onw 3.3.3: background compiles end with onw

CHANGELOG.md

@@ -5,6 +5,11 @@
 日付は日本時間です。各版は Hugging Face の `ryugyosoft/onw` にタグ（`v0.4.2` 以降）付きで残っています。
 onw ウィンドウは、新しい版があるとこのファイルから変更点を表示します（`## v…` の見出しで区切っています）。
 
+## v3.3.3 (2026-10-26)
+- **裏でのコンパイルが onw を閉じた後も続いていたのを修正（モデルの更新は不要）**：画像エンコーダー・融合グラフ・初回ロードの区間を裏でコンパイルする子プロセスが、onw のサーバーを止めたり終了したりした後も動き続けていました（画像エンコーダーは 10 分以上 CPU を使い続け、onw の更新でファイルを置き換えられないこともあった）。起動した onw が終わると、子プロセスも終わるようにしました。途中までの分は、次のロードで続きからコンパイルします。
+- 画像エンコーダーの裏でのコンパイルは、ほかの子プロセスと同じくコンパイルだけを行います（NPU に実行の準備を作らない。同じものを OpenVINO のキャッシュにも書いていたのをやめました）。
+- NPU が受け付けなかった画像エンコーダーのコンパイル結果は、キャッシュから消します（Gemma 4 で約 230 MB）。
+
 ## v3.3.2 (2026-10-25)
 - **不具合の修正（モデルの更新は不要）**
   - この PC で変換するモデル（LFM2.5 24B-A2B）の「ダウンロードして変換」を途中で止めても、変換の処理が裏で動き続けていました（メモリを使ったまま）。止めると変換も終わるようにしました。

CHANGELOG_en.md

@@ -4,6 +4,11 @@
 
 Dates are JST. Every version is on Hugging Face `ryugyosoft/onw`, tagged from `v0.4.2` on.
 
+## v3.3.3 (2026-10-26)
+- **Fix: background compiles went on after onw was closed (no model update needed)**: the child processes that compile the vision encoder, the fused graphs and a first load's segments in the background kept running after the onw server was stopped or exited (the vision encoder for 10+ minutes on the CPU, sometimes keeping an onw update from replacing its files). They now end with the onw that started them; the next load compiles what is missing.
+- The vision encoder's background compile only compiles now, like the other children (no executor on the NPU, and no second copy in OpenVINO's cache).
+- A vision encoder blob the NPU refused is removed from the cache (~230 MB for Gemma 4).
+
 ## v3.3.2 (2026-10-25)
 - **Fixes (no model update needed)**
   - Stopping "Download and convert" (models converted on this PC, LFM2.5 24B-A2B) left the conversion running in the background (holding its memory). Stopping now ends the conversion too.

onw/__init__.py

@@ -2,4 +2,4 @@
 
 A small engine that runs LLMs entirely on the Intel NPU: static segment graphs + host-managed state.
 """
-__version__ = "3.3.2"
+__version__ = "3.3.3"

