Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Studio: hold the forced aligner against a folder delete; match compound audio.cpp size scopes in order (#13633)

Daniel Han89080f0e4be6
diff --git a/studio/backend/core/inference/audio_cpp_models.py b/studio/backend/core/inference/audio_cpp_models.py
index d018233b303..59924d3a77e 100644
--- a/studio/backend/core/inference/audio_cpp_models.py
+++ b/studio/backend/core/inference/audio_cpp_models.py
@@ -2758,12 +2758,16 @@ def _state_key_named(key: str, wanted: str, folder: str) -> bool:
def _scope_word_matches(word: str, scope: set[str], stem: str) -> bool:
- """A scope word is one of ``scope``, or a compound (``v4-small``) whose every part is a word of the
- file stem: its parts can all be folder words, which ``scope`` leaves out."""
+ """A scope word is one of ``scope``, or a compound (``v4-small``) whose parts run, in order, through
+ the file stem: its parts can all be folder words, which ``scope`` leaves out."""
if word in scope:
return True
parts = [p for p in re.split(r"[-_.]+", word) if p]
- return len(parts) > 1 and set(parts) <= set(re.split(r"[-_.]+", stem))
+ if len(parts) < 2:
+ return False
+ # In order and adjacent: `multilingual-ctc` must not name `multilingual-large-ctc`.
+ words = [w for w in re.split(r"[-_.]+", stem) if w]
+ return any(words[i : i + len(parts)] == parts for i in range(len(words) - len(parts) + 1))
def _cached_variants_named(
diff --git a/studio/backend/core/inference/stt_audiocpp_sidecar.py b/studio/backend/core/inference/stt_audiocpp_sidecar.py
index d6d6a65de7a..908084673e8 100644
--- a/studio/backend/core/inference/stt_audiocpp_sidecar.py
+++ b/studio/backend/core/inference/stt_audiocpp_sidecar.py
@@ -37,6 +37,7 @@ from core.inference.audio_cpp_models import (
AudioCppModelError,
CompanionModel,
downloaded_models,
+ family_from_names,
forget,
parse_identifier,
require_runnable,
@@ -648,6 +649,15 @@ class AudioCppSttSidecar:
with self._load_state_lock:
return self._loading_model if self._loading else None
+ @property
+ def holds_timestamp_aligner(self) -> bool:
+ """Whether the forced aligner is open, or may be opened by a load in flight: neither names it
+ as the loaded model, so a folder-scoped delete guard would miss it."""
+ if self._aligned and self._server_alive():
+ return True
+ loading = self.loading_model
+ return bool(loading) and family_from_names([loading]) in _ALIGNED_FAMILIES
+
@property
def keep_alive_seconds(self) -> float:
return self._keep_alive_seconds
diff --git a/studio/backend/hub/services/models/deletion.py b/studio/backend/hub/services/models/deletion.py
index 39d92846847..506010f51cc 100644
--- a/studio/backend/hub/services/models/deletion.py
+++ b/studio/backend/hub/services/models/deletion.py
@@ -1071,7 +1071,7 @@ def _audio_cpp_blocks_delete(repo_id: str) -> bool:
split_variant_ref,
)
from core.inference.orchestrator import peek_inference_backend
- from core.inference.stt_audiocpp_sidecar import get_audio_cpp_stt_sidecar
+ from core.inference.stt_audiocpp_sidecar import QWEN3_ALIGNER, get_audio_cpp_stt_sidecar
wanted = (repo_id or "").strip().lower()
@@ -1094,6 +1094,8 @@ def _audio_cpp_blocks_delete(repo_id: str) -> bool:
active = getattr(backend, "active_model_name", None) if backend is not None else None
loading = tuple(getattr(backend, "loading_models", ()) or ()) if backend is not None else ()
sidecar = get_audio_cpp_stt_sidecar()
+ if wanted == QWEN3_ALIGNER.id.lower() and sidecar.holds_timestamp_aligner:
+ return True
return (
holds(active)
or any(holds(name) for name in loading)
diff --git a/studio/backend/tests/test_audio_cpp_umbrella_ids.py b/studio/backend/tests/test_audio_cpp_umbrella_ids.py
index 1b4db65cd2f..b985bd497fd 100644
--- a/studio/backend/tests/test_audio_cpp_umbrella_ids.py
+++ b/studio/backend/tests/test_audio_cpp_umbrella_ids.py
@@ -209,6 +209,47 @@ def test_folder_guard_matches_the_folder_however_it_was_loaded(monkeypatch, load
assert deletion._audio_cpp_blocks_delete(AUDIO_CPP_REPO) is True
+def test_a_compound_scope_does_not_name_a_larger_sibling(hub):
+ # GigaAM: `multilingual-ctc` named both `...-multilingual-ctc-` and `...-multilingual-large-ctc-`,
+ # so the Hub key resolved to neither and delete / reveal answered 404.
+ folder = "GigaAM-ASR-GGUF"
+ small = acm.AudioCppVariant(
+ "gigaam-multilingual-ctc/F16", (), f"{folder}/gigaam-multilingual-ctc-f16.gguf"
+ )
+ large = acm.AudioCppVariant("large/F16", (), f"{folder}/gigaam-multilingual-large-ctc-f16.gguf")
+ assert acm._cached_variants_named([small, large], "multilingual-ctc/F16", folder) == [small]
+ assert acm._cached_variants_named([small, large], "multilingual-large-ctc/F16", folder) == [
+ large
+ ]
+ assert acm._state_key_named(
+ f"{folder}/gigaam-multilingual-ctc-f16", "multilingual-ctc/F16", folder
+ )
+ assert not acm._state_key_named(
+ f"{folder}/gigaam-multilingual-large-ctc-f16", "multilingual-ctc/F16", folder
+ )
+
+
+def test_a_timestamped_asr_holds_the_forced_aligner_it_opened(monkeypatch):
+ # Qwen3-ASR attaches the aligner only for Timestamps and reports ASR as loaded, so the
+ # folder-scoped guard let a live aligner be deleted.
+ from core.inference import stt_audiocpp_sidecar
+ from hub.services.models import deletion
+
+ aligner = stt_audiocpp_sidecar.QWEN3_ALIGNER.id
+ _idle(monkeypatch)
+ side = stt_audiocpp_sidecar.get_audio_cpp_stt_sidecar()
+ assert deletion._audio_cpp_blocks_delete(aligner) is False
+ monkeypatch.setattr(side, "_server_alive", lambda: True)
+ side._aligned = True
+ assert deletion._audio_cpp_blocks_delete(aligner) is True
+ assert deletion._audio_cpp_blocks_delete(KOKORO) is False
+ side._aligned = False
+ assert deletion._audio_cpp_blocks_delete(aligner) is False
+ # A Qwen3-ASR load in flight may still attach it.
+ _idle(monkeypatch, stt = ASR)
+ assert deletion._audio_cpp_blocks_delete(aligner) is True
+
+
def test_folder_guard_maps_a_legacy_dictation_key_to_its_folder(monkeypatch):
from hub.services.models import deletion