README.md

---
base_model: Steelskull/L3.3-Cu-Mai-R1-70b
tags:
  - fp8
  - vllm
  - compressed-tensors
  - quantized
  - llmcompressor
license: apache-2.0
inference:
  parameters:
    temperature: 0.7
    top_p: 0.9
    max_new_tokens: 2048
library_name: transformers
pipeline_tag: text-generation
---

# L3.3-Cu-Mai-R1-70b - FP8 Dynamic Quantization

This is an FP8 quantized version of [Steelskull/L3.3-Cu-Mai-R1-70b](https://huggingface.co/Steelskull/L3.3-Cu-Mai-R1-70b) using `llmcompressor` with the FP8_DYNAMIC scheme.

## Model Details

- **Base Model**: Steelskull/L3.3-Cu-Mai-R1-70b
- **Quantization**: FP8_DYNAMIC (W8A8)
- **Format**: compressed-tensors (SafeTensors)
- **Memory**: ~50% of original BF16 size
- **Quality**: <1-2% degradation on benchmarks (typical)

## Quick Start

### vLLM (Recommended)

```bash
pip install vllm

# Serve the model
vllm serve REPO_ID \
  --max-model-len 32768 \
  --gpu-memory-utilization 0.95

# Python API
from vllm import LLM
llm = LLM(model="REPO_ID")
outputs = llm.generate("Hello, how are you?")
print(outputs[0].outputs[0].text)
```

### Transformers

```python
from transformers import AutoTokenizer, AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained(
    "REPO_ID",
    device_map="auto",
    torch_dtype="auto"
)
tokenizer = AutoTokenizer.from_pretrained("REPO_ID")

messages = [{'role': 'user', 'content': 'Hello!'}]
inputs = tokenizer.apply_chat_template(messages, return_tensors='pt').to(model.device)
outputs = model.generate(inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0]))
```

## Quantization Details

This model was quantized using:
- **Tool**: [llmcompressor](https://github.com/vllm-project/llm-compressor)
- **Method**: FP8_DYNAMIC (Round-to-Nearest)
- **Targets**: All Linear layers except `lm_head`
- **Scheme**: W8A8 (8-bit weights and activations)


## Performance

### Memory Usage
- **Original BF16**: ~2× size of FP8
- **FP8 Quantized**: ~50% of original
- **Savings**: ~50% VRAM reduction

### Inference Speed
- Expect 1.3-1.8× faster inference vs BF16
- 2× higher throughput (more KV cache available)

## Use Cases

Perfect for:
- ✅ Production inference on limited VRAM
- ✅ Running larger models on single GPU
- ✅ Cost-effective API serving
- ✅ High-throughput applications
- ✅ Extended context lengths (more KV cache)

## Hardware Requirements

**Minimum VRAM** (approximate):
- 70B model: ~40 GB (RTX A6000, A100 40GB)
- 123B model: ~70 GB (A100 80GB, H100, H200)

**Recommended**:
- H100/H200 for best performance
- vLLM for optimized serving
- Enable FP8 KV cache for extended context

## Important Notes

⚠️ **Quantization Trade-offs**:
- Slight quality degradation (typically <1-2%)
- Not suitable for fine-tuning (inference only)
- Best with vLLM (has FP8 kernel optimizations)

✅ **Best Practices**:
- Use `--kv-cache-dtype fp8` for longer contexts
- Set `--gpu-memory-utilization 0.90-0.95`
- Add `--enforce-eager` if you encounter compilation issues

## Citation

If you use this model, please cite:

```bibtex
@misc{model_name-fp8,
  author = {author},
  title = {model_name FP8 Dynamic Quantization},
  year = {2025},
  publisher = {HuggingFace},
  url = {https://huggingface.co/repo_id}
}
```

## License

Inherits license from base model: [Steelskull/L3.3-Cu-Mai-R1-70b](https://huggingface.co/Steelskull/L3.3-Cu-Mai-R1-70b)

## Acknowledgments

- Base model by [Steelskull](https://huggingface.co/Steelskull)
- Quantization via [llmcompressor](https://github.com/vllm-project/llm-compressor)
- Serving optimized for [vLLM](https://github.com/vllm-project/vllm)



---

**Want more FP8 models?** Check out my other quantizations!