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

[Bugfix] Fix LoRA bug (#4032)

Browse Source
This commit is contained in:
Jee Li 2024-04-13 07:56:37 +08:00 committed by GitHub
parent d04973ad54
commit b8aacac31a
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 9 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
vllm/lora/layers.py
View File

@ -32,14 +32,17 @@ if TYPE_CHECKING:
def _get_lora_device(base_layer: nn.Module) -> torch.device: def _get_lora_device(base_layer: nn.Module) -> torch.device:
# code borrowed from https://github.com/fmmoret/vllm/blob/fm-support-lora-on-quantized-models/vllm/lora/layers.py#L34 # code borrowed from https://github.com/fmmoret/vllm/blob/fm-support-lora-on-quantized-models/vllm/lora/layers.py#L34
"""Returns the device for where to place the LoRA tensors.""" """Returns the device for where to place the LoRA tensors."""
# unquantizedLinear
if hasattr(base_layer, "weight"): if hasattr(base_layer, "weight"):
return base_layer.weight.device return base_layer.weight.device
if hasattr(base_layer, "linear_weights") and isinstance( # GPTQ/AWQ/SqueezeLLM
base_layer.linear_weights, dict): elif hasattr(base_layer, "qweight"):
values = list(base_layer.linear_weights.values()) return base_layer.qweight.device
if len(values) and isinstance(values[0], torch.Tensor): # marlin
return values[0].device elif hasattr(base_layer, "B"):
raise ValueError(f"Unsupported base layer: {base_layer}") return base_layer.B.device
else:
raise ValueError(f"Unsupported base layer: {base_layer}")
def _apply_lora( def _apply_lora(