Model Overview

• Model Architecture: DeepSeek-R1-0528
  • Input: Text
  • Output: Text
• Supported Hardware Microarchitecture: AMD MI350/MI355
• ROCm: 7.0
• PyTorch: 2.8.0
• Transformers: 4.56.1
• Operating System(s): Linux
• Inference Engine: SGLang/vLLM

Model Details

In the original modeling_deepseek.py file from the unsloth/DeepSeek-R1-0528-BF16 model, there is no definition or implementation of the MTP (Multi-Token-Predictor) layer. As a result, when you load the original model, there is no MTP layer included, and MTP-specific quantization cannot be performed.

To enable MTP layer loading and quantization, this model is adapted from unsloth/DeepSeek-R1-0528-BF16 by adding an MTP layer in the modeling_deepseek.py file. With this modification, it is possible to use AMD-Quark to quantize the DeepSeek-R1-0528 model with the MTP layer included.

Important Notes:

• When loading this model, you must set trust_remote_code=True to ensure that changes related to the MTP layer in modeling_deepseek.py take effect.
• After loading this model with transformers, evaluation should NOT be performed directly. The reason is that the forward function for the added MTP layer in modeling_deepseek.py is implemented only for calibration during the quantization process, so computation is not guaranteed to be the same as the original DeepSeek-R1-0528.
• Therefore, when quantizing with AMD-Quark, you must add the --skip_evaluation option to skip the evaluation step and only perform quantization.
• To skip quantization for the MTP layers, set exclude_layers="lm_head *self_attn* *mlp.gate *eh_proj *shared_head.head model.layers.61.*".

Below is an example of how to quantize this model:

cd Quark/examples/torch/language_modeling/llm_ptq/
exclude_layers="lm_head *self_attn* *mlp.gate *eh_proj *shared_head.head" 
python3 quantize_quark.py --model_dir $MODEL_DIR \
                          --quant_scheme w_mxfp4_a_mxfp4 \
                          --num_calib_data 32 \
                          --output_dir $output_dir \
                          --exclude_layers $exclude_layers \
                          --dataset pileval \
                          --multi_gpu \
                          --model_export hf_format \
                          --trust_remote_code \
                          --skip_evaluation \
                          --seq_len 512

For further details or issues, please refer to the AMD-Quark documentation or contact the respective developers.

License

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