EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

[Fix] Reduce memory usage for loading llava model & Remove EntryClassRemapping (#1308)

Browse Source
This commit is contained in:
Lianmin Zheng
2024-09-02 21:44:45 -07:00
committed by GitHub
parent a5a134f39f
commit f64eae3a29
17 changed files with 105 additions and 158 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
python/sglang/srt/model_executor/model_runner.py
View File

@@ -606,16 +606,6 @@ def import_model_classes():
assert entry.__name__ not in model_arch_name_to_cls
model_arch_name_to_cls[entry.__name__] = entry
# compat: some models such as chatglm has incorrect class set in config.json
# usage: [ tuple("From_Entry_Class_Name": EntryClass), ]
if hasattr(module, "EntryClassRemapping") and isinstance(
module.EntryClassRemapping, list
):
for remap in module.EntryClassRemapping:
if isinstance(remap, tuple) and len(remap) == 2:
assert remap[0] not in model_arch_name_to_cls
model_arch_name_to_cls[remap[0]] = remap[1]
return model_arch_name_to_cls