Release FunASR 1.4.17 - #3770

FunASR 1.4.17

Release source: 19e33badf1dba6f7606c9410c32c622486b5df13 (tag v1.4.17).

• 487 focused pytest checks passed; two local browser cases were skipped. The final PR-head hosted browser and other CI checks passed separately.
• Real NVIDIA L20Z validation includes CUDA compile checks and 48 CPU/CUDA FP32/BF16 keyword-decoding comparisons. These use synthetic inputs or tiny stand-ins, not pretrained ASR quality evaluation.
• Published wheel and sdist source bytes were checked against the final tag; strict Twine, offline installation, import and CLI checks passed.
• All ten runtime archives (including CUDA Blackwell), wheel and sdist were byte/hash and integrity verified. Eight files were downloaded from this release; four unchanged runtime archives reuse complete verified cache bytes whose full SHA-256 and size equal the new GitHub asset metadata. The new Python packages do not reuse candidate builds. SHA256SUMS covers these 12 payloads. Archive integrity was checked; platform binaries were not executed.

This is focused release validation, not the complete repository test suite or the Nano leaderboard regression.

Main review policy was explicitly updated by the repository administrator to require zero approvals before this normal protected merge. No self-approval or administrator bypass was used. PyPI has a separate final upload and byte-verification gate.

Historical Candidate Evidence

The dated statements below are preserved history; they do not describe the current release state.

October 27 Current Integration Validation

This section supersedes earlier integration-state statements below. The PR is
still a release candidate, not a published release.

• Exact PR head: 1efefd70d3b770d40f38448e11f8acca8a747da0; current main: 940b0d2463141a2e9ace8a9fd965b4aea93f2db5.
• Tested potential merge: 548166f73b3b0132e89ffacb5427b11e8d5d0362, tree ff396e71acf3e36538024333a64ccccc05099076.
  This export is not a real merge or final tag. All 4,153 tracked source blobs
  were verified unchanged before/after testing and again before this update.
• The combined, deduplicated focused pytest results are 444 passed, 2 skipped:
  442 CPU/documentation/package checks and two actual CUDA compile checks.
  The two skips are local browser-rendering cases; a weight-fetching AutoModel
  test was deselected. This is not the entire repository test suite.
• The original Sphinx check failed on a missing extension. Installing the
  missing documentation dependencies only in the isolated validation environment
  made that same check pass, without changing source or assertions.
• Both CUDA compile checks initially failed because the image's Triton 3.6.0
  lacked the interface expected by its PyTorch 2.8 development build. Isolating
  Triton 3.3.0 in the validation environment resolved both failures. The stable
  wheel is not asserted to be the image's exact Triton source-pin build. Global
  image dependencies were not changed; failed logs are retained.
• Five additional real NVIDIA L20Z checks passed: CUDA device selection,
  recursive tensor/NumPy transfer and scalar gathering, FP32 forward/backward
  parity, BF16 finite/tolerance checks, and single-rank NCCL collective execution.
  These and the two CUDA compile tests use synthetic tensors/tiny recording
  stand-ins, not pretrained ASR quality or model inference. Historical
  October 16 H100 ASR evidence below does not cover this newer integration tree.
• Current-main NumPy 1.26.4 CI initially timed out during Ubuntu ffmpeg package
  installation, before project tests. Only that cancelled job was rerun; the
  new job passed,
  including dependency validation and project tests. The successful NumPy 2.4.0
  job was reused, not rerun. Existing successful PR-head CI was not repolled.
• Offline PEP 517 builds, strict Twine checks, offline wheel installation,
  installed import/version and CLI help checks passed. All 439 packaged source
  files and 13 console entry points were verified.
• All ten runtime archives from runtime-llamacpp-v0.2.6, including CUDA
  Blackwell, were recovered and checked against their public sizes and SHA-256
  digests; ZIP/TAR integrity passed. These platform binaries were not executed.
  No model weights were downloaded.
• Candidate wheel SHA-256: fb4e98f6d69774727c28646b6c4b861e2a62eb1047fa5973ec7d7b4f5177216e.
• Candidate normalized sdist SHA-256: 282be2b547f23131fbc0faca4ce7c10c2e0e32048c90c442db276daad9e3f804.
• The local candidate set has 13 files: ten runtime archives, wheel, sdist,
  and SHA256SUMS. All 12 manifest entries passed byte-level verification.
  Manifest SHA-256: 763a520fe76457b34fb23dfa917149ba6e34bd81c8df01808507e5900769e4f4.

These hashes belong to the current candidate, not final merged/tagged
artifacts. No v1.4.17 tag, GitHub Release or PyPI upload was created by this
validation. Genuine independent approval, normal merge, actual final-tag
validation, uploaded-asset verification, then PyPI publication remain required.
The existing independent review request remains in place; this author-account
description update is not an APPROVE and does not bypass branch protection.

Historical Validation And Release Context

Summary

Prepare FunASR 1.4.17. This PR changes eight files only: the package version,
the release-version contract, four README news entries and two repository-role
guides. The documentation explicitly distinguishes the unpublished 1.4.17
candidate from published 1.4.16. No model or runtime implementation changes
are introduced by this PR.

