Repository navigation
[Bugfix][MiniMax-M3] Guard sparse decode sentinel indices - #54597
Merged
benchislett merged 7 commits intoOct 8, 2026
Merged
benchislett merged 7 commits into
benchislett merged 7 commits into
Conversation
kyleliang-nv
force-pushed
the
agent/minimax-m3-sparse-sentinel-safety
branch
from
August 31, 2026 17:30
78b6065 to
991029d
Compare
kyleliang-nv
force-pushed
the
agent/minimax-m3-sparse-sentinel-safety
branch
from
August 31, 2026 22:41
991029d to
f0736aa
Compare
kyleliang-nv
force-pushed
the
agent/minimax-m3-sparse-sentinel-safety
branch
from
September 15, 2026 22:09
f0736aa to
b663838
Compare
kyleliang-nv
marked this pull request as ready for review
September 21, 2026 22:45
kyleliang-nv
requested review from
AndreasKaratzas,
WoosukKwon,
mgoin,
tlrmchlsmth,
yewentao256 and
zyongye
as code owners
September 21, 2026 22:45
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
Contributor
Author
|
/ci run |
|
✅ Triggered Buildkite CI #90309 for commit |
benchislett
enabled auto-merge (squash)
September 22, 2026 03:10
auto-merge was automatically disabled
September 22, 2026 04:46
Head branch was pushed to by a user without write access
Contributor
Author
|
/ci run |
|
❌ This PR is 15 commits behind upstream |
kyleliang-nv
force-pushed
the
agent/minimax-m3-sparse-sentinel-safety
branch
from
September 22, 2026 15:26
9f43f7d to
d3ad848
Compare
Contributor
Author
|
/ci run |
|
✅ Triggered Buildkite CI #90441 for commit |
kyleliang-nv
force-pushed
the
agent/minimax-m3-sparse-sentinel-safety
branch
from
September 22, 2026 16:37
d3ad848 to
1bcbaaa
Compare
Contributor
Author
|
/ci run |
|
✅ Triggered Buildkite CI #90452 for commit |
kyleliang-nv
force-pushed
the
agent/minimax-m3-sparse-sentinel-safety
branch
from
September 23, 2026 15:25
1bcbaaa to
90a990b
Compare
Contributor
Author
|
/ci run |
|
✅ Triggered Buildkite CI #90694 for commit |
kyleliang-nv
force-pushed
the
agent/minimax-m3-sparse-sentinel-safety
branch
from
September 25, 2026 20:47
90a990b to
dafeddc
Compare
Contributor
Author
|
/ci run |
|
✅ Triggered Buildkite CI #91309 for commit |
kyleliang-nv
force-pushed
the
agent/minimax-m3-sparse-sentinel-safety
branch
from
September 26, 2026 16:58
dafeddc to
22d85e2
Compare
Contributor
Author
|
/ci run |
|
✅ Triggered Buildkite CI #91417 for commit |
kyleliang-nv
force-pushed
the
agent/minimax-m3-sparse-sentinel-safety
branch
2 times, most recently
from
September 30, 2026 15:34
ebf2b31 to
1e2fc48
Compare
Contributor
Author
|
/ci run |
|
✅ Triggered Buildkite CI #92185 for commit |
Collaborator
CI selector (shadow): 39 test steps (87 jobs) instead of 51 (87 jobs)Shadow mode: this changes nothing about what CI runs. It shows what the evidence-based selector would pick for this PR, next to today's rules. How it works. Feedback welcome: reply here if it would skip a step this change needs, or runs something unrelated.
Selector would run (39)
Would skip (today's rules run them) (20)
Would add (today's rules do not run them) (8)
AMD mirrors: would skip (19)
AMD mirrors: would add (3)
3 changed files · base |
Map invalid top-k entries to a safe logical block and preserve neutral softmax state without branching in the split-K loop. Assisted-by: OpenAI Codex Signed-off-by: Kyle Liang <kylliang@nvidia.com>
Exercise sentinels and positive out-of-range indices before and after valid entries with both split-K and single-chunk execution. Retain layout and graph coverage after rebasing onto current main. Co-authored-by: OpenAI Codex <codex@openai.com> Signed-off-by: Kyle Liang <kyleliang-nv@users.noreply.github.com>
Compare the poisoned invalid-entry path with a valid-only run on identical Q/K/V, using exact output equality. Preserve the full layout, split-K, and CUDA graph matrix and leave independent reference-accuracy tests unchanged. Co-authored-by: OpenAI Codex <codex@openai.com> Signed-off-by: Kyle Liang <kyleliang-nv@users.noreply.github.com>
Move the existing MiniMax-M3 attention suite out of the fail-fast mixed kernel shard so unrelated failures cannot prevent its regression coverage from running. Keep all other commands and the L4/AMD coverage unchanged. Co-authored-by: OpenAI Codex <codex@openai.com> Signed-off-by: Kyle Liang <kyleliang-nv@users.noreply.github.com>
kyleliang-nv
force-pushed
the
agent/minimax-m3-sparse-sentinel-safety
branch
from
October 5, 2026 16:59
1e2fc48 to
4a4e4ee
Compare
Contributor
Author
|
/ci run |
|
✅ Triggered Buildkite CI #92915 for commit |
benchislett
enabled auto-merge (squash)
October 8, 2026 17:05
Member
|
/ci run |
|
✅ Triggered Buildkite CI #93659 for commit |
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
3 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.
