Potentially superfluous check that disables non gated activations in the cutlass fused moe API

[sidsingh-nvidia]

flashinfer/csrc/fused_moe/cutlass_backend/flashinfer_cutlass_fused_moe_binding.cu

Lines 893 to 904 in 2bb3e9e

	TVM_FFI_ICHECK(
	fc1_weight_block.size(0) == num_experts_on_rank &&
	fc1_weight_block.size(1) ==
	TmaWarpSpecializedGroupedGemmInput::alignToSfDim(
	inter_size, TmaWarpSpecializedGroupedGemmInput::MinNDimAlignmentMXFPX) *
	2 &&
	fc1_weight_block.size(2) * FP8_PER_INT32 *
	TmaWarpSpecializedGroupedGemmInput::MXFPXBlockScaleVectorSize ==
	TmaWarpSpecializedGroupedGemmInput::alignToSfDim(
	hidden_size, TmaWarpSpecializedGroupedGemmInput::MinKDimAlignmentMXFPX))
	<< "fc1 weight block size must be (num_experts_on_rank, inter_size * 2, hidden_size // 4 "
	"// block_scale_vector_size)";

This check is triggered for both non-gated and gated-activations. This should only be triggered if we are using gated activations.

[ianliuy]

What's broken?

When using the CUTLASS fused MoE backend with non-gated activations (e.g., Relu2, Gelu, Silu) and MXFP8 quantization, the fc1 weight shape validation unconditionally rejects the input — even when the shape is correct. Users see a TVM_FFI_ICHECK assertion failure.

Who is affected?

Anyone using the CUTLASS fused MoE path with:

• Quantization: WMxfp8AMxfp8, WMxfp4AFp8, or WMxfp4AMxfp8
• Activation: any non-gated type (Relu2, Gelu, Silu, etc.)

Not affected: gated activations (Swiglu, Geglu, SwigluBias), or other quant modes (NVFP4 already handles this correctly).

When does it trigger?

100% reproducible whenever a non-gated activation is used with any of the 3 MXFP8 quant paths. The shape check runs before any CUDA kernel launch, so it fails immediately at validation time.

Where is the bug?

csrc/fused_moe/cutlass_backend/flashinfer_cutlass_fused_moe_binding.cu, inside getQuantParams() at 3 locations:

Line	Quant path	Bug
~898	isWMxfp8AMxfp8Quant()	fc1_weight_block.size(1) == ... * 2 hardcoded
~1004	isWMxfp4AFp8Quant()	Same hardcoded * 2
~1063	isWMxfp4AMxfp8Quant()	Same hardcoded * 2

Why does it happen?

PR #2581 introduced MXFP8 support for CUTLASS fused MoE. At that time, only gated activations (Swiglu) were supported, so inter_size * 2 was correct — gated activations concatenate gate + up weights, doubling the N dimension.

Later, non-gated activation support was added to the trtllm-gen backend (PR #2707), but the CUTLASS backend's getQuantParams() validation was never updated. The * 2 multiplier should be conditional on isGatedActivation(base_activation_type).

The NVFP4 path in the same file (line ~1131) already handles this correctly with an if (isGatedActivation(...)) guard.

How did we fix it?

For each of the 3 MXFP8 quant branches, we:

1. Extract int const fc1_n_mult = isGatedActivation(base_activation_type) ? 2 : 1;
2. Replace the hardcoded * 2 with * fc1_n_mult
3. Update the error message to print the actual multiplier

This approach:

• Matches the ternary pattern already used in cutlass_fused_moe_kernels.cuh (lines 3961, 4635)
• Addresses the Gemini code review feedback from Fix MXFP8 fc1 weight shape check for non-gated activations #2753 (extract multiplier to local variable)
• Minimizes diff size (~18 lines changed) vs the if/else alternative (~45 lines)

How do we know it works?

Will be filled after implementation.

---

Related

• Builds on the approach identified in Fix MXFP8 fc1 weight shape check for non-gated activations #2753 (stale ~27 days, CI unresolved).
• Addresses the Gemini review feedback from Fix MXFP8 fc1 weight shape check for non-gated activations #2753 by extracting the gated-activation multiplier to a local variable before the validation checks.

I'm working on a fix.

[ianliuy]

Implemented as described above. PR: #3082 (#3082)