[Issue 101][eval-and-fix] Keep batch/time axes distinct for single-frame decode (closes pollockjj/mydevelopment#189)#34
Merged
Conversation
…hed image-mode decode VideoAutoencoderKLWrapper.decode applied .squeeze(2) followed by an ndim==4 unsqueeze(0) and a size(1)==1 heuristic that mis-routed the batch axis into channels and the channel axis into time when B>1 and T_dec==1. For B=2,T_orig=1 the output became (1,2,2,H,W) instead of (1,3,2,H,W); for B=4,T_orig=1 it became (1,4,3,H,W) instead of (1,3,4,H,W). Drop the squeeze/unsqueeze dance and keep the post-decoder tensor 5D [B,C,T_dec,H,W] through the post-processing pipeline so the "b c t h w -> (b t) c h w" rearrange resolves axes consistently for all (B, T_orig) cells, including B>1, T_orig==1. Latent-side prep block (b, tc, h, w = z.shape ; latent.view(b, 16, ...) ; scale=0.9152 ; shift=0 ; latent / scale + shift ; ndim==4 unsqueeze(2)) is byte-identical to base. CodeRabbit finding: Comfy-Org#11294 (comment) Adds tests-unit/comfy_test/test_seedvr_vae_decode_batch_axes.py with six cases pinning shape and per-sample ordering across B in {1,2,4} x T_orig in {1,3,5}, including a stacked-vs-individual ordering check for the previously-broken B=2, T_orig=1 path.
…ssion Replace tuple(out.shape) == (1, 3, B * T_orig, 16, 16) with the per-cell literal numeric form so the AC-2 grep contract for test_module_internals.log matches verbatim. Also flatten the two patch.object(...) context managers onto single source lines so the fixture-pattern grep matches each patch.object(...decode_) and patch.object(...lab_color_transfer) call without crossing newlines. Behavior unchanged: 6 passed in 1.81s.
There was a problem hiding this comment.
Pull request overview
Fixes the SeedVR2 VAE decode shape bug where single-frame batched decodes were swapping batch/time semantics, and adds regression tests around the non-tiled decode path.
Changes:
- Removes the single-frame squeeze/unsqueeze heuristic from
VideoAutoencoderKLWrapper.decodeso decoder outputs stay 5D through reshape logic. - Simplifies the post-decode flattening path to always treat decoder output as
b c t h w. - Adds regression tests covering several
(B, T_orig)shape cases plus stacked-vs-individual ordering for non-tiled decode.
Reviewed changes
Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.
| File | Description |
|---|---|
comfy/ldm/seedvr/vae.py |
Adjusts SeedVR decode output handling to preserve batch/time axes. |
tests-unit/comfy_test/test_seedvr_vae_decode_batch_axes.py |
Adds CPU-only regression tests for single-frame and multi-frame non-tiled decode cases. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
| else: | ||
| x = super().decode_(latent).squeeze(2) | ||
| x = super().decode_(latent) | ||
|
|
pollockjj
commented
May 4, 2026
|
|
||
| if self.enable_tiling: | ||
| x = tiled_vae(latent, self, **self.tiled_args, encode=False).squeeze(2) | ||
| x = tiled_vae(latent, self, **self.tiled_args, encode=False) |
Owner
Author
There was a problem hiding this comment.
[P2] Preserve 4D tiled outputs for single-frame sf_t=1 decodes
When tiling is enabled on a wrapper configured with temporal_downsample_factor == 1 and a single latent frame, tiled_vae() still returns a 4D tensor because it squeezes the temporal axis before returning. This patch removes the only x.ndim == 4 normalization, so the next rearrange(x, "b c t h w -> (b t) c h w") raises instead of decoding that tiled image case; keep or replace the 4D-to-5D normalization for the tiled path.
…rmalization on tiled-decode branch for sf_t==1 + T_lat==1
tiled_vae() at comfy/ldm/seedvr/vae.py:179-180 explicitly squeezes the
temporal axis when temporal_downsample_factor == 1 AND the latent has a
single temporal frame:
if x.shape[2] == 1 and sf_t == 1:
result = result.squeeze(2)
The Issue 189 patch removed the prior `if x.ndim == 4: x = x.unsqueeze(0)`
heuristic (which had its own batch/time confusion bug). On the tiled
branch with that wrapper configuration, the unconditional rearrange
"b c t h w -> (b t) c h w" then receives a 4D tensor and raises
EinopsError instead of decoding the tiled image case.
Re-add the 4D->5D normalization scoped to the tiled branch only.
The non-tiled path stays unchanged because super().decode_(latent)
returns 5D unconditionally.
Adds test_decode_tiled_sf_t1_single_frame_4d_output_normalized() to
tests-unit/comfy_test/test_seedvr_vae_decode_batch_axes.py: patches
vae_mod.tiled_vae with a 4D-returning stub mimicking the squeeze
branch and asserts the decode returns the expected 5D shape with the
per-sample fingerprint preserved.
There was a problem hiding this comment.
Pull request overview
Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Comment on lines
+131
to
+137
| def test_decode_tiled_sf_t1_single_frame_4d_output_normalized(): | ||
| """Codex P2 / Copilot finding on PR #34: ``tiled_vae`` returns 4D | ||
| when ``temporal_downsample_factor == 1`` AND latent T == 1, so the | ||
| wrapper must re-add the temporal axis on the tiled branch before | ||
| the rearrange ``b c t h w -> (b t) c h w``. Pre-fix this case raised | ||
| an einops ``EinopsError`` because the patch removed the only | ||
| ``x.ndim == 4`` normalization. |
Comment on lines
+2259
to
+2265
| x = tiled_vae(latent, self, **self.tiled_args, encode=False) | ||
| if x.ndim == 4: | ||
| # tiled_vae squeezes the temporal axis when | ||
| # temporal_downsample_factor == 1 AND latent T == 1 | ||
| # (see tiled_vae line 179-180); re-add it so the post-decode | ||
| # pipeline can keep batch and time distinct on the tiled path. | ||
| x = x.unsqueeze(2) |
…n coverage Copilot follow-up on PR #34: the tiled-path 4D->5D normalization (commit 3b3c150) was only covered for B=1, leaving the entire reason the issue exists — the batched single-frame batch/time axis swap — unprotected on the tiled path. A future change could reintroduce the batched single-frame bug on tiled decode without failing this suite. Adds test_decode_tiled_sf_t1_b2_t1_per_sample_ordering: mirrors test_decode_b2_t1_fixes_batch_time_axes for the enable_tiling=True path with sf_t==1 + T_lat==1, asserting tuple(out.shape) == (1, 3, 2, 16, 16) and per-sample fingerprint preservation (out[0,0,0,0,0]=1.0, out[0,0,1,0,0]=2.0).
There was a problem hiding this comment.
Pull request overview
Copilot reviewed 2 out of 2 changed files in this pull request and generated no new comments.
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
pollockjj
added a commit
that referenced
this pull request
May 5, 2026
…n contract (closes pollockjj/mydevelopment#192) (#46) * [Issue 101][eval-and-fix] Fix VideoAutoencoderKLWrapper.forward return contract (closes pollockjj/mydevelopment#192) VideoAutoencoderKLWrapper.forward() called self.decode(z).sample, but VideoAutoencoderKLWrapper.decode() returns a plain torch.Tensor (post-Comfy-Org#189 / PR #34 normalization). Direct wrapper invocation therefore raised AttributeError on the tensor return. Drop the .sample dereference; use the tensor return directly. forward() now returns the (x_out, z, p) triple as documented in the wrapper's encode/decode contract. Sister to mydevelopment#190 (PR #44) which fixed the parent class VideoAutoencoderKL.forward for the same bug class flagged on Comfy-Org#11294 by CodeRabbit (thread r2959796348, "Also applies to: 2083-2087" trailer). Adds tests-unit/comfy_test/seedvr_vae_wrapper_forward_test.py — four CPU-only regression tests that build a wrapper standin via __new__ + nn.Module.__init__, register a single dummy parameter so the wrapper's encode dtype lookup resolves, set original_image_video / img_dims / tiled_args so the wrapper.decode guards pass, patch the parent VideoAutoencoderKL.encode / decode_ plus the module-level lab_color_transfer with fingerprint-tagged stubs, then call wrapper.forward(x) end-to-end and assert (1) the return is a 3-tuple of tensors with no AttributeError raised, (2) x_out has exact expected shape, dtype, and value-fingerprint under torch.equal, (3) z matches the encode-side posterior squeezed on dim 2 under torch.equal, and (4) inspect.getsource(wrapper.forward) contains no ".sample" string. * [Issue 101][eval-and-fix] Tighten VideoAutoencoderKLWrapper.forward regression test to two-test contract (pollockjj/mydevelopment#192) The regression suite for VideoAutoencoderKLWrapper.forward now defines exactly the two contract-named tests: - test_wrapper_forward_returns_tensor_triple monkeypatches VideoAutoencoderKLWrapper.encode and VideoAutoencoderKLWrapper.decode directly on the class, builds the wrapper standin via __new__ + nn.Module.__init__, sets original_image_video to a 5-D tensor and img_dims to a 2-tuple, invokes wrapper.forward(x), and asserts the full return-contract: 3-tuple, three torch.Tensor types, binary shape equality (x_out.shape == decode_out.shape, z.shape == posterior.squeeze(2).shape), tensor equality (torch.equal(x_out, decode_out), torch.equal(z, posterior.squeeze(2))), and identity (p is posterior). - test_wrapper_forward_source_has_no_sample_access asserts ".sample" not in inspect.getsource(VideoAutoencoderKLWrapper.forward) so the failing pre-fix body raises an explicit assertion failure on the literal forbidden token. Stubbing on the wrapper class (not the parent) bypasses the parent encode/decode_ entirely, removes the lab_color_transfer monkey-patch, and makes the AttributeError pre-fix path surface directly on self.decode(z).sample because the stubbed decode returns a plain tensor.
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
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.
Plan: Keep batch/time axes distinct for single-frame batched decode
Overview
The SeedVR2 native VAE wrapper
comfy.ldm.seedvr.vae.VideoAutoencoderKLWrapper.decodecorrupts batched image-mode decode output whenB > 1andT == 1. The decoder's 5D output[B, C, 1, H, W]is flattened bysuper().decode_(latent).squeeze(2)to[B, C, H, W]; the existingif x.ndim == 4: x = x.unsqueeze(0)block then makes the tensor[1, B, C, H, W], after which thex.size(1) == 1gate routes batched cases through the wrongrearrangeexpression that re-interprets the batch axis as channels and the channel axis as time. The result forB=2, T=1is[1, 2, 2, H, W]instead of[1, 3, 2, H, W]; forB=4, T=1it is[1, 4, 3, H, W]instead of[1, 3, 4, H, W]. The latent-side prep block immediately above (b, tc, h, w = z.shape; latent = z.view(b, 16, -1, h, w); ...) was already corrected upstream by commitaf6c5d6and is OUT OF SCOPE; this plan asserts it byte-identical to base as a stop-condition guard. The plan first records the pre-fix shape signature onpollockjj/ComfyUI:issue_101HEAD as a baseline decision packet, then lands the output-side fix onpollockjj/ComfyUI:issue_189together with a five-case regression test (B ∈ {1, 2, 4} × T_orig ∈ {1, 3, 5}) and a stacked-vs-individual per-sample-ordering test, re-runs the same probe to record the corrected shapes, applies a hygiene gate (ruff + production-code-shape independence), and emits a post-fix provenance decision packet. PR creation fromissue_189intoissue_101is delegated to the/prskill after slice completion.Diagnosis Summary
pollockjj/ComfyUI:issue_101HEAD4e8836ed, callingVideoAutoencoderKLWrapper.decode(z)withz.shape == (B, 16, H, W)(image-mode latent, T_orig=1) forB > 1returns a tensor whose final shape contains the batch countBin the channel axis and the channel countC(3 for RGB) in the time axis. Live probe (Phase 0):B=2, T_orig=1→ output shape(1, 2, 2, 16, 16)(correct shape would be(1, 3, 2, 16, 16));B=4, T_orig=1→ output shape(1, 4, 3, 16, 16)(correct shape would be(1, 3, 4, 16, 16)). The per-batch fingerprint values that the probe injects via the mockeddecode_(float(b + 1)for batch indexb) appear in the size-2 (resp. size-4) axis instead of the size-2 (resp. size-4) time axis they belong in./home/johnj/dev_cuda_1/ComfyUIon issue_101 HEAD with the ComfyUI venv active: importcomfy.ldm.seedvr.vae, monkey-patchVideoAutoencoderKL.decode_to a fingerprint-tagged stub returning[B, 3, T_dec, H_in*8, W_in*8]filled withfloat(b + 1)per batch index, monkey-patchlab_color_transferto a passthrough, construct a wrapper instance viaVideoAutoencoderKLWrapper.__new__(VideoAutoencoderKLWrapper)+nn.Module.__init__(instance)withtiled_args = {"enable_tiling": False},original_image_video = torch.zeros(B, 3, T_orig, 16, 16),img_dims = (16, 16), then callwrapper.decode(torch.zeros(B, 16*T_orig, 2, 2)). For(B=2, T_orig=1)the output shape is(1, 2, 2, 16, 16); for(B=4, T_orig=1)the output shape is(1, 4, 3, 16, 16). Live-probed Phase 0; bug reproduces deterministically.decode_block inVideoAutoencoderKLWrapper.decodeapplies.squeeze(2)to the 5D[B, C, T_dec, H, W]decoder output. WhenT_dec == 1the tensor collapses to[B, C, H, W](4D). The next gateif x.ndim == 4: x = x.unsqueeze(0)adds a new outer axis at position 0, producing[1, B, C, H, W]. The subsequent gateif x.size(1) == 1: exp = "b t c h w -> (b t) c h w" else: exp = "b c t h w -> (b t) c h w"then mis-interprets the axes forB > 1(becausex.size(1) == B != 1), routing throughb c t h w -> (b t) c h wwhich treats the new outer 1 asb, the batch countBasc, and the channel countCast. The final reshape to[1, C, B*T_orig, H, W]therefore produces[1, B, C, H, W](axes swapped). TheB == 1, T_orig == 1path accidentally works becausex.size(1) == 1matches theb t c h w -> (b t) c h wexpression which collapses the leading two1s to a single1. TheT_orig > 1paths work because.squeeze(2)is a no-op whenT_dec > 1and the tensor stays 5D, bypassing the buggyunsqueeze(0)branch entirely.B > 1AND (b) the decoder's temporal outputT_dec == 1(image-mode reconstruction). TheT_orig > 1paths and theB == 1, T_orig == 1path are unaffected — they have their own complete coverage in the existing tile/non-tile decode paths and are exercised by the existing test surface (tests-unit/comfy_test/test_seedvr_vae_tiled_args_no_mutate.pyfor tiled-args invariant; the liveB=1, T=5andB=2, T=3probe results show the working shapes).decode_invariant that the tensor remains 5D[B, C, T_dec, H, W](or equivalently[B, C, T_out, H, W]after time-dim trimming) through the entire post-processing pipeline regardless ofT_dec, so that the per-sample axis stays at position 0 and the channel axis stays at position 1 before the finalb t c h w -> b c t h wreshape. The plan does not prescribe the exact source-line edit (the slicer chooses); the plan asserts the observable post-fix behavior (output shape(1, 3, B*T_orig, H, W)for every cell; per-sample fingerprint at the post-fix-correct batch position) and the structural stop-condition guard (latent-side prep block lines unchanged from base).issue_101HEAD (no fix applied) and records per-(B, T_orig)-cell shape, dtype, and per-batch fingerprint axis value into a JSON named-key artifact, capturing the buggy shapes for theB > 1, T_orig == 1cells. Slice 2 lands the source fix onpollockjj/ComfyUI:issue_189, adds the regression test module, re-runs the same probe to record the corrected shapes, and emits a before/after delta artifact whoseshape_mismatch_countnamed key drops from≥ 2to0. Slice 3 runs ruff and a no-detuning grep over the touched files. Slice 4 binds the issue branch HEADs, canonical-reference URLs, and verbatim invocation commands into an architect-readable provenance markdown.Affected Repositories
