Release V3.5.2 stabilization and optimization
This commit is contained in:
+77
@@ -1,5 +1,82 @@
|
||||
# Work Log
|
||||
|
||||
## 2026-08-26 - V3.5.2 README final audit and publication authorization
|
||||
|
||||
- Re-audited English/Japanese README chapter order, historical/current-version wording, withdrawn controls, local links, image references, and validation claims before publication.
|
||||
- Moved the Japanese product summary to the document opening and merged the duplicated English V3.5.1 Hotfix/What's New sections into one chronological `Reference Audio & Compatibility Update` section.
|
||||
- Explicitly documented that the withdrawn experimental Last Queued Seed override is not included and that V3.5.2 retains standard ComfyUI Seed behavior.
|
||||
- Kept the user-supplied infographic unchanged and added a precise caption distinguishing its accepted measurement labels from the final ComfyUI 0.33.3 / 172-entry packaging gate.
|
||||
- Regenerated the 172-entry Manifest and verified all local README links. Post-audit full pytest is `485 passed`; Manifest, runtime verifier, node registration, source/runtime hashes, queue state, and `git diff --check` remain PASS.
|
||||
- The user explicitly authorized committing and pushing the complete V3.5.2 release state to the main repository after this audit.
|
||||
|
||||
## 2026-08-26 - V3.5.2 final release gate
|
||||
|
||||
- Completed the final Stabilization & Optimization release gate without adding a generation mode or changing V3.4/V3.5 Sampling, Conditioning payloads, Terminal Merge, Assembly, Seam, Run Storage, or standard ComfyUI Seed behavior.
|
||||
- Created exact pre-change snapshot `pre-v352-release-preparation-exact-20260826_022016` at revision `b8ab766eb364063f8af88781e7be70ee7286fba0`, containing 198 authoritative files plus the four active ComfyUI_W/ComfyUI_WAN metadata files. The standard snapshot attempt stopped only on the known repository `.pytest_cache` ACL.
|
||||
- Ran a final 576x576 / 1x5-second / 20-step T2VA API cache smoke before the release metadata change. MISS reported one encode call and `8.454306s` Prompt/CLIP; HIT reported zero encode calls and `0.000038s`. Both used derived seed `5684542245315818407` and produced identical decoded video MD5 `40e378d7c80e6bebd9d64b322c13e677` and audio PCM MD5 `4fb0a699227c4f834f7a52ba4d395ae8`.
|
||||
- Reused the accepted Phase 4-4 Video Guide GPU A/B because no Production path changed afterward: complete source identity, 124-frame prefix, VAE/Qwen inputs, Sampling seed, decoded video, and audio PCM remain exact between legacy-v3 and optimized preprocessing.
|
||||
- Bumped package, project, and runtime metadata to `3.5.2`; added the Stabilization & Optimization sections to English/Japanese README and Changelog; removed the stale withdrawn Last Queued Seed claim from package information; and updated package validation to the canonical `485 passed` result.
|
||||
- Copied the user-supplied infographic `D:\SD_PIC\Neo\Clip_91.png` to `docs/images/v352-stabilization-optimization.png` bit-exact (SHA-256 `0c1dfd5acff349b98ecdfb57ae50941f6da26a462dfd2aaa9b7c7b0a34abf304`) and embedded it near the top of both READMEs.
|
||||
- Synchronized only `version.py` and `metadata.ini` to ComfyUI_W and ComfyUI_WAN; all four destination hashes match the source. ComfyUI_W restarted successfully and logged `H3 Continuum 3.5.2 loaded` on ComfyUI `0.33.3`.
|
||||
- Final validation: compileall exit `0`; full pytest `485 passed` with only the known `pynvml` FutureWarning; V3.5.2 runtime verifier PASS; all 172 Manifest hashes PASS; five required V3.5 nodes registered; source/runtime hashes PASS; queue `0/0`; and `git diff --check` PASS. Commit, push, and release were intentionally not performed at this gate.
|
||||
|
||||
## 2026-08-26 - Stabilization Phase 4-6A Video Guide memory optimization
|
||||
|
||||
- Completed the read-only Phase 4 Feature Tax & Redundancy Audit and selected only Video Guide long-input memory as an implementation candidate. No Production Sampling speed change, Driving Audio lazy Decode, Audio hash special case, Session/State relaxation, Refine Context change, or V3.4 change was adopted.
|
||||
- The mandatory standard snapshot stopped on the known `.pytest_cache` ACL. Created verified exact snapshot `pre-phase4-6a-video-guide-memory-exact-20260826_012227` with six authoritative/runtime files at revision `b8ab766eb364063f8af88781e7be70ee7286fba0`.
