Release v3.7.0 refinement foundation
This commit is contained in:
+32
-1
@@ -10,7 +10,38 @@
|
||||
- E2-8A Contract, E2-8B1 CPU/static validation, E2-8B2 external LBH GPU PoC, and E2-8C Second Pass GPU acceptance are formally PASS. The E2-8C CPU baseline is `260 passed`.
|
||||
- External LATENT nodes may discard custom dictionary metadata. Identically shaped group reordering therefore cannot be cryptographically detected; count, position, T, geometry, and audio-shape checks are enforced.
|
||||
|
||||
Updated: 2026-08-27
|
||||
Updated: 2026-08-28
|
||||
|
||||
## V3.7.0 Conditioning Adapter and RefineSchedule release candidate (2026-08-30)
|
||||
|
||||
- Post-review frontend/input hardening is complete. `H3ContinuumSamplerV37` now follows the V3.6 compact frontend path for project identity, hidden internal widgets, Run Storage conditional controls, queue normalization, runtime diagnostics/debug/preview settings, and permanent Reference Audio label normalization. Still Image Guide now rejects every IMAGE batch except exactly one image at both public payload creation and defensive source preparation.
|
||||
- Guide/First/Last endpoint collisions are regression-covered at physical frame `0` and `123`: the explicit Still Guide role marker takes priority over frame-number inference, so the owned Guide source is reencoded rather than the First/Last image. Reference Audio remains the accepted V3.5.1 permanent-socket design; the withdrawn frontend `Hidden / Show` widget was not restored.
|
||||
- V3.7 has an internal, versioned `RefineSchedule` contract with `External`, `Full`, `Tail`, and `Partial` modes. `External` preserves the exact incoming SIGMAS tensor; Tail/Partial use exact source slices. The contract records source/effective sigma hashes, evaluation range, video-random/audio-locked noise policy, and is included in `second_pass_contract` identity without invalidating reusable First Pass Run Storage.
|
||||
- Hi-Res Fix and Second Pass expose the schedule only as an optional Python-internal argument. Their public `INPUT_TYPES`, saved-workflow schema, V3.6/V3.5 nodes, Production defaults, and existing SIGMAS socket remain unchanged. Tail 10 and Tail 6 are reproducible from the same 10-evaluation source schedule.
|
||||
- Still Image Guide now owns a CPU clone and SHA-256 identity of the original Guide source image. Only its owner physical group carries that source, and Second Pass rebuilds the marked Guide keyframe from the original image at target latent geometry; First/Last, Terminal Merge ownership, Core PackedLayout/RoPE construction, and First Pass audio bit-exact passthrough remain unchanged.
|
||||
- GPU Correctness PASS: internal Tail 6 at 704x704 used the exact six-evaluation suffix. Under the same Sage path as the earlier external `SplitSigmas` Tail 6 baseline, decoded Video SSIM is `0.982936` and decoded PCM SHA-256 is identical; both are 120 frames / 5.000 seconds. The 1152x1152 Guide run encoded the First Image plus the original 1254x1254 Guide source to 72x72 latent geometry and completed without node/finite/geometry/AV errors. It remains low-motion/freeze-detected, consistent with the already-known hard-anchor quality limitation; Still Image Guide Production promotion therefore remains **HOLD**.
|
||||
- Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v37-pillars-gpu-gate-20260830`. Accepted snapshots: `pre-v37-refine-schedule-conditioning-approved-ecc72da-accepted-20260830_005156` (source, 641 files, zero SHA mismatch) and `pre-v37-refine-schedule-conditioning-runtime-w-approved-accepted-20260830_005202` (ComfyUI_W runtime, 353 files, zero mismatch).
|
||||
- Final local Release Gate is `633 passed`; the 185-entry Manifest, Python/JavaScript syntax, `git diff --check`, ComfyUI_W runtime verifier, live V3.7/V3.6 registration, empty queue, and source/runtime SHA-256 equality for 25 release/Production files pass. ComfyUI_W restarted on 0.34.2 and reports `H3 Continuum 3.7.0 loaded`. Runtime logic and Production defaults are unchanged; commit, push, and release remain pending.
|
||||
|
||||
## Sampling layout-validation measurement (2026-08-28)
|
||||
|
||||
- Phase 1 Correctness PASS snapshots `phase1-correctness-pass-source-accepted-20260828_172802-ecc72da` and `phase1-correctness-pass-runtime-w-accepted-20260828_172802-ecc72da` are the formal optimization A/B baseline.
|
||||
- Detailed Report now measures the existing `validate_native_continuity_layout()` call count, full-validation count, status counts, cumulative host wall including its existing device-to-host transfers/implicit synchronization, physical-group step contribution, and observable accelerator markers. It does not add `torch.cuda.synchronize()` and is inactive outside Detailed Report. Validation, Sampling, payload, schema, workflow, metadata, and public UI contracts are unchanged.
|
||||
- The live ComfyUI_W / RTX 5060 Ti matrix used I2VA + Continuation, Compatibility Reference Context, 2x5 seconds, 384x384, 4 steps, fixed seed `361040`, and identical prompts. Sage and Sage+Sol each made `4` full calls (`1.000`/validated step) at `0.472` and `0.530 ms/step`; Sage+Spectrum and Sage+Sol+Spectrum each made `8` full calls (`2.000`/validated step) at `1.164` and `1.108 ms/step`. Every route reported zero validation failures. Global `--use-sage-attention` is not exposed through the inspected MODEL markers, so Sage route identity is taken from the controlled workflow graph and recorded backend launch arguments rather than inferred from `accelerator_summary()`.