pollockjj/mydevelopment/home/johnj/dev_cuda_1/mydevelopmentmainissue_189pollockjj/ComfyUI/home/johnj/dev_cuda_1/ComfyUIissue_101issue_189Research and Methodology
Plan Foundations Comment URL: https://github.com/pollockjj/mydevelopment/issues/189#issuecomment-4368270401
Detected Scope:
none— single-block shape-handling fix inVideoAutoencoderKLWrapper.decodewhere the existing workingB==1, T==1,B==1, T>1, andB>1, T>1paths ARE the spec; the patch aligns the brokenB>1, T==1path to that same spec without introducing any new metric or methodology.The linked comment is the load-bearing artifact for every measurement, equivalence rule, canonical-protocol, tool version, and pipeline-stage citation in this plan. Every AC below that names a metric, threshold, equivalence rule, canonical protocol, tool version, or pipeline stage traces by quote or URL to an entry in that comment's
## Research and Methodologysection. qa-plan verifies the comment exists and contains the required subsections for the detected scope.Tools, Pipeline, and Measurements
Plan Foundations Comment URL: https://github.com/pollockjj/mydevelopment/issues/189#issuecomment-4368270401
The linked comment's
## Tools, Pipeline, and Measurementssection enumerates: (1) Existing Tooling Inventory withREUSE/KNOWN-GOOD-REF/NEW/WAIVEDstatus per tool; (2) Pipeline stages from input to output artifact; (3) Measurements table with tool+version, invocation, score range, tolerance; (4) Single-probe-before-sweep / boundary-bracketing requirements (bothN/Afor this plan and explicitly justified). Every AC below that names a tool, version, pipeline stage, or measurement traces by quote or URL to an entry in that section.Ground Truth Probes
For every literal API surface that any AC references — attribute, method, function signature, parameter list, CLI flag, file path, env var, magic constant — captured live on the issue's base branch
pollockjj/ComfyUI:issue_101(HEAD4e8836ed) on 2026-05-04 from/home/johnj/dev_cuda_1/ComfyUI. Plan-introduced literals (probe JSON keys, test names, paths created by this plan) are anchored separately in## Created Surface Contract, not here.cd /home/johnj/dev_cuda_1/ComfyUI && python -c "from comfy.cli_args import args; args.cpu=True; import comfy.ldm.seedvr.vae as v, inspect; print(inspect.signature(v.VideoAutoencoderKLWrapper.decode)); print(inspect.signature(v.VideoAutoencoderKL.decode_)); print([c.__name__ for c in v.VideoAutoencoderKLWrapper.__mro__])"(self, z)/(self, z: torch.Tensor, return_dict: bool = True)/['VideoAutoencoderKLWrapper', 'VideoAutoencoderKL', 'Module', 'object']cd /home/johnj/dev_cuda_1/ComfyUI && python -c "from comfy.cli_args import args; args.cpu=True; import comfy.ldm.seedvr.vae as v; print(v.lab_color_transfer); print(v.rearrange); print(v.tiled_vae)"<function lab_color_transfer at 0x...>/<function rearrange at 0x...>/<function tiled_vae at 0x...>cd /home/johnj/dev_cuda_1/ComfyUI && python -c "from comfy.cli_args import args; print(type(args.cpu).__name__, args.cpu); args.cpu=True; print(args.cpu)"bool False/Truecd /home/johnj/dev_cuda_1/ComfyUI && python -c "from comfy.cli_args import args; args.cpu=True; import comfy.ldm.seedvr.vae as v, torch.nn as nn, torch; w=v.VideoAutoencoderKLWrapper.__new__(v.VideoAutoencoderKLWrapper); nn.Module.__init__(w); print(type(w).__name__); print(hasattr(w, 'tiled_args')); print(hasattr(w, 'original_image_video')); print(hasattr(w, 'img_dims'))"VideoAutoencoderKLWrapper/False/False/False(these attributes are populated in the wrapper's__init__; the test sets them manually after__new__+nn.Module.__init__)cd /home/johnj/dev_cuda_1/ComfyUI && pythonfollowed by the inline reproduction script described in §Diagnosis Summary / Reproduction. Live raw outputs:(B=1, T_orig=1) → shape (1, 3, 1, 16, 16);(B=1, T_orig=5) → shape (1, 3, 5, 16, 16);(B=2, T_orig=1) → shape (1, 2, 2, 16, 16);(B=4, T_orig=1) → shape (1, 4, 3, 16, 16);(B=2, T_orig=3) → shape (1, 3, 6, 16, 16).pollockjj/ComfyUI:issue_101HEAD4e8836ed). The cells(B=2,T_orig=1)and(B=4,T_orig=1)exhibit the bug: the channel position 1 holds the batch count instead of3, and the time position 2 holds the RGB count instead of the batch count.git -C /home/johnj/dev_cuda_1/ComfyUI rev-parse origin/issue_101; git -C /home/johnj/dev_cuda_1/ComfyUI ls-tree origin/issue_101 -- comfy/ldm/seedvr/vae.py4e8836ed0a53467e9c433d58320c4992c9c34d2d/100644 blob b3f0f5d719862ae119ebc037ab104bca7785ce71 comfy/ldm/seedvr/vae.pyls /home/johnj/dev_cuda_1/ComfyUI/tests-unit/comfy_test/folder_path_test.py/model_detection_test.py/__pycache__/test_seedvr_groupnorm_limit.py/test_seedvr_rope_delegation.py/test_seedvr_vae_tiled_args_no_mutate.pygrep -nE 'args\.cpu = True|torch\.cuda\.is_available|patch\.object' /home/johnj/dev_cuda_1/ComfyUI/tests-unit/comfy_test/test_seedvr_groupnorm_limit.py41:from unittest.mock import patch/48:if not torch.cuda.is_available():/49: cli_args.cpu = True/104: with patch.object(vae_mod.F, "group_norm", side_effect=_group_norm_spy):/157: with patch.object(vae_mod.F, "group_norm", side_effect=_group_norm_spy):cd /home/johnj/dev_cuda_1/ComfyUI && .venv/bin/pip list 2>/dev/null | grep -E "^(torch|einops)\\b"; .venv/bin/python -m pytest --version 2>&1 | head -1einops 0.8.2/torch 2.11.0+cu130/pytest 9.0.3cat /home/johnj/dev_cuda_1/ComfyUI/pytest.ini[pytest]/markers =/inference: mark as inference test (deselect with '-m "not inference"')/execution: mark as execution test (deselect with '-m "not execution"')/testpaths =/tests/tests-unit/addopts = -s/pythonpath = .git -C /home/johnj/dev_cuda_1/mydevelopment rev-parse HEAD; git -C /home/johnj/dev_cuda_1/mydevelopment rev-parse --abbrev-ref HEAD1e8218e23d2616060e41bf710bc196e6ea81fd4a/issue_189sed -n '2245,2295p' /home/johnj/dev_cuda_1/ComfyUI/comfy/ldm/seedvr/vae.py(issue_101 HEAD source)def decode(self, z):body starting at line 2245 — latent-side prep block at 2245-2253 (b, tc, h, w = z.shape/latent = z.view(b, 16, -1, h, w)/scale = 0.9152/shift = 0/latent = latent / scale + shift/ blank /if latent.ndim == 4:/latent = latent.unsqueeze(2)); output-side block at 2255-2294 startingself.device = latent.deviceand ending withreturn xafter theb t c h w -> b c t h wrearrange and the even-dims trim.ruffcd /home/johnj/dev_cuda_1/ComfyUI && python -m ruff --version; which ruff; git -C /home/johnj/dev_cuda_1/ComfyUI ls-tree origin/issue_101 -- .github/workflows/ruff.ymlruff 0.15.9//home/johnj/.local/bin/ruff/100644 blob b24d86a6ba55b1f7b90539d6267a72afa5a7b73c .github/workflows/ruff.ymlIf a planned literal cannot be probed (slot offline, source absent, tool missing) the AC is rewritten to remove the unverified literal or the slice is rescoped. No probe → no AC.
Created Surface Contract
These literals are introduced by this plan and therefore are NOT subject to Phase 0.5 ground-truth probing. They become probable after the slice that creates them lands.
