Skip to content

Release 2026-08-24 (i4 25433d97) - #216

Open
yy-code-nv wants to merge 2 commits into
mainfrom
release/2026-08-24-25433d97
Open

Release 2026-08-24 (i4 25433d97)#216
yy-code-nv wants to merge 2 commits into
mainfrom
release/2026-08-24-25433d97

Conversation

@yy-code-nv

Copy link
Copy Markdown
Collaborator

Automated release from i4.

_source_commit: 25433d97814a8bc359af6dd27ff5f561f885b2d6-dirty
_dest_commit (base): 5eee9ed574255f017b192161bfbb5a10253d65cf

yy-code-nv and others added 2 commits August 24, 2026 08:07
_source_commit: 25433d97814a8bc359af6dd27ff5f561f885b2d6-dirty
_dest_commit:   5eee9ed
test_context_parallel_attention_two_way builds its global SequencePack with
sequence_pack_from_packed_sequence()'s default cp_world_size=1 and then shards
it over a CP group of 4. That only worked while the packer left the streams
unpadded: the real UND/GEN lengths (1664 / 262144) happen to be divisible by 4.

sequence_pack_from_packed_sequence now always emits a trailing pad segment when
the layout supports one, so the streams grow by one row (1665 / 262145) and
_get_padded_size() cannot round that back up to a CP-divisible length without
knowing the CP world size. get_context_parallel_sharded_sequence() then trips
"text_seq.shape[0] must be divisible by world_size".

Pass cp_world_size through, the way build_packed_sequence() does on the model
path and the way test_sample_lbl_cp_matches_unsharded_baseline already does, so
the pad segment is rounded to 1668 / 262148 and the shards divide evenly.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@lfengad
lfengad enabled auto-merge (squash) August 24, 2026 11:50
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants