Skip to content

Handling input dim size greater than 3 in tuned_gemm.py #482

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 4 commits into
base: main
Choose a base branch
from
Open

Handling input dim size greater than 3 in tuned_gemm.py #482

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
a4eb80f
add case of inp.dim > 3
charlifu Mar 13, 2025
f34cb78
adjust comments
charlifu Mar 13, 2025
d69bc4b
Merge branch 'main' into charlifu/handling_dim_size_greater_than_3
charlifu Mar 17, 2025
907b56c
Merge branch 'main' into charlifu/handling_dim_size_greater_than_3
charlifu Mar 21, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
12 changes: 6 additions & 6 deletions vllm/model_executor/layers/tuned_gemm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -129,10 +129,10 @@ def scaled_mm(
def mm(self, inp, weights, bias=None):
if not support_tuned_gemms:
return F.linear(inp, weights, bias)
# F.Linear can take a 3 dimensional input. vllm
# uses this for linear units. However, sampler
# will use torch.matmul with 2 dimensions only
if inp.dim() == 3:
# F.Linear can take a 3 dimensional (or even larger)
# input. vllm uses this for linear units. However,
# sampler will use torch.matmul with 2 dimensions only
if inp.dim() >= 3:
try:
inp_view = inp.view(-1, inp.size(-1))
batched = True
Expand All @@ -157,7 +157,7 @@ def mm(self, inp, weights, bias=None):
out = self.apply_skinny(m, n, k, inp_view, weights)
if out is not None:
if batched:
out = out.view(inp.shape[0], inp.shape[1], weights.shape[0])
out = out.view(*inp.shape[:-1], weights.shape[0])
if bias is not None:
return out + bias
return out
Expand All @@ -182,7 +182,7 @@ def mm(self, inp, weights, bias=None):
self.tuned_df.to_csv(self.untune_path, index=False)
return F.linear(inp, weights, bias)
if batched:
out = out.view(inp.shape[0], inp.shape[1], weights.shape[0])
out = out.view(*inp.shape[:-1], weights.shape[0])
return out


Expand Down