onw/chat.py

@@ -180,7 +180,10 @@ def export_vision_blob(core, d, device, c):
     bp = vision_blob(d, c)
     if bp is None or os.path.exists(bp) or os.path.exists(bp + ".bad"):
         return
-    first = core.compile_model(os.path.join(d, "vision.xml"), device, c)
+    # [3.3.3] compile only, as the fused graphs' children do: no NPU executor next to the parent's inference (one created
+    # there lost the NPU on Panther Lake) and no CACHE_DIR (OpenVINO's own cache file was a second copy of this blob)
+    nc = {**{k: v for k, v in c.items() if k != "CACHE_DIR"}, **({"NPU_CREATE_EXECUTOR": "0"} if device == "NPU" else {})}
+    first = core.compile_model(os.path.join(d, "vision.xml"), device, nc)
     buf = io.BytesIO()
     first.export_model(buf)
     os.makedirs(os.path.dirname(bp), exist_ok=True)
@@ -335,12 +338,13 @@ class ChatEngine:
             vb = vision_blob(engine_dir, self._venc_args[3])
             if vb is not None and not os.path.exists(vb) and not os.path.exists(vb + ".bad"):
                 import subprocess, sys
+                from .childguard import child_env
                 flags = 0x00004000 if os.name == "nt" else 0      # (BELOW_NORMAL_PRIORITY_CLASS)
                 self._venc_job = subprocess.Popen(
                     [sys.executable, "-m", "onw.visionjob", json.dumps([engine_dir, device, self._venc_args[3]],
                                                                        default=str)],
-                    env={**os.environ, "PYTHONUTF8": "1"}, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL,
-                    creationflags=flags)
+                    env=child_env({**os.environ, "PYTHONUTF8": "1"}), stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL,
+                    creationflags=flags)                  # [3.3.3] (child_env: it ends with this process)
                 print("[vision] encoder compiling in the background", flush=True)
         pld_state = "off (memory-saving mode)" if self.swap_gb is not None else "on" if self.pld else "off"
         print(f"[load] {time.time()-t0:.1f}s on {device}; verify block = {self.block_cost or 0:.1f} single steps, "
@@ -420,6 +424,10 @@ class ChatEngine:
                 err = e
                 if bp:
                     open(bp + ".bad", "w").write(str(e)[-500:])
+                    try:                                  # [3.3.3] the refused blob is never used again (Gemma 4 on
+                        os.remove(bp)                     # Panther Lake: 237 MB left in the cache)
+                    except OSError:
+                        pass
         raise err
 
     # ---- prompt ----

onw/childguard.py

@@ -0,0 +1,41 @@
+"""[3.3.3] The background compile children (onw.visionjob, onw.fusedjob, onw.precompile) end with the onw process that
+started them: they only fill onw's blob store, and the next load starts what is missing again. Before, a vision encoder
+child kept compiling for 10+ minutes after its server had exited (Panther Lake, Gemma 4), holding CPU and the kit's DLLs.
+The parent passes its pid (child_env); the child watches it (watch_parent) - not its own parent process, which on Windows
+is the venv's launcher and outlives onw."""
+import os
+import threading
+import time
+
+ENV = "ONW_PARENT_PID"
+
+
+def child_env(env=None):
+    """The environment for a compile child: env (default os.environ) plus this process's pid."""
+    return {**(os.environ if env is None else env), ENV: str(os.getpid())}
+
+
+def watch_parent(every=2.0):
+    """In a child: end this process (exit code 4) once the onw process named in ONW_PARENT_PID is gone."""
+    pid = os.environ.get(ENV)
+    if not pid:
+        return
+
+    def run():
+        try:
+            import psutil
+        except ImportError:                             # (nothing to watch with: run to the end as before)
+            return
+        try:
+            p = psutil.Process(int(pid))
+        except (psutil.Error, ValueError):              # (the parent is gone already)
+            os._exit(4)
+        while True:
+            time.sleep(every)
+            try:
+                if not p.is_running() or p.status() == psutil.STATUS_ZOMBIE:   # (is_running checks the start time:
+                    os._exit(4)                                                #  a reused pid is not the parent)
+            except psutil.Error:
+                os._exit(4)
+
+    threading.Thread(target=run, daemon=True, name="onw-parent-watch").start()

onw/fusedjob.py

@@ -32,6 +32,8 @@ class _Model:
 
 
 def main():
+    from .childguard import watch_parent
+    watch_parent()                                        # [3.3.3] (ends with the onw process that started it)
     d, device, config, tried_key, built_T, capacity, kt, S, B = json.loads(sys.argv[1])
     from .fused import FusedDecode
     core = npukit.core()