|
||||
- Pre-change verification found that ComfyUI_W alone contained an unrecorded Video Guide preprocess v3, while the repository and ComfyUI_WAN still had v2-equivalent content. With explicit approval, preserved and integrated the active v3 `17k+5` alignment and final-frame padding behavior into the repository before optimizing memory.
|
||||
- Replaced whole-source float32 RGB materialization, finite scan, and one-shot `bytes` hashing with eight-frame normalization/validation/hash updates. The source object now owns only the used prefix. The complete normalized source still contributes to the exact same SHA-256 and unused-tail NaN/Inf detection.
|
||||
- Deterministic 300x256x256 RGB comparison against the original ComfyUI_W v3 retained an identical full contract and hash. Peak RSS above the allocated input changed from `396.758 MiB` to `114.695 MiB`, retained storage from `225.0 MiB` to `93.0 MiB`, and host time from `0.246672s` to `0.257602s`.
|
||||
- Added five CPU regressions for legacy-v3 hash byte order, full-tail identity, unused-tail finite rejection, prefix ownership, frame alignment/padding, and encoding behavior. Focused result is `17 passed`; full result is `485 passed`. Source runtime verification and all 171 Manifest entries pass.
|
||||
- Completed Phase 4-4 GPU A/B through the normal API using the same 15-second / 360-frame 800x800 Video Guide, V3.5 one-chunk workflow, 576x576 output, 20 steps, seed input `46001`, Sage Attention, and Hi-Res Fix disabled. Legacy and optimized paths both used one physical group, guide contract 640x640 / 124 frames, derived Sampling seed `3288707263902886519`, exact-duration 124->120 frames, 24 fps / 5.000 seconds, and 32 kHz stereo audio.
|
||||
- Reloaded the exact VHS input independently and confirmed old/new source SHA-256 and combined identity, retained 124-frame prefix, resized VAE input, and Qwen input are byte-identical. Decoded video MD5 and audio PCM MD5 are also identical (`53cd381aff184c2f8aa5910ab62be342` / `3653c0ec89947a928075d4ba056ff624`). Representative start/middle/end frames show no gross artifact. Phase 4-6A is formally accepted; runtime was restored to the optimized source-identical file and the API queue is empty.
|
||||
|
||||
## 2026-08-26 - Stabilization Phase 3-2 synchronized GPU rebaseline
|
||||
|
||||
- Phase 3-2 is formally PASS. The accepted decision is to make no speculative Phase 3-3 Production optimization: measured Continuum Assembly is negligible, Core VAE/MP4 work is external, and existing Spectrum/Sol-Attn routes provide materially larger gains without changing Continuum's Sampling contract.
|
||||
- Final-pass documentation snapshot: `pre-phase3-2-formal-pass-docs-20260826_004900` (2 files, revision `b8ab766eb364063f8af88781e7be70ee7286fba0`).
|
||||
- Preserved all Phase 3-1 and earlier uncommitted work. Accepted snapshot `pre-stabilization-phase3-2-profile-accepted-20260825_220920` contains 479 files at revision `b8ab766eb364063f8af88781e7be70ee7286fba0`; the standard snapshot attempt stopped only on the known ACL-inaccessible `.pytest_cache`.
|
||||
- Added repository-only API diagnostics that synchronize CUDA before/after measured nodes, separate true whole-workflow allocator peaks from per-node reset peaks, poll lightweight RSS/private and NVML free VRAM during each interval, and sample USS at interval boundaries. Early probes that queried CUDA from a polling thread or called `memory_full_info()` at 10 Hz measurably slowed Sampling and are explicitly excluded.
|
||||
- Rebaselined Phase 3-1 cache-HIT Production behavior after an excluded warm-up and three measured runs. API medians are T2VA `168.069s` and FL2VA `379.765s`; synchronized phase medians show Sampling `120.516s` / `295.216s`, Video Decode `19.154s` / `51.634s`, Audio Decode `0.466s` / `1.040s`, Assembly `0.130s` / `0.619s`, and MP4 save `2.583s` / `7.248s`.
|
||||
- Whole-workflow CUDA allocator peaks are `1.456 GiB` T2VA and `4.129 GiB` FL2VA. Per-node runs independently identify Sampling as the maximum. The FL2VA contract remained 3 logical / 2 physical groups (`[1]`, `[2,3]`), 360 frames, 24 fps, 15.000s; T2VA remained 120 frames / 5.000s.
|
||||
- Compared external attention routes on T2VA with one warm-up excluded and three-run API medians: Sage `168.069s`; Sage + Sol-Attn `162.090s`; Sage + Spectrum `124.830s`; Sage + Sol-Attn + Spectrum `121.147s`. The initial Sol attempt failed only because the diagnostic process could not write its default Triton cache and was interrupted/excluded. The accepted rerun used a writable test cache, logged active sparse kernels, and had zero fallback.
