Model Overview

• Model Architecture: GLM-5.3-Flash
  • Input: Text, Image, Video
  • Output: Text
• Supported Hardware Microarchitecture: AMD MI350/MI355
• ROCm: 7.2.0
• PyTorch: 2.9.1+rocm7.2.0.git7e1940d4
• Transformers: 5.12.1
• Operating System(s): Linux
• Inference Engine: SGLang
• Model Optimizer: AMD-Quark (V0.12)
  • Weight quantization: MOE-only (shared experts quantized), OCP MXFP4, Static
  • Activation quantization: MOE-only, OCP MXFP4, Dynamic

This model was built with GLM-5.3-Flash model by applying AMD-Quark for MXFP4 quantization.

Model Quantization

The model was quantized from zai-org/GLM-5.3-Flash using AMD-Quark. The weights and activations are quantized to MXFP4.

Quantization scripts:

from quark.torch import LLMTemplate, ModelQuantizer

EXCLUDE = [
    "*self_attn*",
    "*mlp.gate",
    "*mlp.gate_proj", "*mlp.up_proj", "*mlp.down_proj",  # dense MLP (layers 0-2)
    "*visual*",
    "*lm_head*",
    "*embed*",
    "model.language_model.layers.45.*",  # MTP block
]

template = LLMTemplate(
    model_type="glm5_next",
    kv_layers_name=["*self_attn.kv_b_proj"],
    q_layer_name="*self_attn.q_b_proj",
    exclude_layers_name=EXCLUDE,
)
LLMTemplate.register_template(template)

quant_config = template.get_config(scheme="mxfp4", exclude_layers=EXCLUDE)

quantizer = ModelQuantizer(quant_config)
quantizer.direct_quantize_checkpoint(
    pretrained_model_path="zai-org/GLM-5.3-Flash",
    save_path=OUTPUT_DIR,
    device="cuda",
    keep_excluded_layers_as_original_model_state=True,
)

Deployment

Use with SGLang

This checkpoint is validated for native MXFP4 deployment on AMD Instinct MI350X and MI355X GPUs (CDNA 4, gfx950). The accuracy measurements below were collected on MI350X. MI355X uses the same gfx950 target, 288 GB HBM capacity per GPU, and ROCm mi35x software path.

The validation used:

• Model revision fc676278b68ba33f6b4724be286af9d7f6c814c2
• SGLang commit 654df43cbee108a81fa1736c34ba8c701f199285, from SGLang PR #36607
• Container image lmsysorg/sglang@sha256:6d68cd19206716cb3f1e31e2ad89cd0852d7ae614a792773c30a4277f8955c72
• ROCm 7.2.0, PyTorch 2.9.1+rocm7.2.0.git7e1940d4, Transformers 5.12.1, and SGLang 0.5.18 with the source commit above mounted into the container

The pinned image predates the GLM-5.3 mixed Quark MXFP4/block-FP8 loader change, so use the exact SGLang source commit above with the image:

git clone https://github.com/sgl-project/sglang.git
git -C sglang checkout 654df43cbee108a81fa1736c34ba8c701f199285

hf auth login
export HF_HOME="${HF_HOME:-$HOME/.cache/huggingface}"

docker run --rm -it \
  --network host \
  --ipc host \
  --device /dev/kfd \
  --device /dev/dri \
  --group-add video \
  --cap-add SYS_PTRACE \
  --security-opt seccomp=unconfined \
  -v "$PWD/sglang:/sgl-workspace/sglang" \
  -v "$HF_HOME:/root/.cache/huggingface" \
  lmsysorg/sglang@sha256:6d68cd19206716cb3f1e31e2ad89cd0852d7ae614a792773c30a4277f8955c72 \
  bash

Inside the container, launch the validated 8-GPU TP8+EP8 configuration:

cd /sgl-workspace/sglang
export PYTHONPATH=/sgl-workspace/sglang/python:${PYTHONPATH:-}
export SGLANG_USE_AITER=1
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
export ROCR_VISIBLE_DEVICES=0,1,2,3,4,5,6,7

python -m sglang.launch_server \
  --model-path amd/GLM-5.3-Flash-Quark-MXFP4 \
  --served-model-name amd/GLM-5.3-Flash-Quark-MXFP4 \
  --tp-size 8 \
  --ep-size 8 \
  --attention-backend dsa \
  --dsa-prefill-backend tilelang \
  --dsa-decode-backend tilelang \
  --linear-attn-backend triton \
  --kv-cache-dtype bfloat16 \
  --moe-runner-backend aiter \
  --disable-shared-experts-fusion \
  --disable-radix-cache \
  --context-length 65536 \
  --max-running-requests 64 \
  --cuda-graph-backend-decode full \
  --cuda-graph-max-bs-decode 64 \
  --mem-fraction-static 0.85 \
  --model-loader-extra-config '{"enable_multithread_load":true,"num_threads":8}' \
  --watchdog-timeout 1200 \
  --trust-remote-code \
  --reasoning-parser glm45 \
  --tool-call-parser glm47 \
  --host 127.0.0.1 \
  --port 30000

For a 4-GPU deployment, use GPUs 0,1,2,3, set --tp-size 4, and omit --ep-size 8. Both configurations were validated with full decode graph capture through batch size 64 (--cuda-graph-backend-decode full).

Evaluation

Accuracy

The full 1,319-example GSM8K test set was evaluated with thinking enabled. These are accuracy results, not throughput or latency benchmarks.

Both runs completed all 1,319 unique examples with a 100% stop finish rate and zero duplicate IDs, empty generations, evaluator errors, truncations, or server faults. MI355X is covered as the equivalent gfx950 deployment target; the reported measurements were collected on MI350X.

Reproduction

The results were produced with sgl-eval at commit a231b7a439b235090ff7baa30778fa2b514309ae:

python -m pip install \
  "git+https://github.com/sgl-project/sgl-eval.git@a231b7a439b235090ff7baa30778fa2b514309ae"

python -m sgl_eval.cli run gsm8k \
  --base-url http://127.0.0.1:30000/v1 \
  --model amd/GLM-5.3-Flash-Quark-MXFP4 \
  --num-examples 1319 \
  --num-threads 64 \
  --max-tokens 32768 \
  --temperature 1.0 \
  --top-p 0.95 \
  --thinking \
  --out-dir ./gsm8k-full

Native MXFP4 execution requires a gfx950 GPU. MI300X and MI325X use gfx942 and are not supported by this native MXFP4 recipe.

License

Modifications Copyright(c) 2026 Advanced Micro Devices, Inc. All rights reserved.