feat(xpu): add dynamic W8A8 INT8 inference for MiniMax-H3

[qiuxin2012]

Summary

This PR adds dynamic W8A8 INT8 inference support for MiniMax-H3 on Intel XPU.

Kernel

• Add Intel XPU dynamic W8A8 INT8 GEMM kernel.
• Add int8-intel-xpu support for MiniMax-H3.
• Add INT8 checkpoint conversion and inference scripts.
• Exposes the following API:

sycl_kernels.onednn_w8a8_int8(
    input,
    qweight,
    weight_scales,
    bias=None,
)

oneDNN verbose confirms that the INT8 GEMM path is used:

  matmul,jit:gemm:any
  src:s8 wei:s8 dst:bf16
  attr-scales:src...+wei...

INT8 checkpoint conversion

Run from the LightX2V repository root:

  SOURCE_DIR=/llm/models/MiniMax-H3/transformer \
  OUTPUT_DIR=/llm/models/MiniMax-H3/quantized/int8 \
  bash tools/convert/examples/convert_minimax_h3_int8_xpu.sh

Equivalent converter command:

  python tools/convert/converter.py \
    --source /llm/models/MiniMax-H3/transformer \
    --output /llm/models/MiniMax-H3/quantized/int8 \
    --output_name minimax_h3_int8_xpu \
    --model_type h3 \
    --quantized \
    --linear_type int8 \
    --device cpu \
    --single_file

The generated checkpoint is:

  /llm/models/MiniMax-H3/quantized/int8/minimax_h3_int8_xpu.safetensors

The converted checkpoint contains 350 INT8 weights and 350 per-channel weight scales. Its size is approximately 32 GB.

Inference

  bash scripts/platforms/intel_xpu/run_minimax_h3_t2av_int8.sh

The default inference configuration uses:

  DTYPE=BF16
  SENSITIVE_LAYER_DTYPE=BF16
  dit_quant_scheme=int8-intel-xpu

Validation

• Kernel tests: 21 passed
• Standalone LightX2V linear relative RMS error: 0.00173
• Full MiniMax-H3 T2AV inference completed successfully on B65 and 4xB60
• Generated video was visually checked and showed no obvious quantization corruption.