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

[Misc][Quark] Upstream Quark format to VLLM (#10765)

Browse Source 
Signed-off-by: kewang-xlnx <kewang@xilinx.com>
Signed-off-by: kewang2 <kewang2@amd.com>
Co-authored-by: kewang2 <kewang2@amd.com>
Co-authored-by: Michael Goin <michael@neuralmagic.com>
This commit is contained in:
kewang-xlnx
2025-01-16 00:05:15 +08:00
committed by GitHub
parent 5ecf3e0aaf
commit de0526f668
32 changed files with 1264 additions and 70 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
vllm/model_executor/models/commandr.py
View File

@@ -437,6 +437,20 @@ class CohereForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
params_dict = dict(self.named_parameters())
loaded_params: Set[str] = set()
for name, loaded_weight in weights:
if (self.quant_config is not None and
(scale_name := self.quant_config.get_cache_scale(name))):
# Loading kv cache scales for quark and
# compressed-tensors quantization
param = params_dict[scale_name]
weight_loader = getattr(param, "weight_loader",
default_weight_loader)
loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else
loaded_weight[0])
weight_loader(param, loaded_weight)
loaded_params.add(scale_name)
continue
for param_name, shard_name, shard_id in stacked_params_mapping:
if shard_name not in name:
continue