 128 | 3D parallel |
## Advanced topics
**FSDP2 configuration**: See [references/fsdp.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\torchtitan/references/fsdp.md) for detailed FSDP2 vs FSDP1 comparison and ZeRO equivalents.
**Float8 training**: See [references/float8.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\torchtitan/references/float8.md) for tensorwise vs rowwise scaling recipes.
**Checkpointing**: See [references/checkpoint.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\torchtitan/references/checkpoint.md) for HuggingFace conversion and async checkpointing.
**Adding custom models**: See [references/custom-models.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\torchtitan/references/custom-models.md) for TrainSpec protocol.
## Resources
- GitHub: https://github.com/pytorch/torchtitan
- Paper: https://arxiv.org/abs/2410.06511
- ICLR 2025: https://iclr.cc/virtual/2025/poster/29620
- PyTorch Forum: https://discuss.pytorch.org/c/distributed/torchtitan/44
website/docs/user-guide/skills/optional/mlops/mlops-training-axolotl.md
+181
View File
@@ -0,0 +1,181 @@
---
title: "Axolotl — Axolotl: YAML LLM fine-tuning (LoRA, DPO, GRPO)"
sidebar_label: "Axolotl"
description: "Axolotl: YAML LLM fine-tuning (LoRA, DPO, GRPO)"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Axolotl
Axolotl: YAML LLM fine-tuning (LoRA, DPO, GRPO).
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/axolotl` |
| Path | `optional-skills/mlops\training\axolotl` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `axolotl`, `torch`, `transformers`, `datasets`, `peft`, `accelerate`, `deepspeed` |
| Platforms | linux, macos |
| Tags | `Fine-Tuning`, `Axolotl`, `LLM`, `LoRA`, `QLoRA`, `DPO`, `KTO`, `ORPO`, `GRPO`, `YAML`, `HuggingFace`, `DeepSpeed`, `Multimodal` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Axolotl Skill
## What's inside
Expert guidance for fine-tuning LLMs with Axolotl — YAML configs, 100+ models, LoRA/QLoRA, DPO/KTO/ORPO/GRPO, multimodal support.
Assistance with axolotl development, generated from official documentation.
## When to Use This Skill
This skill should be triggered when:
- Working with axolotl
- Asking about axolotl features or APIs
- Implementing axolotl solutions
- Debugging axolotl code
- Learning axolotl best practices
## Quick Reference
### Common Patterns
**Pattern 1:** To validate that acceptable data transfer speeds exist for your training job, running NCCL Tests can help pinpoint bottlenecks, for example:
```
./build/all_reduce_perf -b 8 -e 128M -f 2 -g 3
```
**Pattern 2:** Configure your model to use FSDP in the Axolotl yaml. For example:
```
fsdp_version: 2
fsdp_config:
offload_params: true
state_dict_type: FULL_STATE_DICT
auto_wrap_policy: TRANSFORMER_BASED_WRAP
transformer_layer_cls_to_wrap: LlamaDecoderLayer
reshard_after_forward: true
```
**Pattern 3:** The context_parallel_size should be a divisor of the total number of GPUs. For example:
```
context_parallel_size
```
**Pattern 4:** For example: - With 8 GPUs and no sequence parallelism: 8 different batches processed per step - With 8 GPUs and context_parallel_size=4: Only 2 different batches processed per step (each split across 4 GPUs) - If your per-GPU micro_batch_size is 2, the global batch size decreases from 16 to 4
```
context_parallel_size=4
```
**Pattern 5:** Setting save_compressed: true in your configuration enables saving models in a compressed format, which: - Reduces disk space usage by approximately 40% - Maintains compatibility with vLLM for accelerated inference - Maintains compatibility with llmcompressor for further optimization (example: quantization)
```
save_compressed: true
```
**Pattern 6:** Note It is not necessary to place your integration in the integrations folder. It can be in any location, so long as it’s installed in a package in your python env. See this repo for an example: https://github.com/axolotl-ai-cloud/diff-transformer
```
integrations
```
**Pattern 7:** Handle both single-example and batched data. - single example: sample[‘input_ids’] is a list[int] - batched data: sample[‘input_ids’] is a list[list[int]]
```
utils.trainer.drop_long_seq(sample, sequence_len=2048, min_sequence_len=2)
```
### Example Code Patterns
**Example 1** (python):
```python
cli.cloud.modal_.ModalCloud(config, app=None)
```
**Example 2** (python):
```python
cli.cloud.modal_.run_cmd(cmd, run_folder, volumes=None)
```
**Example 3** (python):
```python
core.trainers.base.AxolotlTrainer(
*_args,
bench_data_collator=None,
eval_data_collator=None,
dataset_tags=None,
**kwargs,
)
```
**Example 4** (python):
```python
core.trainers.base.AxolotlTrainer.log(logs, start_time=None)
```
**Example 5** (python):
```python
prompt_strategies.input_output.RawInputOutputPrompter()
```
## Reference Files
This skill includes comprehensive documentation in `references/`:
- **api.md** - Api documentation
- **dataset-formats.md** - Dataset-Formats documentation
- **other.md** - Other documentation
Use `view` to read specific reference files when detailed information is needed.
