The failing "poison" jobs were 800MB+ 4K VR videos: the agent pulled the ENTIRE
file into memory (r.content) just to sample a few frames, which buffered ~1GB in
RAM and — on any slow/contended media store — got cut off mid-download
(ChunkedEncodingError), failed, and re-leased forever. Measured the media read at
~4–6 MB/s (raw off the share, curator out of the path), so no serving-layer tweak
helps; the file simply shouldn't be fully downloaded.
Environment-agnostic fix (works for any deployment, completes even when slow):
- media.sample_frames_from_url(): point ffmpeg straight at curator's /images URL.
It Range-reads only the video index + up to max_frames of content — never the
whole file — and reconnect flags resume a dropped transfer instead of failing.
Generous, env-tunable timeout (FFMPEG_TIMEOUT, default 1200s) = completion over
speed. Removes the bytes-based sample_frames (dead once videos stream).
- worker._download_decode: videos now stream (no fetch_image, no RAM blowup);
stills still download+decode. On an ffmpeg miss, probe curator liveness
(client.is_reachable) → fail the job if curator is up (unprocessable file, stops
the infinite re-lease) vs release if curator is down (transient, survives a
redeploy). Auth header passed so it works whether or not /images is gated.
Build marker 2026-07-01.6. Refs issue #1225.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
"curator unreachable" was printed for every transient error, hiding whether a
single file's transfer stalled (ReadTimeout — curator is up, that stream is slow)
or curator itself is down (ConnectTimeout/ConnectionError) or errored (HTTP 5xx).
Those need completely different fixes, and we've been diagnosing the download
slowness blind.
Add _transient_reason(exc) → a specific label (HTTP <code>, else the exception
class: ReadTimeout / ConnectTimeout / ConnectionError / …) and use it in both
transient paths:
- downloader: "fetch failed job <id> (image <id>, ReadTimeout) — released, backing off"
- consumer: "submit failed job <id> (<reason>) — released, re-lease later"
Now the logs say which failure it actually is (and which image), so we can tell a
slow/stalled transfer apart from an unreachable curator. Build marker 2026-07-01.5.
Refs issue #1225.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
Operator hit an outage after the machine slept overnight: the agent showed
"curator unreachable" in a loop while curator's API (lease) was actually fine and
the browser could still load images — just slowly. Root cause is a feedback loop
in the new pipeline: every download streams a full original through curator's
single Python file-serving path, and the autoscaler grows DOWNLOADERS whenever the
buffer is empty. When downloads are merely SLOW/failing, the buffer is empty for
that reason — so the agent piled on more concurrent large-file GETs, saturating
curator's web workers + NFS, which slowed curator (and its browser) further and
produced more failures → more downloaders. Classic congestion collapse.
- Failure-aware autoscaling: if transient download failures rose since the last
decision, SHRINK the downloader pool toward the floor instead of growing — the
empty buffer is caused by failures, not the GPU starving. It ramps back up only
once downloads succeed again.
- DL_MAX 24 → 8: 24 concurrent large-file downloads through one Python serving
path is too many; 8 keeps a fast GPU fed without stampeding curator.
- fetch_image timeout 180 → (10, 60): the read timeout is between-bytes, so a
large-but-flowing download still completes, but a stuck/dead connection fails in
60s instead of hanging a downloader for 3 min and piling up stuck requests.
Build marker 2026-07-01.4.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
Near-static videos are the dominant GPU load: sampled into up to 64 frames, each
re-runs the whole detect→CCIP→SigLIP chain on ~identical content. Add a CPU
perceptual-hash frame dedup upstream of the GPU so the redundant frames are never
processed at all (not just their embeds).
- media.dedupe_frames() + _dhash(): 8×8 difference-hash (64-bit) per frame; greedy
keep — a frame survives only if its hash differs from every kept frame by
>= min_distance bits (Hamming). A static run collapses to one frame; genuinely
distinct scenes all survive. Order + frame_time preserved.
- Called in worker._download_decode right after sample_frames, so it runs in the
decode stage on the downloader thread (CPU) — the GPU consumers only ever see
deduped frames, and buffered video items shrink (less RAM too).
