refine some typo (#3473)

benchmark/kernels/fused_moe_triton/benchmark_torch_compile_fused_moe.py

@@ -30,7 +30,7 @@ def get_model_config(model_name: str, tp_size: int):
         topk = config.num_experts_per_tok
         intermediate_size = config.moe_intermediate_size
         shard_intermediate_size = 2 * intermediate_size // tp_size
-    elif config.architectures[0] == "DeepseekV2ForCausalLM":
+    elif config.architectures[0] in ["DeepseekV2ForCausalLM", "DeepseekV3ForCausalLM"]:
         E = config.n_routed_experts
         topk = config.num_experts_per_tok
         intermediate_size = config.intermediate_size

python/sglang/srt/layers/moe/fused_moe_triton/fused_moe.py

@@ -1094,7 +1094,7 @@ def fused_moe(
     - num_expert_group: Optional[int]: additional parameter for grouped_topk
     - topk_group: Optional[int]: additional parameter for grouped_topk
     - use_grouped_topk: If True, use grouped_topk instead of fused_topk
-        note: Deepseekv2 model uses grouped_topk
+        note: Deepseek V2/V3/R1 series models use grouped_topk
     - use_fp8_w8a8 (bool): If True, use fp8 arithmetic to compute the inner
         products for w1 and w2. Defaults to False.
     - use_int8_w8a16 (bool): If True, use fp8 arithmetic to compute the inner

python/sglang/srt/layers/moe/topk.py

@@ -75,7 +75,7 @@ def fused_topk(
     return topk_weights, topk_ids
 
 
-# This is used by the Deepseek-V2 model
+# This is used by the Deepseek V2/V3/R1 series models
 @torch.compile(dynamic=True, backend=get_compiler_backend())
 def grouped_topk(
     hidden_states: torch.Tensor,

python/sglang/srt/server_args.py

@@ -795,7 +795,7 @@ class ServerArgs:
         parser.add_argument(
             "--disable-mla",
             action="store_true",
-            help="Disable Multi-head Latent Attention (MLA) for DeepSeek-V2.",
+            help="Disable Multi-head Latent Attention (MLA) for DeepSeek V2/V3/R1 series models.",
         )
         parser.add_argument(
             "--disable-overlap-schedule",