## Working with This Skill
### For Beginners
Start with the getting_started or tutorials reference files for foundational concepts.
### For Specific Features
Use the appropriate category reference file (api, guides, etc.) for detailed information.
### For Code Examples
The quick reference section above contains common patterns extracted from the official docs.
## Resources
### references/
Organized documentation extracted from official sources. These files contain:
- Detailed explanations
- Code examples with language annotations
- Links to original documentation
- Table of contents for quick navigation
### scripts/
Add helper scripts here for common automation tasks.
### assets/
Add templates, boilerplate, or example projects here.
## Notes
- This skill was automatically generated from official documentation
- Reference files preserve the structure and examples from source docs
- Code examples include language detection for better syntax highlighting
- Quick reference patterns are extracted from common usage examples in the docs
## Updating
To refresh this skill with updated documentation:
1. Re-run the scraper with the same configuration
2. The skill will be rebuilt with the latest information
website/docs/user-guide/skills/optional/mlops/mlops-training-trl-fine-tuning.md
+512
View File
@@ -0,0 +1,512 @@
---
title: "Trl Fine Tuning — TRL: SFT, DPO, GRPO, RLOO reward modeling for LLM RLHF"
sidebar_label: "Trl Fine Tuning"
description: "TRL: SFT, DPO, GRPO, RLOO reward modeling for LLM RLHF"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Trl Fine Tuning
TRL: SFT, DPO, GRPO, RLOO reward modeling for LLM RLHF.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/trl-fine-tuning` |
| Path | `optional-skills/mlops\training\trl-fine-tuning` |
| Version | `1.0.1` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `trl`, `transformers`, `datasets`, `peft`, `accelerate`, `torch` |
| Platforms | linux, macos, windows |
| Tags | `Post-Training`, `TRL`, `Reinforcement Learning`, `Fine-Tuning`, `SFT`, `DPO`, `GRPO`, `RLOO`, `RLHF`, `Preference Alignment`, `HuggingFace` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# TRL - Transformer Reinforcement Learning
## Quick start
TRL provides post-training methods for aligning language models with human preferences.
**Installation**:
```bash
pip install trl transformers datasets peft accelerate
```
**Supervised Fine-Tuning** (instruction tuning):
```python
from trl import SFTTrainer
trainer = SFTTrainer(
model="Qwen/Qwen2.5-0.5B",
train_dataset=dataset, # Prompt-completion pairs
)
trainer.train()
```
**DPO** (align with preferences):
```python
from trl import DPOTrainer, DPOConfig
config = DPOConfig(output_dir="model-dpo", beta=0.1)
trainer = DPOTrainer(
model=model,
args=config,
train_dataset=preference_dataset, # chosen/rejected pairs
processing_class=tokenizer
)
trainer.train()
```
## Common workflows
### Workflow 1: Full RLHF pipeline (SFT → Reward Model → RLOO)
Complete pipeline from base model to human-aligned model.
