Release 2026-09-15 (i4 478b5f90) - #258
Open
yy-code-nv wants to merge 6 commits into
Open
yy-code-nv wants to merge 6 commits into
yy-code-nv wants to merge 6 commits into
Conversation
_source_commit: 478b5f9026723776047dd04a8e18caccf85be381-dirty _dest_commit: 2a8339d
test_multiview_dense_attention_runs_a_gen_stream_past_the_varlen_index_limit builds its pack with _multiview_pack, which rounds the GEN stream up to the backend's block -- 526240 real tokens into 526336 on Triton's 128 -- but built its plan without padded_gen_tokens, so the plan described the unpadded 526240 and _check_plan_matches_pack refused the pack before any pass ran. The plan addresses the stream as packed, padding included, which is why every other test here and cosmos3_vfm_network's own plan pass the padded length. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
conftest.pytest_runtest_setup caps an unmarked test at 1/max(worker_count, TEST_MAX_PROCESSES) of the device -- 1/8, i.e. 17.48 GiB of an H200 -- so the case OOMed in its first pass gathering v, at ~17.2 GiB allocated. Its ~20GB of q/k/v and outputs fits no such share of any GPU that clears the 60GB floor it already guards on, so it takes gpus(1) and the whole device. That marker alone was not enough: set_per_process_memory_fraction is process-global and the gpus(n) branch never set it, so a marked test inherited the 1/N budget of whichever unmarked test preceded it -- in a whole-suite run, all of them. The branch now states the full budget it means. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…er takes This release removed get_context_parallel_sharded_sequence's ``attn_implementation`` parameter and the assertion behind it -- the split operates on the SequencePack's und/gen partition and never looked at the attention pattern -- and updated the call sites that passed it positionally. The two that passed it by keyword were left behind, so both tests raised TypeError the moment the 4-rank torchrun step reached them. The three_way case's docstring went with the parameter: what it pins now is the claim that assertion stood in front of, that both streams shard evenly. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Both were needed only because this job ran the case at all: the plan's padded GEN length (which it left to the default and the pack then disagreed with) and the whole-device budget (conftest caps an unmarked test at 1/8 of it, and the case needs ~20GB). The case is L1, and the Unit tests step now selects "not L1" the way i4's own CI selects --L0, so neither applies here. Reverted rather than kept: attention_test.py comes from i4 through .file_mapping.json, so a fix that lives only here is reverted by the next release anyway, and until then it reads as a divergence someone has to explain. Both defects are upstream's -- i4's copy of this test has them too -- and go back as such. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
This suite arrives from i4 through .file_mapping.json and its level markers are i4's, but nothing here read them: the job ran everything, so the L1 cases -- the ones i4 sizes to production geometry and runs nowhere -- executed here for the first time, against a runner they were not written for. Three of them failed for three unrelated environment reasons before any of them said anything about the code. All four pytest invocations in this job now select "not L1", which is what i4's --L0 comes to for these files. Not "L0", so that a test carrying no level marker -- anything added on this side -- still runs. The cost is stated in the comment: an L1 case is now executed in neither repo. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Automated release from i4.
_source_commit:
478b5f9026723776047dd04a8e18caccf85be381-dirty_dest_commit (base):
2a8339d46a6e10e96f26c98509e6080d04ead490