omnius 1.0.816 → 1.0.818

package/dist/index.js

@@ -1,4 +1,4 @@
 [diffend] Oversized file quarantined before diffing.
 name: package/dist/index.js
-size: 34785482 bytes
-sha256: f8f780895f6b7e602bc6b54c06246a210372d0dde51d59977535e8e20ce25d89
+size: 35351446 bytes
+sha256: 10a8553236a1d111a1007db137fd98e738b5410dc573bc10242ff3d5a2604212

package/dist/scripts/asr_cuda_control.py

@@ -0,0 +1,95 @@
+"""Private parent-owned CUDA lease channel. This module imports only stdlib.
+
+The parent acquires before spawning us. It alone holds the negotiator token.
+Every model operation shares one prepare/synchronize/ready critical section.
+"""
+import json
+import os
+import re
+import socket
+import threading
+from contextlib import contextmanager
+
+
+class CudaLeaseFatal(BaseException):
+    """Cannot be swallowed by ordinary ASR/consensus error recovery."""
+
+
+def _terminate(message):
+    os.write(2, ("Live Whisper CUDA lease failed: " + str(message) + "\n").encode())
+    # Includes consensus-thread failures: no other thread may keep using CUDA.
+    os._exit(70)
+
+
+class WhisperCudaControl:
+    def __init__(self, channel, fatal=_terminate):
+        self.channel = channel
+        self.fatal = fatal
+        self.lock = threading.Lock()
+        self.sequence = 0
+        self.active = False
+        self.failed = False
+
+    @classmethod
+    def from_environment(cls):
+        uuid = os.environ.get("CUDA_VISIBLE_DEVICES", "")
+        if (os.environ.get("OMNIUS_ASR_GPU_CONTROL_FD") != "3"
+                or os.environ.get("OMNIUS_ASR_CUDA_DEVICE") != "0"
+                or not re.fullmatch(r"GPU-[0-9a-fA-F]{8}(?:-[0-9a-fA-F]{4}){3}-[0-9a-fA-F]{12}", uuid)):
+            raise CudaLeaseFatal("Live Whisper requires its Omnius scoped CUDA supervisor (FD 3 and one explicit GPU UUID); direct/unleased launch is refused before CUDA imports.")
+        try:
+            channel = socket.socket(fileno=3)
+            # live-whisper may exec its managed venv before loading dependencies.
+            os.set_inheritable(3, True)
+            return cls(channel)
+        except Exception as error:
+            raise CudaLeaseFatal("Live Whisper GPU control FD 3 is unavailable") from error
+
+    def abort(self, error):
+        self.failed = True
+        self.fatal(str(error))
+        raise CudaLeaseFatal(str(error))
+
+    def request(self, action):
+        if self.failed:
+            self.abort("previous control failure")
+        self.sequence += 1
+        try:
+            self.channel.sendall((json.dumps({"id": self.sequence, "action": action, "device": "cuda:0"}) + "\n").encode())
+            data = bytearray()
+            while not data.endswith(b"\n"):
+                part = self.channel.recv(1)
+                if not part or len(data) >= 8192:
+                    raise ValueError("closed or oversized parent response")
+                data.extend(part)
+            reply = json.loads(data)
+            if reply != {"id": self.sequence, "ok": True}:
+                raise ValueError("parent refused CUDA operation")
+        except Exception as error:
+            self.abort(error)
+
+    def loaded(self, synchronize):
+        with self.lock:
+            try:
+                synchronize()
+                self.request("ready")
+                self.active = True
+            except Exception as error:
+                self.abort(error)
+
+    @contextmanager
+    def operation(self, synchronize):
+        with self.lock:
+            if not self.active or self.failed:
+                self.abort("CUDA operation before resident acknowledgement")
+            self.request("prepare")
+            self.active = False
+            try:
+                yield
+            finally:
+                try:
+                    synchronize()
+                    self.request("ready")
+                    self.active = True
+                except Exception as error:
+                    self.abort(error)

package/dist/scripts/live-whisper.py

@@ -30,7 +30,9 @@ Protocol:
     {"type":"error","message":"..."}
 
 Usage:
-  arecord -f S16_LE -r 16000 -c 1 -t raw -q - | python3 live-whisper.py --model base
+  Started by Omnius with a parent-owned scoped GPU lease and private FD 3.
+  Configure OMNIUS_ASR_CUDA_VISIBLE_DEVICES with one capable full GPU UUID.
+  Direct pipes/Nexus launches without this supervisor fail before CUDA imports.
 """
 
 import sys
@@ -42,13 +44,19 @@ import time
 import threading
 from collections import deque
 from pathlib import Path
+from asr_cuda_control import WhisperCudaControl
+
+# Fail before dependency/bootstrap imports can initialize CUDA. The FD remains
+# inheritable across the managed-venv exec, and the parent already owns a grant.
+CUDA_CONTROL = WhisperCudaControl.from_environment()
 
 # ---------------------------------------------------------------------------
 # Configuration
 # ---------------------------------------------------------------------------
 