- Env-tunable FRAME_DEDUPE_DISTANCE (default 8; higher keeps more frames for brief
localized changes an 8×8 hash can miss; 0 disables). Logs `video frames N→M`
when it drops any, so video load reduction is visible.
Complements the spatial per-frame crop dedup (2026-07-01.2); this is the temporal
axis. Build marker 2026-07-01.3.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
Figure boxes are already NMS-merged (iou 0.6) and each YOLO detector self-NMSes,
but the combined per-frame crop pile (figure→concept ∪ anatomy component→concept
∪ panel) was embedded with no cross-proposer dedup — so genuine near-duplicates
slipped through (a figure box ≈ an anatomy component on a solo bust; overlapping
booru head classes on one head), embedding the same region twice and burning a
slot against max_regions.
Add detectors.dedupe_crops(): a greedy, high-IoU (default 0.85), kind-aware pass
over the pending (crop, template) list right before embed_batch — drop boxes that
overlap ≥ iou within the same kind, keep the highest score. The high threshold is
deliberate: it collapses only true near-identical boxes while preserving
intentional nested crops across scopes (a whole figure vs a small head component
sit well below it) and distinct kinds (concept vs panel). Env-tunable DEDUPE_IOU
(≥1.0 disables). Runs on CPU before the GPU work, so it cuts both embed cost and
region count. Temporal (cross-frame) dedup deferred. Build marker 2026-07-01.2.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
The agent workload is download-bound (download 400–5462ms vs GPU ~300–600ms),
so the old N-slot serial chain (each slot: lease→download→decode→GPU→submit)
left the fast GPU idle during every download. Rearchitect worker.py into a
producer/consumer pipeline:
downloader pool (autoscaled by BUFFER OCCUPANCY) → bounded queue → 1–2 GPU
consumers (detect+embed→submit)
- Downloaders are I/O-bound → many overlap; the autoscaler now tunes DOWNLOADER
count by buffer fill (empty = GPU starving → add; full = outpacing GPU → add a
2nd consumer if it has util/VRAM headroom and lifts throughput, else trim).
- Bounded buffer (12) = backpressure: a full buffer blocks downloaders, capping
RAM + lease look-ahead. VRAM pressure sheds a consumer immediately.
- Heartbeat thread keeps every held lease alive (buffered jobs wait on the GPU;
curator's 180s TTL would otherwise reclaim them mid-buffer).
- Preserves all resilience: lease exp-backoff, submit-path retry (#169),
release-on-stop, region caps + video early-exit (#171). Stop drains BOTH pools
and releases every held lease at once (single held-set as source of truth).
- Consumers SHARE one embedder + proposers instance (a 2nd consumer adds
concurrent inference, not N× VRAM — bounds the VRAM creep seen with N slots).
- UI reworked for the pipeline: tiles show downloaders · buffer · on-GPU ·
processed · errors, a buffer-occupancy meter, and a consumers/waited-out line;
the dial now tunes downloaders. Build marker 2026-07-01.1.
Also fix the operator-flagged detector warning: yolo11n + the comic-panel model
threw "'Conv' object has no attribute 'bn'" on every image (ultralytics' load-
time Conv+BN fusion on a version-mismatched graph), silently disabling 2 of 3
crop proposers and spamming the log per image. Disable that fusion (unfused
inference is correct, marginally slower) and permanently self-disable a proposer
on the first inference failure instead of re-throwing forever.
Refs milestone 122.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
Image 81602 turned out to be a 156 MB mp4, not a huge still: the agent samples
up to 64 frames × ~32 regions/frame → ~2000 regions (the 413) and 64 frames of
detect+CCIP+embed (the 38s). The MAX_REGIONS backstop (#171) only truncated the
SUBMIT — the GPU work was already spent. Break out of the frame loop once
accumulated regions reach max_regions, so a long video costs ~a few frames of
GPU (~2-3s), not all 64 (~38s). The whole-image 'embed' task is unaffected (it
mean-pools all frames and returns before this loop).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
Three safety/robustness fixes from the operator's run logs:
- Cap figures per frame (MAX_FIGURES, default 8) like components/panels already
are. Uncapped, a huge/busy image yielded hundreds of figure boxes → hundreds
of per-figure CCIP calls + crops → a 38s job AND a submit too big to accept
(image 81602 looped on 413). This is the acute fix.
