fix: avoid local STT crash on Apple Silicon by AnthonyFrancis · Pull Request #28624 · NousResearch/hermes-agent

AnthonyFrancis · 2026-05-19T08:27:00Z

Summary

Detect macOS Apple Silicon/Rosetta before loading local faster-whisper
Force local STT to CPU/int8 on that path to avoid native device="auto" crashes
Set KMP_DUPLICATE_LIB_OK before importing faster_whisper for the affected path because the crash can happen during import via duplicate Intel OpenMP runtimes
Relax the voice extra NumPy pin to <2 because the current NumPy 2.x pin emits compatibility warnings with the installed torch/ctranslate2 stack
Add regression coverage for Rosetta detection and CPU-only model loading

Fixes #28622

Test Plan

python -m pytest tests/tools/test_transcription_tools.py -q -o 'addopts='
Manual local transcription on affected Mac/Rosetta environment:
- force_cpu True
- Result: {'success': True, 'transcript': 'Hello, one, two, three. Can you hear me?', 'provider': 'local'}
Restarted Hermes gateway and confirmed no new Python crash report was created during restart

austinpickett

use PULL_REQUEST_TEMPLATE.md

Copilot

Pull request overview

This PR mitigates hard-crash failures in the gateway’s local STT (faster-whisper / ctranslate2) path on macOS Apple Silicon—especially when running under Rosetta—by avoiding native device="auto" autodetection and working around Intel OpenMP duplicate-runtime aborts.

Changes:

Add macOS/Rosetta detection and force faster-whisper to load on CPU with compute_type="int8" on affected hosts.
Set KMP_DUPLICATE_LIB_OK prior to importing faster_whisper on the affected path to reduce import-time abort risk.
Adjust the voice extra’s NumPy requirement from a strict NumPy 2.x pin to numpy<2, and add targeted regression tests around the new CPU-forcing behavior.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
tools/transcription_tools.py	Adds Apple Silicon/Rosetta detection and forces CPU/int8 model loading; sets `KMP_DUPLICATE_LIB_OK` before importing faster-whisper on the affected path.
tests/tools/test_transcription_tools.py	Adds regression tests for the CPU-forcing behavior and Rosetta detection; updates existing CUDA fallback tests to explicitly disable the new force-CPU path.
pyproject.toml	Relaxes `voice` extra NumPy constraint to `numpy<2` to avoid NumPy 2.x compatibility warnings with the STT stack.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

+    def test_apple_silicon_forces_cpu_without_auto_probe(self, tmp_path):
+        """Apple Silicon/Rosetta should skip device='auto' to avoid SIGABRT."""
+        audio = tmp_path / "test.ogg"
+        audio.write_bytes(b"fake")
+


teknium1

Thanks for the focused fix. The premise still holds on current main: tools/transcription_tools.py:1101-1103 imports faster_whisper and calls WhisperModel(..., device="auto", compute_type="auto"), which matches the crash path described in #28622.

Problems

The dependency-pin change is incomplete for Hermes' actual install paths. pyproject.toml's voice extra is one path, but local STT is also lazy-installed via tools/lazy_deps.py:110-114, and that still pins numpy==2.4.3 on current main.
The pyproject.toml hunk is stale against current main: gh pr diff 28624 --patch | git apply --check --verbose - fails only for pyproject.toml, while the transcription code and tests apply with offsets.

Suggested changes

Update tools/lazy_deps.py:110-114 to match the NumPy constraint chosen for the voice extra, so gateway lazy installs get the same fix.
Resolve the pyproject.toml optional-dependency conflict against current main's voice block at pyproject.toml:148-154, and update lock metadata if maintainers require uv.lock to track optional-dependency specifiers.

Automated hermes-sweeper review.

teknium1 · 2026-06-15T13:18:36Z

-  "numpy==2.4.3",
+  "numpy<2",
 ]
 pty = [


This relaxes the voice extra, but local STT is also installed through the lazy dependency table on current main (tools/lazy_deps.py:110-114), which still pins numpy==2.4.3. Please update that sibling install path too so gateway lazy installs receive the same compatibility fix.

teknium1 · 2026-07-13T08:32:02Z

Thanks for the focused Apple Silicon/Rosetta mitigation. The premise remains present on current main: tools/transcription_tools.py:1102-1104 imports faster_whisper and calls WhisperModel(..., device="auto", compute_type="auto").

