Skip to content

Bump version to 0.2.10 - #923

Merged
Ying1123 merged 1 commit into
mainfrom
bump_version
Aug 4, 2024
Merged

Ying1123 merged 1 commit into
mainfrom
bump_version

Conversation

@Ying1123

@Ying1123 Ying1123 commented Aug 4, 2024

Copy link
Copy Markdown
Contributor

No description provided.

@Ying1123
Ying1123 merged commit 141e8c7 into main Aug 4, 2024
@Ying1123
Ying1123 deleted the bump_version branch August 4, 2024 23:52
timethink pushed a commit to timethink/sglang that referenced this pull request Mar 9, 2025
efschu pushed a commit to efschu/htsglang that referenced this pull request Aug 27, 2026
…ner memo outlived its cutover, sgl-project#924 a Mamba slot returned twice) into the flip train
efschu pushed a commit to efschu/htsglang that referenced this pull request Aug 27, 2026
…gl-project#924 added, re-prove the sgl-project#718 stub

`--verify` named three things, and they are not the same kind: TWO new modules
(test_hicache_owner_ctx_cutover_923.py, test_mamba_double_free_924.py) plus ONE
whose bytes moved and whose proof therefore expired
(test_hicache_phase_guard_718.py -- its controller stub gained the new prologue
methods). All three got a solo probe; nothing else in the table moved.

MEASURED on this tip, hermetic (CUDA_VISIBLE_DEVICES=""):
  * fresh serial reference, 187 modules, one process, one order:
    2 failed / 2798 passed / 927 skipped, 169.5s. 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 three new probes.
  Demotions: ZERO, counted by diffing the table's rows before and after.

  before  PARALLEL 172  RANKS 13  UNCLASSIFIED 2 (+1 stale)
  after   PARALLEL 174  RANKS 13  SERIAL 0  EXCLUDED 0   -> VERIFY OK

THE DELTA RECONCILES EXACTLY, measured per module rather than inferred:
2798 - 2761 = 37 = 20 (sgl-project#923) + 17 (sgl-project#924). The sgl-project#718 stub contributes 0: 12
passed on this tip and 12 on 38bd92f, so the stub fix changed what the
double models without changing how many tests it carries.

AND THE STUB FIX IS A CO-REQUISITE, NOT COSMETIC -- proven rather than
asserted. The 2d stub, checked out into this tree against the merged
controller, fails: `1 failed, 11 passed`
(TestPhaseGuard::test_write_runs_normally_when_the_flip_is_not_routing). So the
stub had to move with sgl-project#923's prologue, and shipping the controller change
without it would have gone red at the desk. Tree restored to HEAD afterwards.

gate_partition.tsv (managers) unchanged and not rebuilt -- no managers test
module moved bytes, and the 35-module gap's reasoning stands from Train-2c.
gate_partition_scheduler.tsv untouched for the same reason.

NAMED, NOT FIXED: kv_reshard.py carries 6 UP037 that the pre-commit hook would
select. They are PRE-EXISTING -- the same 6 on 38bd92f, at lines this merge
does not touch -- so they are not this train's regression and are recorded here
rather than swept into a merge commit as unrelated churn.
efschu pushed a commit to efschu/htsglang that referenced this pull request Sep 24, 2026
…e path (xsn422 anchor 409-604 ms per chunk), under --p-host-overlap

weg2xsn422 (--p-host-overlap + #PGAP, median over 141 4096-chunks per stage):
the host still sat 409/421/480 ms per chunk in plan_parts anchor, i.e. almost
the whole forward. py-spy PP0 (park-test prefill, 60 s, 50 Hz): 1731 of 2357
samples on ONE line, MambaSlotAllocator._do_alloc
`self.slot_used[select_index] = True`. A Python scalar assigned into a CUDA
tensor is copied host->device with non_blocking=False, which PyTorch
completes with cudaStreamSynchronize on the CURRENT stream -- the schedule
stream, which carries the device fence of the forward still running. The D2D
anchor copy itself never blocked.

SGLANG_WEG2_P_NOSYNC=1 (set for group P by --p-host-overlap; unset = the
stock lines, byte-identical):
* MambaSlotAllocator: index_fill_ (device scalar) instead of the scalar
  assignment in alloc/free; the sgl-project#924 double-free answer is copied to PINNED
  host memory on the stream and read once its event is complete.
* HiCacheController._refuse_unaddressable_kv_rows (sgl-project#923): for identity rows
  (no DCP owner context) the bound is computed on the device and read at the
  next check through pinned memory; a hit stops the rank by name (the copy is
  already issued). Owner-ruled (compact) pools keep the host refusal.
* HiCacheController.move_indices (direct + layer_first): device indices stay
  on the device, permuted there (permutation sent pinned, non-blocking) --
  the arena writers take them as they are; `.cpu()` waited for the stream.
  (xsn422 PP1: rowcheck 58 ms, publish 72 ms per chunk after the launch.)

Tests (desk, CPU-only, test27b.sh, one pytest per file), b1f9b55 +
5fd4e77 + this commit:
  test_weg2_p_nosync.py 9 passed (env, launcher carries it, index_fill_ ledger
    equals the scalar ledger, double free still refused, drain reads the
    pinned answer and keeps an incomplete one, sgl-project#923 deferred check raises on a
    hit only once complete, gate = identity rows on the device, move_indices
    branch shape, CPU pairs unchanged)
  test_weg2_p_host_overlap.py 21, test_weg2_mamba_slot_trail_1467 4,
  test_hicache_owner_ctx_cutover_923 20, test_mamba_pool_floor 23: passed.
efschu pushed a commit to efschu/htsglang that referenced this pull request Oct 20, 2026
… turn-end copy, no displacement/disk I/O in the mamba claim, rank-uniform issue, sgl-project#923 a named refusal

A1: writing_check(write_back=True) waits only for the acks it needs: never an S108 DoneEvent (polled,
    kept in order); K21 and K15 drain their own node's ack only (_s108_drain_own). The flush re-polls
    and reset() joins the copier. sgl-project#1465 line: waited= skipped_turn_end= skipped_not_needed= own=.
A2: under _s108_turn_end_active the mamba anchor claims FREE room only (claim_needs_room on the mamba
    arena; _pdflip_mamba_claim skips H19 displacement and the park spill) -- no drop_unreferenced ->
    _secure_before_drop -> arena_secure_to_disk in a turn end. Full = KV-only, anchor deferred by name.
A3: no wall-clock budget (clock=None), no TP0-only mamba-full stop in a turn-end sweep; a full mamba
    arena never refuses the node on TP0 alone (KV goes out on every rank).
A4: a copier failure (sgl-project#923 row, raising pool) REFUSES the op: DoneEvent ready with .refused, the ack
    path (_s108_ack) aborts the claim, clears the host values, releases the lock, names it
    (S108 TURN-END-REFUSED) and reports through _1421_refused (R12) -- no exception in the scheduler.

Tests (serial, MemoryMax=2G, CUDA_VISIBLE_DEVICES=""): test_s108 18/18 green incl. REAL sweep ->
write_backup -> _pdflip_direct_claim -> controller (mamba full/room), drain-only-own, A4 undo, A3
structure. Review tests 66db115b2f: R1 green, R2 green, R3 (characterisation of the old rank death)
now fails as intended. Mutants red: drain waits on S108, mamba guard off, wall budget back, refused
raises, refused ack completes. Regression vs rc27 1a9b2d9 (47 files): equal; 1465 source-order test
adapted (no empty-set assert any more).
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