- Global per-JOB backstop (MAX_REGIONS, default 128): if total regions still
exceed the cap (long video), keep the highest-scoring and log the drop, so a
submit body can never blow past curator's limit.
- Stale "active" meter: stop() now resets _active to 0 (no slots remain, so the
meter must read 0 at once), and _bump clamps at 0 so a slot finishing after the
reset can't drive it negative.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
Trusted local library, not an upload surface, so a legitimately large image
(90–95M px, operator-flagged) must load. PIL only WARNS at the 89M-px default but
RAISES DecompressionBombError at ~179M px, which would fail those jobs. Set
Image.MAX_IMAGE_PIXELS = None. (The agent works off individual extracted files —
curator's archive_extractor unpacks zip/cbz/rar/7z at import — so this is about
big single images, not archives.)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
Bump the GPU-agent base image from 12.4.1-cudnn-runtime-ubuntu22.04 (Python 3.10,
CUDA 12.4, early-2024) to 12.9.2-cudnn-runtime-ubuntu24.04:
- Ubuntu 24.04 LTS → Python 3.12 — one modern runtime, no more 3.10.
- CUDA 12.9 + cuDNN 9 — current within the CUDA-12 / cuDNN-9 line that the
default onnxruntime-gpu wheel AND torch cu124 are built against. NOT CUDA 13:
ONNX Runtime's CUDA-13 support is still nascent (separate wheels + open
"Unsupported CUDA version: 13" reports), and torch bundles cu124 anyway. The
GPU (Ampere/Ada, 12 GB) is fine on either — this is a library-alignment call,
not a hardware limit.
- PIP_BREAK_SYSTEM_PACKAGES=1: 24.04 marks system Python externally-managed
(PEP 668); a single-purpose container owns its environment, so global installs
are fine and simplest.
- agent/ruff.toml pinned to py312 (was py310) so CI lints against the real
runtime; from __future__ import annotations stays (PEP 649 lazy annotations
are 3.14, so self-refs still evaluate on 3.12).
CI builds the image but has no GPU — validate on the desktop after pull that it
starts and loads CUDAExecutionProvider (not CPU fallback).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
The agent container (CUDA base, Python 3.10) crashed on startup with
`NameError: name 'Config' is not defined` — an earlier `ruff --fix` unquoted the
`from_env(cls) -> Config` self-reference, which is safe on CI's Python 3.14
(PEP 649 lazy annotations) but is evaluated at class-definition time on 3.10.
CI lint/compile run on 3.14, so it slipped through.
- config.py: `from __future__ import annotations` so the self-referential
annotation is a string, never evaluated — works on 3.10 and every version.
- agent/ruff.toml: pin the agent to `target-version = "py310"` (its real runtime)
and inherit the root rules. Ruff now flags exactly this class as F821, so CI's
lint lane catches it instead of shipping a broken image. (CI otherwise lints on
3.14, masking 3.10 issues.)
- client.py: submit path now retries in-place. A dedicated session with a
urllib3 Retry (connect/read/status, 0.5s backoff, 500/502/503/504, POST) so a
momentary blip after the GPU work is done doesn't discard it and force a full
re-download + recompute elsewhere. A duplicate submit after a lost response is
a harmless 409 no-op. Lease/fetch keep the plain session + loop-level backoff.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
Addresses operator reports: Stop never finishes, the agent polls curator
constantly, and stale-cached pages get mistaken for a failed deploy.
- Stop is prompt: flip _running BEFORE any lock so /status + worker loops see
"stopped" immediately, and add a stop/shrink checkpoint in _process (after
decode, before the expensive detect+embed) that releases the job and bails —
so a Stop doesn't wait out heavy GPU work.
- Lazy curator polling: the queue snapshot is fetched only while a browser is
actually watching (a /status hit within UI_IDLE_GRACE) and on a 5s cadence,
not a constant background loop. The work loop's own lease/submit is curator's
only visitor otherwise — nothing polls just to poll.
- Build marker: VERSION is embedded in the page and reported on /status; the UI
shows a "reload" banner when they differ, so a browser-cached page can't be
mistaken for "the new image didn't deploy" (complements the no-store header).
