chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

[piecewise] Refactor VLM to support input embed buffer and remove external embedder hack (#14155)

Browse Source
This commit is contained in:
Byron Hsu
2025-11-30 21:43:09 -08:00
committed by GitHub
parent 0b9dbea593
commit 0825d7f4c6
7 changed files with 145 additions and 439 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
python/sglang/srt/model_executor/model_runner.py
View File

@@ -97,11 +97,6 @@ from sglang.srt.layers.sampler import Sampler
from sglang.srt.layers.torchao_utils import apply_torchao_config_to_model
from sglang.srt.lora.lora_manager import LoRAManager
from sglang.srt.lora.lora_registry import LoRARef
from sglang.srt.managers.mm_utils import (
external_mm_preprocess_routine,
resolve_external_mm_data_embedding_funcs,
should_use_external_mm_preprocess,
)
from sglang.srt.mem_cache.allocator import (
BaseTokenToKVPoolAllocator,
PagedTokenToKVPoolAllocator,
@@ -2548,15 +2543,6 @@ class ModelRunner:
skip_attn_backend_init: bool = False,
pp_proxy_tensors=None,
) -> Union[LogitsProcessorOutput, PPProxyTensors, EmbeddingPoolerOutput]:
if self.is_multimodal and should_use_external_mm_preprocess(self.model):
data_embedding_funcs = resolve_external_mm_data_embedding_funcs(self.model)
forward_batch = external_mm_preprocess_routine(
forward_batch=forward_batch,
multimodal_model=self.model,
data_embedding_funcs=data_embedding_funcs,
)
kwargs = {}
if self.support_pp:
kwargs["pp_proxy_tensors"] = pp_proxy_tensors