Set GPU_QUEUE_THROTTLE to low (#24088)

send_alert()'s short circuit for skipping a no-op "update" push only
compared object and zone counts between before/after, never severity.
Both underlying collections are cumulative and deduplicated (objects
is a set of labels, zones only appends a zone not already present),
so a segment being promoted from detection to alert can leave both
counts unchanged, silently dropping the single most notable
transition in a review's life. Add a severity comparison to the same
check so a detection -> alert promotion always notifies.

docs/docs/configuration/record.md

@@ -9,7 +9,7 @@ import NavPath from "@site/src/components/NavPath";
 
 Recordings can be enabled and are stored at `/media/frigate/recordings`. The folder structure for the recordings is `YYYY-MM-DD/HH/<camera_name>/MM.SS.mp4` in **UTC time**. These recordings are written directly from your camera stream without re-encoding. Each camera supports a configurable retention policy. Frigate chooses the largest matching retention value between the recording retention and the tracked object retention when determining if a recording should be removed.
 
-New recording segments are written from the camera stream to cache, they are only moved to disk if they match the setup recording retention policy.
+New recording segments are written from the camera stream to cache, they are only moved to disk if they pass a validation check and match the setup recording retention policy.
 
 :::tip
 

docs/docs/troubleshooting/common_errors.md

@@ -66,17 +66,19 @@ An FFmpeg message meaning it probed the stream but never saw enough decodable vi
 
 ## Recording
 
-<FaqItem id="no-new-recording-segments" question="No new recording segments were created for <camera> in the last 120s">
+<FaqItem id="no-new-recording-segments" question="No new recording segments were created (or: No new valid recording segments were created / No valid segments created since last invalid segment) for <camera> in the last 120s">
 
-Frigate's record watchdog is restarting the record FFmpeg process because no valid segment has reached the cache. This means the record stream is not connecting or the segments are being rejected (see the audio-codec entry below).
+Frigate's record watchdog is restarting the record FFmpeg process because the camera stopped producing usable recordings. The wording distinguishes the cases: `No new recording segments` means no new segment file reached the cache, so ffmpeg isn't getting video out of the record stream; the two `valid` variants mean recordings are arriving but keep failing validation. Either way the fault is on the camera or network side, and the restart is Frigate trying to recover.
 
-See [Recordings: the record stream isn't connecting](/troubleshooting/recordings#the-record-stream-isnt-connecting).
+See [Recordings: no new recording segments were created](/troubleshooting/recordings#no-new-recording-segments-were-created).
 
 </FaqItem>
 
-<FaqItem id="invalid-or-missing-video-stream-in-segment" question="Invalid or missing video stream in segment. Discarding.">
+<FaqItem id="invalid-or-missing-video-stream-in-segment" question="Invalid or missing video stream in segment. Discarding. / Discarding a corrupt recording segment / Failed to probe corrupt segment / Invalid recording segment detected">
 
-A cached recording segment failed validation (no readable video stream) and was deleted. The most common cause is a segment that was truncated because the record FFmpeg process was killed mid-write, so this often appears alongside, and as a consequence of, the record-stream restarts above. A segment containing only audio triggers it too.
+A cached recording segment failed validation and was deleted, either because it had no readable video stream or because its length was impossible. This nearly always means the camera stopped sending usable video partway through the segment: a camera that rebooted, dropped the connection, or ran out of simultaneous connections, or an unreliable link such as WiFi or a failing switch port. Broken camera timestamps (a "Smart Codec" / H.264+ mode) cause the corrupt-segment variants. The same stream failure trips the record watchdog, so the restarts above usually appear alongside these messages.
+
+See [Recordings: invalid or missing video stream in segment](/troubleshooting/recordings#invalid-or-missing-video-stream-in-segment).
 