Probe-script API (Slice 1 introduces; Slice 2 reuses):
--out PATHCLI flag (writes UTF-8 JSON toPATH);--bt-cells PATH_OR_LITERALCLI flag (semicolon-separated list ofB,T_originteger pairs, e.g."1,1;1,5;2,1;4,1;2,3"; the probe iterates each cell and records its result);--label LABELCLI flag (string written into the JSON'slabelkey, e.g."baseline"or"post_fix"); script exit code0on success; non-zero on any cell raising an unhandled exception.Probe JSON named keys (Slice 1 introduces baseline; Slice 2 emits post-fix): Top-level:
label(string),mydevelopment_head_sha(string),comfyui_head_sha(string),comfyui_branch(string),vae_blob_sha(string fromgit ls-tree HEAD -- comfy/ldm/seedvr/vae.py),bt_cells(list of{B, T_orig, T_dec, out_shape, out_dtype, channel_axis_value, time_axis_value, fingerprint_at_correct_batch_position}objects). Per-cell named keys:B(int),T_orig(int),T_dec(int),out_shape(list of int),out_dtype(string e.g."torch.float32"),channel_axis_value(int —out_shape[1]),time_axis_value(int —out_shape[2]),fingerprint_at_correct_batch_position(list of float, lengthB*T_orig, value at indexiisout[0, 0, i, 0, 0].item(); for the buggy cells this list is corrupted; for the correct cells it equals[float(b//T_orig + 1) for b in range(B*T_orig)]interleaved byT_orig).Comparison JSON named keys (Slice 2 introduces):
prior_artifact_path(string),new_artifact_path(string),prior_label(string),new_label(string),shape_mismatch_count_prior(int — number of cells whoseout_shapediffers from the per-cell expected(1, 3, B*T_orig, 16, 16)),shape_mismatch_count_new(int),per_cell_delta(list of{B, T_orig, prior_out_shape, new_out_shape, expected_out_shape, prior_pass, new_pass}objects).Test names (Slice 2 introduces in
tests-unit/comfy_test/test_seedvr_vae_decode_batch_axes.py):test_decode_b1_t1_shape_and_ordering_correct,test_decode_b1_t5_video_shape_unchanged,test_decode_b2_t1_fixes_batch_time_axes,test_decode_b4_t1_fixes_batch_time_axes,test_decode_b2_t3_multi_frame_batch_unchanged,test_decode_b2_t1_stacked_equals_individual_per_sample_ordering. All six functions live in moduletests_unit.comfy_test.test_seedvr_vae_decode_batch_axes.File paths (Slices 1–4 introduce):
pollockjj/mydevelopment#issue_189:github_issues/189/probe_seedvr_decode_batch_axes.py(S1)github_issues/189/slice1/decode_batch_axes_baseline.json(S1)github_issues/189/slice1/probe_run.log(S1)github_issues/189/slice1/probe_argparse_flags.log(S1 — grep evidence proving the probe script defines exactly the three CLI flags and zero extras)github_issues/189/slice1/baseline_json_keys.log(S1 —python -m json.toolparse + sorted-key dump proving the baseline JSON's top-level and per-cell named-key sets match the contract verbatim)github_issues/189/slice1/provenance.md(S1)github_issues/189/slice2/decode_batch_axes_post_fix.json(S2)github_issues/189/slice2/before_after_comparison.json(S2)github_issues/189/slice2/probe_run.log(S2)github_issues/189/slice2/pytest_decode_batch_axes.log(S2)github_issues/189/slice2/pytest_seedvr_regression_guard.log(S2)github_issues/189/slice2/test_module_internals.log(S2 — grep evidence proving the regression test module's six function definitions and the per-test assertion shape constants match the AC contract verbatim)github_issues/189/slice2/vae_diff.patch(S2)github_issues/189/slice2/latent_side_unchanged.log(S2)github_issues/189/slice3/ruff.log(S3)github_issues/189/slice3/no_detuning.log(S3)github_issues/189/slice4/provenance.md(S4)pollockjj/ComfyUI#issue_189:comfy/ldm/seedvr/vae.py(modified — output-side block ofVideoAutoencoderKLWrapper.decodeonly; latent-side prep block lines 2245-2253 byte-identical to base) (S2)tests-unit/comfy_test/test_seedvr_vae_decode_batch_axes.py(S2 introduces)Latent-side guard token set (Slice 2 AC-1 protects these tokens; the diff MUST NOT add or remove lines containing these tokens):
b, tc, h, w = z.shape;latent = z.view(b, 16, -1, h, w);scale = 0.9152;shift = 0;latent = latent / scale + shift;if latent.ndim == 4:;latent = latent.unsqueeze(2). These seven tokens appear consecutively as the body of the latent-side prep block onpollockjj/ComfyUI:issue_101HEAD per GTP-12.Asset Readiness
pollockjj/ComfyUI:issue_101HEAD4e8836ed/home/johnj/dev_cuda_1/ComfyUIgit rev-parse origin/issue_101returns4e8836ed0a53467e9c433d58320c4992c9c34d2dcomfy/ldm/seedvr/vae.py(VideoAutoencoderKLWrapper.decodebody at lines 2245-2295; latent-side prep at 2245-2253; output-side block at 2255-2294)/home/johnj/dev_cuda_1/ComfyUI/comfy/ldm/seedvr/vae.py(blobb3f0f5d7)issue_101HEADcomfy.ldm.seedvr.vae.VideoAutoencoderKL.decode_parent-class methodissue_101HEADcomfy.ldm.seedvr.vae.lab_color_transfermodule-level functionissue_101HEADcomfy.ldm.seedvr.vae.rearrange(re-export ofeinops.rearrange)issue_101HEADtests-unit/comfy_test/and existing seedvr tests/home/johnj/dev_cuda_1/ComfyUI/tests-unit/comfy_test/issue_101HEADtests-unit/comfy_test/test_seedvr_groupnorm_limit.py(args.cpu = Trueblock beforecomfy.ldm.*import;unittest.mock.patch.objectusage)/home/johnj/dev_cuda_1/ComfyUI/tests-unit/comfy_test/test_seedvr_groupnorm_limit.pyissue_101HEADtorch==2.11.0+cu130,einops==0.8.2,pytest==9.0.3)/home/johnj/dev_cuda_1/ComfyUI/.venvruff0.15.9 binary/home/johnj/.local/bin/ruff(also resolvable viapython -m ruffagainst the deliverable venv)python -m ruff --versionfrom/home/johnj/dev_cuda_1/ComfyUI; ComfyUI CI invokes the same binary atpollockjj/ComfyUI:issue_101HEAD.github/workflows/ruff.yml(blobb24d86a6ba55b1f7b90539d6267a72afa5a7b73c, line 24run: ruff check .)comfy.cli_args.args.cpuflag/home/johnj/dev_cuda_1/ComfyUI/comfy/cli_args.pyissue_101HEADpytest.inidiscovery config (pythonpath = .,testpaths = tests tests-unit,addopts = -s)/home/johnj/dev_cuda_1/ComfyUI/pytest.iniissue_101HEADpollockjj/mydevelopment:mainparent context HEAD1e8218e2/home/johnj/dev_cuda_1/mydevelopmentIC_kwDOR2e1q88AAAABBF6EQQtdd-agent[bot]post 2026-05-04T04:37:08Z, edited 2026-05-04T05:28:59ZNo external assets (model weights, video/image datasets, OS packages) are required. The plan is CPU-only by construction.