 SCRIPT_DIR = Path(__file__).resolve().parent
-VENV = SCRIPT_DIR / ".whisper-venv"
+VENV = Path(os.environ.get("OMNIUS_ASR_PYTHON_VENV", "").strip() or
+            str(Path.home() / ".omnius" / "runtimes" / "asr" / "whisper-venv")).expanduser()
 PY = VENV / "bin" / "python"
 PIP = VENV / "bin" / "pip"
 
@@ -238,7 +246,7 @@ def _in_venv() -> bool:
 
 
 def _ensure_venv():
-    if VENV.exists():
+    if PY.exists():
         return
     emit_status("Creating Python venv for Whisper...")
     import venv
@@ -264,7 +272,7 @@ def _ensure_deps():
     if need:
         emit_status(f"Installing: {', '.join(need)}...")
         subprocess.check_call(
-            [str(PIP), "install", *need],
+            [sys.executable, "-m", "pip", "install", *need],
             stdout=subprocess.DEVNULL, stderr=subprocess.PIPE,
         )
         for mod in ["numpy", "whisper"]:
@@ -278,7 +286,7 @@ def _ensure_deps():
     except ImportError:
         try:
             subprocess.check_call(
-                [str(PIP), "install", "pyusb"],
+                [sys.executable, "-m", "pip", "install", "pyusb"],
                 stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL,
             )
         except Exception:
@@ -534,6 +542,8 @@ def main():
     import whisper
 
     device = select_whisper_device()
+    if device != "cuda:0":
+        CUDA_CONTROL.abort("Scoped live Whisper requires logical cuda:0; CPU fallback is refused")
     emit_status(f"Loading Whisper {args.model} model on {device}...")
     try:
         model_dir = whisper_model_dir()
@@ -568,6 +578,8 @@ def main():
         emit_status("ReSpeaker DSP active: hardware VAD, echo cancellation, DOA.")
 
     fp16 = str(device).startswith("cuda")
+    import torch
+    CUDA_CONTROL.loaded(lambda: torch.cuda.synchronize(device))
     emit({
         "type": "ready",
         "model": args.model,
@@ -583,13 +595,15 @@ def main():
         silent, _ = digitally_silent_pcm16(samples)
         if silent:
             return {"text": "", "segments": []}
-        return m.transcribe(
-            samples,
-            fp16=fp16,
-            language=args.language,
-            no_speech_threshold=0.6,
-            condition_on_previous_text=False,
-        )
+        # Primary, partial and consensus calls share the same lease/lock.
+        with CUDA_CONTROL.operation(lambda: torch.cuda.synchronize(device)):
+            return m.transcribe(
+                samples,
+                fp16=fp16,
+                language=args.language,
+                no_speech_threshold=0.6,
+                condition_on_previous_text=False,
+            )
 
     # Shared PCM intake
     audio_buf = np.zeros(0, dtype=np.float32)

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "omnius",
-  "version": "1.0.816",
+  "version": "1.0.818",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "omnius",
-      "version": "1.0.816",
+      "version": "1.0.818",
       "bundleDependencies": [
         "image-to-ascii"
       ],
@@ -2602,9 +2602,9 @@
       }
     },
     "node_modules/bare-fs": {
-      "version": "4.8.2",
-      "resolved": "https://registry.npmjs.org/bare-fs/-/bare-fs-4.8.2.tgz",
-      "integrity": "sha512-+ZI68KHMUvosXfKbg/UOHK0tbCdRnegbvPEdEcZ3Nd6TetieQsJPRXBRXPdLyy8+3VSEbPXtsumTpEtt78xv9w==",
+      "version": "4.8.3",
+      "resolved": "https://registry.npmjs.org/bare-fs/-/bare-fs-4.8.3.tgz",
+      "integrity": "sha512-KQi9ShyPS/nOCtKpma7At8wbq4ORSNUmD67hsZrPcaWJcEmL8PHyS2xVWT308y8/LvHFOUTytTTDsD8dInx6bA==",
       "license": "Apache-2.0",
       "optional": true,
       "dependencies": {
@@ -5522,9 +5522,9 @@
       "license": "MIT"
     },
     "node_modules/readable-stream": {
-      "version": "4.7.0",
-      "resolved": "https://registry.npmjs.org/readable-stream/-/readable-stream-4.7.0.tgz",
-      "integrity": "sha512-oIGGmcpTLwPga8Bn6/Z75SVaH1z5dUut2ibSyAMVhmUggWpmDn2dapB0n7f8nwaSiRtepAsfJyfXIO5DCVAODg==",
+      "version": "4.8.0",
+      "resolved": "https://registry.npmjs.org/readable-stream/-/readable-stream-4.8.0.tgz",
+      "integrity": "sha512-2jhTTU3Eilp/2DMT71gbcoqC35eQHasMJH0RHkYxwlxIv3+FR6Y5xAFZIwEGLxKkjZYEFlsiMNIdwUqjK2lLHQ==",
       "license": "MIT",
       "dependencies": {
         "abort-controller": "^3.0.0",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omnius",
-  "version": "1.0.816",
+  "version": "1.0.818",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/library.js",