CI: the lint lane now also `ruff check`s agent/ and compileall-parses it, so the
GPU agent is linted + syntax-checked before its image builds (build.yml only
`docker build`s it). Fixed the agent's pre-existing UP037/B905 so it passes.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
Instrument the job pipeline so we can see where wall-clock actually goes and
decide — on data, not theory — whether a download/compute split is worth
building. Each stage is timed per job and a rolling breakdown is logged every
30s to the agent console, e.g.:
timing/30s — lease 8ms · download 310ms · decode 40ms · gpu 165ms · submit 70ms | wall/job 585ms (214 jobs)
- lease timed around client.lease() in the slot loop (per batch).
- download = fetch_image; decode = image/frame decode; gpu = detect + CCIP +
batched embed; submit = the results POST. One-time model load is excluded
from the gpu figure.
- Thread-safe accumulator (stage -> [sum, count]) summarised + reset by a small
daemon reporter thread; logs only when there was work.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
Operator: the status tiles (state/active/processed) and the Start/Stop buttons
freeze while the GPU meters stay live. Root cause: /status made an INLINE
blocking curator call (queue_status) on every poll, and with curator buried
under a 112k-job backlog that call stalled — freezing the whole status refresh
(the GPU bars survived because /gpu is a lock-free local read). Made worse by the
old util-band autoscaler, which grew workers toward the 32 cap forever because
util plateaus ~50% on this IO-bound load and never hit the 70 grow threshold —
piling load onto curator and the agent process.
- /status is now a pure in-memory read: worker.status() is lock-free, and the
curator queue snapshot is refreshed by a background poller (never inline).
- Autoscaler replaced with a smoothed, throughput-aware climb that SETTLES:
samples util every 2s and EWMA-smooths it (raw util swings 0↔99), then every
~24s grows by one only while each grow keeps lifting smoothed jobs/s; when a
grow stops helping it backs off one and holds, re-probing occasionally. No
runaway, no flopping.
- GPU util bar now shows a smoothed value: the agent's own EWMA (util_smooth,
exposed on /gpu) when running, else smoothed client-side — so it glides
instead of bouncing 0↔99.
- act() aborts a slow Start/Stop POST after 8s so the buttons can't stick; the
now-always-fast /status refresh recovers state regardless.
- Log pane: bound the page to the viewport (height:100vh) so the Logs card
scrolls INTERNALLY instead of overflowing off-screen; cap the ring buffer at
400 lines.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
The control page is a static string served with no cache headers, so after
pulling a fresh agent image the browser kept showing the OLD UI until a hard
refresh (operator-flagged). Add a no-store middleware covering the page and the
status/gpu/logs polls.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
Operator: the buttons fire but the status view doesn't reflect the change. Cause:
act() ignored the POST's own status response and waited on the separate /status
poll (which lags behind the curator queue call). Now:
- act() applies the POST's returned status immediately for instant feedback, and
shows an optimistic "starting"/"stopping" state (pulsing, buttons disabled)
the moment it's clicked.
- A stop that still has in-flight jobs draining shows "stopping" until active
hits 0, then resolves to "stopped" on its own.
- applyStatus() guards the /status-only fields (connection pill + queue) so the
lean action response can't blank them — the Start/Stop path deliberately skips
the slow curator call to stay snappy.
Also de-duplicate GPU reads: read_gpu() now caches (1s TTL) with one probe at a
time, and /status no longer spawns its own nvidia-smi — so the fast /gpu poll +
autoscaler + /status share a single subprocess instead of piling up in the
server thread pool (which was what made clicks feel dead under load).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
Two operator-reported issues with the GPU agent:
1. Worker count flopped almost every cycle, spiking the GPU. The hill-climb
probed +1, judged it over a too-short noisy throughput window, saw no clear
gain and reverted -1 — every tick. Replace it with a GPU-utilization-band
controller: HOLD while smoothed util sits in a healthy band, grow only on
clear spare capacity (util below the low mark + VRAM headroom), shrink under
saturation or memory pressure. Util is EWMA-smoothed and decisions are spaced
(DECIDE_EVERY samples), so a noisy nvidia-smi reading can't move the pool.
Load stays consistent instead of probe/reverting.
