jiaxwang/GLM-4.7-MXFP4

Model Overview

• Model Architecture: GLM-4.7
  • Input: Text
  • Output: Text
• Supported Hardware Microarchitecture: AMD MI350/MI355
• ROCm: 7.0
• Operating System(s): Linux
• Inference Engine: vLLM
• Model Optimizer: AMD-Quark
  • moe
    • Weight quantization: MOE-only, OCP MXFP4, Static
    • Activation quantization: MOE-only, OCP MXFP4, Dynamic
  • KV cache quantization: OCP FP8, Static
• Calibration Dataset: Pile

This model was built with GLM-4.7 model by applying AMD-Quark for MXFP4 quantization.

Model Quantization

The model was quantized from zai-org/GLM-4.7 using AMD-Quark. The weights and activations are quantized to MXFP4. AMD-Quark has been installed from source code inside the Docker image rocm/vllm-private:vllm_dev_base_mxfp4_20260122.

Quantization scripts:

Step1: Creat the quantize_glm.py

import runpy
from quark.torch import LLMTemplate

# Register GLM-4 MoE template
glm4_moe_template = LLMTemplate(
    model_type="glm4_moe",
    kv_layers_name=["*k_proj", "*v_proj"],
    q_layer_name="*q_proj",
    exclude_layers_name=["lm_head","*mlp.gate","*self_attn*","*shared_experts.*","*mlp.down_proj","*mlp.gate_proj","*mlp.up_proj"],
)
LLMTemplate.register_template(glm4_moe_template)
print(f"[INFO]: Registered template '{glm4_moe_template.model_type}'")

# Run quantize_quark.py
# Get the absolute path to the quantize_quark.py script
quantize_script = "/app/Quark/examples/torch/language_modeling/llm_ptq/quantize_quark.py"

runpy.run_path(quantize_script, run_name="__main__")

Step1: Quantize with the quantize_glm.py

export CUDA_VISIBLE_DEVICES=0,1,2,3
export MODEL_DIR=zai-org/GLM-4.7
export output_dir=amd/GLM-4.7-MXFP4

exclude_layers="*self_attn* *mlp.gate lm_head *mlp.gate_proj *mlp.up_proj *mlp.down_proj *shared_experts.*"
python3 quantize_glm.py --model_dir $MODEL_DIR \
                        --quant_scheme mxfp4 \
                        --num_calib_data 128 \
                        --exclude_layers $exclude_layers \
                        --kv_cache_dtype fp8 \
                        --model_export hf_format \
                        --output_dir $output_dir \
                        --multi_gpu

Deployment

Use with vLLM

This model can be deployed efficiently using the vLLM backend.

Evaluation

The model was evaluated on GSM8K benchmarks.

Accuracy

Benchmark	GLM-4.7	GLM-4.7-MXFP4(this model)	Recovery
GSM8K	94.16	93.63	99.44%

Reproduction

The GSM8K results were obtained using the lm-evaluation-harness framework, based on the Docker image rocm/vllm-private:vllm_dev_base_mxfp4_20260122, with vLLM, lm-eval and amd-quark compiled and installed from source inside the image.

Launching server

vllm serve amd/GLM-4.7-MXFP4 \
    --tensor-parallel-size 4 \
    --tool-call-parser glm47 \
    --reasoning-parser glm45 \
    --enable-auto-tool-choice \
    --kv_cache_dtype fp8

Evaluating model in a new terminal

lm_eval \
  --model local-completions \
  --model_args "model=amd/GLM-4.7-MXFP4,base_url=http://0.0.0.0:8000/v1/completions,tokenized_requests=False,tokenizer_backend=None,num_concurrent=32" \
  --tasks gsm8k \
  --num_fewshot 5 \
  --batch_size 1

License

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