--- license: mit tags: - deepseek - fp8 - vllm base_model: deepseek-ai/DeepSeek-R1-Distill-Llama-70B library_name: transformers --- # DeepSeek-R1-Distill-Llama-70B-FP8-dynamic ## Model Overview - **Model Architecture:** LlamaForCausalLM - **Input:** Text - **Output:** Text - **Model Optimizations:** - **Weight quantization:** FP8 - **Activation quantization:** FP8 - **Release Date:** 2/1/2025 - **Version:** 1.0 - **Model Developers:** Neural Magic Quantized version of [DeepSeek-R1-Distill-Llama-70B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B). ### Model Optimizations This model was obtained by quantizing the weights and activations of [DeepSeek-R1-Distill-Llama-70B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B) to FP8 data type. This optimization reduces the number of bits per parameter from 16 to 8, reducing the disk size and GPU memory requirements by approximately 50%. Only the weights and activations of the linear operators within transformers blocks are quantized. Weights are quantized using a symmetric per-channel scheme, whereas quantizations are quantized using a symmetric per-token scheme. [LLM Compressor](https://github.com/vllm-project/llm-compressor) is used for quantization. ## Use with vLLM This model can be deployed efficiently using the [vLLM](https://docs.vllm.ai/en/latest/) backend, as shown in the example below. ```python from transformers import AutoTokenizer from vllm import LLM, SamplingParams number_gpus = 2 model_name = "neuralmagic/DeepSeek-R1-Distill-Llama-70B-FP8-dynamic" tokenizer = AutoTokenizer.from_pretrained(model_name) sampling_params = SamplingParams(temperature=0.6, max_tokens=256, stop_token_ids=[tokenizer.eos_token_id]) llm = LLM(model=model_name, tensor_parallel_size=number_gpus, trust_remote_code=True) messages_list = [ [{"role": "user", "content": "Who are you? Please respond in pirate speak!"}], ] prompt_token_ids = [tokenizer.apply_chat_template(messages, add_generation_prompt=True) for messages in messages_list] outputs = llm.generate(prompt_token_ids=prompt_token_ids, sampling_params=sampling_params) generated_text = [output.outputs[0].text for output in outputs] print(generated_text) ``` vLLM also supports OpenAI-compatible serving. See the [documentation](https://docs.vllm.ai/en/latest/) for more details. ## Creation This model was created with [llm-compressor](https://github.com/vllm-project/llm-compressor) by running the code snippet below. ```python from transformers import AutoModelForCausalLM, AutoTokenizer from llmcompressor.modifiers.quantization import QuantizationModifier from llmcompressor.transformers import oneshot from llmcompressor.transformers.compression.helpers import calculate_offload_device_map import os # Load model model_stub = "deepseek-ai/DeepSeek-R1-Distill-Llama-70B" model_name = model_stub.split("/")[-1] device_map = calculate_offload_device_map( model_stub, reserve_for_hessians=True, num_gpus=2, torch_dtype="auto", ) model = AutoModelForCausalLM.from_pretrained( model_stub, device_map=device_map, torch_dtype="auto", ) tokenizer = AutoTokenizer.from_pretrained(model_stub) # Configure the quantization algorithm and scheme recipe = QuantizationModifier( targets="Linear", scheme="FP8_DYNAMIC", ignore=["lm_head"], ) # Apply