Skip to content
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

Remove CPU sync before Sampler #414

Merged
merged 1 commit into from
Oct 22, 2024
Merged

Remove CPU sync before Sampler #414

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
24 changes: 12 additions & 12 deletions vllm/model_executor/layers/sampler.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -200,13 +200,13 @@ def _init_sampling_tensors(
self._do_penalties = do_penalties
self._do_top_p_top_k = do_top_p_top_k
self._do_min_p = do_min_p
self._top_p_scalar = sampling_tensors.top_ps[0].item()
self._top_k_scalar = sampling_tensors.top_ks[0].item()
self._top_p_scalar = sampling_tensors.top_ps[0]
self._top_k_scalar = sampling_tensors.top_ks[0]
scalar_p = torch.all(sampling_tensors.top_ps == self._top_p_scalar)
scalar_k = torch.all(sampling_tensors.top_ks == self._top_k_scalar)
self._scalar_p_and_k = (scalar_p and scalar_k).item()
if self._scalar_p_and_k and self._do_top_p_top_k:
self._apply_top_k_top_p_opt = ApplyToppTopkScalar(5)
self._scalar_p_and_k = torch.logical_and(scalar_p, scalar_k)

self._apply_top_k_top_p_opt = ApplyToppTopkScalar(5)

def forward(
self,
Expand Down Expand Up @@ -266,13 +266,13 @@ def forward(
logits.div_(sampling_tensors.temperatures.unsqueeze(dim=1))

if do_top_p_top_k and flashinfer_top_k_top_p_sampling is None:
if self._scalar_p_and_k:
logits = self._apply_top_k_top_p_opt(logits,
self._top_p_scalar,
self._top_k_scalar)
else:
logits = _apply_top_k_top_p(logits, sampling_tensors.top_ps,
sampling_tensors.top_ks)
# If we have a scalar p and k, we can use the optimized version.
logits = torch.where(
self._scalar_p_and_k,
self._apply_top_k_top_p_opt(logits, self._top_p_scalar,
self._top_k_scalar),
_apply_top_k_top_p(logits, sampling_tensors.top_ps,
sampling_tensors.top_ks))

if do_min_p:
logits = _apply_min_p(logits, sampling_tensors.min_ps)
Expand Down
Loading