The candidate started from e7e61293d308f4342ab9307895e710e4b51b949c.
Main has since advanced to 28a889cba5a00333dfb2dee115bd1cfab51f25e1, including
#3772 (independent automatic batch keys) and #3774 (UTF-8 dataset-list I/O).
These main-only changes are not covered by the old candidate validation below.
The actual final merged/tagged tree must be validated before publication.

Exact Candidate Validation

• Current signed+DCO head: 1efefd70d3b770d40f38448e11f8acca8a747da0.
• Tree: 4374e90d9b49f330ee82c5d5319d5b49552994aa.
• The first candidate violated the existing three-news-item README contract.
  The appended fix preserves all three items and both release-status statements;
  the contract itself was not weakened.
• October 10 exact-head focused checks: 78 passed, 1 optional Sphinx check
  skipped, 22 subtests passed. Two rendering cases were deselected locally.
  This is not the complete repository suite or GPU/model validation.
• The six exact-head hosted checks subsequently completed successfully,
  including build-and-validate and browser. Unchanged terminal CI was not rerun
  for this documentation reconciliation.
• The old PEP 517/Twine/offline-install artifacts were built at first head
  241c1efb4dfa36139ef82b28ca6ee900d892c280. They are historical evidence only,
  not validated artifacts for the current head or final release tag.

October 16 Current Integration Preflight

The current synthetic merge b84975986ab2a08d722c88e30eb5cb0cff4d0a25
combines main 28a889cba5a00333dfb2dee115bd1cfab51f25e1 and this PR's exact
head 1efefd70d3b770d40f38448e11f8acca8a747da0; its tree is
34d832415b4714d2f8c0a528ed0ffbfd89fee89d. This does not merge the PR.

• Fourteen related CPU/unit/documentation/package test files: 214 passed,
  1 optional Sphinx check skipped, 30 subtests passed. The suite emitted
  75 warnings, primarily existing NumPy copy-argument and deprecated audio-module
  warnings. All 4,151 source blobs were verified unchanged before and after.

• Offline PEP 517 wheel and sdist builds and Twine metadata checks passed on
  that integration tree. Required runtime package files match source bytes.

• All ten runtime-llamacpp-v0.2.6 archives, including CUDA Blackwell, were
  checked from existing cached bytes against the current release's SHA-256
  digests and sizes. Archive integrity was checked; platform binaries were not
  executed. No model weights were downloaded.

• Wheel SHA-256: 83953731ac6e0c219f5cbe591f05b5af5d2f922992c0946faf88fec8a8498931.

• Normalized sdist SHA-256: 1f2bb05fc0e0a68891bdd6e965339a4c6c68b5267cc787b09c406be9844f0932.

• The candidate checksum manifest covers the ten runtime archives and two
  Python packages; its SHA-256 is
  ff02a9f2bc72d946602514533ce83e4ad2debfaa58ac4f1220042139122b2f37.

• The built candidate wheel was then installed offline into an isolated target
  and exercised on an actual NVIDIA H100 using existing cached SenseVoiceSmall,
  FSMN-VAD and punctuation weights. All three modules executed CUDA operations
  (2,256 ASR, 60 VAD and 686 punctuation Linear calls).

• The existing 80.589-second public recording produced exactly matching text,
  sentence information, timestamps and words. Three distinct unkeyed audio slices
  also produced three unique automatic keys. They were processed with supported
  VAD batch_size=1; this is not a claim of tensor batch size three.

• The first GPU check failed when its test harness passed batch_size=3 to
  FSMN-VAD, which explicitly requires one. That failed receipt is retained.
  Correcting only the harness call gave exit 0 in 40.36 seconds; no repository
  implementation or expected output was changed to obtain the pass.

These are integration-preflight artifacts, not final merged/tagged artifacts,
the complete test suite, other-platform runtime validation, or a fresh hosted CI result.
Nothing has been uploaded or published by this preflight. Rebuild and verify
the actual final tag and uploaded bytes before PyPI publication. An initial
review request has been sent to repository administrator @pengzhendong; the
required genuine independent approval is not replaced by the author review.

Review And Publication State

An agent-assisted substantive review inspected all eight changed paths and the
new main-only delta. This is a COMMENT review by the PR-author account, not an
independent APPROVE, and does not satisfy the required approving review.

The observed main protection still requires one approving review and enables
Code Owner review. No approval is fabricated, no administrator bypass is used,
and protection is not changed. Repository auto-merge is disabled.

As of the October 16 review, this PR is open and unmerged. Published GitHub and
PyPI version 1.4.16 already exists; 1.4.17 has not been tagged or published.
The existing pre-release backup and original audit artifacts remain preserved.

Before release: resolve the real review gate, lease exact head/main and discussion
state, validate the actual merge/tag tree, build and inspect its real wheel/sdist
bytes, reconcile release assets and verify the final manifest. The current
runtime contract is runtime-llamacpp-v0.2.6 with ten archives including CUDA
Blackwell: thirteen final assets including wheel, sdist and checksum manifest.
The current contract must be rechecked at publication; do not substitute stale
artifact hashes or omit a runtime. PyPI remains the final publishing step.