--- license: mit tags: - deepseek - int4 - vllm - llmcompressor base_model: deepseek-ai/DeepSeek-R1-Distill-Qwen-32B library_name: transformers --- # DeepSeek-R1-Distill-Qwen-32B-quantized.w4a16 ## Model Overview - **Model Architecture:** Qwen2ForCausalLM - **Input:** Text - **Output:** Text - **Model Optimizations:** - **Weight quantization:** INT4 - **Release Date:** 2/4/2025 - **Version:** 1.0 - **Model Developers:** Neural Magic Quantized version of [DeepSeek-R1-Distill-Qwen-32B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B). ### Model Optimizations This model was obtained by quantizing the weights of [DeepSeek-R1-Distill-Qwen-32B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B) to INT4 data type. This optimization reduces the number of bits per parameter from 16 to 4, reducing the disk size and GPU memory requirements by approximately 75%. Only the weights of the linear operators within transformers blocks are quantized. Weights are quantized using a symmetric per-group scheme, with group size 128. The [GPTQ](https://arxiv.org/abs/2210.17323) algorithm is applied for quantization, as implemented in the [llm-compressor](https://github.com/vllm-project/llm-compressor) library. ## 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 = 1 model_name = "neuralmagic/DeepSeek-R1-Distill-Qwen-32B-quantized.w4a16" 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.modifiers.smoothquant import SmoothQuantModifier from llmcompressor.transformers import oneshot # Load model model_stub = "deepseek-ai/DeepSeek-R1-Distill-Qwen-32B" model_name = model_stub.split("/")[-1] num_samples = 2048 max_seq_len = 8192 tokenizer = AutoTokenizer.from_pretrained(model_stub) model = AutoModelForCausalLM.from_pretrained( model_stub, device_map="auto", torch_dtype="auto", ) def preprocess_fn(example): return {"text": tokenizer.apply_chat_template(example["messages"], add_generation_prompt=False, tokenize=False)} ds = load_dataset("neuralmagic/LLM_compression_calibration", split="train") ds = ds.map(preprocess_fn) # Configure the quantization algorithm and scheme recipe = QuantizationModifier( targets="Linear", scheme="W4A16", ignore=["lm_head"], dampening_frac=0.1, ) # Apply quantization oneshot( model=model, dataset=ds, recipe=recipe, max_seq_length=max_seq_len, num_calibration_samples=num_samples, ) # Save to disk in compressed-tensors format save_path = model_name + "-quantized.w4a16 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-Qwen-32B-quantized.w4a16",dtype=auto,max_model_len=4096,tensor_parallel_size=1,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-Qwen-32B-quantized.w4a16",dtype=auto,max_model_len=4096,tensor_parallel_size=1,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-Qwen-32B neuralmagic/DeepSeek-R1-Distill-Qwen-32B-quantized.w4a16 Recovery
OpenLLM V1 ARC-Challenge (Acc-Norm, 25-shot) 64.59 63.65 98.6%
GSM8K (Strict-Match, 5-shot) 82.71 84.38 102.0%
HellaSwag (Acc-Norm, 10-shot) 83.80 83.32 99.4%
MMLU (Acc, 5-shot) 81.12 80.91 99.7%
TruthfulQA (MC2, 0-shot) 58.41 58.54 100.2%
Winogrande (Acc, 5-shot) 76.40 75.14 98.4%
Average Score 74.51 74.32 99.8%