2. GPU util/VRAM bars only updated on manual refresh. They rode the /status
poll, which blocks on the curator queue call (slow when curator is busy), so
the meters froze between refreshes. Give them a dedicated /gpu endpoint
(local nvidia-smi only, no curator round-trip) polled every 1.5s, and drop
the curator queue-status timeout 15s -> 5s so /status itself stays snappy.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
Operator meant the LOG section should fill down the viewport (vertical), not the
whole page going full-width horizontally. Restore the centered column (820px),
make .wrap a full-height flex column, and let the Logs card flex to fill the
remaining height to the bottom (drop the fixed 230px log-pane cap).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
Two issues surfaced by the live logs (GPU pegged at ~0% util, 0.5 jobs/s,
truncated-image failures):
- BATCH the SigLIP embeds: collect all of an image's crops (figure + booru_yolo
components + panels) and embed them in ONE forward pass instead of one
forward+lock per crop. The per-crop path serialised every crop through the
inference lock and starved the GPU (≈0% util, autoscaler stuck oscillating);
batching gives a real GPU-bound workload + far higher throughput. CCIP still
runs per figure inline.
- LOAD_TRUNCATED_IMAGES in the agent (matches the server embedder): slightly-
truncated scraped images now load instead of failing the job 3× then erroring
("image file is truncated (N bytes not processed)").
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
- Page fills the viewport horizontally (drop the 780px cap).
- Copy button on the Logs card → copies the console (clipboard API on localhost,
textarea-execCommand fallback), with a brief "Copied" confirmation.
- Silence httpx/httpcore/huggingface_hub/urllib3/filelock/uvicorn.access/
ultralytics to WARNING so the console shows agent activity (detector loads,
job errors, autoscale moves) instead of per-request HF-download spam.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
- logbuf.py: bounded in-memory log ring buffer + a logging.Handler on the root
logger; GET /logs serves it; the control page polls it into a console pane —
so runs are monitorable without `docker logs`. worker now logs autoscale moves
(one line per change, with jobs/s + util + VRAM) and job failures (job + image
+ reason); detectors already log load/disable.
- Restyled the whole control page: a proper dark layout with a header + live
connection pill, cards (Control / Status / Logs), a styled Auto switch +
worker stepper, status tiles, separate GPU-util and VRAM meters, and the log
console. No longer feels like an afterthought; all the existing control hooks
are preserved.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
The new per-job workload (3 detectors + several SigLIP embeds) is far more
GPU-bound than the old I/O-bound CCIP pass, so the right worker count shifted and
is hard to guess. Add an Auto mode (default ON) that finds it:
- _control_loop samples jobs/sec + GPU util/VRAM every ~6s and hill-climbs the
target: grow while throughput keeps improving and VRAM stays under budget,
revert a step that doesn't help, back off under memory pressure (VRAM >= 90%),
then settle and periodically re-probe (the GPU/IO balance shifts over a run).
- A manual concurrency set is an override → leaves Auto; an "Auto" toggle in the
control UI re-enables it. status() reports `auto`; the dial reflects the
auto-chosen count (read-only) while Auto is on.
- AUTO_SCALE env (default on) + compose doc. Agent py-compiled (outside CI).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
Better region PROPOSERS feeding the existing crop→SigLIP→max-over-bag heads (no
change to the learned-tagging approach; no per-tag cost — propose once, embed
each region, all heads in one matmul).
- detectors.py: lazy ultralytics YOLO wrapper, each proposer independently
optional + guarded (a bad weight spec / inference error self-disables that one,
logged, never breaks the worker). Weights resolve from an ultralytics name |
http(s) URL | "hf_repo::file", cached under HF_HOME. NMS merge so a figure two
detectors both find collapses to one crop.
- worker: figure boxes = imgutils detect_person ∪ general COCO person (merged)
→ CCIP + concept (anime + Western/realistic coverage); booru_yolo anatomy
components (head/cat-head/anatomy/…) → concept crops; comic panels → kind=
'panel' concept crops. Capped per frame (MAX_COMPONENTS/MAX_PANELS).
- config + compose: PERSON_WEIGHTS (default yolo11n.pt, works OOB),
ANATOMY_WEIGHTS + PANEL_WEIGHTS (operator sets booru_yolo URL + mosesb panel
hf::file; empty = off). ultralytics added to requirements.