 </FaqItem>
 

docs/docs/troubleshooting/recordings.md

@@ -209,6 +209,50 @@ If the record stream uses a "Smart Codec"/H.264+ mode or changes encoding parame
 
 </FaqItem>
 
+<FaqItem id="invalid-or-missing-video-stream-in-segment" question="I see the message: WARNING : Invalid or missing video stream in segment ... Discarding.">
+
+Every recording segment is validated before it leaves the cache. Frigate probes each finished `.mp4` in `/tmp/cache` and requires a readable video stream and a valid duration before moving to storage. A segment that fails is deleted, so those ~10 seconds of footage are lost. Three messages come from this check:
+
+- `Invalid or missing video stream in segment <path>. Discarding.` The segment holds no video, or could not be read at all.
+- `Failed to probe corrupt segment <path>` followed by `Discarding a corrupt recording segment: <path>`. The segment was read, but its length could not be determined.
+- `Discarding a corrupt recording segment: <path>` on its own. The segment's length is impossible (empty, or longer than ten minutes), which points at broken timestamps coming from the camera.
+
+For each one, the camera watchdog also logs `Invalid recording segment detected for <camera> at <timestamp>`.
+
+:::warning
+
+This is almost always a **camera or network problem**, not a Frigate one. A segment is only complete once ffmpeg has finished writing it, so anything that interrupts the stream partway through leaves behind a file that cannot be saved. Frigate is reporting the interruption, not causing it.
+
+:::
+
+#### Start with the camera and the network
+
+- **The camera dropped the connection.** Cameras reboot, reinitialize their stream when switching to night mode, and cut clients off when they are overloaded or out of simultaneous connections. Count everything pulling from the camera at once: Frigate's detect and record streams, go2rtc, a phone app, and any other NVR each use one. Routing all roles through a single [RTSP restream](/configuration/restream#reduce-connections-to-camera) so the camera only ever sees one connection often resolves this by itself.
+- **The link to the camera is unreliable.** WiFi cameras, powerline adapters, a saturated uplink, a failing switch port, or a marginal cable all produce this pattern, and usually only on one camera at a time. WiFi cameras are [not recommended](https://ipcamtalk.com/threads/multiple-cameras-high-bandwidth.77100/#post-861110).
+- **The camera cannot reliably send what it is being asked for.** A high bitrate 4K stream can be more than the camera's own hardware can encode and push out under load. Lower the bitrate, or record a lower-resolution profile.
+- **The camera is using a "Smart Codec", H.264+, or H.265+ mode.** These change encoding parameters mid-stream and produce the broken timestamps behind the corrupt-segment variant. Turn the mode off and set the camera's keyframe interval equal to its frame rate. See [Segments are only ~1 second long](#segments-are-only-1-second-long).
+
+Read the rest of the Frigate and/or go2rtc log around the **first** occurrence. When the camera or the network is at fault, other messages show up with it, such as `No frames received from <camera> in 20 seconds`, `Non-monotonic DTS`, `RTP: PT=xx: bad cseq`, `error while decoding MB`, or a connection timeout. Each of those is explained in [Common error messages](/troubleshooting/common_errors). To confirm the camera is the source, open its stream in the [go2rtc web interface](/troubleshooting/go2rtc) on port `1984` or play the same URL in VLC, and leave it running long enough for the failures to happen again.
+
+#### If the camera and network check out
+
+- **Audio the recording cannot store.** Some cameras send G.711 audio, which cannot be saved in an MP4 and stops segments from finalizing. See [Incompatible audio codec](#incompatible-audio-codec-recordings-silently-fail-to-save).
+- **Frigate itself was stopped or restarted.** A single warning per camera around a restart is expected and needs no action.
+- **The system ran out of room or memory.** A full `/tmp/cache`, or the host killing Frigate for using too much memory, cuts off the segment being written. Both leave other errors in the log alongside this one. See [No space left on device](#errno-28-no-space-left-on-device).
+
+</FaqItem>
+
+<FaqItem id="no-new-recording-segments-were-created" question="I see the message: ERROR : No new recording segments were created for <camera> in the last 120s. Restarting the ffmpeg record process...">
+
+When a camera stops producing usable recordings for two minutes, Frigate restarts that camera's record process to try to recover. The wording tells you how far the recordings got:
+
+- **`No new recording segments were created`**: no new segment file showed up in the cache at all, so ffmpeg isn't getting video out of the record stream. The camera is unreachable or refusing the connection, the stream URL, path, or credentials are wrong, or the camera accepted the connection and then sent nothing. See [The record stream isn't connecting](#the-record-stream-isnt-connecting).
+- **`No new valid recording segments were created`** and **`No valid segments created since last invalid segment`**: recordings are arriving, but they keep failing validation, so the camera is sending video that cannot be saved. See [Invalid or missing video stream in segment](#invalid-or-missing-video-stream-in-segment) above.
+
+The restart is Frigate recovering from a problem, not causing one. One of these after a camera reboot or a brief network drop is normal. Seeing them repeat every couple of minutes means the camera or the network is still failing, and the restarts can extend the damage, because each one cuts off the segment that was being written. Work from the earliest failure in that camera's log rather than from the restarts.
+
+</FaqItem>
+
 <FaqItem id="i-see-the-message-warning--unable-to-keep-up-with-recording-segments-in-cache-for-camera-keeping-the-5-most-recent-segments-out-of-6-and-discarding-the-rest" question="I see the message: WARNING : Unable to keep up with recording segments in cache for camera. Keeping the 5 most recent segments out of 6 and discarding the rest...">
 
