Skip to content

vllm.model_executor.layers.quantization.utils.flashinfer_fp4_moe

Utility helpers for NVFP4 + FlashInfer fused-MoE path

is_flashinfer_fp4_cutlass_moe_available 

is_flashinfer_fp4_cutlass_moe_available() -> bool

Return True when FlashInfer CUTLASS NV-FP4 kernels can be used.

Source code in vllm/model_executor/layers/quantization/utils/flashinfer_fp4_moe.py 
36
37
38
39
40
41
42
43
def is_flashinfer_fp4_cutlass_moe_available() -> bool:
    """Return `True` when FlashInfer CUTLASS NV-FP4 kernels can be used."""
    return (
        envs.VLLM_USE_FLASHINFER_MOE_FP4
        and has_flashinfer_cutlass_fused_moe()
        and current_platform.is_cuda()
        and current_platform.has_device_capability(100)
    )

reorder_w1w3_to_w3w1 

reorder_w1w3_to_w3w1(
    weight: Tensor, scale: Tensor, dim: int = -2
) -> tuple[Tensor, Tensor]

Re-order the concatenated [w1, w3] tensors to [w3, w1]

Source code in vllm/model_executor/layers/quantization/utils/flashinfer_fp4_moe.py 
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
def reorder_w1w3_to_w3w1(
    weight: torch.Tensor, scale: torch.Tensor, dim: int = -2
) -> tuple[torch.Tensor, torch.Tensor]:
    """Re-order the concatenated `[w1, w3]` tensors to `[w3, w1]`"""
    size = weight.size(dim)
    assert size % 2 == 0, f"Expected even size in dim {dim}, got {size}"
    half = size // 2

    w1, w3 = weight.split(half, dim=dim)
    s1, s3 = scale.split(half, dim=dim)

    return (
        torch.cat([w3, w1], dim=dim).contiguous(),
        torch.cat([s3, s1], dim=dim).contiguous(),
    )