- backend: image_region 'kind' doc notes 'panel'; no migration (free String,
and the bag scorer keys on a non-null siglip_embedding, not the kind, so any
SigLIP region joins the bag automatically).
Agent is outside CI — py-compiled here; operator tests on the GPU and checks
Western-vs-anime crop quality via /api/ccip observability.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
Make the SigLIP embedder an operator choice (drop-in to SigLIP 2:
google/siglip2-so400m-patch16-512 is a verified 1152-d model at 512px → no
schema change, better small-cue fidelity). A swap = set model + re-embed +
retrain, all operator-driven; the GPU agent does the re-embed so it's fast.
- settings: embedder_model_name is now a setting (migration 0065) alongside the
existing embedder_model_version; both editable + validated (non-empty) in the
ml admin API. The server embedder loads by HF name (AutoImageProcessor/Model,
model-agnostic), preferring the pre-downloaded local dir for the default so
existing deploys don't re-download; rebuilds on a name change.
- agent: new 'embed' job = whole-image SigLIP embedding (mean-pool video frames)
under the lease-announced model → POST /jobs/submit_embedding writes
image_record.siglip_embedding + siglip_model_version. The lease now announces
the model FROM THE SETTING (not a constant).
- re-embed routing: enqueue_gpu_backfill('embed') selects unembedded + stale-
version images; 'siglip' now re-embeds concept crops whose version != current
(so a swap re-triggers crops, not just the never-embedded back-catalogue). The
CPU ml-worker backfill no longer re-embeds on a version mismatch (it can't
churn the library at 512px) — the GPU agent owns version re-embeds. Daily
'embed' + 'siglip' beats self-heal.
- scoring: score_image only bags embeddings in the CURRENT model's space (whole-
image gated by siglip_model_version, concept regions by embedding_version) so a
mid-swap stale vector isn't scored by new-space heads; legacy NULL = current.
- UI: GpuAgentCard "Embedding model (advanced)" — edit name/version, Save, and
"Re-embed library (GPU)" (queues embed + siglip); points at SigLIP 2.
Tests: lease announces model + submit_embedding round-trip; enqueue 'embed'
selects stale/unembedded; stale-version excluded from scoring; embedder model
settable + empty rejected; siglip gate updated to current-version concept.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
For redeploying curator while away with nobody to restart the agent:
- _process now distinguishes a TRANSPORT error (curator down/redeploying, 5xx,
401/403/408/409/429, or our lease reclaimed mid-flight) from a genuine job
fault. On a transport error it hands the job back (best effort) and signals
the loop to back off — instead of calling fail(), which would burn the job's
server-side attempt budget (MAX_ATTEMPTS=3) and permanently error good jobs
across a redeploy. Job-specific 4xx (404 image gone) still fail so they don't
re-lease forever.
- lease loop retries with capped exponential backoff (poll_idle → 60s) and
resets on the first successful lease, so a long outage is gentle and recovery
is automatic within ≤60s of curator returning. Sleeps are interruptible so
Stop / pool-shrink stays responsive.
- AUTO_START env (default on in compose) resumes the worker on container start,
so a host reboot / crash-restart (restart: unless-stopped) self-heals with
nobody at the desktop.
- control UI shows a "waited out" counter + an "curator unreachable, holding
work" banner so the recovering state reads as recovery, not failure.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
Lift recall on small/local concepts (glasses, cum, stomach-bulge, xray,
lactation) that the whole-image SigLIP vector washes out: the GPU agent now
embeds figure crops with SigLIP too, stored as kind='concept' regions, and the
suggestion rail scores each image as a BAG (whole-image + every concept crop),
taking each head's MAX over the bag. The whole-image vector is always in the
bag, so this can never score lower than before.
Model-agnostic by construction: the server ANNOUNCES the embedding model
(HF name + version) in the lease, so the agent loads whatever the heads were
trained in and stays in lock-step — a model swap is a server setting + a
re-embed migration, never an agent change.
- agent: model-agnostic CropEmbedder (torch/transformers get_image_features,
fp16 on CUDA, inference-locked); worker branches on job.task — 'ccip' emits
figure(CCIP)+concept(SigLIP) in one pass, 'siglip' emits concept-only so the
back-catalogue backfill never churns figure/CCIP regions; torch cu124 +
transformers in the image.