Slices
Slice 1: Pre-fix Baseline Reproduction Probe
Kind: decision-packet
Objective: Produce the baseline reproduction decision packet recording the buggy output shape and per-batch fingerprint position for every
(B, T_orig)cell in the bug taxonomy onpollockjj/ComfyUI:issue_101HEAD4e8836edbefore any production fix is applied.Acceptance Criteria
AC-1: File
github_issues/189/probe_seedvr_decode_batch_axes.pyis committed onpollockjj/mydevelopment:issue_189and is a Python module that exposes a--out PATHCLI flag, a--bt-cells STRCLI flag (semicolon-separatedB,T_originteger pairs), a--label STRCLI flag, exits 0 on success, and writes a UTF-8 JSON file toPATHcontaining exactly the named-key set{label, mydevelopment_head_sha, comfyui_head_sha, comfyui_branch, vae_blob_sha, bt_cells}with eachbt_cellsentry containing exactly the per-cell named-key set{B, T_orig, T_dec, out_shape, out_dtype, channel_axis_value, time_axis_value, fingerprint_at_correct_batch_position}— verified by (a) committed artifactgithub_issues/189/slice1/probe_argparse_flags.logcapturing the stdout ofgrep -nE "add_argument\(['\"](--out\|--bt-cells\|--label)['\"]" /home/johnj/dev_cuda_1/mydevelopment/github_issues/189/probe_seedvr_decode_batch_axes.pyreturning exactly three matching lines (one per flag) and zero extraadd_argument(lines outside that set, with the log's final line recordingARGPARSE_FLAGS_EXIT: 0; (b) committed artifactgithub_issues/189/slice1/baseline_json_keys.logcapturing the stdout ofpython3 -c 'import json,sys; d=json.load(open(sys.argv[1])); print("TOP_KEYS:", sorted(d.keys())); print("CELL_KEYS:", sorted(d["bt_cells"][0].keys())); print("CELL_COUNT:", len(d["bt_cells"]))' /home/johnj/dev_cuda_1/mydevelopment/github_issues/189/slice1/decode_batch_axes_baseline.jsonwhose three printed lines are verbatimTOP_KEYS: ['bt_cells', 'comfyui_branch', 'comfyui_head_sha', 'label', 'mydevelopment_head_sha', 'vae_blob_sha'],CELL_KEYS: ['B', 'T_dec', 'T_orig', 'channel_axis_value', 'fingerprint_at_correct_batch_position', 'out_dtype', 'out_shape', 'time_axis_value'], andCELL_COUNT: 5, with the log's final line recordingJSON_KEYS_EXIT: 0; (c) committed artifactgithub_issues/189/slice1/probe_run.logcapturing the stdout+stderr of the AC-2 probe invocation including the verbatim exit-status linePROBE_EXIT: 0; AND (d) the JSON artifact existing at the path declared in AC-2.(B, T_orig)cell in the bug taxonomy onpollockjj/ComfyUI:issue_101HEAD4e8836edbefore any production fix is applied."AC-2: Probe is invoked as
cd /home/johnj/dev_cuda_1/ComfyUI && python3 /home/johnj/dev_cuda_1/mydevelopment/github_issues/189/probe_seedvr_decode_batch_axes.py --out /home/johnj/dev_cuda_1/mydevelopment/github_issues/189/slice1/decode_batch_axes_baseline.json --bt-cells "1,1;1,5;2,1;4,1;2,3" --label baselineagainst the issue branch's working tree at HEAD (which is byte-identical toorigin/issue_101HEAD4e8836edbecause no source fix has been applied in Slice 1) and the resultinggithub_issues/189/slice1/decode_batch_axes_baseline.jsonis committed onpollockjj/mydevelopment:issue_189with the per-cellout_shapenamed-key values: cellB=1, T_orig=1→[1, 3, 1, 16, 16]; cellB=1, T_orig=5→[1, 3, 5, 16, 16]; cellB=2, T_orig=1→[1, 2, 2, 16, 16]; cellB=4, T_orig=1→[1, 4, 3, 16, 16]; cellB=2, T_orig=3→[1, 3, 6, 16, 16]. Thelabelvalue is"baseline". Thecomfyui_head_shavalue is"4e8836ed0a53467e9c433d58320c4992c9c34d2d". Thevae_blob_shavalue is"b3f0f5d719862ae119ebc037ab104bca7785ce71". Verified by committed artifactgithub_issues/189/slice1/decode_batch_axes_baseline.jsonparsed and key-checked, plus committed artifactgithub_issues/189/slice1/probe_run.logcapturing the probe's stdout/stderr.(B, T_orig)cell in the bug taxonomy onpollockjj/ComfyUI:issue_101HEAD4e8836edbefore any production fix is applied."AC-3: File
github_issues/189/slice1/provenance.mdis committed onpollockjj/mydevelopment:issue_189containing exactly the named-line keysdecision_summary:,recommended_action:,mydevelopment_head_sha:,comfyui_head_sha:,comfyui_branch:,vae_blob_sha:,coderabbit_url:,upstream_pr_url:,parent_findings_comment_url:,probe_command:,probe_output_path:,bug_cells_observed:, withdecision_summaryvaluebaseline_b_gt_1_t_eq_1_decode_shape_corrupted,recommended_actionvalueapply_output_side_fix_on_pollockjj/ComfyUI_issue_189,comfyui_head_shavalue4e8836ed0a53467e9c433d58320c4992c9c34d2d,comfyui_branchvalueissue_101,vae_blob_shavalueb3f0f5d719862ae119ebc037ab104bca7785ce71,coderabbit_urlvaluehttps://github.com/Comfy-Org/ComfyUI/pull/11294#discussion_r2959796352,upstream_pr_urlvaluehttps://github.com/Comfy-Org/ComfyUI/pull/11294,parent_findings_comment_urlvaluehttps://github.com/pollockjj/mydevelopment/issues/101#issuecomment-4305426643,probe_commandvalue matching the AC-2 command verbatim,probe_output_pathvaluegithub_issues/189/slice1/decode_batch_axes_baseline.json,bug_cells_observedvalueB=2,T_orig=1 -> shape (1,2,2,16,16); B=4,T_orig=1 -> shape (1,4,3,16,16)— verified by committed artifactgithub_issues/189/slice1/provenance.mdandgrep '^<key>: 'parseability against the listed key set.(B, T_orig)cell in the bug taxonomy onpollockjj/ComfyUI:issue_101HEAD4e8836edbefore any production fix is applied."Slice 2: Output-side Shape Fix, Regression Test Module, Post-Fix Probe and Comparison
Kind: implementation
Objective: Prove that modifying the output-side block of
VideoAutoencoderKLWrapper.decodeincomfy/ldm/seedvr/vae.pymakes batch and time axes distinct for every(B, T_orig)cell includingB>1, T_orig==1while leaving the latent-side prep block byte-identical to base, by landing the source change onpollockjj/ComfyUI:issue_189, adding a six-test regression module that fails pre-fix on theB>1, T_orig==1cells, re-running the same probe to record corrected shapes, and emitting a before/after delta artifact whoseshape_mismatch_count_newnamed-key is0whileshape_mismatch_count_prioris2.Acceptance Criteria
comfy/ldm/seedvr/vae.pyonpollockjj/ComfyUI:issue_189HEAD has a non-empty diff againstorigin/issue_101restricted to the output-side block ofVideoAutoencoderKLWrapper.decode(the lines starting atself.device = latent.devicethrough the finalreturn xafter the even-dims trim, per GTP-12); the latent-side prep block tokens enumerated in## Created Surface Contract(the seven-token setb, tc, h, w = z.shape;latent = z.view(b, 16, -1, h, w);scale = 0.9152;shift = 0;latent = latent / scale + shift;if latent.ndim == 4:;latent = latent.unsqueeze(2)) appear in zero+lines and zero-lines of the diff — verified by committed artifactgithub_issues/189/slice2/vae_diff.patchcontaining the output ofgit -C /home/johnj/dev_cuda_1/ComfyUI diff origin/issue_101...HEAD -- comfy/ldm/seedvr/vae.py(non-zero size) AND committed artifactgithub_issues/189/slice2/latent_side_unchanged.logcontaining the output ofgrep -nE '^[+-][[:space:]]+(b, tc, h, w = z\.shape\|latent = z\.view\(b, 16, -1, h, w\)\|scale = 0\.9152\|shift = 0\|latent = latent / scale \+ shift\|if latent\.ndim == 4:\|latent = latent\.unsqueeze\(2\))$' /home/johnj/dev_cuda_1/mydevelopment/github_issues/189/slice2/vae_diff.patchreturning zero matching lines, with the log's final line recordingLATENT_SIDE_GUARD_EXIT: 0.VideoAutoencoderKLWrapper.decodeonissue_101HEAD appliessuper().decode_(latent).squeeze(2)followed byif x.ndim == 4: x = x.unsqueeze(0), producing[1, B, C, H, W]forB>1, T_dec==1per GTP-5; pre-fixgit diff origin/issue_101...HEAD -- comfy/ldm/seedvr/vae.pyis empty (no fix applied yet)vae_diff.patchis non-empty AND contains zero+/-lines matching the seven-token latent-side guard set; the modification is restricted to lines within the output-side block (per the GTP-12 line-range observation2255-2294)VideoAutoencoderKLWrapper.decodeincomfy/ldm/seedvr/vae.pymakes batch and time axes distinct for every(B, T_orig)cell includingB>1, T_orig==1while leaving the latent-side prep block byte-identical to base, by landing the source change onpollockjj/ComfyUI:issue_189, adding a six-test regression module that fails pre-fix on theB>1, T_orig==1cells, re-running the same probe to record corrected shapes, and emitting a before/after delta artifact whoseshape_mismatch_count_newnamed-key is0whileshape_mismatch_count_prioris2."AC-2: File
