Implements .dev/docs/keep-fix5-implement1.txt against the 9e6c50f tree:
Phase 1 - the paste pipeline: the face upscaler joins the batched core
(the _batched_paste/_batched_paste_multiface aborts are deleted), mode
is forced chunked with a bg model, face_upscale_model threads through
the paste chunk bodies (parse masks stay on the 512 faces, the ROI
grid_sample warps the upscaled faces), upscale_faces_batch_gpu (B=1,
FP16 autocast on CUDA), and _tiled_scale_and_resize_batch replaces
_tiled_scale_batch + the round-12 bg grouping (B=1 FP16 loop, one
bicubic-antialias resize straight to the paste canvas; the no-model
branch keeps _resize_bg_frames). bg_chunk_size and its pins are removed
as dead machinery. parse_soft_masks and _run_upscaler run under
autocast(enabled=(device.type == 'cuda')).
Phase 2 - facelib: RetinaFace caches priors by (h, w, device) and
detect_faces filters by confidence and runs torchvision.ops.nms on the
device before the single CPU transfer (output format preserved);
read_image gains check_gray=True, only the three sequence call sites
pass False.
Phase 3 - archs: SDPA in CrossAttention._attention, AttnBlock and both
gmflow attention forms; get_flow pinned to float32 with autocast
disabled; pos_emb cast to query_emb.dtype; keep_net under autocast at
both sequence call sites.
Phase 4 - the vendored basicsr pruning: data/, losses/, models/,
train.py, test.py, 23 unused archs, 7 unused utils deleted (78 files);
DCNv2Pack removed from arch_util; the three __init__.py rewrites;
enumeration manifests regenerated. metrics/ and ops/ stay on disk
byte-identical, off the import path. 4.2 skipped (explicitly optional).
Autocast is CUDA-only everywhere, so the CPU gate keeps exercising the
FP32 paths; no model is loaded. Gate: 556 passed, 0 failed (embedded
python; 557 -> 556 = the dead bg_chunk_size pins out, new pins in).
GPU throughput and the FP16/SDPA numerics stay owner-measured.
The tracked files still pointed at the development documents, which now live
in the untracked .dev tree: the usage headings promised the embedded
interpreter while the examples are interpreter agnostic, and the provenance
notes named a local document. The headings are reworded and the provenance
notes name the tool that produces the record instead.
The audit artifacts and the plan moved to the local .dev/docs tree, so the
references in the code, the tools and the tests follow them there. The
license file names the tool that produces the record instead of a path.
The five review findings of 2026-10-23 for phase P5:
1. The P5.5 fallback branch (a checkpoint without flownet.* keys) died with
a bare RuntimeError at the strict load: the arch always owns the flownet
(keep_arch.py:882), so the state dict of the net always carries the
flownet.* keys - and the GMFlow weights the arch had just loaded were
discarded. The flownet tensors of the built net (the trained ones - the
arch was built with flownet_path) are now merged into the checkpoint
state dict before load_state_dict(..., strict=True). The load stays
strict (an unexpected key still raises) and the trained flow weights
survive. The stub arches of the tests behave like nn.Module now - a
no-op recorder had hidden the defect.
2. P5.4's acceptance is no longer unmet: the GPU acceptance run was
executed (real weights, real ComfyUI modules, RTX 4070 Ti SUPER) and is
green (5 passed, exit code 0; the run log is in the audit doc section 5).
KEEP_MODEL_DIR now really decides where the weights come from - it is the
models root of the run (folder_paths.models_dir), the harness fixture
no longer isolates the gpu_acceptance tier, and the run needs
KEEP_NO_COMFY_MOCKS=1 so the real ComfyUI modules are in charge.
3. The "the flow weights are not random" assertion of the acceptance test
could not distinguish random weights from trained ones. It now compares
the flownet tensors with the tensors of their source (the KEEP checkpoint
or the GMFlow file) bit for bit.
4. The superseded download helper load_file_from_url_comfy is gone
(src/keep_paths.py only resolves paths; the download lives in
src/keep_download.py since P5.2). The stale docstrings, the stale
comment in the vendored facelib and the two tests that called the dead
helper are fixed.
5. The risk list mentioned a pending deletion of assets/keep-preview.png
that does not exist - the asset is committed and present (93279 bytes).
The rule in the implementation workflow now states the truth.
Checks: the gate (pytest -q -m "unit or mock_integration") is green (234
passed, 1 skipped, 1 xfailed); the opt-in acceptance run is green (5 passed);
the acceptance tier still skips cleanly without its environment variables.
Five review findings of the P2 phase, each fixed at its root:
1. (medium) The P2.6 guard could not see an undeclared third-party import:
the closure's imports were filtered through a hard-coded 10-name map, so
every module outside it was dropped silently (a module-level `import av`
in a copy of the repo still gave "1 passed"). The map is now a *rename*
map (cv2 -> opencv-python, PIL -> Pillow, yaml -> pyyaml); every other
module is mapped to its own name, so an unknown module shows up as an
undeclared dependency. The guard is split by import time: every
third-party module imported at import time must be a declared
distribution, every lazy (function-level) one must be allowlisted with a
reason (dlib, gdown, mc, wandb - all unreachable from the node).
2. (low) The P2.1 (c) progression test was vacuous: it returned before its
assertions and probed a second, top-level copy of the vendored tree
(`sys.path.insert(modules/deps)`). It now imports the packaged
architecture, asserts that FeedForward / AdaLayerNorm come from the
vendored _diffusers_compat and that no top-level `basicsr` shadows an
installed one.
3. (low) Two swallow-and-stub fallbacks remained in the vendored tree
(facelib/detection, facelib/parsing - their stubs called an unimported
`urlparse` and would have raised NameError). Both now import the vendored
symbol directly; a guard forbids a fallback stub in the vendored tree.
4. (low) `basicsr/utils/file_client.py` still appended to `sys.path` (the
optional `sys_path` of MemcachedBackend), which the plan's §3 rule
forbids anywhere in the package. The mutation and the parameter are gone
and guard (d) now covers the whole package (the vendored tree included).
5. (low) The unreachable body of `KEEPFaceProcessor.process_video` still
called the deleted VideoReader / VideoWriter. The 141 dead lines are
removed; an AST guard keeps the references out.
Found while fixing the findings: the CI job did not install `requests`,
which the vendored `basicsr/utils/download_util.py` imports at module level
(P2.6 added it to requirements.txt but not to the workflow) - a clean CI
runner died with "No module named 'requests'". The install list is now
completed and a test derives it from requirements.txt.
Gate: pytest -m "unit or mock_integration" -> 109 passed, 1 skipped,
4 xfailed; plan's local gate (+ headless) -> 109 passed, 3 skipped,
4 xfailed; tools/scan_compat.py -> finding_count 0 (exit 0); ruff check . ->
All checks passed. A mutation check re-introduces six defects into a copy of
the repository: the matching guard fails in 6/6 cases.
WildAi