[Bug] modelopt_quant.py: NVFP4 input_scale not sliced to local experts with EP > 1

[vroomfondel]

Bug Description

ModelOptNvFp4FusedMoEMethod.process_weights_after_loading() in modelopt_quant.py crashes with EP > 1 in the else branch (when none of enable_flashinfer_cutlass_moe, enable_flashinfer_trtllm_moe, enable_flashinfer_cutedsl_moe is active).

w13_input_scale and w2_input_scale are allocated globally (num_experts) but multiplied against EP-local w13_weight_scale_2 (num_local_experts), causing a shape mismatch.

The cutedsl branch handles this correctly via _slice_scale(), but that helper is scoped inside the elif block and not reachable from else.

Reproduction

• Model: nvidia/MiniMax-M2.5-NVFP4 (256 experts)
• Config: TP=2, EP=2, no explicit MoE runner backend (hits the else branch)
• SGLang version: 0.5.9-dev2 (commit acab24a), also reproducible on current main

Error

File ".../sglang/srt/layers/quantization/modelopt_quant.py", line 1560, in process_weights_after_loading
    (w13_input_scale * w13_weight_scale_2).to(torch.float32),
RuntimeError: The size of tensor a (256) must match the size of tensor b (128) at non-singleton dimension 0

Suggested Fix

Add EP-aware slicing in the else branch, same logic as _slice_scale():

        else:
            w13_input_scale = layer.w13_input_scale.max(dim=-1).values.to(torch.float32)
            w2_input_scale = layer.w2_input_scale
            # EP-aware slicing (no-op when ep_size=1)
            if layer.moe_ep_size > 1:
                _ep_start = layer.moe_ep_rank * layer.num_local_experts
                _ep_end = _ep_start + layer.num_local_experts
                w13_input_scale = w13_input_scale[_ep_start:_ep_end]
                w2_input_scale = w2_input_scale[_ep_start:_ep_end]

Note

PR #20963 (Nvidia modelopt refactoring) is currently migrating this code as-is into modelopt/schemes/modelopt_fp4.py — the bug will carry over unless fixed.