quantization oneshot( model=model, recipe=recipe, ) # Save to disk in compressed-tensors format save_path = model_name + "-FP8-dynamic model.save_pretrained(save_path) tokenizer.save_pretrained(save_path) print(f"Model and tokenizer saved to: {save_path}") ``` ## Evaluation The model was evaluated on OpenLLM Leaderboard [V1](https://huggingface.co/spaces/open-llm-leaderboard-old/open_llm_leaderboard) and [V2](https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/), using the following commands: OpenLLM Leaderboard V1: ``` lm_eval \ --model vllm \ --model_args pretrained="neuralmagic/DeepSeek-R1-Distill-Llama-70B-FP8-dynamic",dtype=auto,max_model_len=4096,tensor_parallel_size=2,enable_chunked_prefill=True \ --tasks openllm \ --write_out \ --batch_size auto \ --output_path output_dir \ --show_config ``` OpenLLM Leaderboard V2: ``` lm_eval \ --model vllm \ --model_args pretrained="neuralmagic/DeepSeek-R1-Distill-Llama-70B-FP8-dynamic",dtype=auto,max_model_len=4096,tensor_parallel_size=2,enable_chunked_prefill=True \ --apply_chat_template \ --fewshot_as_multiturn \ --tasks leaderboard \ --write_out \ --batch_size auto \ --output_path output_dir \ --show_config ``` ### Accuracy
Category Metric deepseek-ai/DeepSeek-R1-Distill-Llama-70B neuralmagic/DeepSeek-R1-Distill-Llama-70B-FP8-dynamic Recovery
OpenLLM V1 ARC-Challenge (Acc-Norm, 25-shot) 63.65 63.05 99.1%
GSM8K (Strict-Match, 5-shot) 93.03 93.03 100.0%
HellaSwag (Acc-Norm, 10-shot) 84.85 84.71 99.8%
MMLU (Acc, 5-shot) 78.04 77.45 99.3%
TruthfulQA (MC2, 0-shot) 56.67 56.62 99.9%
Winogrande (Acc, 5-shot) 78.22 78.45 100.3%
Average Score 75.74 75.55 99.8%
OpenLLM V2 IFEval (Inst Level Strict Acc, 0-shot) 43.15 42.75 99.1%
BBH (Acc-Norm, 3-shot) 64.32 64.31 100.0%
Math-Hard (Exact-Match, 4-shot) 35.04 36.59 104.4%
GPQA (Acc-Norm, 0-shot) 37.15 37.46 100.8%
MUSR (Acc-Norm, 0-shot) 42.89 42.76 99.7%
MMLU-Pro (Acc, 5-shot) 47.22 46.64 98.8%
Average Score 44.96 45.08 100.3%
Coding HumanEval (pass@1) 81.10 81.00 99.9%
HumanEval (pass@10) 87.60 88.60 101.1%
HumanEval+ (pass@10) 75.20 75.50 100.4%
HumanEval+ (pass@10) 83.10 84.30 101.4%
## Inference Performance This model achieves up to 1.4x speedup in single-stream deployment and up to 3.0x speedup in multi-stream asynchronous deployment, depending on hardware and use-case scenario. The following performance benchmarks were conducted with [vLLM](https://docs.vllm.ai/en/latest/) version 0.7.2, and [GuideLLM](https://github.com/neuralmagic/guidellm).
Benchmarking Command ``` guidellm --model neuralmagic/DeepSeek-R1-Distill-Llama-70B-FP8-dynamic --target "http://localhost:8000/v1" --data-type emulated --data "prompt_tokens=,generated_tokens=" --max seconds 360 --backend aiohttp_server ```
### Single-stream performance (measured with vLLM version 0.7.2)
Instruction Following
256 / 128		 Multi-turn Chat
512 / 256		 Docstring Generation
768 / 128		 RAG
1024 / 128		 Code Completion
256 / 1024		 Code Fixing
1024 / 1024		 Large Summarization
4096 / 512		 Large RAG
10240 / 1536
GPU class Number of GPUs Model Average cost reduction Latency (s) QPD Latency (s) QPD Latency (s) QPD Latency (s) QPD Latency (s) QPD Latency (s) QPD Latency (s) QPD Latency (s) QPD
A6000 4 deepseek-ai/DeepSeek-R1-Distill-Llama-70B --- 7.4 152 14.9 76 7.5 149 7.7 146 57.2 20 58.9 19 31.9 35 98.4 11
