amd/DeepSeek-R1-0528-MXFP4-ASQ

Model Overview

• Model Architecture: DeepSeek-R1-0528
  • Input: Text
  • Output: Text
• Supported Hardware Microarchitecture: AMD MI350/MI355
• ROCm: 7.0
• Operating System(s): Linux
• Inference Engine: SGLang
• Model Optimizer: AMD-Quark
  • Weight quantization: OCP MXFP4, Static
  • Activation quantization: OCP MXFP4, Dynamic
• Calibration Dataset: Pile

This model was built with deepseek-ai DeepSeek-R1-0528 model by applying AMD-Quark for MXFP4 quantization.

Model Quantization

The model was quantized from deepseek-ai/DeepSeek-R1-0528 using AMD-Quark. Both weights and activations were quantized to MXFP4 format, and the AutoSmoothQuant algorithm was applied to enhance accuracy.

Preprocessing requirement:

Before executing the quantization script below, the original FP8 model must first be dequantized to BFloat16. You can either perform the dequantization manually using this conversion script, or use the pre-converted BFloat16 model available at unsloth/DeepSeek-R1-0528-BF16.

Quantization scripts:

cd Quark/examples/torch/language_modeling/llm_ptq/
exclude_layers="*self_attn* *mlp.gate.* *lm_head"
python3 quantize_quark.py --model_dir $MODEL_DIR \
                          --quant_scheme w_mxfp4_a_mxfp4 \
                          --num_calib_data 128 \
                          --exclude_layers $exclude_layers \
                          --multi_gpu \
                          --quant_algo autosmoothquant \
                          --model_export hf_format \
                          --output_dir amd/DeepSeek-R1-0528-MXFP4-ASQ

Deployment

Use with SGLang

This model can be deployed efficiently using the SGLang backend.

Evaluation

The model was evaluated on AIME24, GPQA Diamond, MATH-500, and GSM8K benchmarks. The tasks of AIME24, GPQA Diamond, and MATH-500 were conducted using the lighteval framework while GSM8K using lm-eval-harness.

Accuracy

Benchmark	DeepSeek-R1-0528	DeepSeek-R1-0528-MXFP4-ASQ(this model)	Recovery
AIME24	88.00	87.67	99.62%
GPQA Diamond	79.90	79.65	99.69%
MATH-500	97.06	96.90	99.84%
GSM8K	95.30	95.18	99.87%

Reproduction

The results can be reproduced using vLLM emulation docker: rocmshared/pytorch:vllm-gfx950-mxfp4-mxfp6-v3. The results of AIME24, MATH-500, and GPQA Diamond, were obtained using vLLM while GSM8K was obtained using SGLang. For AIME24, MATH-500, and GPQA Diamond, we took 10 rounds with different random seeds for reliable performance.

MODEL_ARGS="model_name=amd/DeepSeek-R1-0528-MXFP4-ASQ,dtype=bfloat16,tensor_parallel_size=8,max_model_length=71536,max_num_batched_tokens=32768,gpu_memory_utilization=0.85,generation_parameters={max_new_tokens:65536,temperature:0.6,top_p:0.95,seed:$SEED}"
OUTPUT_DIR="results/DeepSeek-R1-0528-MXFP4-ASQ-Seed"
LOG="logs/deepseek_0528_maxfp4.log"

for i in $(seq 1 10); do
    # seed in [0, 2**30 - 1]
    SEED=$(shuf -i 0-1073741823 -n 1)

    lighteval vllm $MODEL_ARGS "custom|aime24_single|0|0,custom|math_500_single|0|0,custom|gpqa:diamond_single|0|0" \
        --use-chat-template \
        --output-dir "$OUTPUT_DIR/seed_$SEED" \
        2>&1 | tee -a "$LOG"

The result of GSM8K was obtained using lm-eval-harness and the following commands.

MODEL_ARGS="model=amd/DeepSeek-R1-0528-MXFP4-ASQ,base_url=http://localhost:8000/v1/completions,num_concurrent=999999,timeout=999999,tokenized_requests=False,max_length=38768,temperature=0.6,top_p=0.95,add_bos_token=True,seed=$SEED"
lm_eval \
    --model local-completions \
    --model_args $MODEL_ARGS \
    --tasks gsm8k \
    --num_fewshot 8 \
    --batch_size auto

License

Modifications Copyright(c) 2025 Advanced Micro Devices, Inc. All rights reserved.