[Feature gap][LLMC]: missing args

[chensuyue]

Problem Description

low_gpu_mem_usage not supported in AutoRoundModifier, maybe other args missing as well. Please enhance the feature gap.
Some model quantize will be OOM on XPU without this config.

Reproduction Steps

low_gpu_mem_usage=True

Environment Information

No response

Error Logs

Additional Context

No response

[yiliu30]

Hi @chensuyue, I tested Qwen3-30B-A3B-Instruct-2507 with both the tuning case and the iters=0 case.

• Tuning case
  It failed on both AR and LLMC.
• iters=0 case
  It passed on both AR and LLMC.

import torch
import os
from auto_round import AutoRound
scheme = "MXFP8"
# Quantize the model
model_name = "/home/yiliu7/workspace/auto-round/scripts/Qwen3-30B-A3B-Instruct-2507"
basename = os.path.basename(model_name)
autoround = AutoRound(
    # "/media/hf_models/Qwen/Qwen-Image",
    model_name,
    scheme=scheme,
    iters=4,
    disable_opt_rtn=True,
    # static_kv_dtype="fp8",
)

# Save the quantized model
output_dir = f"{basename}-{scheme}"

# Currently loading the quantized diffusion model is not supported, so use fake format
autoround.quantize_and_save(output_dir, format="auto_round", inplace=True)

• Test envs:

llmcompressor      0.11.1.dev7+ge183fed8   /home/yiliu7/workspace/llm-compressor
auto-round         0.14.0.dev443+g30f72bda /home/yiliu7/workspace/auto-round

[chensuyue]

Tuning saved with LLMC format: https://inteltf-jenk.sh.intel.com/job/Intel-GPU-key-model-test/627/artifact/quantization_Qwen3-30B-A3B-Instruct-2507_mxfp4.log