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

Support pinning adapter via server args. (#9249)

Browse Source
This commit is contained in:
Lifu Huang
2025-08-20 16:25:01 -07:00
committed by GitHub
parent 24eaebeb4b
commit b0980af89f
8 changed files with 162 additions and 55 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
python/sglang/srt/managers/tokenizer_manager.py
View File

@@ -298,7 +298,7 @@ class TokenizerManager:
# The registry dynamically updates as adapters are loaded / unloaded during runtime. It
# serves as the source of truth for available adapters and maps user-friendly LoRA names
# to internally used unique LoRA IDs.
self.lora_registry = LoRARegistry(self.server_args.lora_paths or {})
self.lora_registry = LoRARegistry(self.server_args.lora_paths)
# Lock to serialize LoRA update operations.
# Please note that, unlike `model_update_lock`, this does not block inference, allowing
# LoRA updates and inference to overlap.