agent-cli 0.105.0__tar.gz → 0.105.1__tar.gz

{agent_cli-0.105.0 → agent_cli-0.105.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: agent-cli
-Version: 0.105.0
+Version: 0.105.1
 Summary: A suite of AI-powered command-line tools for text correction, audio transcription, and voice assistance.
 Project-URL: Homepage, https://github.com/basnijholt/agent-cli
 Author-email: Bas Nijholt <bas@nijho.lt>

{agent_cli-0.105.0 → agent_cli-0.105.1}/agent_cli/server/whisper/backends/transformers.py

@@ -6,6 +6,7 @@ import asyncio
 import logging
 import tempfile
 import time
+import traceback
 import wave
 from concurrent.futures import ProcessPoolExecutor
 from dataclasses import dataclass
@@ -428,8 +429,8 @@ def _transcribe_with_generate(
     )
 
 
-def _transcribe_in_subprocess(kwargs: dict[str, Any]) -> dict[str, Any]:
-    """Run transcription in subprocess. Reuses model from _state."""
+def _transcribe_with_loaded_model(kwargs: dict[str, Any]) -> dict[str, Any]:
+    """Dispatch transcription using the model retained in this worker."""
     if _state.model is None or _state.processor is None:
         msg = "Model not loaded in subprocess. Call _load_model_in_subprocess first."
         raise RuntimeError(msg)
@@ -479,6 +480,44 @@ def _transcribe_in_subprocess(kwargs: dict[str, Any]) -> dict[str, Any]:
     )
 
 
+def _clear_exception_frames(exc: BaseException) -> None:
+    """Release helper locals retained by an exception and its chained failures."""
+    pending = [exc]
+    seen: set[int] = set()
+    while pending:
+        current = pending.pop()
+        if id(current) in seen:
+            continue
+        seen.add(id(current))
+        traceback.clear_frames(current.__traceback__)
+        pending.extend(
+            error for error in (current.__cause__, current.__context__) if error is not None
+        )
+        if isinstance(current, BaseExceptionGroup):
+            pending.extend(current.exceptions)
+
+
+def _transcribe_in_subprocess(kwargs: dict[str, Any]) -> dict[str, Any]:
+    """Transcribe and return unused CUDA memory while keeping model weights loaded."""
+    if not _state.device or not _state.device.startswith("cuda"):
+        return _transcribe_with_loaded_model(kwargs)
+
+    import torch  # noqa: PLC0415
+
+    try:
+        return _transcribe_with_loaded_model(kwargs)
+    except BaseException as exc:
+        # Failed helper frames otherwise keep temporary tensors live until the
+        # executor serializes the exception, which happens after our finally.
+        # Clear locals without losing the traceback's function names and lines.
+        _clear_exception_frames(exc)
+        raise
+    finally:
+        # Helper frames have returned, so their temporary tensors are now free.
+        # Only unused allocator blocks are released; model weights stay resident.
+        torch.cuda.empty_cache()
+
+
 class TransformersWhisperBackend:
     """ASR backend using HuggingFace transformers.
 

{agent_cli-0.105.0 → agent_cli-0.105.1}/docs/commands/server/whisper.md

@@ -55,6 +55,7 @@ Run a local ASR server with automatic backend selection based on your platform:
 - **OpenAI-compatible API** at `/v1/audio/transcriptions` - drop-in replacement for OpenAI's Whisper API
 - **Wyoming protocol** for [Home Assistant](https://www.home-assistant.io/) voice integration (Wyoming is the standard protocol for local voice services)
 - **TTL-based memory management** - models unload after idle period, freeing RAM/VRAM
+- **CUDA cache cleanup** - the transformers backend releases unused GPU cache after each transcription while keeping model weights loaded
 - **Multiple models** - run different model sizes with independent TTLs
 - **Background preloading** - downloads start at startup without blocking; use `--preload` to wait
 - **Multi-platform support** - automatically uses the optimal backend for your hardware (`auto` switches to `nemo` for Parakeet models)

{agent_cli-0.105.0 → agent_cli-0.105.1}/macos/appcast.xml

@@ -4,6 +4,22 @@
         <title>Agent CLI Updates</title>
         <link>https://github.com/basnijholt/agent-cli</link>
         <description>Signed Agent CLI macOS app updates.</description>
+        <item>
+            <title>0.105.0</title>
+            <pubDate>Tue, 08 Sep 2026 10:29:52 +0000</pubDate>
+            <link>https://github.com/basnijholt/agent-cli</link>
+            <sparkle:version>204</sparkle:version>
+            <sparkle:shortVersionString>0.105.0</sparkle:shortVersionString>
+            <sparkle:minimumSystemVersion>13.0</sparkle:minimumSystemVersion>
+            <sparkle:hardwareRequirements>arm64</sparkle:hardwareRequirements>
+            <description sparkle:format="markdown"><![CDATA[## What’s Changed
+
+* Prevent Qwen long-audio truncation (#641) @basnijholt
+* Publish diarization Docker image (#640) @basnijholt
+
+]]></description>
+            <enclosure url="https://github.com/basnijholt/agent-cli/releases/download/v0.105.0/AgentCLI.zip" length="18454913" type="application/octet-stream" sparkle:edSignature="Z3253GuNAV6GC2V4Qib0Bm7m8bSdlWHPiXeHIZNa06qAJaY+qRXzrSPRyT8iZW5GnkZOJHciIbqvkkoJ9/DoCw=="/>
+        </item>
         <item>
             <title>0.104.0</title>
             <pubDate>Mon, 07 Sep 2026 21:38:19 +0000</pubDate>
@@ -193,20 +209,5 @@
 ]]></description>
             <enclosure url="https://github.com/basnijholt/agent-cli/releases/download/v0.100.0/AgentCLI.zip" length="22908645" type="application/octet-stream" sparkle:edSignature="2xn/3yQOqyqwCQ5871EAgFra/D/z6UrDl1B+ZxUh7PnOxJpL4Oe44UzQgjFB+LJXzqrY6Kmqvu9GBmk+3aZ1Ag=="/>
         </item>
-        <item>
-            <title>0.99.0</title>
-            <pubDate>Fri, 05 Jun 2026 22:57:28 +0000</pubDate>
-            <link>https://github.com/basnijholt/agent-cli</link>
-            <sparkle:version>194</sparkle:version>
-            <sparkle:shortVersionString>0.99.0</sparkle:shortVersionString>
-            <sparkle:minimumSystemVersion>13.0</sparkle:minimumSystemVersion>
-            <sparkle:hardwareRequirements>arm64</sparkle:hardwareRequirements>
-            <description sparkle:format="markdown"><![CDATA[## What’s Changed
-
-* Add live transcription preview (#571) @basnijholt
-
-]]></description>
-            <enclosure url="https://github.com/basnijholt/agent-cli/releases/download/v0.99.0/AgentCLI.zip" length="22924494" type="application/octet-stream" sparkle:edSignature="COI8nqDd0cBtOO9jw/QRmsucsxE0SPiSgNF/o+s/nv3FCjSXy0C+gZX/bIFtCYyz/DYkff+IaHaLzhw8Z5aEBQ=="/>
-        </item>
     </channel>
 </rss>

