huggingface/text-generation-inference
5
0
Fork 0
mirror of https://github.com/huggingface/text-generation-inference.git synced 2025-09-10 11:54:52 +00:00
Code Issues Packages Projects Releases Wiki Activity

revert vllm change

Browse Source
This commit is contained in:
OlivierDehaene 2023-07-28 16:51:33 +02:00
parent 19dc7d31b9
commit 51203e4087
2 changed files with 21 additions and 16 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
server/Makefile-vllm
View File

@ -1,4 +1,4 @@
vllm_commit := 084ca75d4271f8f67be731bc58e0d41d8e0afd3a
vllm_commit := d284b831c17f42a8ea63369a06138325f73c4cf9
vllm:
# Clone vllm

35
server/text_generation_server/utils/layers.py
View File

@ -219,31 +219,36 @@ class TensorParallelHead(SuperLayer):
)
def forward(self, input: torch.Tensor) -> torch.Tensor:
if not self.should_gather:
return super().forward(input)
world_size = self.process_group.size()
# Fast branch for single requests
if (
self.should_gather
and len(input.shape) == 2
and isinstance(self.linear, FastLinear)
and input.shape[0] == 1
):
if len(input.shape) == 2 and isinstance(self.linear, FastLinear):
out_dim = self.linear.weight.shape[0]
world_out = input.new_empty(1, out_dim * world_size)
local_out = input.new_empty(1, out_dim)
if input.shape[0] == 1:
world_out = input.new_empty(1, out_dim * world_size)
local_out = input.new_empty(1, out_dim)
gather_input = local_out
else:
world_out = input.new_empty(out_dim * world_size, input.shape[0])
gather_input = input.new_empty(out_dim, input.shape[0])
local_out = gather_input.T
torch.mm(input, self.linear.weight.T, out=local_out)
torch.distributed.all_gather_into_tensor(
world_out, local_out, group=self.process_group
world_out, gather_input, group=self.process_group
)
return world_out
if input.shape[0] == 1:
return world_out
return world_out.T
output = super().forward(input)
if not self.should_gather:
return output
world_output = [torch.empty_like(output) for _ in range(world_size)]
world_output = [
torch.empty_like(output) for _ in range(self.process_group.size())
]
torch.distributed.all_gather(world_output, output, group=self.process_group)
world_output = torch.cat(world_output, dim=-1)
return world_output