wassname/vllm
1
0
Fork 0
mirror of https://github.com/wassname/vllm.git synced 2026-09-09 11:40:01 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Misc] Add BNB support to GLM4-V model (#12184)

Browse Source 
Signed-off-by: Isotr0py <2037008807@qq.com>
This commit is contained in:
Isotr0py
2025-01-19 19:49:22 +08:00
committed by GitHub
parent 936db119ed
commit edaae198e7
3 changed files with 60 additions and 53 deletions
Download Patch File Download Diff File Expand all files Collapse all files
vllm/model_executor/model_loader/loader.py
+11 -4
View File
@@ -1105,15 +1105,22 @@ class BitsAndBytesModelLoader(BaseModelLoader):
weight_name,
index,
) in self.modules_mapping.inverse_packed_mapping.items():
shard_pos = quant_param_name.find(shard_name)
# Some models, such as MiniCPM V2.5/2.6, contain both
# module names 'kv_proj' and 'qkv_proj'. To prevent 'kv_proj'
# from being incorrectly identified as being present in
# 'vpm.encoder.layers.0.self_attn.qkv_proj.weight
if shard_pos > 0 and quant_param_name[shard_pos - 1] == ".":
shard_pos = quant_param_name.find(shard_name)
can_correct_rename = (shard_pos > 0) and (
quant_param_name[shard_pos - 1] == ".")
# If the quant_param_name is packed, it won't occur in the
# param_dict before renaming.
new_quant_param_name = quant_param_name.replace(
shard_name, weight_name)
need_rename = (quant_param_name not in param_dict) \
and (new_quant_param_name in param_dict)
if can_correct_rename and need_rename:
shard_index = index
quant_param_name = quant_param_name.replace(
shard_name, weight_name)
quant_param_name = new_quant_param_name
break
# Models like Clip/Siglip may skip some layers in initialization,