Fix flashinfer cutlass MoE output shape for non-FP4-packed inputs

[alisonshao]

Summary

Fixes nightly-test-perf-4-gpu-b200 failure caused by ValueError: Invalid shape of output: expected (512, 7168), got torch.Size([512, 14336]) when starting DeepSeek-V3-FP4 with flashinfer_cutlass MoE backend.

Root Cause

PR #13327 introduced a regression in ModelOptNvFp4FusedMoEMethod.apply() when refactoring the MoE dispatcher implementation. The output tensor allocation was changed from:

symm_output = torch.empty(x.shape[0], original_col, ...)

to:

symm_output = torch.empty(x.shape[0], x.shape[1] * 2, ...)

The * 2 multiplier was intended for when x is FP4-packed (where each byte holds 2 FP4 values, so x.shape[1] is half the original hidden_size). However, when should_use_flashinfer_cutlass_moe_fp4_allgather() returns False (as in the failing test with --tp 4 --ep 4), the hidden_states are NOT FP4-packed and already have the full hidden_size. This caused the output to be allocated with double the expected size (7168 * 2 = 14336 instead of 7168).

Fix

The output size is now conditional based on whether x_sf (hidden_states_scale) is provided:

• When x_sf is not None: hidden_states are FP4-packed, use x.shape[1] * 2
• When x_sf is None: hidden_states are not packed, use x.shape[1]

Test plan

• Syntax check passes
• nightly-test-perf-4-gpu-b200 should pass with this fix

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[alisonshao]

/tag-and-rerun-ci

[alisonshao]

https://github.com/sgl-project/sglang/actions/runs/19721366821/job/56504438983