biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Model] support bitsandbytes quantization with minicpm3 model (#10682)

Browse Source 
Signed-off-by: Ubuntu <zixuanzhang@bytedance.com>
This commit is contained in:
zixuanzhang226
2024-11-27 23:58:02 -08:00
committed by GitHub
parent cb4e1c3f3a
commit 70dc14fbd0
1 changed files with 6 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
vllm/model_executor/models/minicpm3.py
View File

@@ -241,6 +241,12 @@ class MiniCPM3ForCausalLM(MiniCPMForCausalLM):
# `embedding_modules` and `embedding_padding_modules` # `embedding_modules` and `embedding_padding_modules`
# are inherited from MiniCPMForCausalLM # are inherited from MiniCPMForCausalLM
bitsandbytes_stacked_params_mapping = {
# shard_name, weight_name, index
"gate_proj": ("gate_up_proj", 0),
"up_proj": ("gate_up_proj", 1),
}
def _init_model(self, *, vllm_config: VllmConfig, prefix: str = ""): def _init_model(self, *, vllm_config: VllmConfig, prefix: str = ""):
self.model = MiniCPM3Model(vllm_config=vllm_config, self.model = MiniCPM3Model(vllm_config=vllm_config,
prefix=maybe_prefix(prefix, "model")) prefix=maybe_prefix(prefix, "model"))