Overview
Guard MiniMax-M3's Triton sparse-decode block-table dereference against negative sentinel and out-of-range logical block indices. The independent metadata-producer correction remains #54604.
Rebased on October 5, 2026 onto public main
54d93af9fb89d0bb8392bdb72db42bb6a17cb4ca; head4a4e4ee62ddc232d930a2666403356cad436eefc. All four PR patches are unchanged according togit range-diff. Main still has the unguarded dereference. Ready for code review, with the remaining validation below explicit.Claims
Validation
pre-commit run --from-ref 54d93af9fb89d0bb8392bdb72db42bb6a17cb4ca --to-ref HEADandgit diff --checkpassed.1e2fc48cc515, Buildkite #92185's standalone B200 job passed: 121 passed, 45 skipped. The overall build failed in the AMD sleep-mode job, which exited with code 139; the full build is not claimed as green.dafeddcc063fpassed all 16 regression cases on each of B200, L4, and MI300 (48 total) in Buildkite #91309. Historical results validate those exact commits, not the October 5 head./ci runstarted Buildkite #92915 for the October 5 head4a4e4ee62ddc232d930a2666403356cad436eefc. Current-head GPU results are pending.Reproduce the focused suite on a supported accelerator:
.venv/bin/python -m pytest tests/kernels/attention/test_minimax_m3.py \ -k 'decode_sparse_attention or decode_wrong_layout' -vDetails
Triton's
bt_row + blkuses raw pointer arithmetic. Forblk == -1, the unsafe access reads immediately before the request's row, not its final element. The regression uses contiguous rows and runs request 1 with a poisoned final entry in row 0, making this address deterministic.The kernel maps invalid logical blocks to safe block 0, masks their K/V and scale loads, and preserves softmax state for those entries. The upper bound is this query token's
ceil(kv_len / 128): a block that becomes valid for a later speculative token contributes nothing to the earlier token because it was already fully causally masked. This does not redefine all-empty query output in the merge kernel.The regression covers negative and large positive invalid indices, invalid-before/after-valid orderings, split-K and single-chunk execution, both KV layouts, eager execution, capture, and three graph replays. It requires finite output and exact parity with a valid-only run on identical inputs; independent FP32-reference tests and their tolerances are unchanged. Run unfixed-kernel controls in a dedicated process because an illegal access can poison the device context.
Open-PR searches were refreshed on October 5. Related #56932 adds specialized decode paths but does not replace the general-loop guard; #57420 concerns prefill, #49054 empty-row merge NaNs, and #52060 scalar KV scales. This PR does not duplicate those changes.
The checklist review found no new rebase-related code issue. The scope remains a local kernel safety fix plus regression/CI coverage; it adds no model-runner or scheduler policy.
Pull Request Checklist
I used vLLM's
/pr-checklistskill. (Mandatory for agents, optional for humans).AI assistance was used during the creation of this PR.
Design Fit: Minimizes impact on core components, reuses existing functionality, and justifies added complexity.
Testing and Validation: Validates the change and ensures any added tests are meaningful and reliable, with CI coverage or documented CI resource constraints and validation performed outside CI.
Code Quality and Style: Keeps code and comments clear and concise, and updates relevant documentation and examples.
Pull Request Contents: Includes a brief summary and relevant links, supports claims with evidence, explains root causes and implementation trade-offs, and follows the contributing guide.
Testing and Validation remains unchecked because the GPU/end-to-end items above have not all been completed. OpenAI Codex assisted with implementation, review, rebases, and tests. The human submitter must review every changed line and run/confirm the relevant validation before merge.
BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing (anything written below this line will be removed by GitHub Actions)