haoyang-amd/ts

Note that the MTP layers of this model are also PTPC-quantized.

Model Overview

• Model Architecture: DeepSeek-V3.2-Speciale
  • Input: Text
  • Output: Text
• Supported Hardware Microarchitecture: AMD MI350/MI355
• ROCm: 7.0
• Operating System(s): Linux
• Inference Engine: SGLang/vLLM
• Model Optimizer: AMD-Quark (V0.10)
  • Weight quantization: Perchannel, FP8E4M3, Static
  • Activation quantization: Pertoken, FP8E4M3, Dynamic
• Calibration Dataset: Pile

This model was built with deepseek-ai/DeepSeek-V3.2-Speciale model by applying AMD-Quark for FP8E4M3 PTPC quantization.

Model Quantization

The model was quantized from deepseek-ai/DeepSeek-V3.2-Speciale using AMD-Quark. The weights are quantized to FP8 and activations are quantized to FP8.

Accuracy

Benchmark	DeepSeek-V3.2-Speciale	DeepSeek-V3.2-Speciale-ptpc(this model)
gsm8k	96.00	95.75

Reproduction

vllm version: 0.11.2.dev521+gad32e3e19.rocm710

aiter version: 0.1.6.post2.dev55+g59bd8ff2c

lm_eval version: 0.4.9.2

export VLLM_USE_V1=1  
export SAFETENSORS_FAST_GPU=1
export VLLM_ROCM_USE_AITER=1
export VLLM_ROCM_USE_AITER_MOE=1
model_path="/model_path/deepseek-ai/DeepSeek-V3.2-Speciale-ptpc"
vllm serve $model_path \
  --tensor-parallel-size 8 \
  --data-parallel-size 1 \
  --max-num-batched-tokens 32768 \
  --trust-remote-code \
  --no-enable-prefix-caching \
  --disable-log-requests \
  --kv-cache-dtype bfloat16 \
  --gpu_memory_utilization 0.85 \
  --compilation-config '{"cudagraph_mode": "FULL_AND_PIECEWISE"}' \
  --block-size 1

lm_eval \
  --model local-completions \
  --tasks gsm8k \
  --model_args model=/model_path/deepseek-ai/DeepSeek-V3.2-Speciale-ptpc,base_url=http://127.0.0.1:8000/v1/completions \
  --batch_size auto \
  --limit 400

Deployment

This model can be deployed efficiently using the vLLM backends.

License

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