2 neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w8a8 1.93 7.7 292 15.2 148 7.8 287 8.0 282 60.7 37 60.2 37 32.3 70 104.0 22
2 neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w4a16 2.83 4.9 457 10.0 225 5.5 411 5.8 389 38.9 58 39.2 57 23.7 95 76.6 29
A100 2 deepseek-ai/DeepSeek-R1-Distill-Llama-70B --- 6.4 157 12.8 79 6.6 153 6.7 151 50.4 20 50.8 20 27.0 37 85.4 12
2 neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w8a8 1.48 4.1 245 8.2 123 4.2 238 4.3 235 32.4 31 32.8 31 17.6 57 90.8 11
1 neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w4a16 2.69 4.6 440 9.2 220 4.9 407 5.2 389 35.3 57 36.3 55 21.2 95 68.1 30
H100 2 deepseek-ai/DeepSeek-R1-Distill-Llama-70B --- 3.8 149 7.6 74 3.9 146 3.9 144 30.0 19 30.4 19 16.1 35 56.5 10
2 neuralmagic/DeepSeek-R1-Distill-Llama-70B-FP8-dynamic 1.39 2.7 210 5.3 106 2.7 207 2.8 203 21.1 27 21.4 26 11.5 49 47.2 12
1 neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w4a16 1.83 4.0 277 7.9 138 4.1 266 4.2 262 31.2 35 31.8 34 17.8 61 61.4 18
**Use case profiles: prompt tokens / generation tokens **QPD: Queries per dollar, based on on-demand cost at [Lambda Labs](https://lambdalabs.com/service/gpu-cloud) (observed on 2/18/2025). ### Multi-stream asynchronous performance (measured with vLLM version 0.7.2)
Instruction Following
256 / 128		 Multi-turn Chat
512 / 256		 Docstring Generation
768 / 128		 RAG
1024 / 128		 Code Completion
256 / 1024		 Code Fixing
1024 / 1024		 Large Summarization
4096 / 512		 Large RAG
10240 / 1536
Hardware Model Average cost reduction Maximum throughput (QPS) QPD Maximum throughput (QPS) QPD Maximum throughput (QPS) QPD Maximum throughput (QPS) QPD Maximum throughput (QPS) QPD Maximum throughput (QPS) QPD Maximum throughput (QPS) QPD Maximum throughput (QPS) QPD
A6000x4 deepseek-ai/DeepSeek-R1-Distill-Llama-70B --- 3.65 4102 1.56 1757 1.90 2143 1.48 1665 0.44 493 0.34 380 0.22 245 0.05 55
neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w8a8 1.76 5.89 6625 2.94 3307 3.36 3775 2.59 2916 0.74 828 0.53 601 0.35 398 0.11 120
neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w4a16 1.48 4.91 5528 2.01 2259 2.03 2280 1.12 1255 1.11 1251 0.76 852 0.24 267 0.07 81
A100x4 deepseek-ai/DeepSeek-R1-Distill-Llama-70B --- 10.41 5235 5.10 2565 5.50 2766 4.36 2193 1.49 751 1.21 607 0.89 447 0.19 98
neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w8a8 1.63 18.11 9103 8.90 4477 9.41 4730 7.42 3731 2.44 1229 1.89 948 1.26 631 0.30 149
neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w4a16 1.12 12.63 6353 5.32 2673 5.58 2804 4.27 2144 2.30 1158 1.45 729 0.76 381 0.22 110
H100x4 deepseek-ai/DeepSeek-R1-Distill-Llama-70B --- 14.04 2113 10.85 1634 12.25 1844 9.93 1494 3.68 554 2.82 425 1.81 273 0.35 52
neuralmagic/DeepSeek-R1-Distill-Llama-70B-FP8-dynamic 1.78 41.44 6236 19.64 2956 21.03 3166 16.72 2516 6.01 904 4.46 672 2.55 383 0.49 74
neuralmagic/DeepSeek-R1-Distill-Llama-70B-quantized.w4a16 1.45 36.61 5509 15.12 2275 16.24 2443 13.22 1990 5.48 825 3.01 453 2.07 312 0.43 64
**Use case profiles: prompt tokens / generation tokens **QPS: Queries per second. **QPD: Queries per dollar, based on on-demand cost at [Lambda Labs](https://lambdalabs.com/service/gpu-cloud) (observed on 2/18/2025).