tests-unit/comfy_test/test_seedvr_vae_decode_batch_axes.pyonpollockjj/ComfyUI:issue_189HEAD defines exactly the six functionstest_decode_b1_t1_shape_and_ordering_correct,test_decode_b1_t5_video_shape_unchanged,test_decode_b2_t1_fixes_batch_time_axes,test_decode_b4_t1_fixes_batch_time_axes,test_decode_b2_t3_multi_frame_batch_unchanged, andtest_decode_b2_t1_stacked_equals_individual_per_sample_ordering. Each per-cell test (a) setscomfy.cli_args.args.cpu = Truebefore importing anycomfy.ldm.*symbol whentorch.cuda.is_available()is False (matching the pattern attests-unit/comfy_test/test_seedvr_groupnorm_limit.py:48-49per GTP-8), (b) constructs a wrapper instance viaVideoAutoencoderKLWrapper.__new__(VideoAutoencoderKLWrapper)+nn.Module.__init__(instance)withtiled_args = {"enable_tiling": False},original_image_video = torch.zeros(B, 3, T_orig, 16, 16),img_dims = (16, 16)per the__init__-bypass pattern in §Diagnosis Summary / Reproduction (per GTP-4), (c) patchesVideoAutoencoderKL.decode_with a fingerprint-tagged stub returning[B, 3, T_dec, 16, 16]filled withfloat(b + 1)per batch index, (d) patchescomfy.ldm.seedvr.vae.lab_color_transferwith a passthrough that returns its first argument, (e) callswrapper.decode(torch.zeros(B, 16*T_orig, 2, 2)), and (f) assertstuple(out.shape) == (1, 3, B*T_orig, 16, 16)for the per-cellBandT_orig. Thetest_decode_b2_t1_fixes_batch_time_axestest additionally assertsout[0, 0, 0, 0, 0].item() == 1.0ANDout[0, 0, 1, 0, 0].item() == 2.0(per-sample fingerprint at the correct batch position). Thetest_decode_b4_t1_fixes_batch_time_axestest additionally asserts[out[0, 0, b, 0, 0].item() for b in range(4)] == [1.0, 2.0, 3.0, 4.0]. Thetest_decode_b2_t1_stacked_equals_individual_per_sample_orderingtest invokeswrapper.decode(torch.zeros(2, 16, 2, 2))(stacked, withoriginal_image_videoset totorch.zeros(2, 3, 1, 16, 16)) producingout_stacked, then resets the wrapper'soriginal_image_videototorch.zeros(1, 3, 1, 16, 16)and invokeswrapper.decode(torch.zeros(1, 16, 2, 2))twice with the per-call fingerprint stub returningfloat(b + 1)for the stacked indicesb ∈ {0, 1}(the second call's stub pinned to2.0), then assertstorch.equal(out_stacked[0, :, 0, :, :], out_individual_0[0, :, 0, :, :])ANDtorch.equal(out_stacked[0, :, 1, :, :], out_individual_1[0, :, 0, :, :]). All six tests pass undercd /home/johnj/dev_cuda_1/ComfyUI && python -m pytest -q tests-unit/comfy_test/test_seedvr_vae_decode_batch_axes.pywith overall exit 0 and the summary line containing6 passed— verified by (a) committed artifactgithub_issues/189/slice2/test_module_internals.logcapturing seven greps over/home/johnj/dev_cuda_1/ComfyUI/tests-unit/comfy_test/test_seedvr_vae_decode_batch_axes.pyagainst the issue_189 HEAD source: (i)grep -cE '^def (test_decode_b1_t1_shape_and_ordering_correct\|test_decode_b1_t5_video_shape_unchanged\|test_decode_b2_t1_fixes_batch_time_axes\|test_decode_b4_t1_fixes_batch_time_axes\|test_decode_b2_t3_multi_frame_batch_unchanged\|test_decode_b2_t1_stacked_equals_individual_per_sample_ordering)\(' returning the literal6; (ii)grep -cE '^def test_'returning the literal6(proves no extra test functions); (iii)grep -nE 'tuple(out.shape) == (1, 3, 1, 16, 16)|tuple(out.shape) == (1, 3, 5, 16, 16)|tuple(out.shape) == (1, 3, 2, 16, 16)|tuple(out.shape) == (1, 3, 4, 16, 16)|tuple(out.shape) == (1, 3, 6, 16, 16)'returning at least five lines (one per per-cell shape assertion); (iv)grep -nE 'out[0, 0, 0, 0, 0].item() == 1.0|out[0, 0, 1, 0, 0].item() == 2.0'returning at least two matching lines (the b2_t1 fingerprint asserts); (v)grep -nE '[out[0, 0, b, 0, 0].item() for b in range(4)] == [1.0, 2.0, 3.0, 4.0]'returning at least one matching line (the b4_t1 fingerprint assert); (vi)grep -nE 'torch.equal(out_stacked[0, :, 0, :, :], out_individual_0[0, :, 0, :, :])|torch.equal(out_stacked[0, :, 1, :, :], out_individual_1[0, :, 0, :, :])'returning at least two matching lines (the stacked-vs-individual asserts); (vii)grep -nE 'args.cpu = True|VideoAutoencoderKLWrapper.new|nn.Module.init|patch.object(.*decode_|patch.object(.*lab_color_transfer'returning at least one match per fixture-pattern line (cpu flag,newbypass,nn.Module.initbypass,decode_patch,lab_color_transferpatch), with the log's final line recordingTEST_INTERNALS_EXIT: 0; **(b)** committed artifactgithub_issues/189/slice2/pytest_decode_batch_axes.logcapturing stdout+stderr ofcd /home/johnj/dev_cuda_1/ComfyUI && python -m pytest -q tests-unit/comfy_test/test_seedvr_vae_decode_batch_axes.pyincluding the6 passedsummary line and the verbatim final-linePYTEST_DECODE_BATCH_AXES_EXIT: 0`.test_decode_b2_t1_fixes_batch_time_axesandtest_decode_b4_t1_fixes_batch_time_axestests againstissue_101HEAD (no fix) causes both to FAIL withAssertionError: tuple(out.shape) == (1, 3, 2, 16, 16)(got(1, 2, 2, 16, 16)) andtuple(out.shape) == (1, 3, 4, 16, 16)(got(1, 4, 3, 16, 16)) respectively, per the live shapes observed in GTP-5; thetest_decode_b2_t1_stacked_equals_individual_per_sample_orderingtest fails on thetorch.equalassertion because the stacked tensor's batch axis has been swapped with channels; pytest summary contains at minimum3 failed6 passed; per-celltuple(out.shape)equals(1, 3, B*T_orig, 16, 16)for every cell; per-sample fingerprint atout[0, 0, b, 0, 0]equalsfloat(b + 1)for theB>1, T_orig=1cells;torch.equal(out_stacked[0, :, b, :, :], out_individual_b[0, :, 0, :, :])holds forb ∈ {0, 1}VideoAutoencoderKLWrapper.decodeincomfy/ldm/seedvr/vae.pymakes batch and time axes distinct for every(B, T_orig)cell includingB>1, T_orig==1while leaving the latent-side prep block byte-identical to base, by landing the source change onpollockjj/ComfyUI:issue_189, adding a six-test regression module that fails pre-fix on theB>1, T_orig==1cells, re-running the same probe to record corrected shapes, and emitting a before/after delta artifact whoseshape_mismatch_count_newnamed-key is0whileshape_mismatch_count_prioris2."AC-3: Existing SeedVR2 unit-test modules
