[Perf][MoE] Skip the TRTLLM fp8-block activation-scale transpose copy (#61773)

Signed-off-by: Thien Tran <gau.nernst@yahoo.com.sg>
This commit is contained in:
Thien Tran
2026-10-15 12:09:07 -07:00
committed by GitHub
parent 4d58fa254a
commit 82b9ccaa79
5 changed files with 77 additions and 14 deletions
@@ -249,6 +249,8 @@ class FusedMoEQuantConfig:
_w1: FusedMoEQuantDesc
_w2: FusedMoEQuantDesc
is_scale_swizzled: bool = True
# Emit block-quant activation scales column-major (physical [G, M] layout).
is_scale_transposed: bool = False
# MXFP4-specific TRTLLM parameters for SwiGLU activation clamping.
# These correspond to gemm1_alpha, gemm1_beta, gemm1_clamp_limit
@@ -503,6 +505,7 @@ class FusedMoEQuantConfig:
w2_zp: torch.Tensor | None = None,
weight_dtype: torch.dtype | str | None = None,
is_scale_swizzled: bool = True,
is_scale_transposed: bool = False,
gemm1_alpha: float | None = None,
gemm1_beta: float | None = None,
gemm1_clamp_limit: float | None = None,
@@ -537,6 +540,9 @@ class FusedMoEQuantConfig:
- is_scale_swizzled: Whether the activation scale-factor layout is
swizzled. Pass through to the underlying quantization kernel for
dtypes that distinguish layouts (nvfp4, mxfp8). Defaults to True.
- is_scale_transposed: Whether block-quantized activation scales are
emitted column-major (physical [num_groups, num_tokens] layout) by
the quantize step. Defaults to False.
- gemm1_alpha: Optional MXFP4 TRTLLM SwiGLU alpha parameter.
- gemm1_beta: Optional MXFP4 TRTLLM SwiGLU beta parameter.
- gemm1_clamp_limit: Optional MXFP4 TRTLLM SwiGLU clamp limit.
@@ -573,6 +579,7 @@ class FusedMoEQuantConfig:
weight_dtype, w_shape, w2_scale, g2_alphas, w2_zp, w2_bias
),
is_scale_swizzled=is_scale_swizzled,
is_scale_transposed=is_scale_transposed,
gemm1_alpha=gemm1_alpha,
gemm1_beta=gemm1_beta,
gemm1_clamp_limit=gemm1_clamp_limit,
@@ -597,6 +604,7 @@ def fp8_w8a8_moe_quant_config(
a2_gscale: torch.Tensor | None = None,
g1_alphas: torch.Tensor | None = None,
g2_alphas: torch.Tensor | None = None,
is_scale_transposed: bool = False,
gemm1_alpha: float | None = None,
gemm1_beta: float | None = None,
gemm1_clamp_limit: float | None = None,
@@ -617,6 +625,7 @@ def fp8_w8a8_moe_quant_config(
per_act_token_quant=per_act_token_quant,
per_out_ch_quant=per_out_ch_quant,
block_shape=block_shape,
is_scale_transposed=is_scale_transposed,
gemm1_alpha=gemm1_alpha,
gemm1_beta=gemm1_beta,
gemm1_clamp_limit=gemm1_clamp_limit,