{agent_cli-0.105.0 → agent_cli-0.105.1}/tests/test_transformers_backend.py

@@ -3,7 +3,9 @@
 from __future__ import annotations
 
 import sys
+import traceback
 import wave
+import weakref
 from contextlib import nullcontext
 from types import SimpleNamespace
 from typing import TYPE_CHECKING
@@ -433,3 +435,103 @@ def test_transcribe_dispatches_qwen3_asr_to_native_adapter(
     )
 
     assert result == expected
+
+
+@pytest.mark.parametrize("device", ["cuda", "cuda:0", "cpu", "mps"])
+@pytest.mark.parametrize(
+    "failure", [None, "direct", "interrupt", "cause", "context", "group", "cycle"]
+)
+def test_transcription_releases_unused_cuda_buffers(
+    monkeypatch: pytest.MonkeyPatch,
+    tmp_path: Path,
+    device: str,
+    failure: str | None,
+) -> None:
+    """A completed or failed request must not retain CUDA's peak working set.
+
+    PyTorch's caching allocator keeps freed tensor storage reserved until
+    empty_cache(), which releases unused storage without unloading live tensors:
+    https://docs.pytorch.org/docs/stable/notes/cuda.html#memory-management
+    The small allocator double models that external boundary without a GPU.
+    """
+
+    class Buffer:
+        def __init__(self, size: int) -> None:
+            self.size = size
+
+    class Allocator:
+        def __init__(self) -> None:
+            self.live: weakref.WeakSet[Buffer] = weakref.WeakSet()
+            self.reserved = 0
+
+        def allocate(self, size: int) -> Buffer:
+            buffer = Buffer(size)
+            self.live.add(buffer)
+            self.reserved += size
+            return buffer
+
+        def empty_cache(self) -> None:
+            if not device.startswith("cuda"):
+                pytest.fail("CPU/MPS transcription must not initialize CUDA")
+            self.reserved = sum(buffer.size for buffer in self.live)
+
+    allocator = Allocator()
+    weights = allocator.allocate(100)
+    monkeypatch.setitem(sys.modules, "torch", SimpleNamespace(cuda=allocator))
+    monkeypatch.setattr(
+        backend,
+        "_state",
+        backend._SubprocessState(
+            model=weights,
+            processor=object(),
+            device=device,
+            is_qwen3_asr=True,
+        ),
+    )
+    wav_path = tmp_path / "request.wav"
+    with wave.open(str(wav_path), "wb") as wav_file:
+        wav_file.setnchannels(1)
+        wav_file.setsampwidth(2)
+        wav_file.setframerate(16000)
+        wav_file.writeframes(b"\x00\x00")
+    monkeypatch.setattr(backend, "_load_qwen_audio", lambda _: object())
+
+    def infer_inner() -> dict[str, object]:
+        temporary = allocator.allocate(900)
+        if failure:
+            # Its traceback holds `temporary` until the worker clears frame locals.
+            message = "inference failed"
+            error_type = KeyboardInterrupt if failure == "interrupt" else RuntimeError
+            raise error_type(message)
+        return {"text": "transcribed", "working_set": temporary.size}
+
+    def infer(**_kwargs: object) -> dict[str, object]:
+        try:
+            return infer_inner()
+        except RuntimeError as exc:
+            message = "inference failed"
+            if failure == "cause":
+                raise ValueError(message) from exc
+            if failure == "context":
+                raise ValueError(message) from None
+            if failure == "group":
+                raise ExceptionGroup(message, [exc]) from None
+            if failure == "cycle":
+                wrapped = ValueError(message)
+                exc.__cause__ = wrapped
+                raise wrapped from exc
+            raise
+
+    monkeypatch.setattr(backend, "_transcribe_qwen3_asr", infer)
+    if failure:
+        with pytest.raises(BaseException, match="inference failed") as error:
+            backend._transcribe_in_subprocess({"wav_path": str(wav_path)})
+        assert "infer" in [frame.name for frame in traceback.extract_tb(error.value.__traceback__)]
+    else:
+        assert backend._transcribe_in_subprocess({"wav_path": str(wav_path)}) == {
+            "text": "transcribed",
+            "working_set": 900,
+        }
+
+    assert allocator.reserved == (100 if device.startswith("cuda") else 1000)
+    assert backend._state.model is weights