|
||||
- The representative Sage run used 576x576, 2x5 seconds, 20 steps, and the same seed. It made `20` full calls, `1.000` call/validated step, with `14.289 ms` cumulative validator wall (`0.714 ms/step`) versus `321.632848 s` measured Sampling and `382.055317 s` API elapsed. The validator therefore accounts for approximately `0.00444%` of Sampling and `0.00374%` of API time.
|
||||
- Decision: do not implement a layout-validation cache or fast-path. Even eliminating the measured validator completely has a maximum representative saving of about `14 ms`; it is not a material Sampling optimization target. Spectrum doubles the validator call count in this matrix but remains about `1.1 ms/step`. No Sampling optimization was started. Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\layout-validation-profile-phase`.
|
||||
|
||||
## V3.6.1 Run Storage correctness hotfix (unreleased, 2026-08-28)
|
||||
|
||||
- Absent Last Frame hashes are canonicalized as `""` for every new chunk contract. Candidate revisions may bypass a hash mismatch only when both top-level Last Frame values are absent and normalizing the per-chunk `last_frame_hash` is the sole contract difference.
|
||||
- This narrow path reuses legacy V3.6.1 final chunks stored with `"none"`; any real Last Frame or any other field difference remains incompatible. Run Storage schema and Sampling Contract versions are unchanged. FL2VA Long Terminal Merge atomicity remains unchanged.
|
||||
- Compact UI now resets `Regenerate From = Auto` and `Variation Nonce = 0` whenever Run Storage is Off, including graph reload and final API serialization. Backend validation remains enabled.
|
||||
- CPU/static regression expects normal no-Last-Frame I2VA 2→3 at 2/1, legacy `none` at 2/1, real Last Frame at 1/2, Terminal Merge at 1/2, exact resume at full reuse, and frontend Off/reload/queue normalization. Focused validation is `64 passed`; full validation is `565 passed` with only the known `pynvml` warning.
|
||||
- Actual ComfyUI_W 0.34.2 / RTX 5060 Ti 16 GiB GPU regression copied only legacy revision `9c35c64b529a5c2c` into isolated test Run `run_95c2caec1a0846a09cf3570acb29e27f`. Prompt `bf619577-2057-4635-8fef-2537a40295c4` completed as `2 reused, 1 generated`, first regenerated Chunk 3, output `MiniMax_H3_00008_.mp4`. Exact 2-chunk resume prompt `e593b06a-9a37-4cc1-932c-0d13e55979c1` completed as `2 reused, 0 generated` with no regenerated chunk.
|
||||
- G3 is repaired without changing Run Storage identity or Sampling behavior: the `FLF_STRATEGY` compatibility guard now applies only when Terminal Merge is actually enabled. Non-Terminal FL2VA therefore retains a prefix already validated by Run Storage; Terminal Merge still rejects a missing/mismatched strategy before its atomic pair handling. Focused validation is `63 passed`; full validation is `568 passed` with only the known `pynvml` warning.
|
||||
- **Phase 1 GPU Acceptance G1-G6 is PASS.** G1 canonical no-Last-Frame I2VA 2→3 and G2 an isolated copy of actual legacy V3.6.1 `"none"` cache both completed at `2 reused / 1 generated`, Sampling 1. G3 real-Last-Frame non-Terminal FL2VA completed at `1 / 2`, Sampling 2, and no longer reports `FLF chunk contract changed`. G4 Terminal Merge remains `1 / 2` with one physical `77T Video / 433T Audio` Sampling call. G6 Run Storage Off completed with one Sampling call and no queue error.
|
||||
- G5 fresh 3-chunk and matched 2→3 extension remain bit-exact: all three raw Video latent SHA-256 values, all three raw Audio latent SHA-256 values, decoded RGB24 SHA, decoded PCM SHA, 360 frames, and 480256 PCM samples/channel are identical. Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\phase1-gpu-acceptance-20260828\20260828_170034\gate_summary.json`.
|
||||
- Verified pre-G3 snapshots are source `pre-g3-fl2va-prefix-correctness-source-accepted-20260828_165400-ecc72da` (632 files excluding `.git` and the inaccessible `.pytest_cache`) and ComfyUI_W runtime `pre-g3-fl2va-prefix-correctness-runtime-w-accepted-20260828_165400-ecc72da` (351 files excluding the inaccessible `.pytest_cache`), both with zero SHA-256 mismatch. Final accepted snapshots are `phase1-correctness-pass-source-accepted-20260828_172802-ecc72da` and `phase1-correctness-pass-runtime-w-accepted-20260828_172802-ecc72da`. All 180 Manifest entries, source/runtime SHA-256, runtime verifier, Python/JavaScript syntax, empty queue, and `git diff --check` pass. Public UI, metadata, workflows, Reference/Masked transport identity, schema/version, and optimizations are unchanged; Sampling hot-path optimization remains a separate later phase.
|
||||
|
||||
## V3.6.1 Continuity fallback hotfix (2026-08-27)
|
||||
|
||||
|
||||
Reference in New Issue
Block a user