DeepSeek-v4 Integrate TileKernels

[Zhichenzzz]

No description provided.

[Zhichenzzz]

@yueming-yuan This pr is to merge tilekernels (mhc and quant) into radixark#1045. thanks for reviewing!

[yiakwy-xpu-ml-framework-team]

@Zhichenzzz both UE8M0 and FP32 should be supported.

see sglang 's impl:

def _dequant_fp8(weight: torch.Tensor, scale: torch.Tensor) -> torch.Tensor:
    from einops import rearrange

    assert (
        weight.dtype == torch.float8_e4m3fn
    ), f"expected fp8_e4m3fn, got {weight.dtype}"
    assert scale.dtype in (
        torch.float8_e8m0fnu,
        torch.float32,
    ), f"expected fp8_e8m0fnu or float32, got {scale.dtype}"
    if envs.SGLANG_DEBUG_SANITY_CHECK_CONFIG.get() and not is_large_dummy_model():
        assert weight.shape == (8192, 4096), f"unexpected weight shape {weight.shape}"
        assert scale.shape == (64, 32), f"unexpected scale shape {scale.shape}"

    weight_f32 = rearrange(
        weight.float(), "(sn bn) (sk bk) -> sn bn sk bk", bn=128, bk=128
    )
    result = rearrange(
        weight_f32 * scale.float()[:, None, :, None], "sn bn sk bk -> (sn bn) (sk bk)"
    )
    if envs.SGLANG_DEBUG_SANITY_CHECK_CONFIG.get() and not is_large_dummy_model():
        assert result.shape == (8192, 4096)

    return result.to(torch.bfloat16)

cc @yueming-yuan

[yiakwy-xpu-ml-framework-team]

@Zhichenzzz Note DeepSeek uses fp32 scale by default :

https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/inference/kernel.py#L105

[yueming-yuan]

@yiakwy-xpu-ml-framework-team Thank you for your comment! This is a good catch. We will make it aligned with SGLang's actual precision, depending on the runtime kernel choice.