> **Note (TRL 1.x):** PPO has been **removed** from TRL — `PPOTrainer`, `PPOConfig`, and
> `python -m trl.scripts.ppo` no longer exist. Use an online-RL trainer TRL still ships:
> **RLOO** (`RLOOTrainer` / `trl rloo`) is the closest drop-in for a reward-model-driven
> RLHF pipeline, and **GRPO** (`GRPOTrainer` / `trl grpo`, see Workflow 3) is the
> memory-efficient alternative. The step below uses RLOO.
Copy this checklist:
```
RLHF Training:
- [ ] Step 1: Supervised fine-tuning (SFT)
- [ ] Step 2: Train reward model
- [ ] Step 3: RLOO reinforcement learning
- [ ] Step 4: Evaluate aligned model
```
**Step 1: Supervised fine-tuning**
Train base model on instruction-following data:
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
from trl import SFTTrainer, SFTConfig
from datasets import load_dataset
# Load model
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-0.5B")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-0.5B")
# Load instruction dataset
dataset = load_dataset("trl-lib/Capybara", split="train")
# Configure training
training_args = SFTConfig(
output_dir="Qwen2.5-0.5B-SFT",
per_device_train_batch_size=4,
num_train_epochs=1,
learning_rate=2e-5,
logging_steps=10,
save_strategy="epoch"
)
# Train
trainer = SFTTrainer(
model=model,
args=training_args,
train_dataset=dataset,
processing_class=tokenizer
)
trainer.train()
trainer.save_model()
```
**Step 2: Train reward model**
Train model to predict human preferences:
```python
from transformers import AutoModelForSequenceClassification
from trl import RewardTrainer, RewardConfig
# Load SFT model as base
model = AutoModelForSequenceClassification.from_pretrained(
"Qwen2.5-0.5B-SFT",
num_labels=1 # Single reward score
)
tokenizer = AutoTokenizer.from_pretrained("Qwen2.5-0.5B-SFT")
# Load preference data (chosen/rejected pairs)
dataset = load_dataset("trl-lib/ultrafeedback_binarized", split="train")
# Configure training
training_args = RewardConfig(
output_dir="Qwen2.5-0.5B-Reward",
per_device_train_batch_size=2,
num_train_epochs=1,
learning_rate=1e-5
)
# Train reward model
trainer = RewardTrainer(
model=model,
args=training_args,
processing_class=tokenizer,
train_dataset=dataset
)
trainer.train()
trainer.save_model()
```
**Step 3: RLOO reinforcement learning**
Optimize policy using the reward model. PPO was removed in TRL 1.x; use the RLOO CLI
(`trl rloo`) with the trained reward model passed via `--reward_model_name_or_path`:
```bash
trl rloo \
--model_name_or_path Qwen2.5-0.5B-SFT \
--reward_model_name_or_path Qwen2.5-0.5B-Reward \
--dataset_name trl-internal-testing/descriptiveness-sentiment-trl-style \
--output_dir Qwen2.5-0.5B-RLOO \
--learning_rate 3e-6 \
--per_device_train_batch_size 64 \
--num_generations 4
```
Equivalent Python (`RLOOTrainer` / `RLOOConfig`):
```python
from trl import RLOOTrainer, RLOOConfig
from transformers import AutoModelForSequenceClassification, AutoTokenizer
reward_model = AutoModelForSequenceClassification.from_pretrained(
"Qwen2.5-0.5B-Reward", num_labels=1
)
config = RLOOConfig(
output_dir="Qwen2.5-0.5B-RLOO",
per_device_train_batch_size=64,
learning_rate=3e-6,
num_generations=4,
)
trainer = RLOOTrainer(
model="Qwen2.5-0.5B-SFT",
reward_funcs=reward_model, # a reward model (or a callable reward function)
args=config,
train_dataset=dataset, # prompt-only dataset
processing_class=tokenizer,
)
trainer.train()
```
**Step 4: Evaluate**
```python
from transformers import pipeline
# Load aligned model
generator = pipeline("text-generation", model="Qwen2.5-0.5B-RLOO")
# Test
prompt = "Explain quantum computing to a 10-year-old"
output = generator(prompt, max_length=200)[0]["generated_text"]
print(output)
```
### Workflow 2: Simple preference alignment with DPO
Align model with preferences without reward model.