 This warning means the recording maintainer cannot move recording segments from the RAM cache to disk fast enough. When the cache fills up, Frigate discards the oldest segments to avoid running out of memory and crashing, so you lose recorded footage. This is almost always a storage throughput or system resource problem. Work through the steps below to identify which.

frigate/comms/webpush.py

@@ -421,6 +421,7 @@ class WebPushClient(Communicator):
         # Don't notify if message is an update and important fields don't have an update
         if (
             state == "update"
+            and payload["before"]["severity"] == payload["after"]["severity"]
             and len(payload["before"]["data"]["objects"])
             == len(payload["after"]["data"]["objects"])
             and len(payload["before"]["data"]["zones"])

frigate/detectors/detection_runners.py

@@ -25,25 +25,31 @@ def is_arm64_platform() -> bool:
     return machine in ("aarch64", "arm64", "armv8", "armv7l")
 
 
-def get_ort_session_options(
-    is_complex_model: bool = False,
-) -> ort.SessionOptions | None:
+def get_ort_session_options(model_type: str | None = None) -> ort.SessionOptions | None:
     """Get ONNX Runtime session options with appropriate settings.
 
     Args:
-        is_complex_model: Whether the model needs basic optimization to avoid graph fusion issues.
+        model_type: Model being loaded, used to pin its graph optimization level.
 
     Returns:
-        SessionOptions with appropriate optimization level, or None for default settings.
+        SessionOptions with a pinned optimization level, or None for default settings.
     """
-    if is_complex_model:
-        sess_options = ort.SessionOptions()
-        sess_options.graph_optimization_level = (
-            ort.GraphOptimizationLevel.ORT_ENABLE_BASIC
-        )
-        return sess_options
+    # Import here to avoid circular imports
+    from frigate.embeddings.types import EnrichmentModelTypeEnum
 
-    return None
+    if model_type == EnrichmentModelTypeEnum.jina_v2.value:
+        # below EXTENDED the CUDA EP returns an identical vector for every image,
+        # and ORT_ENABLE_ALL fails to build on CPU with a SimplifiedLayerNormFusion error
+        level = ort.GraphOptimizationLevel.ORT_ENABLE_EXTENDED
+    elif model_type == EnrichmentModelTypeEnum.jina_v1.value:
+        # aggressive optimizations create or expect nodes that don't exist
+        level = ort.GraphOptimizationLevel.ORT_ENABLE_BASIC
+    else:
+        return None
+
+    sess_options = ort.SessionOptions()
+    sess_options.graph_optimization_level = level
+    return sess_options
 
 
 # Import OpenVINO only when needed to avoid circular dependencies
@@ -115,21 +121,6 @@ class BaseModelRunner(ABC):
 class ONNXModelRunner(BaseModelRunner):
     """Run ONNX models using ONNX Runtime."""
 
