DeepSeek-R1-Distill-Llama-8B-quantized.w4a16

Model Overview

Model Architecture: LlamaForCausalLM
- Input: Text
- Output: Text
Model Optimizations:
- Weight quantization: INT4
Release Date: 2/1/2025
Version: 1.0
Model Developers: Neural Magic

Quantized version of DeepSeek-R1-Distill-Llama-8B.

Model Optimizations

This model was obtained by quantizing the weights of DeepSeek-R1-Distill-Llama-8B 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 algorithm is applied for quantization, as implemented in the llm-compressor library.

Use with vLLM

This model can be deployed efficiently using the vLLM backend, as shown in the example below.

from transformers import AutoTokenizer
from vllm import LLM, SamplingParams

number_gpus = 1
model_name = "neuralmagic/DeepSeek-R1-Distill-Llama-8B-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 for more details.

Creation

This model was created with llm-compressor by running the code snippet below.

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-Llama-8B"
model_name = model_stub.split("/")[-1]

num_samples = 1024
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 and V2, using the following commands:

OpenLLM Leaderboard V1:

lm_eval \
  --model vllm \
  --model_args pretrained="neuralmagic/DeepSeek-R1-Distill-Llama-8B-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-Llama-8B-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-Llama-8B	neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w4a16	Recovery
OpenLLM V1	ARC-Challenge (Acc-Norm, 25-shot)	45.05	42.92	95.3%
	GSM8K (Strict-Match, 5-shot)	62.77	63.08	100.5%
	HellaSwag (Acc-Norm, 10-shot)	76.78	74.62	97.2%
	MMLU (Acc, 5-shot)	55.65	52.11	93.6%
	TruthfulQA (MC2, 0-shot)	50.55	48.57	96.1%
	Winogrande (Acc, 5-shot)	68.51	68.43	99.9%
	Average Score	59.88	58.29	97.3%
OpenLLM V2	IFEval (Inst Level Strict Acc, 0-shot)	38.34	37.85	98.7%
	BBH (Acc-Norm, 3-shot)	38.19	38.19	100.0%
	Math-Hard (Exact-Match, 4-shot)	0.00	0.00	---
	GPQA (Acc-Norm, 0-shot)	28.87	24.15	83.7%
	MUSR (Acc-Norm, 0-shot)	33.31	34.63	104.0%
	MMLU-Pro (Acc, 5-shot)	20.10	19.78	98.4%
	Average Score	26.47	25.77	97.4%
Coding	HumanEval (pass@1)	49.90	49.40	99.0%
	HumanEval (pass@10)	68.90	69.30	100.6%
	HumanEval+ (pass@10)	44.10	45.00	102.0%
	HumanEval+ (pass@10)	62.90	63.40	100.8%

Inference Performance

This model achieves up to 2.6x speedup in single-stream deployment and up to 1.3x speedup in multi-stream asynchronous deployment, depending on hardware and use-case scenario. The following performance benchmarks were conducted with vLLM version 0.7.2, and GuideLLM.

Benchmarking Command

guidellm --model neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w4a16 --target "http://localhost:8000/v1" --data-type emulated --data "prompt_tokens=<prompt_tokens>,generated_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-Llama-8B	---	3.0	1511	6.0	755	3.0	1483	3.1	1462	23.6	191	24.0	188	12.7	353	41.1	110
	neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w8a8	1.53	1.9	2356	3.8	1175	2.0	2291	2.0	2207	15.2	297	15.5	290	8.5	531	28.6	157
	neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w4a16	2.35	1.2	3870	2.3	1918	1.3	3492	1.3	3335	9.1	492	9.5	472	5.8	771	22.7	198
A100x1	deepseek-ai/DeepSeek-R1-Distill-Llama-8B	---	1.5	1308	3.1	657	1.6	1274	1.6	1263	12.1	166	12.4	162	6.5	308	25.6	78
	neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w8a8	1.30	1.1	1763	2.3	882	1.2	1716	1.2	1698	9.0	223	9.2	218	4.9	409	25.7	78
	neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w4a16	1.76	0.8	2501	1.6	1236	0.9	2350	0.9	2287	6.4	316	6.6	306	3.7	544	24.7	82
H100x1	deepseek-ai/DeepSeek-R1-Distill-Llama-8B	---	1.0	1146	1.9	574	1.0	1128	1.0	1111	7.6	144	7.7	142	4.1	266	16.3	67
	neuralmagic/DeepSeek-R1-Distill-Llama-8B-FP8-dynamic	1.25	0.7	1567	1.4	758	0.7	1484	0.7	1462	5.7	191	5.8	189	3.2	347	22.5	49
	neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w4a16	1.30	0.7	1527	1.4	768	0.7	1495	0.7	1463	5.6	194	5.7	190	3.1	350	14.7	74

**Use case profiles: prompt tokens / generation tokens

**QPD: Queries per dollar, based on on-demand cost at Lambda Labs (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
A6000x1	deepseek-ai/DeepSeek-R1-Distill-Llama-8B	---	12.6	56742	5.7	25687	6.5	29349	5.2	23259	1.6	7250	1.2	5181	0.8	3445	0.1	616
	neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w8a8	1.34	17.4	78101	7.6	34351	8.8	39790	7.0	31532	2.3	10405	1.5	6960	1.0	4355	0.2	785
	neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w4a16	0.91	10.9	48964	5.1	22989	4.8	21791	3.8	17039	2.2	9726	1.2	5434	0.6	2544	0.1	578
A100x1	deepseek-ai/DeepSeek-R1-Distill-Llama-8B	---	24.5	49296	11.3	22657	13.0	26047	10.5	21020	3.5	7029	2.5	4995	1.7	3503	0.3	659
	neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w8a8	1.27	30.8	62042	14.1	28419	17.2	34554	13.8	27719	4.6	9299	3.1	6215	2.2	4331	0.4	807
	neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w4a16	0.97	22.7	45708	10.5	21216	11.1	22353	8.9	17939	3.9	7758	2.6	5241	1.6	3196	0.4	718
H100x1	deepseek-ai/DeepSeek-R1-Distill-Llama-8B	---	49.0	53593	22.6	24750	28.3	30971	22.9	25035	7.2	7912	5.1	5561	3.6	3939	0.6	703
	neuralmagic/DeepSeek-R1-Distill-Llama-8B-FP8-dynamic	1.14	57.1	62517	26.0	28440	34.5	37781	28.7	31360	7.2	7877	5.4	5923	4.3	4697	0.7	782
	neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w4a16	1.01	49.8	54452	22.9	25035	28.5	31162	23.0	25200	6.8	7493	5.0	5431	3.7	4079	0.7	787