Copy this checklist:
```
DPO Training:
- [ ] Step 1: Prepare preference dataset
- [ ] Step 2: Configure DPO
- [ ] Step 3: Train with DPOTrainer
- [ ] Step 4: Evaluate alignment
```
**Step 1: Prepare preference dataset**
Dataset format:
```json
{
"prompt": "What is the capital of France?",
"chosen": "The capital of France is Paris.",
"rejected": "I don't know."
}
```
Load dataset:
```python
from datasets import load_dataset
dataset = load_dataset("trl-lib/ultrafeedback_binarized", split="train")
# Or load your own
# dataset = load_dataset("json", data_files="preferences.json")
```
**Step 2: Configure DPO**
```python
from trl import DPOConfig
config = DPOConfig(
output_dir="Qwen2.5-0.5B-DPO",
per_device_train_batch_size=4,
num_train_epochs=1,
learning_rate=5e-7,
beta=0.1, # KL penalty strength
max_prompt_length=512,
max_length=1024,
logging_steps=10
)
```
**Step 3: Train with DPOTrainer**
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
from trl import DPOTrainer
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-0.5B-Instruct")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-0.5B-Instruct")
trainer = DPOTrainer(
model=model,
args=config,
train_dataset=dataset,
processing_class=tokenizer
)
trainer.train()
trainer.save_model()
```
**CLI alternative**:
```bash
trl dpo \
--model_name_or_path Qwen/Qwen2.5-0.5B-Instruct \
--dataset_name argilla/Capybara-Preferences \
--output_dir Qwen2.5-0.5B-DPO \
--per_device_train_batch_size 4 \
--learning_rate 5e-7 \
--beta 0.1
```
### Workflow 3: Memory-efficient online RL with GRPO
Train with reinforcement learning using minimal memory.
For in-depth GRPO guidance — reward function design, critical training insights (loss behavior, mode collapse, tuning), and advanced multi-stage patterns — see **[references/grpo-training.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\training\trl-fine-tuning/references/grpo-training.md)**. A production-ready training script is in **[templates/basic_grpo_training.py](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\training\trl-fine-tuning/templates/basic_grpo_training.py)**.
Copy this checklist:
```
GRPO Training:
- [ ] Step 1: Define reward function
- [ ] Step 2: Configure GRPO
- [ ] Step 3: Train with GRPOTrainer
```
**Step 1: Define reward function**
```python
def reward_function(completions, **kwargs):
"""
Compute rewards for completions.