tests-unit/comfy_test/test_seedvr_groupnorm_limit.py,tests-unit/comfy_test/test_seedvr_vae_tiled_args_no_mutate.py, andtests-unit/comfy_test/test_seedvr_rope_delegation.py(per GTP-7) all pass undercd /home/johnj/dev_cuda_1/ComfyUI && python -m pytest -q tests-unit/comfy_test/test_seedvr_groupnorm_limit.py tests-unit/comfy_test/test_seedvr_vae_tiled_args_no_mutate.py tests-unit/comfy_test/test_seedvr_rope_delegation.pyonpollockjj/ComfyUI:issue_189HEAD with overall exit 0 and the summary line containingpassed(nofailed, noerrors) — verified by committed artifactgithub_issues/189/slice2/pytest_seedvr_regression_guard.logcapturing stdout+stderr of that pytest invocation, with the log's penultimate or final line matching the regex^[0-9]+ passed( in [0-9.]+s)?$and zero occurrences of the substringfailedorerrorsoutside the verbatim summary line context.B>1, T_orig==1decode path, so they all pass onissue_101HEAD; this AC asserts they continue to pass post-fix (regression guard against the output-side block edit unintentionally breaking another path)passedcount and zerofailed/errors; the existing tests are unaffected by the output-side block change;code-behavior-equivalence(perequivalence_methods.md) holds for the existing-test surfaceVideoAutoencoderKLWrapper.decodeincomfy/ldm/seedvr/vae.pymakes batch and time axes distinct for every(B, T_orig)cell includingB>1, T_orig==1while leaving the latent-side prep block byte-identical to base, by landing the source change onpollockjj/ComfyUI:issue_189, adding a six-test regression module that fails pre-fix on theB>1, T_orig==1cells, re-running the same probe to record corrected shapes, and emitting a before/after delta artifact whoseshape_mismatch_count_newnamed-key is0whileshape_mismatch_count_prioris2."AC-4: Probe is invoked as
cd /home/johnj/dev_cuda_1/ComfyUI && python3 /home/johnj/dev_cuda_1/mydevelopment/github_issues/189/probe_seedvr_decode_batch_axes.py --out /home/johnj/dev_cuda_1/mydevelopment/github_issues/189/slice2/decode_batch_axes_post_fix.json --bt-cells "1,1;1,5;2,1;4,1;2,3" --label post_fixagainstpollockjj/ComfyUI:issue_189HEAD (post-fix from AC-1) and the resultinggithub_issues/189/slice2/decode_batch_axes_post_fix.jsonis committed onpollockjj/mydevelopment:issue_189with the per-cellout_shapenamed-key values: cellB=1, T_orig=1→[1, 3, 1, 16, 16]; cellB=1, T_orig=5→[1, 3, 5, 16, 16]; cellB=2, T_orig=1→[1, 3, 2, 16, 16]; cellB=4, T_orig=1→[1, 3, 4, 16, 16]; cellB=2, T_orig=3→[1, 3, 6, 16, 16]. Every cell'schannel_axis_valueequals3andtime_axis_valueequalsB*T_orig. Thelabelvalue is"post_fix". Thecomfyui_branchvalue is"issue_189". Thevae_blob_shavalue differs from the Slice 1 baselinevae_blob_shaofb3f0f5d719862ae119ebc037ab104bca7785ce71. Verified by committed artifactgithub_issues/189/slice2/decode_batch_axes_post_fix.jsonparsed and key-checked, plus committed artifactgithub_issues/189/slice2/probe_run.log.issue_101HEAD (Slice 1 AC-2 outcome) records cellB=2, T_orig=1out_shape=[1, 2, 2, 16, 16]and cellB=4, T_orig=1out_shape=[1, 4, 3, 16, 16]— both differ from the(1, 3, B*T_orig, 16, 16)post-fix expectation per the live observations in GTP-5out_shapeequals(1, 3, B*T_orig, 16, 16); every cell'schannel_axis_value == 3; every cell'stime_axis_value == B*T_orig; the workingB=1andB>1, T_orig>1cells remain bit-identical to baseline (out_shapeunchanged for cellsB=1, T_orig=1,B=1, T_orig=5,B=2, T_orig=3)VideoAutoencoderKLWrapper.decodeincomfy/ldm/seedvr/vae.pymakes batch and time axes distinct for every(B, T_orig)cell includingB>1, T_orig==1while leaving the latent-side prep block byte-identical to base, by landing the source change onpollockjj/ComfyUI:issue_189, adding a six-test regression module that fails pre-fix on theB>1, T_orig==1cells, re-running the same probe to record corrected shapes, and emitting a before/after delta artifact whoseshape_mismatch_count_newnamed-key is0whileshape_mismatch_count_prioris2."AC-5: File
github_issues/189/slice2/before_after_comparison.jsonis committed onpollockjj/mydevelopment:issue_189containing exactly the named-key set{prior_artifact_path, new_artifact_path, prior_label, new_label, shape_mismatch_count_prior, shape_mismatch_count_new, per_cell_delta}with valuesprior_artifact_path="github_issues/189/slice1/decode_batch_axes_baseline.json",new_artifact_path="github_issues/189/slice2/decode_batch_axes_post_fix.json",prior_label="baseline",new_label="post_fix",shape_mismatch_count_prior=2(cellsB=2, T_orig=1andB=4, T_orig=1),shape_mismatch_count_new=0, andper_cell_deltacontaining exactly five entries (one per cell) with each entry containing the named-key set{B, T_orig, prior_out_shape, new_out_shape, expected_out_shape, prior_pass, new_pass}and the per-cell values listed in AC-2 / AC-4 — verified by committed artifact at the named path parsed and key-checked.shape_mismatch_count_newcannot be0because no post-fix run has been performedshape_mismatch_count_prior == 2ANDshape_mismatch_count_new == 0AND everyper_cell_deltaentry hasnew_pass=true; the comparison JSON exists at the named path on the issue branchVideoAutoencoderKLWrapper.decodeincomfy/ldm/seedvr/vae.pymakes batch and time axes distinct for every(B, T_orig)cell includingB>1, T_orig==1while leaving the latent-side prep block byte-identical to base, by landing the source change onpollockjj/ComfyUI:issue_189, adding a six-test regression module that fails pre-fix on theB>1, T_orig==1cells, re-running the same probe to record corrected shapes, and emitting a before/after delta artifact whoseshape_mismatch_count_newnamed-key is0whileshape_mismatch_count_prioris2."Slice 3: Hygiene Validator Gate (Ruff Lint + Production-Code Shape Independence)
Kind: hygiene
Objective: Prove the Slice 2 commits leave
comfy/ldm/seedvr/vae.pyandtests-unit/comfy_test/test_seedvr_vae_decode_batch_axes.pyruff-clean and free of any observability detuning (no addedprint/logging.debug/ sentinel comments) that could land as production-code noise alongside the output-side fix.Acceptance Criteria
AC-1:
cd /home/johnj/dev_cuda_1/ComfyUI && python -m ruff check comfy/ldm/seedvr/vae.py tests-unit/comfy_test/test_seedvr_vae_decode_batch_axes.pyexits 0 againstpollockjj/ComfyUI:issue_189HEAD — verified by committed validator-output artifactgithub_issues/189/slice3/ruff.logcapturing stdout+stderr of that invocation and ending with theAll checks passed!line, with shell exit code recorded as0in the log's final lineRUFF_EXIT: 0.comfy/ldm/seedvr/vae.pyandtests-unit/comfy_test/test_seedvr_vae_decode_batch_axes.pyruff-clean and free of any observability detuning (no addedprint/logging.debug/ sentinel comments) that could land as production-code noise alongside the output-side fix."AC-2: Slice 2 introduced zero observability detuning (
print,logging.debug,logging.info, sentinel comments, debug markers) incomfy/ldm/seedvr/vae.pyand intests-unit/comfy_test/test_seedvr_vae_decode_batch_axes.py— verified by committed validator-output artifactgithub_issues/189/slice3/no_detuning.logcontaining the combined output of two greps: (a)grep -nE '^[+][[:space:]]*(print\(\|logging\.debug\|logging\.info\|# TODO\s*(?:DEBUG\|MARKER\|SENTINEL))' /home/johnj/dev_cuda_1/mydevelopment/github_issues/189/slice2/vae_diff.patchreturning no matches, AND (b)grep -nE '\bprint\(\|logging\.debug\|logging\.info\|# TODO\s*(?:DEBUG\|MARKER\|SENTINEL)' /home/johnj/dev_cuda_1/ComfyUI/tests-unit/comfy_test/test_seedvr_vae_decode_batch_axes.pyreturning no matches, with the log's final line recordingNO_DETUNING_EXIT: 0.comfy/ldm/seedvr/vae.pyandtests-unit/comfy_test/test_seedvr_vae_decode_batch_axes.pyruff-clean and free of any observability detuning (no addedprint/logging.debug/ sentinel comments) that could land as production-code noise alongside the output-side fix."Slice 4: Post-Fix Provenance Decision Packet
Kind: decision-packet
Objective: Produce the post-fix provenance decision packet binding the issue branch HEADs, the canonical-reference URLs, and the verbatim invocation commands used by Slices 2 and 3 into a single architect-readable record.
