[Model][gpt-oss] Support DP+EP for GPT-OSS with FlashInfer trtllm-gen MoE (#23819)

Signed-off-by: Po-Han Huang <pohanh@nvidia.com>

vllm/model_executor/layers/quantization/mxfp4.py

@@ -623,8 +623,6 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
 
         if should_use_flashinfer_mxfp4():
             from flashinfer import mxfp8_quantize, trtllm_fp4_block_scale_moe
-            assert not self.moe.use_ep, (
-                "EP is not supported for flashinfer mxfp4 moe backend yet.")
             if _should_use_flashinfer_mxfp4_bf16():
                 assert x.dtype == torch.bfloat16
                 x_quant = x
@@ -650,12 +648,12 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
                 None,  # output1_scale_scalar
                 None,  # output1_scale_gate_scalar
                 None,  # output2_scale_scalar
-                self.num_experts,
+                global_num_experts,
                 top_k,
                 None,  # n_group
                 None,  # topk_group
                 self.intermediate_size,  # padded to multiple of 256
-                0,  # local_expert_offset
+                layer.ep_rank * layer.local_num_experts,  # local_expert_offset
                 self.num_experts,  # local num experts
                 None,
                 self._get_tile_tokens_dim(x, top_k),