Args:
completions: List of generated texts
Returns:
List of reward scores (floats)
"""
rewards = []
for completion in completions:
# Example: reward based on length and unique words
score = len(completion.split()) # Favor longer responses
score += len(set(completion.lower().split())) # Reward unique words
rewards.append(score)
return rewards
```
Or use a reward model:
```python
from transformers import pipeline
reward_model = pipeline("text-classification", model="reward-model-path")
def reward_from_model(completions, prompts, **kwargs):
# Combine prompt + completion
full_texts = [p + c for p, c in zip(prompts, completions)]
# Get reward scores
results = reward_model(full_texts)
return [r["score"] for r in results]
```
**Step 2: Configure GRPO**
```python
from trl import GRPOConfig
config = GRPOConfig(
output_dir="Qwen2-GRPO",
per_device_train_batch_size=4,
num_train_epochs=1,
learning_rate=1e-5,
num_generations=4, # Generate 4 completions per prompt
max_new_tokens=128
)
```
**Step 3: Train with GRPOTrainer**
```python
from datasets import load_dataset
from trl import GRPOTrainer
# Load prompt-only dataset
dataset = load_dataset("trl-lib/tldr", split="train")
trainer = GRPOTrainer(
model="Qwen/Qwen2-0.5B-Instruct",
reward_funcs=reward_function, # Your reward function
args=config,
train_dataset=dataset
)
trainer.train()
```
**CLI**:
```bash
trl grpo \
--model_name_or_path Qwen/Qwen2-0.5B-Instruct \
--dataset_name trl-lib/tldr \
--output_dir Qwen2-GRPO \
--num_generations 4
```
## When to use vs alternatives
**Use TRL when:**
- Need to align model with human preferences
- Have preference data (chosen/rejected pairs)
- Want to use reinforcement learning (RLOO, GRPO)
- Need reward model training
- Doing RLHF (full pipeline)
**Method selection**:
- **SFT**: Have prompt-completion pairs, want basic instruction following
- **DPO**: Have preferences, want simple alignment (no reward model needed)
- **RLOO**: Have a reward model, want online RL (the reward-model-driven RLHF path; PPO was removed in TRL 1.x)
- **GRPO**: Memory-constrained, want online RL with reward functions
- **Reward Model**: Building RLHF pipeline, need to score generations
**Use alternatives instead:**
- **HuggingFace Trainer**: Basic fine-tuning without RL
- **Axolotl**: YAML-based training configuration
- **LitGPT**: Educational, minimal fine-tuning
- **Unsloth**: Fast LoRA training
## Common issues
**Issue: OOM during DPO training**
Reduce batch size and sequence length:
```python
config = DPOConfig(
per_device_train_batch_size=1, # Reduce from 4
max_length=512, # Reduce from 1024
gradient_accumulation_steps=8 # Maintain effective batch
)
```
Or use gradient checkpointing:
```python
model.gradient_checkpointing_enable()
```
**Issue: Poor alignment quality**
Tune beta parameter:
```python
# Higher beta = more conservative (stays closer to reference)
config = DPOConfig(beta=0.5) # Default 0.1
# Lower beta = more aggressive alignment
config = DPOConfig(beta=0.01)
```
**Issue: Reward model not learning**
Check loss type and learning rate:
```python
config = RewardConfig(
learning_rate=1e-5, # Try different LR
num_train_epochs=3 # Train longer
)
```
Ensure preference dataset has clear winners:
```python
# Verify dataset
print(dataset[0])
# Should have clear chosen > rejected
```
**Issue: Online RL (RLOO/GRPO) training unstable**
Adjust the KL/beta regularization toward the reference policy:
```python
from trl import RLOOConfig
config = RLOOConfig(
beta=0.05, # KL coefficient toward the reference model (increase for stability)
num_generations=4, # more samples per prompt = lower-variance advantage estimates
)
```
## Advanced topics
**SFT training guide**: See [references/sft-training.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\training\trl-fine-tuning/references/sft-training.md) for dataset formats, chat templates, packing strategies, and multi-GPU training.
**DPO variants**: See [references/dpo-variants.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\training\trl-fine-tuning/references/dpo-variants.md) for IPO, cDPO, RPO, and other DPO loss functions with recommended hyperparameters.
**Reward modeling**: See [references/reward-modeling.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\training\trl-fine-tuning/references/reward-modeling.md) for outcome vs process rewards, Bradley-Terry loss, and reward model evaluation.
**Online RL methods**: See [references/online-rl.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\training\trl-fine-tuning/references/online-rl.md) for PPO, GRPO, RLOO, and OnlineDPO with detailed configurations.
**GRPO deep dive**: See [references/grpo-training.md](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\training\trl-fine-tuning/references/grpo-training.md) for expert-level GRPO patterns — reward function design philosophy, training insights (why loss increases, mode collapse detection), hyperparameter tuning, multi-stage training, and troubleshooting. Production-ready template in [templates/basic_grpo_training.py](https://github.com/NousResearch/hermes-agent/blob/main/optional-skills/mlops\training\trl-fine-tuning/templates/basic_grpo_training.py).
