--- license: mit tags: - deepseek - int4 - vllm - llmcompressor base_model: deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B library_name: transformers --- # DeepSeek-R1-Distill-Qwen-1.5B-quantized.w4a16 ## Model Overview - **Model Architecture:** Qwen2ForCausalLM - **Input:** Text - **Output:** Text - **Model Optimizations:** - **Weight quantization:** INT4 - **Release Date:** 2/7/2025 - **Version:** 1.0 - **Model Developers:** Neural Magic Quantized version of [DeepSeek-R1-Distill-Qwen-1.5B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B). ### Model Optimizations This model was obtained by quantizing the weights of [DeepSeek-R1-Distill-Qwen-1.5B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B) 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-1.5B-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-1.5B" 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-1.5B-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-1.5B-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-1.5B neuralmagic/DeepSeek-R1-Distill-Qwen-1.5B-quantized.w4a16 Recovery
OpenLLM V1 ARC-Challenge (Acc-Norm, 25-shot) 37.20 35.84 96.3%
GSM8K (Strict-Match, 5-shot) 69.98 68.01 97.2%
HellaSwag (Acc-Norm, 10-shot) 43.86 42.38 96.6%
MMLU (Acc, 5-shot) 37.38 36.98 98.9%
TruthfulQA (MC2, 0-shot) 45.21 46.68 103.3%
Winogrande (Acc, 5-shot) 54.30 55.49 102.2%
Average Score 47.99 47.56 99.1%
OpenLLM V2 IFEval (Inst Level Strict Acc, 0-shot) 34.63 34.21 98.8%
BBH (Acc-Norm, 3-shot) 2.40 2.82 ---
Math-Hard (Exact-Match, 4-shot) 0.00 0.00 ---
GPQA (Acc-Norm, 0-shot) 0.93 1.19 ---
MUSR (Acc-Norm, 0-shot) 1.26 1.37 ---
MMLU-Pro (Acc, 5-shot) 1.32 1.31 ---
Average Score 6.80 6.82 ---
Coding HumanEval (pass@1) 37.90 35.70 94.2%
HumanEval (pass@10) 61.30 61.40 100.2%
HumanEval+ (pass@10) 33.00 31.90 96.7%
HumanEval+ (pass@10) 55.90 55.50 99.3%
## Inference Performance This model achieves up to 1.5x speedup in single-stream 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-1.5B-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
Hardware 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
A6000x1 deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B --- 0.8 5667 1.6 2776 0.8 5515 0.8 5466 6.4 705 6.5 697 3.5 1295 18.3 246
neuralmagic/DeepSeek-R1-Distill-Qwen-1.5B-quantized.w8a8 1.14 0.7 6635 1.3 3340 0.7 6396 0.7 6343 5.3 845 5.4 832 2.9 1547 21.3 211
neuralmagic/DeepSeek-R1-Distill-Qwen-1.5B-quantized.w4a16 1.38 0.5 8293 1.1 4184 0.6 7976 0.6 7504 4.3 1051 4.4 1033 2.5 1819 21.1 213
A100x1 deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B --- 0.6 3359 1.2 1654 0.6 3286 0.6 3241 4.7 424 4.9 411 2.6 778 21.1 95
neuralmagic/DeepSeek-R1-Distill-Qwen-1.5B-quantized.w8a8 1.05 0.6 3531 1.1 1807 0.6 3427 0.6 3480 4.5 448 4.5 447 2.4 842 23.5 86
neuralmagic/DeepSeek-R1-Distill-Qwen-1.5B-quantized.w4a16 1.03 0.6 3469 1.1 1751 0.6 3403 0.6 3407 4.5 447 4.6 435 2.5 815 23.3 86
H100x1 deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B --- 0.4 2604 0.8 1299 0.4 2543 0.4 2551 3.3 330 3.4 326 1.8 612 14.0 78
neuralmagic/DeepSeek-R1-Distill-Qwen-1.5B-FP8-dynamic 1.04 0.4 2694 0.8 1364 0.4 2670 0.4 2639 3.2 347 3.2 341 1.6 673 14.1 78
neuralmagic/DeepSeek-R1-Distill-Qwen-1.5B-quantized.w4a16 0.84 0.5 2111 1.0 1065 0.5 2068 0.5 2119 4.1 270 4.1 265 2.1 530 15.1 73
**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).