OpenLLM V2 IFEval (Inst Level Strict Acc, 0-shot) 42.87 72.48 99.1%
BBH (Acc-Norm, 3-shot) 57.96 57.54 99.3%
Math-Hard (Exact-Match, 4-shot) 0.00 0.00 ---
GPQA (Acc-Norm, 0-shot) 26.95 26.41 98.0%
MUSR (Acc-Norm, 0-shot) 43.95 44.34 100.9%
MMLU-Pro (Acc, 5-shot) 49.82 48.43 97.2%
Average Score 36.92 %
Coding HumanEval (pass@1) 86.00 86.00 100.0%
HumanEval (pass@10) 92.50 92.60 100.1%
HumanEval+ (pass@10) 82.00 80.60 98.3%
HumanEval+ (pass@10) 88.70 89.70 101.1%
## Inference Performance This model achieves up to 3.4x speedup in single-stream deployment and up to 2.0x 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-Qwen-32B-quantized.w4a16 --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 2 deepseek-ai/DeepSeek-R1-Distill-Qwen-32B --- 6.3 359 12.8 176 6.5 347 6.6 342 49.9 45 50.8 44 26.6 85 83.4 27
1 neuralmagic/DeepSeek-R1-Distill-Qwen-32B-quantized.w8a8 1.81 6.9 648 13.8 325 7.2 629 7.2 622 54.8 82 55.6 81 30.0 150 94.8 47
1 neuralmagic/DeepSeek-R1-Distill-Qwen-32B-quantized.w4a16 3.07 3.9 1168 7.8 580 4.3 1041 4.6 975 29.7 151 30.9 146 19.3 233 61.4 73
A100 1 deepseek-ai/DeepSeek-R1-Distill-Qwen-32B --- 5.6 361 11.1 180 5.7 350 5.8 347 44.0 46 44.7 45 23.6 85 73.7 27
1 neuralmagic/DeepSeek-R1-Distill-Qwen-32B-quantized.w8a8 1.50 3.7 547 7.3 275 3.8 536 3.8 528 29.0 69 29.5 68 15.7 128 53.1 38
1 neuralmagic/DeepSeek-R1-Distill-Qwen-32B-quantized.w4a16 2.30 2.2 894 4.5 449 2.4 831 2.5 798 17.4 116 18.0 112 10.5 191 49.5 41
H100 1 deepseek-ai/DeepSeek-R1-Distill-Qwen-32B --- 3.3 327 6.7 163 3.4 320 3.4 317 26.6 41 26.9 41 14.3 77 47.8 23
1 neuralmagic/DeepSeek-R1-Distill-Qwen-32B-FP8-dynamic 1.52 2.2 503 4.3 252 2.2 490 2.3 485 17.3 63 17.5 63 9.5 116 33.4 33
1 neuralmagic/DeepSeek-R1-Distill-Qwen-32B-quantized.w4a16 1.61 2.1 532 4.1 268 2.1 516 2.1 513 16.1 68 16.5 66 9.1 120 31.9 34
**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
A6000x2 deepseek-ai/DeepSeek-R1-Distill-Qwen-32B --- 6.2 13940 1.9 4348 2.7 6153 2.1 4778 0.6 1382 0.4 930 0.3 685 0.1 124
neuralmagic/DeepSeek-R1-Distill-Qwen-32B-quantized.w8a8 1.80 8.7 19492 4.2 9474 4.1 9290 3.0 6802 1.2 2734 0.9 1962 0.5 1177 0.1 254
neuralmagic/DeepSeek-R1-Distill-Qwen-32B-quantized.w4a16 1.30 5.9 13366 2.5 5733 2.4 5409 1.6 3525 1.2 2757 0.7 1663 0.3 676 0.1 214
A100x2 deepseek-ai/DeepSeek-R1-Distill-Qwen-32B --- 12.9 13016 5.8 5848 6.3 6348 5.1 5146 2.0 1988 1.5 1463 0.9 869 0.2 192
neuralmagic/DeepSeek-R1-Distill-Qwen-32B-quantized.w8a8 1.52 21.4 21479 8.9 8948 10.6 10611 8.2 8197 3.0 3018 2.0 2054 1.2 1241 0.3 264
neuralmagic/DeepSeek-R1-Distill-Qwen-32B-quantized.w4a16 1.09 13.5 13568 6.5 6509 6.0 6075 4.7 4754 2.8 2790 1.6 1651 0.9 862 0.2 225
H100x2 deepseek-ai/DeepSeek-R1-Distill-Qwen-32B --- 25.5 14392 12.5 7035 14.0 7877 11.3 6364 3.6 2041 2.7 1549 1.9 1057 0.4 200
neuralmagic/DeepSeek-R1-Distill-Qwen-32B-FP8-dynamic 1.46 46.7 25538 20.3 11082 23.3 12728 18.4 10049 5.3 2881 3.7 2097 2.6 1445 0.5 256
neuralmagic/DeepSeek-R1-Distill-Qwen-32B-quantized.w4a16 1.23 36.9 20172 17.4 9500 18.0 9822 14.2 7755 5.3 2900 3.3 1867 2.3 1265 0.4 241
**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).