## Hardware requirements
- **GPU**: NVIDIA (CUDA required)
- **VRAM**: Depends on model and method
- SFT 7B: 16GB (with LoRA)
- DPO 7B: 24GB (stores reference model)
- RLOO 7B: 40GB (policy + reward model)
- GRPO 7B: 24GB (more memory efficient)
- **Multi-GPU**: Supported via `accelerate`
- **Mixed precision**: BF16 recommended (A100/H100)
**Memory optimization**:
- Use LoRA/QLoRA for all methods
- Enable gradient checkpointing
- Use smaller batch sizes with gradient accumulation
## Resources
- Docs: https://huggingface.co/docs/trl/
- GitHub: https://github.com/huggingface/trl
- Papers:
- "Training language models to follow instructions with human feedback" (InstructGPT, 2022)
- "Direct Preference Optimization: Your Language Model is Secretly a Reward Model" (DPO, 2023)
- "Group Relative Policy Optimization" (GRPO, 2024)
- Examples: https://github.com/huggingface/trl/tree/main/examples/scripts
website/docs/user-guide/skills/optional/mlops/mlops-training-unsloth.md
+98
View File
@@ -0,0 +1,98 @@
---
title: "Unsloth — Unsloth: 2-5x faster LoRA/QLoRA fine-tuning, less VRAM"
sidebar_label: "Unsloth"
description: "Unsloth: 2-5x faster LoRA/QLoRA fine-tuning, less VRAM"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Unsloth
Unsloth: 2-5x faster LoRA/QLoRA fine-tuning, less VRAM.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/unsloth` |
| Path | `optional-skills/mlops\training\unsloth` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `unsloth`, `torch`, `transformers`, `trl`, `datasets`, `peft` |
| Platforms | linux, macos |
| Tags | `Fine-Tuning`, `Unsloth`, `Fast Training`, `LoRA`, `QLoRA`, `Memory-Efficient`, `Optimization`, `Llama`, `Mistral`, `Gemma`, `Qwen` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Unsloth Skill
Assistance with unsloth development, generated from official documentation.
## When to Use This Skill
This skill should be triggered when:
- Working with unsloth
- Asking about unsloth features or APIs
- Implementing unsloth solutions
- Debugging unsloth code
- Learning unsloth best practices
## Quick Reference
### Common Patterns
*Quick reference patterns will be added as you use the skill.*
## Reference Files
This skill includes full documentation in `references/`:
- **llms-txt.md** - Llms-Txt documentation
Use `view` to read specific reference files when detailed information is needed.
## Working with This Skill
### For Beginners
Start with the getting_started or tutorials reference files for foundational concepts.
### For Specific Features
Use the appropriate category reference file (api, guides, etc.) for detailed information.
### For Code Examples
The quick reference section above contains common patterns extracted from the official docs.
## Resources
### references/
Organized documentation extracted from official sources. These files contain:
- Detailed explanations
- Code examples with language annotations
- Links to original documentation
- Table of contents for quick navigation
### scripts/
Add helper scripts here for common automation tasks.
### assets/
Add templates, boilerplate, or example projects here.
## Notes
- This skill was automatically generated from official documentation
- Reference files preserve the structure and examples from source docs
- Code examples include language detection for better syntax highlighting
- Quick reference patterns are extracted from common usage examples in the docs
## Updating
To refresh this skill with updated documentation:
1. Re-run the scraper with the same configuration
2. The skill will be rebuilt with the latest information
<!-- Trigger re-upload 1763621536 -->
website/docs/user-guide/skills/optional/mlops/mlops-whisper.md
+336
View File
@@ -0,0 +1,336 @@
---
title: "Whisper — Transcribe and translate speech in 99 languages"
sidebar_label: "Whisper"
description: "Transcribe and translate speech in 99 languages"
---
{/* This page is auto-generated from the skill's SKILL.md by website/scripts/generate-skill-docs.py. Edit the source SKILL.md, not this page. */}
# Whisper
Transcribe and translate speech in 99 languages.
