[Performance] Use CuTe-DSL for FlashInfer MXFP4 quantization (#48417)
Signed-off-by: BWAAEEEK <jooho414@gmail.com>
This commit is contained in:
@@ -56,7 +56,9 @@ class FlashInferMxFp4LinearKernel(MxFp4LinearKernel):
|
||||
out_shape = x.shape[:-1] + (layer.output_size_per_partition,)
|
||||
x_2d = x.reshape(-1, x.shape[-1])
|
||||
|
||||
x_fp4, x_scale = flashinfer_mxfp4_quantize(x_2d.contiguous())
|
||||
x_fp4, x_scale = flashinfer_mxfp4_quantize(
|
||||
x_2d.contiguous(), backend="cute-dsl"
|
||||
)
|
||||
out = flashinfer_scaled_fp4_mm(
|
||||
x_fp4,
|
||||
weight,
|
||||
|
||||
Reference in New Issue
Block a user