|
||||
- The fastest combined route additionally passed one measured FL2VA run at `288.548s` API with 3 logical / 2 physical groups, terminal seed/trim unchanged, 640x640, 360 frames, 24 fps, 15.000s, and audio. `ffprobe` validated every selected output; contact sheets showed no gross persistent artifact. External accelerator outputs are intentionally not required to be MD5-identical.
|
||||
- No Continuum Production optimization was adopted. Sampling remains contract-sensitive, Core VAE Decode and MP4 save are external, and Assemble + Seam is below the 1% threshold. The diagnostic wrapper was removed from the running process and the normal ComfyUI API was restored at `127.0.0.6:8188` with output `D:\output\video\comfy_video`. No Production runtime synchronization was required. Evidence is under `D:\Codex\_test_results\ComfyUI-H3-Continuum\stabilization-phase3-2-20260825`; commit/push were not performed.
|
||||
- Focused diagnostics are `13 passed`; full pytest is `480 passed`. Diagnostic `py_compile`, all 170 Manifest hashes, source runtime verification, and `git diff --check` pass with only the known `pynvml` FutureWarning.
|
||||
|
||||
## 2026-08-25 - Stabilization Phase 3-1 Prompt/CLIP cross-run cache
|
||||
|
||||
- Created exact source/runtime snapshot `pre-stabilization-phase3-1-prompt-cache-20260825_211303` with 19 files and preserved all earlier uncommitted stabilization changes.
|
||||
- Added a V3.5-only private Prompt/CLIP conditioning cache attached to the exact CLIP object. It is capped at 16 LRU entries and invalidates on prompt, exact resized First/Last pixels/shape/dtype, Last Frame presence, patch UUID, or CLIP layer. V3.4, public schemas, Sampling, Terminal Merge, Assembly, Seam, and Run Storage were not changed.
|
||||
- Conservatively bypassed Reference Image/Audio/Video/Timeline and CLIP schedule/tokenizer/hook states. Cache exceptions and CUDA-resident outputs fall back to uncached encode. Tensor payloads are shared only when CPU-resident; metadata dict/list/tuple containers are copied per use.
|
||||
- Added Detailed Report evidence with real `hits/misses/bypasses/encode_calls`. Dedicated and integrated CPU result is `25 passed`; full pytest is `467 passed` with only the known `pynvml` and inaccessible repository pytest-cache warnings.
|
||||
- Synchronized only `v2/h3_builder.py`, `v2/sequence.py`, `v2/nodes.py`, `v3/nodes.py`, and `v3/driving_nodes.py` to ComfyUI_W; every SHA-256 matched. ComfyUI_W was restarted with its recorded arguments before GPU tests.
|
||||
- GPU/API T2VA A/B PASS: 576x576, 120 frames / 5.000s. Prompt/CLIP changed from `5.898398s` (`misses=1`, `encode_calls=1`) to `0.000028s` (`hits=1`, `encode_calls=0`). Decoded video frame MD5 sequence and decoded audio PCM MD5 are identical. A separate base-seed-only change produced a new derived Sampling seed while retaining one HIT / zero encode calls.
|
||||
- GPU/API FL2VA Long Terminal Merge A/B PASS: 640x640, 360 frames / 15.000s, 3 logical / 2 physical groups (`[1]`, `[2,3]`). Prompt/CLIP changed from `21.642687s` (`misses=4`, `encode_calls=4`) to `0.007128s` (`hits=4`, `encode_calls=0`). Decoded video frame MD5 sequence and decoded audio PCM MD5 are identical. A separate base-seed-only change retained four HITs / zero encode calls, changed both normal and terminal physical seeds, and preserved the group contract.
|
||||
- No CUDA retention regression was observed. FL2VA warm completion RSS was 6095.8 MiB MISS versus 6117.1 MiB HIT, a bounded approximately 21.3 MiB CPU cost for four retained conditioning entries. Evidence is under `D:\Codex\_test_results\ComfyUI-H3-Continuum\stabilization-phase3-1-20260825`. Commit and push were not performed.
|
||||
|
||||
## 2026-08-25 - Stabilization Phase 2-2 Conditioning attribution
|
||||
|
||||
- Created exact source/runtime snapshot `pre-stabilization-phase2-2-conditioning-20260825_200800` with 15 files including snapshot metadata. Existing uncommitted Phase 2-1 and all prior V3.5.1 work were preserved.
|
||||
- Split Detailed Report Conditioning timing into identity video VAE, Reference Image/Audio VAE, Driving Audio VAE, Video/Timeline Reference VAE, and Prompt/CLIP. The instrumentation is inactive outside the existing V3.5 Detailed Report path and never synchronizes CUDA.
