| --- |
| tags: |
| - fp8 |
| - quantized |
| - mistral |
| - roleplay |
| - creative-writing |
| - reasoning |
| base_model: TheDrummer/Behemoth-R1-123B-v2 |
| library_name: transformers |
| pipeline_tag: text-generation |
| license: apache-2.0 |
| --- |
| |
| # Behemoth-R1-123B-v2 FP8 Dynamic |
|
|
| FP8 Dynamic quantization of [TheDrummer/Behemoth-R1-123B-v2](https://huggingface.co/TheDrummer/Behemoth-R1-123B-v2) using llmcompressor. |
|
|
| ## Model Details |
|
|
| - **Base Model**: TheDrummer/Behemoth-R1-123B-v2 (Mistral Large 2411 finetune) |
| - **Quantization**: FP8 Dynamic (W8A8) via llmcompressor |
| - **Scheme**: FP8_DYNAMIC, lm_head excluded |
| - **Size**: ~123 GB (vs 246 GB FP16) |
| - **Format**: SafeTensors with compressed-tensors metadata |
|
|
| ## Usage with vLLM |
|
|
| ```bash |
| python3 -m vllm.entrypoints.openai.api_server \ |
| --model Irvollo/Behemoth-R1-123B-v2-FP8-Dynamic \ |
| --quantization compressed-tensors \ |
| --dtype bfloat16 \ |
| --max-model-len 32768 \ |
| --gpu-memory-utilization 0.95 \ |
| --enable-prefix-caching \ |
| --trust-remote-code |
| ``` |
|
|
| ## Reasoning / Thinking |
|
|
| Supports native reasoning via `<think>` tag prefill: |
|
|
| ```json |
| { |
| "messages": [ |
| {"role": "user", "content": "Your question"}, |
| {"role": "assistant", "content": "<think>\n"} |
| ], |
| "continue_final_message": true, |
| "add_generation_prompt": false |
| } |
| ``` |
|
|
| ## Hardware Requirements |
|
|
| - **Single GPU**: H200 NVL (141 GB) — tight with ~18 GB KV cache |
| - **Recommended**: 2x A100 80GB or H100 for comfortable KV headroom |
|
|
| ## Quantization Details |
|
|
| - Quantized on 2x NVIDIA B200 (358 GB VRAM) |
| - Calibration: 616 linear layers in <1 second |
| - Total pipeline: ~11 minutes |
| - Tool: [llmcompressor](https://github.com/vllm-project/llm-compressor) |
|
|
| ## Credits |
|
|
| - Original model by [TheDrummer](https://huggingface.co/TheDrummer) |
| - FP8 quantization by [Irvollo](https://huggingface.co/Irvollo) |
|
|
