DeepSeek-R1-Distill-Qwen-14B-quantized.w8a8

Model Overview

Model Architecture: Qwen2ForCausalLM
- Input: Text
- Output: Text
Model Optimizations:
- Weight quantization: INT8
- Activation quantization: INT8
Release Date: 2/4/2025
Version: 1.0
Model Developers: Neural Magic

Quantized version of DeepSeek-R1-Distill-Qwen-14B.

Model Optimizations

This model was obtained by quantizing the weights and activations of DeepSeek-R1-Distill-Qwen-14B to INT8 data type. This optimization reduces the number of bits used to represent weights and activations from 16 to 8, reducing GPU memory requirements (by approximately 50%) and increasing matrix-multiply compute throughput (by approximately 2x). Weight quantization also reduces disk size 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. 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-Qwen-14B-quantized.w8a8"

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-Qwen-14B"
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 = [
    SmoothQuantModifier(smoothing_strength=0.8),
    QuantizationModifier(
        targets="Linear",
        scheme="W8A8",
        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.w8a8
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-Qwen-14B-quantized.w8a8",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-14B-quantized.w8a8",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-14B	neuralmagic/DeepSeek-R1-Distill-Qwen-14B-quantized.w8a8	Recovery
OpenLLM V1	ARC-Challenge (Acc-Norm, 25-shot)	58.79	57.85	98.4%
	GSM8K (Strict-Match, 5-shot)	87.04	87.79	100.9%
	HellaSwag (Acc-Norm, 10-shot)	81.51	81.04	99.4%
	MMLU (Acc, 5-shot)	74.46	74.26	99.7%
	TruthfulQA (MC2, 0-shot)	54.77	54.94	100.3%
	Winogrande (Acc, 5-shot)	69.38	70.48	101.6%
	Average Score	70.99	71.06	100.1%
OpenLLM V2	IFEval (Inst Level Strict Acc, 0-shot)	43.05	42.78	99.4%
	BBH (Acc-Norm, 3-shot)	47.16	47.53	100.8%
	Math-Hard (Exact-Match, 4-shot)	0.00	0.00	---
	GPQA (Acc-Norm, 0-shot)	35.07	37.22	106.2%
	MUSR (Acc-Norm, 0-shot)	45.14	43.56	96.5%
	MMLU-Pro (Acc, 5-shot)	34.86	33.63	96.5%
	Average Score	34.21	34.12	99.7%
Coding	HumanEval (pass@1)	78.90	78.40	99.4%
	HumanEval (pass@10)	89.80	90.10	100.3%
	HumanEval+ (pass@10)	72.60	72.40	99.7%
	HumanEval+ (pass@10)	84.90	84.90	100.0%

Inference Performance

This model achieves up to 1.6x speedup in both single-stream and 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-Qwen-14B-quantized.w8a8 --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-Qwen-14B	---	5.4	837	10.7	419	5.5	813	5.6	805	42.2	107	42.8	105	22.9	197	71.7	63
	neuralmagic/DeepSeek-R1-Distill-Qwen-14B-quantized.w8a8	1.59	3.3	1345	6.7	673	3.4	1315	3.5	1296	26.5	170	26.8	168	14.5	310	48.3	93
	neuralmagic/DeepSeek-R1-Distill-Qwen-14B-quantized.w4a16	2.51	2.0	2275	4.0	1127	2.2	2072	2.3	1945	15.3	294	15.9	283	9.9	456	36.6	123
A100x1	deepseek-ai/DeepSeek-R1-Distill-Qwen-14B	---	2.6	765	5.2	383	2.7	746	2.7	732	20.8	97	21.2	95	11.3	179	36.7	55
	neuralmagic/DeepSeek-R1-Distill-Qwen-14B-quantized.w8a8	1.34	1.9	1072	3.8	533	1.9	1045	1.9	1032	14.8	136	15.2	132	8.1	248	39.6	51
	neuralmagic/DeepSeek-R1-Distill-Qwen-14B-quantized.w4a16	1.93	1.2	1627	2.5	810	1.3	1530	1.4	1474	9.7	208	10.2	197	5.8	348	37.6	53
H100x1	deepseek-ai/DeepSeek-R1-Distill-Qwen-14B	---	1.6	672	3.3	334	1.7	662	1.7	652	12.8	85	13.0	84	7.0	155	25.2	43
	neuralmagic/DeepSeek-R1-Distill-Qwen-14B-FP8-dynamic	1.33	1.2	925	2.3	467	1.2	908	1.2	896	9.3	118	9.5	115	5.2	210	23.9	46
	neuralmagic/DeepSeek-R1-Distill-Qwen-14B-quantized.w4a16	1.37	1.2	944	2.3	474	1.2	931	1.2	907	9.1	121	9.2	119	5.1	214	22.5	49

**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-Qwen-14B	---	13.7	30785	5.5	12327	6.5	14517	5.1	11439	2.0	4434	1.3	2982	0.6	1462	0.2	371
	neuralmagic/DeepSeek-R1-Distill-Qwen-14B-quantized.w8a8	1.44	21.4	48181	8.2	18421	9.8	22051	7.8	17462	2.8	6281	1.7	3758	1.0	2335	0.2	419
	neuralmagic/DeepSeek-R1-Distill-Qwen-14B-quantized.w4a16	0.98	12.7	28540	5.7	12796	5.4	12218	3.7	8401	2.5	5583	1.3	2987	0.7	1489	0.2	368
A100x1	deepseek-ai/DeepSeek-R1-Distill-Qwen-14B	---	15.6	31306	7.1	14192	7.7	15435	6.0	11971	2.4	4878	1.6	3298	0.9	1862	0.2	355
	neuralmagic/DeepSeek-R1-Distill-Qwen-14B-quantized.w8a8	1.31	20.8	41907	9.3	18724	10.5	21043	8.4	16886	3.0	5975	1.9	3917	1.2	2481	0.2	464
	neuralmagic/DeepSeek-R1-Distill-Qwen-14B-quantized.w4a16	0.94	14.0	28146	6.5	13042	6.5	12987	5.1	10194	2.6	5269	1.5	2925	0.9	1849	0.2	382
H100x1	deepseek-ai/DeepSeek-R1-Distill-Qwen-14B	---	31.4	34404	14.1	15482	16.6	18149	13.3	14572	4.7	5099	2.6	2849	1.9	2060	0.3	347
	neuralmagic/DeepSeek-R1-Distill-Qwen-14B-FP8-dynamic	1.31	40.9	44729	18.5	20260	22.1	24165	18.1	19779	5.7	6246	3.4	3681	2.5	2746	0.4	474
	neuralmagic/DeepSeek-R1-Distill-Qwen-14B-quantized.w4a16	1.12	33.3	36387	15.0	16453	17.6	19241	14.2	15576	4.6	5034	3.0	3292	2.2	2412	0.4	481