|
||||
- Focused validation is `29 passed`; full validation is `453 passed` using a writable explicit pytest `--basetemp`. The initial full attempt produced setup errors only because the default `pytest-of-links` directory is ACL-inaccessible. Manifest, syntax, and related regressions pass.
|
||||
- Synchronized only `v2/sequence.py`, `v3/memory_attribution.py`, the two related test files, and `MANIFEST.sha256` to ComfyUI_W; source/runtime SHA-256 matched before GPU execution.
|
||||
- GPU/API 1x5 T2VA accepted result: 576x576, 120 frames / 5.000s, 158.701s API, 135.970s Sampler host wall, 129.935s Sampling, 6.004392s Conditioning, and 6.004374s Prompt/CLIP. Warm-up independently measured 6.283131s Conditioning and 6.283109s Prompt/CLIP. One 3.398s execution-cache replay was excluded.
|
||||
- GPU/API 3x5 FL2VA Long Terminal Merge accepted result: 640x640, 3 logical / 2 physical groups (`[1]`, `[2,3]`), 360 frames / 15.000s, 410.727s API, 346.197s Sampler host wall, 322.335s Sampling, 23.644903s Conditioning, 22.873708s Prompt/CLIP, and 0.771176s First/Last image VAE. Decode and Assemble completed with Exact Duration `362 -> 360`.
|
||||
- No optimization was implemented. Prompt/CLIP accounts for effectively all T2VA Conditioning and 96.7% of FL2VA Conditioning; the remaining measured Continuum-side work is not an optimization target. Reference-specific branches remain instrumented but require separate connected GPU evidence before any conclusion.
|
||||
- Evidence is under `D:\Codex\_test_results\ComfyUI-H3-Continuum\stabilization-phase2-2-20260825`. Commit and push were not performed.
|
||||
|
||||
## 2026-08-25 - Stabilization Phase 2-1 performance attribution instrumentation
|
||||
|
||||
- Attempted the mandatory full snapshot first; it stopped only on the known ACL-inaccessible `.pytest_cache`. Accepted exact-file snapshot `pre-stabilization-phase2-exact-20260825_190303` contains nine source/runtime target files at revision `b8ab766eb364063f8af88781e7be70ee7286fba0`, with source/runtime Production hashes equal before modification.
|
||||
- Extended the existing V3.5 Detailed Report collector with non-synchronizing host timings for Preparation/Conditioning, physical-group preparation/Sampling/CPU commit plus finite validation/Refine Context capture, and Finalization. No new node, socket, widget, schema, or normal-report overhead was added.
|
||||
- Terminal Merge regression confirms exactly two performance group records for logical `[1]` and `[2,3]`. Dedicated tests also prove the collector never calls CUDA synchronization.
|
||||
- Focused CPU/static validation is `27 passed`; full validation is `451 passed`; only the known `pynvml` FutureWarning remains. Runtime verification, Manifest validation, `py_compile`, `git diff --check`, and source/runtime SHA-256 equality pass.
|
||||
- The first API run accidentally reached the pre-existing 18:46 ComfyUI listener; the newly started process had exited on port conflict. That result is explicitly excluded. After confirming an empty queue, the exact old listener PID was stopped and the synchronized runtime was restarted with the recorded Stability Matrix arguments.
|
||||
- Accepted GPU/API measurements: warm 1x5 T2VA 576x576 = 143.888s API / 134.736s Sampler host wall (Sampling 128.737s); 3x5 FL2VA Long Terminal Merge 640x640 = 396.477s API / 348.047s Sampler host wall (Sampling 323.835s, physical `[1]` + `[2,3]`); 1x5 Hi-Res 1.2x = 330.005s and 704x704. All outputs are 24 fps with exact 5.000s or 15.000s duration and 32 kHz audio.
|
||||
- Warm baseline versus Hi-Res peak Working Set was 6.882 versus 7.374 GiB; peak Private Bytes was 20.655 versus 21.998 GiB. The 3x5 peaks were 9.409 GiB Working Set and 21.159 GiB Private Bytes. Global VRAM used peaked around 15.4-15.6 GiB and is recorded only as a process-external observation.
|
||||
- Sampling accounts for 93.0-95.5% of measured Sampler time. CPU commit/finite validation, Refine Context, and group preparation are below the adoption threshold and will not be optimized. Conditioning is 4.4-6.9% and is eligible only for finer attribution, not an optimization change. Evidence is under `D:\Codex\_test_results\ComfyUI-H3-Continuum\stabilization-phase2-20260825`.
|
||||
|
||||
## 2026-08-25 - Stabilization Phase 0 local CI portability fix
|
||||
|
||||
- Froze V3.5.1 feature work and began stabilization with a test-only change. Production Sampling, Layout, Session, Run Storage, Refine Context, UI, and runtime files were not changed.
|
||||
|
||||
Reference in New Issue
Block a user