File size: 5,360 Bytes
4ec41f2 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 | # Implementation and source mapping
Five instructions were read completely from /private/prompt/. The actual server HEAD is recorded in reports/source_code_mapping.json. The server already has a latent-aware stateplay/pipeline.py. The implementation uses the actual diffsynth StatePlayModel and model_fn_stateplay, restores omitted frozen patch weights from local Wan shards, and strictly restores the trained transition tokenizer. No generic SFT switch_train_mode is used to overwrite the RL freeze list.
Official DiffusionNFT was cloned outside the project at /tmp/stateplay_rl_research/DiffusionNFT and read at commit cbb14f84b8312b620390dfcbe2fab69c1383b104. The local 22-page v2 paper was read, including Equation 5, Algorithm 1, section 3.3 and Appendix C. Migration differences are Wan sigma/time convention, first-observation masking, four-video condition groups instead of prompt-text grouping, score-point std floor/guard, independent SFT anchor weighting, one sampled timestep per visit rather than the full official sampling grid by default, and small shared-backbone parameter snapshots. Full-schedule and velocity-MSE alternatives are executable.
The original checkpoint identity and all source-model fingerprints are recorded before/after. Only new rl_posttrain code and outputs were edited. Existing dirty modifications were preserved. Temporary output paths are protected against aliasing the original SFT checkpoint.
Real model tests and mock tests are separate. Full 480x832/f25 memory tests, actual 101-frame/f26 inference, real GPT-6.1-sol one-request 124-image grading, real 9B full-group backward and seven-GPU LoRA updates, source-parameter freeze comparisons, original-entry pixel parity and real optimizer-boundary restart are evidenced in the reports. API concurrency/failures are local mock HTTP tests; the paid API budget remains one complete group. No long formal training was launched.
Numerical parity debugging identified autocast LayerNorm promotion, not a different checkpoint. Sampling now uses the existing entry's bf16 inference without autocast and restores exact FP32 master storage. The initial already-graded debug group remains immutable with its actual pre-fix rendering and its actual clean latents. A real NFT optimizer test consumes those saved scores/latents without another API call.
Full Qwen input is 124 images and 43,610 prefix tokens. The initial full backward OOM was resolved by disabling PEFT full-input FP32 casting and offloading checkpoint input activations. The model includes actual vision and hybrid language branches. Only language LoRA targets update. Save/reload checks compare all 400 adapter tensors exactly. Frozen buffers and parameters are protected. Native Greedy JSON scoring returns four finite numbers, and malformed output is a hard failure.
Quality is deliberately not certified: one debug teacher condition is insufficient, the permutation test may fail, episode IDs are absent, and independent SAM2.1/Grounding DINO/ClipAttackNet assets are unavailable. Supplied 100/200 test archives are excluded from training by video hash. The paper's Move/Att reductions are implemented and tested on extractor artifacts but no fabricated real benchmark scores are reported. Formal C requires actual quality metrics and manual source/shortcut review.
The reviewer follow-up corrects the StatePlay source paper to arXiv:2607.26754v1 (the named Lin PDF). The numeric filename 2605.15256v1.pdf is ReactiveGWM. D is the extra high-level Super Art attempt/macro channel. Only the local test collector supplies its exact 28-step schedule; that schedule must not be assumed for released clips. The immutable v1 teacher group is quarantined; formal v2 teacher data and deterministic v3 renderings are separate identities.
Prepared study_v2 includes 10,000 clips, both holdout archives, five temporal fingerprints and informative action-window comparisons. One two-clip training component was detected, with no train/holdout overlaps at the chosen thresholds. This is an automated leakage screen, not evidence of episode independence. Source/manual review remains mandatory for formal RM quality. A has no RM quality dependency.
The revised C audit uses deterministic backend flags and one name-sorted FP32 gradient reduction per optimizer update, mathematically equal to DDP's global mean. It reproduces input/noise/timestep/anchor/EMA traces, parameters, optimizer and policy-bank state exactly over six real world ranks. Failed historical numerical reports remain preserved. Qwen's formal seven-rank accum4 recovery is tested separately; the source-bound acceptance gate requires its bitwise comparison to pass.
The revised full Qwen accum4 test passes bitwise on all seven ranks after pinning the upstream FLA/Triton configuration choice. Its three optimizer steps take 787 seconds and 36.50 GiB allocated per card. Short suffix/full-logit equivalence additionally uses a real FP32 language oracle (temporary test-only casts, frozen bytes restored), with identical losses and maximum gradient difference 1.9758e-6. The matched BF16/offload comparison has a 1.2755e-4 maximum gradient difference from GEMM geometry; it is recorded, not called bitwise equality. Production training precision is unchanged. Formal v2's actual complete processor prefix has 43,867 tokens and all 124 images.
|