LLMC Upstream Tasks

[yiliu30]

• MXFP4 Support [AutoRound] Add MXFP4 Support vllm-project/llm-compressor#2276

---

Others

• Embedding model [Bug][Autoround]: Error during sample generation: 'Linear' object has no attribute 'weight' vllm-project/llm-compressor#2245 (comment)
• W8A16_FP [WIP][CT][XPU] Add W8A16_FP8 Linear Support vllm-project/vllm#32578

[yiliu30]

Intel GPU Quantization Scheme Support Status In vLLM/Compressed-Tensor

Scheme	Module Type	Kernel Readiness	Note	CT Scheme	Status/Plan
W4A16_INT	Linear	Ready, vllm-xpu-kernel	Using IPEX for now; need to switch to vllm‑xpu‑kernel	CompressedTensorsWNA16	Supported
W8A16_FP	Linear	Ready, vllm-xpu-kernel		CompressedTensorsW8A16Fp8	P0 WW03
W8A8_FP	Linear	Ready, vllm-xpu-kernel(onednn)	Dequantized to higher precision internal	CompressedTensorsW8A8Fp8	Supported (vLLM team)
W4A16_INT	MoE	Ready, vllm-xpu-kernel		CompressedTensorsWNA16MoEMethod	P1 WW06
W8A16_FP	MoE	Ready, vllm-xpu-kernel		CompressedTensorsWNA16MoEMethod	P0 WW04
W4A16_MXFP	MoE	Ready, vllm-xpu-kernel		Not supported yet	P2 WW07

cc @hshen14 @thuang6 @rogerxfeng8 @jikunshang