profile: add vlm prefill profiler ranges (#30871)

python/sglang/srt/managers/mm_utils.py

@@ -1074,6 +1074,10 @@ def general_mm_embed_routine(
                 if forward_batch.mm_inputs[i] is not None
             ]
             server_args = get_server_args()
+            # Makes VLM profiles directly attributable: this range includes
+            # encoder/ViT execution and multimodal feature placement, while
+            # the language model range below excludes both.
+            with torch.profiler.record_function("sglang.vlm.mm_embedding"):
                 if server_args and server_args.enable_adaptive_dispatch_to_encoder:
                     # Split by precomputed vs non-precomputed so get_embedding_and_mask only sees uniform batches
                     input_embeds, other_info = _embed_mm_inputs_with_split(
@@ -1143,6 +1147,7 @@ def general_mm_embed_routine(
     else:
         input_embeds = None
 
+    with torch.profiler.record_function("sglang.vlm.language_model_prefill"):
         hidden_states = language_model(
             input_ids=None,
             forward_batch=forward_batch,