docs+gizmo: Free VRAM button, quarantine broken variants, README rewrite#5
Open
semx2a wants to merge 5 commits into
Open
docs+gizmo: Free VRAM button, quarantine broken variants, README rewrite#5semx2a wants to merge 5 commits into
semx2a wants to merge 5 commits into
Conversation
The package had no Nuke-level init.py, so ./gizmos was never added to the plugin path (H2_SamViT: "Unknown command") and the H2_SamViT_Gizmo package itself was not importable from gizmo knob callbacks. Add an init.py that registers ./gizmos and puts the package's parent on the plugin path (which Nuke also adds to sys.path). Also ignore .codegraph/ tool artifacts.
Replace the ad-hoc pip installer (which failed on sam2/sam3 source builds: missing setuptools/bdist_wheel under --no-build-isolation) with uv: - pyproject.toml declares the full dependency set; torch/torchvision pinned to the CUDA 12.6 wheel index; sam2/sam3 build via [tool.uv.extra-build-dependencies] (setuptools/wheel injected into their isolated builds); triton-windows for sam3's `import triton` on Windows. - uv.lock committed for byte-identical installs across machines. - install.py becomes a thin driver: bootstraps uv, runs `uv sync`, records the venv in env_config.json, verifies imports. - --venv <path> installs a machine-local env (faster cold starts); the shared ./venv next to the plugin remains the zero-setup default.
The mode-based viewer click workflow was fully wired but dead:
- viewer_events imported PySide2 only (NameError on import under Nuke 16's
PySide6) and was never installed by anything — install it from menu.py.
- Qt6 wraps the viewer GL surface as a plain QWidget (QGLWidget is gone),
so the class-name gate rejected every click; detect the surface via the
"Viewer.N" ancestor objectName + size instead, and use event.position()
with a pos() fallback.
- Clicks now fall back to the node that entered an edit mode, instead of
silently requiring the H2 node to stay selected in the DAG.
- ui_overlay: forward slashes in Read-node file knobs (Windows backslashes
are mangled as escape sequences) and serial-numbered overlay files so the
file-knob change triggers the reload — knob("reload").execute() from a
click event raised "I'm already executing something else" on every point.
- callbacks: the legacy add_fg/bg_point checkbox shim now only fires for
Boolean knobs (on current gizmos they are PyScript buttons, where
setValue(False) is setCommand → TypeError spam).
Two hard constraints make in-process torch impossible: 1. Nuke bundles its own libtorch (c10.dll/torch_cpu.dll for Inference/ CopyCat), so importing pip-torch inside Nuke fails with WinError 1114. 2. Endpoint-security (EDR) tooling in some environments additionally blocks torch's c10.dll in ANY descendant of Nuke.exe — proven by elimination (identical env and module lists; the same venv loads torch from PowerShell or a WMI-detached process, and fails in every Nuke child regardless of env/console/cwd/job-breakaway). Architecture: - worker.py is the only place torch is imported. On Windows it is spawned OUTSIDE Nuke's process tree via WMI (parent: WmiPrvSE.exe, pythonw.exe) and serves a JSON-lines protocol over a token-protected localhost socket advertised in a %TEMP% rendezvous file; logs go to <rendezvous>.log and it self-destructs after 30 min idle. POSIX keeps a plain piped child. - worker_client.py manages spawn/connect/reuse, a cancellable progress bar (cancel kills the worker), busy-guard, log-tail error reports. - inference.py becomes the Nuke-side broker: render frame to PNG → send request → apply returned mask (temporal consistency + Read update stay in-process). The nuke-free mask pipeline (finalize_mask) is shared with the worker. Frame render and mask Read paths use forward slashes (Windows file knobs) and masks are written next to the saved .nk (h2_samvit_masks/) instead of %TEMP% so they persist with the comp. - model_manager: unwrap the official SAM2 checkpoints' top-level "model" key before load_state_dict. - env_bootstrap: venv site-packages are APPENDED to sys.path (Nuke's own bundled packages keep priority; torch-family removed from the in-process check), no more PATH/VIRTUAL_ENV mutation of the Nuke process. - callbacks: Run Inference / Process Sequence gated on the checkpoint being downloaded (clear refusal instead of a mid-flight surprise); cancelling one frame stops the whole sequence; Free VRAM action.
- H2_SamViT.gizmo: add a Free VRAM button (unloads the model from GPU via the worker; next inference reloads the model only). - Quarantine H2_SamViT_v2/v3.gizmo into gizmos_disabled/: v2 calls a function that no longer exists (run_sam_inference) and v3's sequence path imports torch in-process, which cannot work here. - README rewritten to match reality: mode-based click workflow (the old Ctrl+Click table described a removed interaction scheme), fit-the-viewer caveat, zero-setup shared venv + optional local venv, the detached worker architecture and why it must not be "simplified", mask storage next to the saved script, model-download gating, and troubleshooting entries pointing at the worker log.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Part 5 of the series — builds on #4 (review the last commit only). Closes out the stack.
H2_SamViT.gizmo: Free VRAM button (unloads the model from GPU via the worker; next inference reloads the model only).H2_SamViT_v2/v3.gizmointogizmos_disabled/: v2 callsrun_sam_inferencewhich no longer exists, and v3's sequence path imports torch in-process, which cannot work (see feat: run all torch inference in a detached out-of-process worker #4).