[Quantization] Remove the dead ocp_mx_scheme branch from moe_kernel_quantize_input (#52603)
Signed-off-by: xuebwang-amd <xuebwang@amd.com> Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
xuebwang-amd committed
Aug 19, 2026 at 05:42 UTC
d3fafe0c27f9666a06675858738aaeab949da0f5
1 file changed
-16
vllm/model_executor/layers/fused_moe/utils.py
-16
@@ -365,25 +365,9 @@ def moe_kernel_quantize_input(
365
per_act_token_quant: bool,
366
block_shape: list[int] | None = None,
367
is_scale_swizzled: bool = True,
368
- ocp_mx_scheme: str | None = None,
368
quantization_emulation: bool = False,
369
mx_alignment: int = 0,
370
) -> tuple[torch.Tensor, torch.Tensor | None]:
372
- # Handle OCP MX scheme that requires QDQ (quantize-dequantize) for emulation
373
- if ocp_mx_scheme is not None:
374
- if ocp_mx_scheme in {"w_mxfp4", "w_mxfp4_a_mxfp4"}:
375
- pass # No QDQ needed for these schemes
376
- elif ocp_mx_scheme.endswith("a_fp8"):
377
- # Perform QDQ (quantize and dequantize) on activation for emulation
378
- # purpose, because there is no native kernel for weight in ocp_mx_scheme
379
- # and activation in FP8. The implementation is based on existing
380
- # non-emulation ops.
381
- # TODO: Remove this `ocp_mx_scheme is not None` block and rely solely
382
- # on `quantization_emulation`.
383
- return _fp8_quantize_dequantize(A, A_scale)
384
- # else: For other schemes (e.g., *_a_mxfp6_e3m2, *_a_mxfp6_e2m3),
385
- # weights are already dequantized, and we proceed with normal
386
- # activation quantization below.
371
if quant_dtype == current_platform.fp8_dtype():
372
if quantization_emulation:
373
return _fp8_quantize_dequantize(A, A_scale)