- server: lease announces embed_model_name/embed_version; score_image is
max-over-bag (version-filtered region embeddings); enqueue_gpu_backfill
'siglip' gates on a missing concept region (drains the back-catalogue,
retries failures, no double-enqueue); daily siglip-backfill beat; UI button;
/api/ccip/overview reports images_with_concept_siglip.
- v1 scope: suggestion rail only — auto-apply stays whole-image (conservative;
heads' thresholds were calibrated on whole-image). Bulk-apply bag = follow-up.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
At 8 workers the GPU sat at ~5% util / <5GB VRAM — the pipeline is I/O-bound
(downloading + decoding images over HTTP), so the GPU starves until many workers
overlap that I/O. Raise MAX_CONCURRENCY 8→32 and make the UI worker control a
number input (reaching 32 by ±1 was tedious); the cap is reported via /status so
the UI clamps to it. Also size the shared requests pool (pool_maxsize=64) — the
default 10 would have throttled 32 workers + spammed "connection pool is full".
Verified by running; watch GPU util/VRAM climb as you dial up.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
A naive convert('RGB') on a palette-with-transparency image (common: character
PNGs on a clear background) lets PIL guess the transparent pixels — black-ish
artifacts that bleed into the crop + the CCIP embedding (and the "should be
converted to RGBA" warning). to_rgb() composites over white first for a clean,
consistent background; used by both stills and video frames.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
Control UI gains what the operator asked for:
- GPU load (nvidia-smi): util %, VRAM used/total + bar, temp — so you can see how
hard the card is working while you're at the desktop.
- Worker count is now a live − / + control (POST /concurrency), not just an env:
the worker is a pool of independent slots (shared model, so slots add concurrent
inference, not N× VRAM). Dial up for speed, down to free the card. Replaces
pause/resume with Start/Stop + the worker dial.
- Graceful release on stop / pool-shrink: a slot hands its still-leased jobs back
via client.release() so they're re-picked immediately (pairs with the server
recovery sweep).
Not CI-tested (agent/ outside CI) — verified by running.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
onnxruntime-gpu needs cuDNN 9; the plain cuda:12.4.1-runtime image lacks it
(libcudnn.so.9 missing → CUDAExecutionProvider falls back to CPU). Switch to
the -cudnn-runtime variant which bundles cuDNN 9.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
compose file (pull the published image, GPU reservation, model-cache volume,
.env for the token) so the agent runs with `docker compose up -d` instead of a
long docker run. A copy + .env template also placed in ~/Documents/fc-gpu-agent.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
Build + push fabledcurator-agent alongside web/ml (own CUDA + onnxruntime-gpu
image, context=agent/, same tag cadence: main → :main/:latest/:c-<sha>, tag →
:<version>). So the operator PULLS + runs it on the GPU machine instead of
building locally. README switched to docker pull.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
The last piece: a Dockerised desktop-GPU worker that talks to FC ONLY over HTTP
(lease → fetch pixels → detect figures + CCIP-embed → submit), so Redis/Postgres
stay private. New top-level agent/ (outside CI scope — verified by running it):
- fc_agent/worker.py: the lease/compute/submit loop, concurrency 1, start/pause/
stop (stop frees the card; unprocessed leases expire + re-queue).
- fc_agent/models.py: imgutils wrappers — detect_person (figures) + CCIP embed.
The two API seams to verify against the installed dghs-imgutils (flagged).
- fc_agent/media.py: stills + video frame sampling (ffmpeg) at FC's cadence →
per-frame instances (the bag).
- fc_agent/crops.py: vendored crop primitive. client.py: the FC HTTP client.
- fc_agent/app.py: FastAPI localhost control UI (start/pause/stop + progress +
queue depth). Dockerfile (CUDA + onnxruntime-gpu + ffmpeg) + requirements +
README (token → build → run --gpus all → Start; CPU-fallback path).
This completes the CCIP pipeline end to end: agent produces region CCIP vectors →
RegionService stores → matcher suggests characters → rail. Verified by running on
the desktop (not CI). README calls out the imgutils API + model-string checks.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
