Fix Nightly NV CI (#33564)

Co-authored-by: Brayden Zhong <brayden@radixark.ai>
Co-authored-by: Baizhou Zhang <sobereddiezhang@gmail.com>

python/sglang/srt/arg_groups/overrides.py

@@ -1725,6 +1725,9 @@ def _deepseek_moe_quant_resolution(view: Any) -> dict:
         if (
             view.moe_a2a_backend == "none"
             and view.moe_runner_backend == "auto"
+            # LongCat top-k spans the zero-expert logits, which trtllm-gen's
+            # fused routing cannot see.
+            and not model_arch.startswith("LongcatFlash")
             and (
                 quantization
                 in ["fp8", "modelopt_fp8", "modelopt_fp4", "modelopt_mixed"]

python/sglang/srt/layers/n_gram_embedding.py

@@ -49,7 +49,7 @@ class NgramEmbedding(torch.nn.Module):
                 + int(over_embedding_m + i * 2 + 1)
             )
         self.oe_embeder = VocabParallelEmbedding(
-            num_embeddings=self.exclusive_oe_embedder_size_sums[-1],
+            num_embeddings=int(self.exclusive_oe_embedder_size_sums[-1]),
             embedding_dim=oe_hidden_dim,
             use_attn_tp_group=use_attn_tp_group,
         )