mudler / mudler/vllm.cpp

Four SACRED gates carry #2805's mute switch: the self-determinism licence is inside an existence guard

Open
#2,987 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
C++
Stars
423
Forks
53
Avg merge
20h 26m
Merged PRs (30d)
310

Description

Row: -

Four SACRED token-exact gates carry the mute switch #2805 names, in the form
#2805 describes, and #2805 covers only the OPT one.

The shape is a self-determinism licence read from greedy_dist.npy inside an
existence guard, with a -1 sentinel that reaches the report:

int64_t multi_cells = -1;
if (fs::exists(gdir / "greedy_dist.npy")) {
  ...
  CHECK(multi_cells == 0);
}

Deleting the licence file removes the licence instead of failing the gate, and
the STRICT bar keeps running unlicensed.

Sites, found by a fresh reviewer of #2981 while verifying the OPT repair:

  • tests/vllm/models/test_deepseek_v2_paged_engine.cpp:229-252
  • tests/vllm/models/test_glm4_moe_lite_paged_engine.cpp:141-162
  • tests/vllm/models/test_qwen3_32b_nvfp4a16_paged_engine.cpp:251-272
  • tests/vllm/models/test_qwen3coder_paged_engine.cpp:158 (guard without the
    sentinel -- same consequence, one symptom fewer)

None carries a DK floor either, so the degenerate K=1 capture that #2981
showed is the reachable form of this defect passes on all four.

Why this is filed rather than fixed in #2981. Each of these is another
row's SACRED gate. Scoping the OPT fix to the OPT row was correct; widening it
would have put four rows' correctness bars in one pull request.

What closes this

Each of the four gets the #2981 treatment, on its owning row: the licence
fails closed when its file is absent or unreadable, a DK floor, and the
section placed so it executes where re-captures actually arrive (#2981 moved
it above the device-only checkpoint guard, because behind that guard the
licence is unexecuted on every host that skips, including CI).

Each needs its own red-before: remove the licence file and confirm the gate
reds rather than passing at a reduced assertion count.

Evidence

#2981 measured the OPT case end to end: with greedy_dist.npy removed the
pre-fix binary printed
STRICT correctness gate: 6/6 prompts token-exact (96/96 tokens) ... (vLLM self-determinism: -1 multi-valued cells) and exited 0 at 34 assertions.

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start with the SACRED gates in tests/vllm/models/test_deepseek_v2_paged_engine.cpp:229-252, test_glm4_moe_lite_paged_engine.cpp:141-162, test_qwen3_32b_nvfp4a16_paged_engine.cpp:251-272, and test_qwen3coder_paged_engine.cpp:158, using #2981's OPT repair as the reference. Confirm each gate fails closed for an absent or unreadable greedy_dist.npy, adds a DK floor, and executes where re-captures arrive. Remove each licence file as a red-before check and verify the gate fails rather than passing with fewer assertions.

Written by the indexing model from the issue text.

Assessment

Tech stack
cpp
Domain
testing-qa
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Active
Clarity
Clearly specified
Newbie friendliness
55/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.