[CI][ROCm][Bugfix] DeepSeek V4.1: don't pass expert intermediate padding to AITER

vllm/models/deepseek_v4/amd/model.py

@@ -578,6 +578,9 @@ def _validate_aiter_mega_moe_config(vllm_config: VllmConfig) -> None:
 
 
 class DeepseekV4MoE(nn.Module):
+    # None: use the padding vLLM added.
+    experts_intermediate_pad: int | None = None
+
     def __init__(
         self,
         vllm_config: VllmConfig,
@@ -726,6 +729,7 @@ def __init__(
             top_k=config.num_experts_per_tok,
             hidden_size=config.hidden_size,
             intermediate_size=config.moe_intermediate_size,
+            intermediate_pad=self.experts_intermediate_pad,
             renormalize=config.norm_topk_prob,
             quant_config=quant_config,
             prefix=f"{prefix}.experts",

vllm/models/deepseek_v41/amd/model.py

@@ -77,6 +77,10 @@
 
 
 class DeepseekV4MoE(DeepseekV4MoEBase):
+    # TODO: revert once https://github.com/ROCm/aiter/pull/6598 is in the
+    # pinned AITER; with the padding set, its FlyDSL MoE can output NaN.
+    experts_intermediate_pad = 0
+
     def __init__(
         self,
         vllm_config: VllmConfig,