## Skill metadata
| | |
|---|---|
| Source | Optional — install with `hermes skills install official/mlops/whisper` |
| Path | `optional-skills/mlops\whisper` |
| Version | `1.0.0` |
| Author | Orchestra Research |
| License | MIT |
| Dependencies | `openai-whisper`, `transformers`, `torch` |
| Platforms | linux, macos |
| Tags | `Whisper`, `Speech Recognition`, `ASR`, `Multimodal`, `Multilingual`, `OpenAI`, `Speech-To-Text`, `Transcription`, `Translation`, `Audio Processing` |
## Reference: full SKILL.md
:::info
The following is the complete skill definition that Hermes loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.
:::
# Whisper - Robust Speech Recognition
OpenAI's multilingual speech recognition model.
## When to use Whisper
**Use when:**
- Speech-to-text transcription (99 languages)
- Podcast/video transcription
- Meeting notes automation
- Translation to English
- Noisy audio transcription
- Multilingual audio processing
**Metrics**:
- **72,900+ GitHub stars**
- 99 languages supported
- Trained on 680,000 hours of audio
- MIT License
**Use alternatives instead**:
- **AssemblyAI**: Managed API, speaker diarization
- **Deepgram**: Real-time streaming ASR
- **Google Speech-to-Text**: Cloud-based
## Quick start
### Installation
```bash
# Requires Python 3.8-3.11
pip install -U openai-whisper
# Requires ffmpeg
# macOS: brew install ffmpeg
# Ubuntu: sudo apt install ffmpeg
# Windows: choco install ffmpeg
```
### Basic transcription
```python
import whisper
# Load model
model = whisper.load_model("base")
# Transcribe
result = model.transcribe("audio.mp3")
# Print text
print(result["text"])
# Access segments
for segment in result["segments"]:
print(f"[{segment['start']:.2f}s - {segment['end']:.2f}s] {segment['text']}")
```
## Model sizes
```python
# Available models
models = ["tiny", "base", "small", "medium", "large", "turbo"]
# Load specific model
model = whisper.load_model("turbo") # Fastest, good quality
```
| Model | Parameters | English-only | Multilingual | Speed | VRAM |
|-------|------------|--------------|--------------|-------|------|
| tiny | 39M | ✓ | ✓ | ~32x | ~1 GB |
| base | 74M | ✓ | ✓ | ~16x | ~1 GB |
| small | 244M | ✓ | ✓ | ~6x | ~2 GB |
| medium | 769M | ✓ | ✓ | ~2x | ~5 GB |
| large | 1550M | ✗ | ✓ | 1x | ~10 GB |
| turbo | 809M | ✗ | ✓ | ~8x | ~6 GB |
**Recommendation**: Use `turbo` for best speed/quality, `base` for prototyping
## Transcription options
### Language specification
```python
# Auto-detect language
result = model.transcribe("audio.mp3")
# Specify language (faster)
result = model.transcribe("audio.mp3", language="en")
# Supported: en, es, fr, de, it, pt, ru, ja, ko, zh, and 89 more
```
### Task selection
```python
# Transcription (default)
result = model.transcribe("audio.mp3", task="transcribe")
# Translation to English
result = model.transcribe("spanish.mp3", task="translate")
# Input: Spanish audio → Output: English text
```
### Initial prompt
```python
# Improve accuracy with context
result = model.transcribe(
"audio.mp3",
initial_prompt="This is a technical podcast about machine learning and AI."
)
# Helps with:
# - Technical terms
# - Proper nouns
# - Domain-specific vocabulary
```
### Timestamps
```python
# Word-level timestamps
result = model.transcribe("audio.mp3", word_timestamps=True)
for segment in result["segments"]:
for word in segment["words"]:
print(f"{word['word']} ({word['start']:.2f}s - {word['end']:.2f}s)")
```
### Temperature fallback
```python
# Retry with different temperatures if confidence low
result = model.transcribe(
"audio.mp3",
temperature=(0.0, 0.2, 0.4, 0.6, 0.8, 1.0)
)
```
## Command line usage
```bash
# Basic transcription
whisper audio.mp3
# Specify model
whisper audio.mp3 --model turbo
# Output formats
whisper audio.mp3 --output_format txt # Plain text
whisper audio.mp3 --output_format srt # Subtitles
whisper audio.mp3 --output_format vtt # WebVTT
whisper audio.mp3 --output_format json # JSON with timestamps
# Language
whisper audio.mp3 --language Spanish
# Translation
whisper spanish.mp3 --task translate
```
## Batch processing
```python
import os
audio_files = ["file1.mp3", "file2.mp3", "file3.mp3"]
for audio_file in audio_files:
print(f"Transcribing {audio_file}...")
