Skip to content

Make API Key OpenAI-compatible - #917

Merged
Ying1123 merged 6 commits into
mainfrom
api-key
Aug 4, 2024
Merged

Ying1123 merged 6 commits into
mainfrom
api-key

Conversation

@Ying1123

@Ying1123 Ying1123 commented Aug 4, 2024 •

Copy link
Copy Markdown
Contributor

@Ying1123
Ying1123 merged commit 0d4f3a9 into main Aug 4, 2024
@Ying1123
Ying1123 deleted the api-key branch August 4, 2024 20:35
timethink pushed a commit to timethink/sglang that referenced this pull request Mar 9, 2025
zhsurpass added a commit to zhsurpass/sglang that referenced this pull request Jul 2, 2026
Co-authored-by: Kurkur <mccllm@qq.com>
efschu pushed a commit to efschu/htsglang that referenced this pull request Aug 27, 2026
…n bracket and sgl-project#760 drain honesty: instruments, no sizing change) into the flip train
efschu pushed a commit to efschu/htsglang that referenced this pull request Aug 27, 2026
sgl-project#917/sgl-project#918 added, clear 2 UP037

`--verify` named exactly what moved: THREE new mem_cache modules
(test_row_reference_not_owner_916.py, test_quiesce_reports_what_happened_917.py,
test_reservation_span_derivation_918.py) and not one changed sha256 among the
181 already in the table.

MEASURED on this tip, hermetic (CUDA_VISIBLE_DEVICES=""):
  * fresh serial reference, 184 modules, one process, one order:
    2 failed / 2745 passed / 927 skipped, 169.1s. 2745 = the previous 2683
    plus exactly the 62 the three new modules contribute, and the 2 failures
    are the same two test_acceptance_emitters_758.py::RefillTiming names,
    compared as a LIST, not as a count.
  * the 181 solo logs carried forward, plus three new solo probes.

sgl-project#918 rewrites 141 lines of memory_pool.py, so carrying proofs forward is a
claim that needs evidence rather than an assumption. The builder supplies it:
admission is `solo failure set == serial failure set`, and the SERIAL half was
just re-measured against the new bytes. Any module whose failure set the
reservation-span derivation moved would no longer match and would be DEMOTED
with the reason recorded. Demotions: ZERO -- which is the measurement that
sgl-project#918 perturbs no module's outcome, not a hope that it does not.

  before  PARALLEL 168  RANKS 13  UNCLASSIFIED 3
  after   PARALLEL 171  RANKS 13  SERIAL 0  EXCLUDED 0   -> VERIFY OK

gate_partition.tsv (managers) DELIBERATELY NOT REBUILT. sgl-project#917 adds one module
there, and that table already carries a 35-module gap whose original solo logs
(/tmp/868_solo) no longer exist. Rebuilding it now would hand 270 proven
PARALLEL rows the reason `unproven:no_solo_log` and demote them all -- a table
made worse to make one row look better. The new module therefore lands in the
serial lane by the table's own default-for-the-unknown rule, the runner names
it "unclassified: not in the partition table" in its report, and the serial
lane's own arithmetic shows it ran: 606 -> 617 passed, exactly its 11 tests.

gate_partition_scheduler.tsv untouched: no scheduler test module moved bytes.

UP037 x2 in kv_vmm_backing.py, the same hygiene as Train-2 and 2b. The file
carries `from __future__ import annotations`, so unquoting is inert at runtime;
re-imported and the sgl-project#918 suite re-run (20 passed) to prove it rather than
assume it.
efschu pushed a commit to efschu/htsglang that referenced this pull request Aug 27, 2026
`--verify` named exactly one new module,
test/registered/unit/mem_cache/test_pool_invariant_live_double_912b.py, and
no sha256 among the 184 already in the table moved. `test_seam_extent_
contract_783.py` DID change bytes, but it sits in the managers table's
long-standing unclassified set, so no fast-lane proof went stale with it --
PARALLEL stayed at 270 there, checked rather than assumed.

MEASURED on this tip, hermetic (CUDA_VISIBLE_DEVICES=""):
  * fresh serial reference, 185 modules, one process, one order:
    2 failed / 2761 passed / 927 skipped, 164.8s. 2761 = 2745 + exactly the
    16 the new module contributes; the 2 failures are the same two
    test_acceptance_emitters_758.py::RefillTiming names, compared as a LIST.
  * the 184 solo logs carried forward, plus one new solo probe.
  Demotions: ZERO, counted by diffing the table's own rows before and after.

  before  PARALLEL 171  RANKS 13  UNCLASSIFIED 1
  after   PARALLEL 172  RANKS 13  SERIAL 0  EXCLUDED 0   -> VERIFY OK

gate_partition.tsv (managers) again NOT rebuilt, for the reason recorded in
Train-2c: its 35-module gap has no surviving solo logs, so a rebuild would
demote 270 proven rows to make two new ones look better. sgl-project#920's module and
sgl-project#917's before it therefore run in the serial lane by the default-for-the-
unknown rule, both named in the runner's report, and the lane's arithmetic
shows it: 617 -> 630 passed, which is exactly the new module's 13. The
adapted sgl-project#783 contract contributes 0 to that delta -- 11 passed on this tip
and 11 on 3502a9a, measured on both, so pinning the DECISION instead of
the literal changed what the tests assert without changing how many there
are.

gate_partition_scheduler.tsv untouched: seam_census_double.py is a harness,
not a test_*.py, so it carries no table row.
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.

1 participant