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

fix lora target all + csgmv backend (#14796)

Browse Source
This commit is contained in:
b8zhong
2025-12-10 11:42:51 -08:00
committed by GitHub
parent 0e54a69548
commit da9b801eb7
1 changed files with 11 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/server_args.py
+11
View File
@@ -4297,6 +4297,17 @@ class ServerArgs:
), "If 'all' is specified in --lora-target-modules, it should be the only module specified." ), "If 'all' is specified in --lora-target-modules, it should be the only module specified."
self.lora_target_modules = set(SUPPORTED_LORA_TARGET_MODULES) self.lora_target_modules = set(SUPPORTED_LORA_TARGET_MODULES)
# When using the chunked SGMV backend, skip embedding / lm_head layers for now,
# since it does not support these yet (TODO: implement embedding / lm_head support)
if self.lora_backend == "csgmv":
logger.warning(
"LoRA backend 'csgmv' does not yet support embedding or lm_head layers; "
"dropping 'embed_tokens' and 'lm_head' from --lora-target-modules=all. "
"To apply LoRA to these, use --lora-backend triton."
)
self.lora_target_modules.discard("embed_tokens")
self.lora_target_modules.discard("lm_head")
# Ensure sufficient information is provided for LoRA initialization. # Ensure sufficient information is provided for LoRA initialization.
assert self.lora_paths or ( assert self.lora_paths or (
self.max_lora_rank and self.lora_target_modules self.max_lora_rank and self.lora_target_modules