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

Enabled BnB NF4 inference on Gaudi (#20172)

Browse Source 
Signed-off-by: Ruheena Suhani Shaik <rsshaik@habana.ai>
This commit is contained in:
Ruheena Suhani Shaik
2025-07-15 08:56:08 +05:30
committed by GitHub
parent 80305c1b24
commit 016b8d1b7f
2 changed files with 18 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
vllm/model_executor/model_loader/bitsandbytes_loader.py
View File

@@ -199,6 +199,10 @@ class BitsAndBytesModelLoader(BaseModelLoader):
if self.pre_quant:
if self.load_8bit:
if current_platform.is_hpu():
raise ValueError(
"currently hpu supports 4bit quantization only")
return self._quantized_8bit_generator(
hf_weights_files, use_safetensors,
quant_state_dict), quant_state_dict
@@ -302,6 +306,10 @@ class BitsAndBytesModelLoader(BaseModelLoader):
in temp_state_dict):
quant_state = _parse_quant_state(mapped_weight_name,
temp_state_dict)
if current_platform.is_hpu():
assert quant_state.quant_type == "nf4", (
"currently hpu supports nf4 quant_type only")
quant_state_dict[mapped_weight_name] = quant_state
yield org_weight_name, weight_tensor
else:
@@ -372,10 +380,12 @@ class BitsAndBytesModelLoader(BaseModelLoader):
...]
# bitsandbytes requires data in GPU
if weight_sub_tensor.is_cuda:
if (weight_sub_tensor.is_cuda
or weight_sub_tensor.device.type == "hpu"):
loaded_weight = weight_sub_tensor
else:
loaded_weight = weight_sub_tensor.cuda()
loaded_weight = weight_sub_tensor.to(
device=current_platform.device_type)
# remove the following after the issue is fixed:
# https://github.com/bitsandbytes-foundation/bitsandbytes/issues/1342