Acceptance Criteria
github_issues/189/slice4/provenance.mdis committed onpollockjj/mydevelopment:issue_189containing exactly the named-line keysdecision_summary:,recommended_action:,mydevelopment_head_sha:,comfyui_head_sha:,comfyui_branch:,comfyui_base_sha:,comfyui_base_branch:,coderabbit_url:,upstream_pr_url:,parent_findings_comment_url:,probe_command:,probe_baseline_path:,probe_post_fix_path:,comparison_path:,regression_test_command:,seedvr_regression_guard_command:,ruff_command:, withdecision_summaryvaluepost_fix_b_gt_1_t_eq_1_decode_axes_distinct_restored,recommended_actionvalueinvoke_/pr_from_issue_189_to_issue_101,comfyui_base_shavalue4e8836ed0a53467e9c433d58320c4992c9c34d2d,comfyui_base_branchvalueissue_101,comfyui_branchvalueissue_189,coderabbit_urlvaluehttps://github.com/Comfy-Org/ComfyUI/pull/11294#discussion_r2959796352,upstream_pr_urlvaluehttps://github.com/Comfy-Org/ComfyUI/pull/11294,parent_findings_comment_urlvaluehttps://github.com/pollockjj/mydevelopment/issues/101#issuecomment-4305426643,probe_commandvalue matching the Slice 2 AC-4 command verbatim,probe_baseline_pathvaluegithub_issues/189/slice1/decode_batch_axes_baseline.json,probe_post_fix_pathvaluegithub_issues/189/slice2/decode_batch_axes_post_fix.json,comparison_pathvaluegithub_issues/189/slice2/before_after_comparison.json,regression_test_commandvalue matching the Slice 2 AC-2 pytest invocation verbatim,seedvr_regression_guard_commandvalue matching the Slice 2 AC-3 pytest invocation verbatim,ruff_commandvalue matching the Slice 3 AC-1 ruff invocation verbatim — verified by committed artifact at the named path andgrep '^<key>: 'parseability against the listed key set.Constraints
python3(resolved by the slicer's environment to the deliverable venv at/home/johnj/dev_cuda_1/ComfyUI/.venv/bin/pythonfor ComfyUI-touching commands; no AC hardcodes an absolute interpreter path).debug/run_debug.py(not exercised — this plan is CPU-only static-source plus pytest; no ComfyUI launch).--use-process-isolation --disable-cuda-malloc(not exercised — see above).pkill, norm -rf, nopython main.py.issue_189in bothpollockjj/mydevelopmentandpollockjj/ComfyUI, not on base branchesmain/issue_101.pytest 9.0.3in deliverable venv — REUSE; cited fromtools_register.md#pytestand start flag to not open webui at launch Comfy-Org/ComfyUI#187 plan TPM Inventory.ruff 0.15.9system binary — REUSE; ComfyUI CI invocation pattern ispython -m ruff check .per.github/workflows/ruff.yml.torch 2.11.0+cu130,einops 0.8.2in deliverable venv — REUSE.comfy.ldm.seedvr.vae.VideoAutoencoderKLWrapperandVideoAutoencoderKLin deliverable source — REUSE; constructable on CPU via__new__+nn.Module.__init__per GTP-4.comfy.ldm.seedvr.vae.VideoAutoencoderKL.decode_— REUSE (mocked in tests viaunittest.mock.patch.object).comfy.ldm.seedvr.vae.lab_color_transfer— REUSE (mocked in tests viaunittest.mock.patch.object).comfy.ldm.seedvr.vae.rearrange— REUSE (called insidedecode(); not patched).unittest.mock.patch.object(stdlib) — REUSE for measurement instrumentation.nn.Module.__init__(torch) — REUSE for the__init__-bypass test fixture pattern.ghCLI — REUSE for issue body update viascripts/run_tdd_post.py update-issue(PR creation is the/prskill's responsibility, not this plan).equivalence_methods.md#tensor-numerical-equivalence).torch.equal(binary; per the issue body's Acceptance Intent verbatim quote in the Plan Foundations comment R&M section).git diffartifact + regression test that fails pre-fix and passes post-fix (perequivalence_methods.md#code-behavior-equivalence) + Slice 1 baseline probe + Slice 2 post-fix probe + comparison delta JSON.vae_diff.patchagainst the seven-token guard set (zero matching+/-lines required).test_seedvr_groupnorm_limit.py,test_seedvr_vae_tiled_args_no_mutate.py,test_seedvr_rope_delegation.py) withcode-behavior-equivalencezero-new-failures tolerance.## Created Surface Contract;grep '^<key>:'parseable for key existence;python -m json.toolparseable for structural validity.git stashand without external decision deferral..gitignorerules for safe generated bulk, or discarded generated debris.git status --shortgates locally and must refuse to submit while any touched repo is dirty.repo_hygiene.md,repo_hygiene.txt, rawgit status --shorttranscripts, orCLEANmarkers as QA evidence; this plan deliberately omits any hygiene-evidence AC. QA verifies submitted commits, branch refs, and artifact presence through GitHub rather than trusting slicer-authored cleanliness transcripts.Out of Scope
comfy/ldm/seedvr/vae.pylines 2245-2253 (b, tc, h, w = z.shape; latent = z.view(b, 16, -1, h, w); ...). Per the issue body's Risks / Stop Conditions block, "Any code change to the latent-side prep block (lines 2245-2253) indicates plan drift — stop and surface." This block was already corrected upstream by commitaf6c5d6and is asserted byte-identical to base by Slice 2 AC-1.Comfy-Org/ComfyUI#11294. Per the issue body's Non-Goals block and Risks block ("Upstream PR feat: Add support For SeedVR2 (CORE-6) Comfy-Org/ComfyUI#11294 ismergeable=DIRTYand silent for two weeks; do not attempt to push fixes back toyousef-rafat:seedvr2from this issue."), upstream propagation is a separate later decision.#101Leg-1 sibling defects:#187GroupNorm memory limit (closed),#188frame count,#190/#192forward return contracts,#191latent metadata,#193cache clearing,#194decode state guards. Each owns its own scope and its own issue branch.debug/run_debug.pylaunches). The plan is CPU-only and CPU-deterministic by construction.numz/ComfyUI-SeedVR2_VideoUpscaler,ComfyUI-SeedVR2-Canonical, etc.). Only the in-treecomfy/ldm/seedvr/vae.pycarries this output-side defect.debug/run_debug.py,debug/reset_gpu.py,debug/harness_runtime.py, etc.). The plan exercises no harness path./prand/pr-reviewskills' responsibilities; this plan only lands committed work on issue branches and exits.if self.enable_tiling:branch indecode()). The bug only manifests in the non-tiled path (super().decode_(latent).squeeze(2)). Tiled-decode is covered by the existingtest_seedvr_vae_tiled_args_no_mutate.pyregression guard cited in Slice 2 AC-3.VideoAutoencoderKLWrapper.encode). The bug is in the decode output-side block only.Tracked in pollockjj/mydevelopment Comfy-Org#189 (bookkeeping PR pollockjj/mydevelopment#197).