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

[V0 deprecation] Remove V0 HPU backend (#21131)

Browse Source 
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
This commit is contained in:
Woosuk Kwon
2025-07-17 16:37:36 -07:00
committed by GitHub
parent ac9fb732a5
commit 4de7146351
27 changed files with 10 additions and 3926 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
vllm/model_executor/layers/vocab_parallel_embedding.py
View File

@@ -388,20 +388,8 @@ class VocabParallelEmbedding(torch.nn.Module):
# Copy the data. Select chunk corresponding to current shard.
loaded_weight = loaded_weight.narrow(output_dim, start_idx, shard_size)
if current_platform.is_hpu():
# FIXME(kzawora): Weight copy with slicing bugs out on Gaudi here,
# so we're using a workaround. Remove this when fixed in
# HPU PT bridge.
padded_weight = torch.cat([
loaded_weight,
torch.zeros(param.shape[0] - loaded_weight.shape[0],
*loaded_weight.shape[1:])
])
param.data.copy_(padded_weight)
else:
param[:loaded_weight.shape[0]].data.copy_(loaded_weight)
param[loaded_weight.shape[0]:].data.fill_(0)
param[:loaded_weight.shape[0]].data.copy_(loaded_weight)
param[loaded_weight.shape[0]:].data.fill_(0)
def forward(self, input_):
if self.tp_size > 1: