minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

profile: add vlm prefill profiler ranges (#30871)

Browse Source
This commit is contained in:
Mick
2026-07-12 14:07:10 +08:00
committed by GitHub
parent bce3fc987d
commit f1c247edf9
1 changed files with 37 additions and 32 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/managers/mm_utils.py
+37 -32
View File
@@ -1074,32 +1074,36 @@ def general_mm_embed_routine(
if forward_batch.mm_inputs[i] is not None
]
server_args = get_server_args()
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(
mm_inputs_list=mm_inputs_list,
extend_prefix_lens=extend_prefix_lens,
extend_seq_lens=extend_seq_lens,
input_ids=input_ids,
forward_batch=forward_batch,
input_embedding=embed_tokens,
multimodal_model=multimodal_model,
data_embedding_func_mapping=data_embedding_funcs,
placeholder_tokens=placeholder_tokens,
use_deepstack=use_deepstack,
)
else:
input_embeds, other_info = embed_mm_inputs(
mm_inputs_list=mm_inputs_list,
extend_prefix_lens=extend_prefix_lens,
extend_seq_lens=extend_seq_lens,
input_ids=input_ids,
input_embedding=embed_tokens,
multimodal_model=multimodal_model,
data_embedding_func_mapping=data_embedding_funcs,
placeholder_tokens=placeholder_tokens,
use_deepstack=use_deepstack,
)
# 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(
mm_inputs_list=mm_inputs_list,
extend_prefix_lens=extend_prefix_lens,
extend_seq_lens=extend_seq_lens,
input_ids=input_ids,
forward_batch=forward_batch,
input_embedding=embed_tokens,
multimodal_model=multimodal_model,
data_embedding_func_mapping=data_embedding_funcs,
placeholder_tokens=placeholder_tokens,
use_deepstack=use_deepstack,
)
else:
input_embeds, other_info = embed_mm_inputs(
mm_inputs_list=mm_inputs_list,
extend_prefix_lens=extend_prefix_lens,
extend_seq_lens=extend_seq_lens,
input_ids=input_ids,
input_embedding=embed_tokens,
multimodal_model=multimodal_model,
data_embedding_func_mapping=data_embedding_funcs,
placeholder_tokens=placeholder_tokens,
use_deepstack=use_deepstack,
)
# add for qwen3_vl deepstack
if use_deepstack:
@@ -1143,12 +1147,13 @@ def general_mm_embed_routine(
else:
input_embeds = None
hidden_states = language_model(
input_ids=None,
forward_batch=forward_batch,
input_embeds=input_embeds,
**kwargs,
)
with torch.profiler.record_function("sglang.vlm.language_model_prefill"):
hidden_states = language_model(
input_ids=None,
forward_batch=forward_batch,
input_embeds=input_embeds,
**kwargs,
)
return hidden_states