result = model.transcribe(audio_file)
# Save to file
output_file = audio_file.replace(".mp3", ".txt")
with open(output_file, "w") as f:
f.write(result["text"])
```
## Real-time transcription
```python
# For streaming audio, use faster-whisper
# pip install faster-whisper
from faster_whisper import WhisperModel
model = WhisperModel("base", device="cuda", compute_type="float16")
# Transcribe with streaming
segments, info = model.transcribe("audio.mp3", beam_size=5)
for segment in segments:
print(f"[{segment.start:.2f}s -> {segment.end:.2f}s] {segment.text}")
```
## GPU acceleration
```python
import whisper
# Automatically uses GPU if available
model = whisper.load_model("turbo")
# Force CPU
model = whisper.load_model("turbo", device="cpu")
# Force GPU
model = whisper.load_model("turbo", device="cuda")
# 10-20× faster on GPU
```
## Integration with other tools
### Subtitle generation
```bash
# Generate SRT subtitles
whisper video.mp4 --output_format srt --language English
# Output: video.srt
```
### With LangChain
```python
from langchain.document_loaders import WhisperTranscriptionLoader
loader = WhisperTranscriptionLoader(file_path="audio.mp3")
docs = loader.load()
# Use transcription in RAG
from langchain_chroma import Chroma
from langchain_openai import OpenAIEmbeddings
vectorstore = Chroma.from_documents(docs, OpenAIEmbeddings())
```
### Extract audio from video
```bash
# Use ffmpeg to extract audio
ffmpeg -i video.mp4 -vn -acodec pcm_s16le audio.wav
# Then transcribe
whisper audio.wav
```
## Best practices
1. **Use turbo model** - Best speed/quality for English
2. **Specify language** - Faster than auto-detect
3. **Add initial prompt** - Improves technical terms
4. **Use GPU** - 10-20× faster
5. **Batch process** - More efficient
6. **Convert to WAV** - Better compatibility
7. **Split long audio** - &lt;30 min chunks
8. **Check language support** - Quality varies by language
9. **Use faster-whisper** - 4× faster than openai-whisper
10. **Monitor VRAM** - Scale model size to hardware
## Performance
| Model | Real-time factor (CPU) | Real-time factor (GPU) |
|-------|------------------------|------------------------|
| tiny | ~0.32 | ~0.01 |
| base | ~0.16 | ~0.01 |
| turbo | ~0.08 | ~0.01 |
| large | ~1.0 | ~0.05 |
*Real-time factor: 0.1 = 10× faster than real-time*
## Language support
Top-supported languages:
- English (en)
- Spanish (es)
- French (fr)
- German (de)
- Italian (it)
- Portuguese (pt)
- Russian (ru)
- Japanese (ja)
- Korean (ko)
- Chinese (zh)
Full list: 99 languages total
## Limitations
1. **Hallucinations** - May repeat or invent text
2. **Long-form accuracy** - Degrades on >30 min audio
3. **Speaker identification** - No diarization
4. **Accents** - Quality varies
5. **Background noise** - Can affect accuracy
6. **Real-time latency** - Not suitable for live captioning
## Resources
- **GitHub**: https://github.com/openai/whisper ⭐ 72,900+
- **Paper**: https://arxiv.org/abs/2212.04356
- **Model Card**: https://github.com/openai/whisper/blob/main/model-card.md
- **Colab**: Available in repo
- **License**: MIT