frozenleaves/vllm
1
0
Fork 0
mirror of https://github.com/vllm-project/vllm.git synced 2025-10-20 14:53:52 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] Fix broken deepseek fp8 TP weights loading (#24367)

Browse Source 
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
This commit is contained in:
Isotr0py
2025-09-07 00:23:12 +08:00
committed by GitHub
parent 0eadaeff7e
commit 00a4e56d8d
2 changed files with 5 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm/model_executor/layers/linear.py
View File

@ -262,7 +262,7 @@ class LinearBase(CustomOp):
self.tp_size = (get_tensor_model_parallel_world_size()
if not disable_tp else 1)
def __post_init__(self):
def update_param_tp_status(self):
for param in self.parameters():
if isinstance(param, BasevLLMParameter):
param.tp_rank = self.tp_rank
@ -459,6 +459,7 @@ class ColumnParallelLinear(LinearBase):
})
else:
self.register_parameter("bias", None)
self.update_param_tp_status()
def weight_loader(self, param: Parameter, loaded_weight: torch.Tensor):
@ -1250,6 +1251,7 @@ class RowParallelLinear(LinearBase):
})
else:
self.register_parameter("bias", None)
self.update_param_tp_status()
def weight_loader(self, param: Parameter, loaded_weight: torch.Tensor):
input_dim = getattr(param, "input_dim", None)

3
vllm/model_executor/layers/quantization/fp8.py
View File

@ -270,7 +270,8 @@ class Fp8LinearMethod(LinearMethodBase):
layer.weight_block_size = None
if self.block_quant:
tp_size = get_tensor_model_parallel_world_size()
tp_size = getattr(layer, "tp_size",
get_tensor_model_parallel_world_size())
assert self.quant_config.weight_block_size is not None
layer.weight_block_size = self.quant_config.weight_block_size
block_n, block_k = (