-    @staticmethod
-    def is_cpu_complex_model(model_type: str) -> bool:
-        """Check if model needs basic optimization level to avoid graph fusion issues.
-
-        Some models (like Jina-CLIP) have issues with aggressive optimizations like
-        SimplifiedLayerNormFusion that create or expect nodes that don't exist.
-        """
-        # Import here to avoid circular imports
-        from frigate.embeddings.types import EnrichmentModelTypeEnum
-
-        return model_type in [
-            EnrichmentModelTypeEnum.jina_v1.value,
-            EnrichmentModelTypeEnum.jina_v2.value,
-        ]
-
     @staticmethod
     def is_migraphx_complex_model(model_type: str) -> bool:
         # Import here to avoid circular imports
@@ -323,6 +314,12 @@ class OpenVINOModelRunner(BaseModelRunner):
         if device in ["GPU", "AUTO", "NPU"]:
             self.ov_core.set_property(device, {"PERFORMANCE_HINT": "LATENCY"})
 
+        if device in ["GPU", "AUTO"]:
+            try:
+                self.ov_core.set_property("GPU", {"GPU_QUEUE_THROTTLE": "LOW"})
+            except Exception as e:
+                logger.debug(f"GPU_QUEUE_THROTTLE not supported: {e}")
+
         if device == "NPU" and OpenVINOModelRunner.is_detection_model(model_type):
             try:
                 self.ov_core.set_property(device, {"NPU_TURBO": "YES"})
@@ -626,9 +623,7 @@ def get_optimized_runner(
     return ONNXModelRunner(
         ort.InferenceSession(
             model_path,
-            sess_options=get_ort_session_options(
-                ONNXModelRunner.is_cpu_complex_model(model_type)
-            ),
+            sess_options=get_ort_session_options(model_type),
             providers=providers,
             provider_options=options,
         ),

frigate/test/test_detection_runners.py

@@ -0,0 +1,41 @@
+"""Tests for ONNX Runtime session option selection."""
+
+import unittest
+
+import onnxruntime as ort
+
+from frigate.detectors.detection_runners import get_ort_session_options
+from frigate.detectors.detector_config import ModelTypeEnum
+from frigate.embeddings.types import EnrichmentModelTypeEnum
+
+
+class TestGetOrtSessionOptions(unittest.TestCase):
+    def test_jina_v2_uses_extended(self):
+        """jina-clip-v2 returns an identical vector for every image on the CUDA
+        execution provider at anything below EXTENDED."""
+        options = get_ort_session_options(EnrichmentModelTypeEnum.jina_v2.value)
+
+        self.assertIsNotNone(options)
+        self.assertEqual(
+            options.graph_optimization_level,
+            ort.GraphOptimizationLevel.ORT_ENABLE_EXTENDED,
+        )
+
+    def test_jina_v1_uses_basic(self):
+        options = get_ort_session_options(EnrichmentModelTypeEnum.jina_v1.value)
+
+        self.assertIsNotNone(options)
+        self.assertEqual(
+            options.graph_optimization_level,
+            ort.GraphOptimizationLevel.ORT_ENABLE_BASIC,
+        )
+
+    def test_other_models_use_defaults(self):
+        for model_type in [
+            None,
+            EnrichmentModelTypeEnum.paddleocr.value,
+            EnrichmentModelTypeEnum.arcface.value,
+            ModelTypeEnum.rfdetr.value,
+        ]:
+            with self.subTest(model_type=model_type):
+                self.assertIsNone(get_ort_session_options(model_type))

frigate/util/config.py

@@ -533,7 +533,7 @@ def migrate_018_0(config: dict[str, dict[str, Any]]) -> dict[str, dict[str, Any]
     genai = new_config.get("genai")
 
     if genai and genai.get("provider"):
-        genai["roles"] = ["embeddings", "descriptions", "chat"]
+        genai["roles"] = ["descriptions", "chat"]
         new_config["genai"] = {"default": genai}
 
     # Remove deprecated sync_recordings from global record config