onw/precompile.py

@@ -1,20 +1,22 @@
-"""A first load's precompile child (runtime.compile_jobs): build the segments named in ONW_PRECOMPILE_ONLY exactly as
-the parent would, so they land in the compile cache (marked as the parent's run, ONW_NPUW_SESSION), then exit.
-usage (internal): python -m onw.precompile '[MODEL_DIR, DEVICE, CONFIG, SIZES, CONTEXT]'"""
-import json, sys
-import openvino as ov
-from . import npukit
-
-
-def main():
-    d, device, config, sizes, context = json.loads(sys.argv[1])
-    from .runtime import Precompiled, SegmentedModel
-    try:
-        SegmentedModel(npukit.core(), d, device, config, sizes=sizes, context=context)
-    except Precompiled:
-        return
-    sys.exit(3)                                           # (not reached: the model raises once compiled)
-
-
-if __name__ == "__main__":
-    main()
+"""A first load's precompile child (runtime.compile_jobs): build the segments named in ONW_PRECOMPILE_ONLY exactly as
+the parent would, so they land in the compile cache (marked as the parent's run, ONW_NPUW_SESSION), then exit.
+usage (internal): python -m onw.precompile '[MODEL_DIR, DEVICE, CONFIG, SIZES, CONTEXT]'"""
+import json, sys
+import openvino as ov
+from . import npukit
+
+
+def main():
+    from .childguard import watch_parent
+    watch_parent()                                        # [3.3.3] (ends with the onw process that started it)
+    d, device, config, sizes, context = json.loads(sys.argv[1])
+    from .runtime import Precompiled, SegmentedModel
+    try:
+        SegmentedModel(npukit.core(), d, device, config, sizes=sizes, context=context)
+    except Precompiled:
+        return
+    sys.exit(3)                                           # (not reached: the model raises once compiled)
+
+
+if __name__ == "__main__":
+    main()

onw/runtime.py

@@ -1869,7 +1869,8 @@ class SegmentedModel:
                 arg = json.dumps([c["d"], c["device"], cfg, c["tried"].key, c["built_T"], self.kv.capacity, kt, S, B],
                                  default=str)
                 flags = 0x00004000 if os.name == "nt" else 0      # (BELOW_NORMAL_PRIORITY_CLASS: answers go first)
-                p = subprocess.Popen([sys.executable, "-m", "onw.fusedjob", arg], env={**env, "ONW_FUSED_SEG": seg},
+                from .childguard import child_env                  # [3.3.3] (ends with this process)
+                p = subprocess.Popen([sys.executable, "-m", "onw.fusedjob", arg], env=child_env({**env, "ONW_FUSED_SEG": seg}),
                                      stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL, creationflags=flags)
                 running.append(p)
                 with _FUSED_KIDS_LOCK:
@@ -2063,8 +2064,9 @@ class SegmentedModel:
             while waiting or running:
                 while waiting and len(running) < jobs:
                     n = waiting.pop(0)
+                    from .childguard import child_env              # [3.3.3] (ends with this process)
                     running.append((n, subprocess.Popen([sys.executable, "-m", "onw.precompile", args],
-                                                        env={**env, "ONW_PRECOMPILE_ONLY": n}, stdout=subprocess.DEVNULL,
+                                                        env=child_env({**env, "ONW_PRECOMPILE_ONLY": n}), stdout=subprocess.DEVNULL,
                                                         stderr=subprocess.DEVNULL, **kw)))
                 time.sleep(0.5)
                 for n, p in [x for x in running if x[1].poll() is not None]:

onw/visionjob.py

@@ -7,6 +7,8 @@ from . import npukit
 
 
 def main():
+    from .childguard import watch_parent
+    watch_parent()                                        # [3.3.3] (ends with the onw process that started it)
     d, device, cfg = json.loads(sys.argv[1])
     from .chat import export_vision_blob
     from .runtime import enable_compiler

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "onw"
-version = "3.3.2"
+version = "3.3.3"
 description = "Ore no NPU ga konna ni ugoku wake nai - LLMs entirely on the Intel NPU"
 readme = "README.md"
 license = {text = "Apache-2.0"}