Problems

The NumPy part of the fix only changes the explicit voice extra. Local STT also lazy-installs through tools/transcription_tools.py:208-229; its actual dependency table at tools/lazy_deps.py:132-136 still requests numpy==2.4.3, so that install path misses the compatibility change.
The pyproject.toml hunk is stale: gh pr diff 28624 --patch | git apply --check --verbose - fails for that file, while the code and test hunks apply with offsets. The current voice block is pyproject.toml:173-179.
numpy<2 also needs a supported lower bound and a regenerated lockfile under the dependency policy in AGENTS.md:561-576.

Suggested changes

Re-resolve the voice extra and uv.lock, and mirror the selected NumPy requirement in tools/lazy_deps.py:132-136.
Keep the CPU/int8 loading and Rosetta tests when salvaging the source change.

Automated hermes-sweeper review.

Force CPU (int8) for faster-whisper on Apple Silicon / Rosetta, where ctranslate2's device=auto path can hard-abort in native code. Salvaged from PR #28624 without the numpy pin change (main already moved on). (cherry picked from commit 7edf2d5, pyproject.toml hunk dropped)

teknium1 · 2026-07-29T01:24:09Z

Merged into main via consolidated salvage PR #73510 (merge c0c5dac531). Your Apple Silicon / Rosetta local STT crash guard (CPU int8 + KMP_DUPLICATE_LIB_OK on Darwin arm64) was cherry-picked with your authorship; only the stale numpy-pin hunk was dropped.

Your contribution is credited to you in git history. Thank you! Closing this PR as merged-via-salvage.

Force CPU (int8) for faster-whisper on Apple Silicon / Rosetta, where ctranslate2's device=auto path can hard-abort in native code. Salvaged from PR NousResearch#28624 without the numpy pin change (main already moved on). (cherry picked from commit 7edf2d5, pyproject.toml hunk dropped)

fix: avoid local STT crash on Apple Silicon

7edf2d5

AnthonyFrancis requested a review from a team May 19, 2026 08:27

alt-glitch added type/bug Something isn't working P2 Medium — degraded but workaround exists tool/tts Text-to-speech and transcription comp/gateway Gateway runner, session dispatch, delivery labels May 19, 2026

austinpickett requested a review from Copilot May 19, 2026 10:36

austinpickett requested changes May 19, 2026

View reviewed changes

Copilot started reviewing on behalf of austinpickett May 19, 2026 10:36 View session

Copilot AI reviewed May 19, 2026

View reviewed changes

Comment thread tests/tools/test_transcription_tools.py

Comment on lines +512 to +516

def test_apple_silicon_forces_cpu_without_auto_probe(self, tmp_path):

"""Apple Silicon/Rosetta should skip device='auto' to avoid SIGABRT."""

audio = tmp_path / "test.ogg"

audio.write_bytes(b"fake")

teknium1 reviewed Jun 15, 2026

View reviewed changes

teknium1 added sweeper:risk-compatibility Sweeper risk: may break existing users, config, migrations, defaults, or upgrades sweeper:blast-moderate Sweeper blast radius: moderate — a subsystem or single platform labels Jul 13, 2026

GottZ mentioned this pull request Jul 28, 2026

Gateway voice notes crash Python on Apple Silicon/Rosetta local STT #28622

Closed

teknium1 mentioned this pull request Jul 28, 2026

fix(stt): provider robustness — device config, CUDA fallbacks, container normalization (SILK/m4a), lazy-install recovery #73510

Merged

teknium1 closed this Jul 29, 2026

Navigation Menu

Uh oh!

fix: avoid local STT crash on Apple Silicon - #28624

fix: avoid local STT crash on Apple Silicon#28624
AnthonyFrancis wants to merge 1 commit into
NousResearch:mainfrom
AnthonyFrancis:fix/apple-silicon-local-stt-crash

AnthonyFrancis commented May 19, 2026

Uh oh!

austinpickett left a comment

Uh oh!

Copilot AI left a comment

Uh oh!

teknium1 left a comment

Uh oh!

teknium1 Jun 15, 2026

Uh oh!

teknium1 commented Jul 13, 2026

Uh oh!

teknium1 commented Jul 29, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

Uh oh!

Conversation

AnthonyFrancis commented May 19, 2026

Summary

Test Plan

Uh oh!

austinpickett left a comment

Choose a reason for hiding this comment

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

teknium1 left a comment

Choose a reason for hiding this comment

Problems

Suggested changes

Uh oh!

teknium1 Jun 15, 2026

Choose a reason for hiding this comment

Uh oh!

teknium1 commented Jul 13, 2026

Problems

Suggested changes

Uh oh!

teknium1 commented Jul 29, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants