openai-oss / README.md
davanstrien's picture
davanstrien HF Staff
Add reasoning_effort parameter to minimal script
a4ee9cd
|
raw
history blame
5.87 kB

πŸš€ OpenAI GPT OSS Models - Works on Regular GPUs!

Generate synthetic datasets with transparent reasoning using OpenAI's GPT OSS models. No H100s required - works on L4, A100, A10G, and even T4 GPUs!

πŸŽ‰ Key Discovery

The models work on regular datacenter GPUs! Transformers automatically handles MXFP4 β†’ bf16 conversion, making these models accessible on standard hardware.

🌟 Quick Start

Test Locally (Single Prompt)

uv run gpt_oss_transformers.py --prompt "Write a haiku about mountains"

Run on HuggingFace Jobs (No GPU Required!)

# Generate haiku with reasoning (~$1.50/hr on A10G)
hf jobs uv run --flavor a10g-small \
    https://huggingface.co/datasets/uv-scripts/openai-oss/raw/main/gpt_oss_transformers.py \
    --input-dataset davanstrien/haiku_dpo \
    --output-dataset username/haiku-reasoning \
    --prompt-column question \
    --max-samples 50

πŸ’‘ What You Get

The models output structured reasoning in separate channels:

Raw Output:

analysisI need to write a haiku about mountains. Haiku: 5-7-5 syllable structure...
assistantfinalSilent peaks climb high,
Echoing winds trace stone's breath,
Dawn paints them gold bright.

Parsed Dataset:

{
  "prompt": "Write a haiku about mountains",
  "think": "[Analysis] I need to write a haiku about mountains. Haiku: 5-7-5 syllable structure...",
  "content": "Silent peaks climb high,\nEchoing winds trace stone's breath,\nDawn paints them gold bright.",
  "reasoning_level": "high",
  "model": "openai/gpt-oss-20b"
}

πŸ–₯️ GPU Requirements

βœ… Confirmed Working GPUs

GPU Memory Status Notes
L4 24GB βœ… Tested Works perfectly!
A100 40/80GB βœ… Works Great performance
A10G 24GB βœ… Recommended Best value at $1.50/hr
T4 16GB ⚠️ Limited May need 8-bit for 20B
RTX 4090 24GB βœ… Works Consumer GPU support

Memory Requirements

  • 20B model: ~40GB VRAM when dequantized (use A100-40GB or 2xL4)
  • 120B model: ~240GB VRAM when dequantized (use 4xA100-80GB)

🎯 Examples

Creative Writing with Reasoning

# Process haiku dataset with high reasoning
uv run gpt_oss_transformers.py \
    --input-dataset davanstrien/haiku_dpo \
    --output-dataset my-haiku-reasoning \
    --prompt-column question \
    --reasoning-level high \
    --max-samples 100

Math Problems with Step-by-Step Solutions

# Generate math solutions with reasoning traces
uv run gpt_oss_transformers.py \
    --input-dataset gsm8k \
    --output-dataset math-with-reasoning \
    --prompt-column question \
    --reasoning-level high

Test Different Reasoning Levels

# Compare reasoning levels
for level in low medium high; do
    echo "Testing: $level"
    uv run gpt_oss_transformers.py \
        --prompt "Explain gravity to a 5-year-old" \
        --reasoning-level $level \
        --debug
done

πŸ“‹ Script Options

Option Description Default
--input-dataset HuggingFace dataset to process -
--output-dataset Output dataset name -
--prompt-column Column with prompts prompt
--model-id Model to use openai/gpt-oss-20b
--reasoning-level Reasoning depth: low/medium/high high
--max-samples Limit samples to process None
--temperature Sampling temperature 0.7
--max-tokens Max tokens to generate 512
--prompt Single prompt test (skip dataset) -
--debug Show raw model output False

πŸ”§ Technical Details

Why It Works Without H100s

  1. Automatic MXFP4 Handling: When your GPU doesn't support MXFP4, you'll see:

    MXFP4 quantization requires triton >= 3.4.0 and triton_kernels installed, 
    we will default to dequantizing the model to bf16
    
  2. No Flash Attention 3 Required: FA3 needs Hopper architecture, but models work fine without it

  3. Simple Loading: Just use standard transformers:

    model = AutoModelForCausalLM.from_pretrained(
        "openai/gpt-oss-20b",
        torch_dtype=torch.bfloat16,
        device_map="auto"
    )
    

Channel Output Format

The models use a simplified channel format:

  • analysis: Chain of thought reasoning
  • commentary: Meta operations (optional)
  • final: User-facing response

Reasoning Control

Control reasoning depth via system message:

messages = [
    {
        "role": "system", 
        "content": f"...Reasoning: {level}..."
    },
    {"role": "user", "content": prompt}
]

🚨 Best Practices

  1. Token Limits: Use 1000+ tokens for detailed reasoning
  2. Security: Never expose reasoning channels to end users
  3. Batch Size: Keep at 1 for memory efficiency
  4. Reasoning Levels:
    • low: Quick responses
    • medium: Balanced reasoning
    • high: Detailed chain-of-thought

πŸ› Troubleshooting

Out of Memory

  • Use larger GPU flavor: --flavor a100-large
  • Reduce batch size to 1
  • Try 8-bit quantization for smaller GPUs

No GPU Available

  • Use HuggingFace Jobs (no local GPU needed!)
  • Or use cloud instances with GPU support

Empty Reasoning

  • Increase --max-tokens to 1500+
  • Ensure prompts trigger reasoning

πŸ“š References

πŸŽ‰ The Bottom Line

You don't need H100s! These models work great on regular datacenter GPUs. Just run the script and start generating datasets with transparent reasoning.


Last tested: 2025-08-05 on NVIDIA L4 GPUs - Working perfectly!