Skip to content

[Refactor] Record what a bound stage path does instead of reading it back #43453

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

Merged
ch-wan merged 2 commits into from
Oct 10, 2026
+121 −207
Merged

[Refactor] Record what a bound stage path does instead of reading it back #43453

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
9da7a2f
Bind a single arriving capability on stage boundaries
ch-wan Oct 9, 2026
08e64b9
Record what a bound stage boundary path does instead of reading it back
ch-wan Oct 9, 2026
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
14 changes: 0 additions & 14 deletions python/sglang/srt/layers/layer_boundary/adapters/attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -55,7 +55,6 @@ def __init__(
self.hidden_states_local = hidden_states
self.forward_batch = forward_batch
self.qkv_latent_func = qkv_latent_func
self.hidden_states_ = None
self.qkv_latent_ = None
# When True, hidden_states_local is already attn_tp-gathered upstream
# (e.g. by the input-scattered attention input step for DSA). fetch_* must NOT gather again.
Expand All @@ -72,14 +71,6 @@ def fetch_qkv_latent(self):
self.qkv_latent_ = tp_gather(self.qkv_latent_, self.forward_batch)
return self.qkv_latent_

def fetch_hidden_states(self):
if self.hidden_states_ is not None:
return self.hidden_states_
self.hidden_states_ = self.hidden_states_local
if get_attn_tp_context().input_scattered and not self.is_pre_gathered:
self.hidden_states_ = tp_gather(self.hidden_states_, self.forward_batch)
return self.hidden_states_


class AttnTpContext:
def __init__(self):
Expand Down Expand Up @@ -131,11 +122,6 @@ def fetch_qkv_latent(self):
assert attn_inputs is not None
return attn_inputs.fetch_qkv_latent()

def fetch_hidden_states(self):
attn_inputs = get_forward().attn_inputs
assert attn_inputs is not None
return attn_inputs.fetch_hidden_states()

def clear_attn_inputs(self) -> None:
get_forward().set("attn_inputs", None)

Expand Down
Loading
Loading