Commit Graph
934 Commits
Author SHA1 Message Date
SolitaryThinker 0b332ff4e3 [fix]: renoise sigma must be non-0-dim so mixing runs in fp32 like main (0-dim tensors are scalars in promotion); parameterize fastwan anchor for vsa 2026-07-22 20:11:13 -07:00
SolitaryThinker f0b9c1773a [fix]: quantize fp8 weights lazily on the serving device (main converts post-GPU; CPU fp8 casts can round differently); ledger GateResult 2026-07-22 20:02:56 -07:00
SolitaryThinker 70beba1bc8 [feat]: FastWan VSA port — vendored tile math + kernel wrapper (fail-closed), VSA block/model, card; capture shim covers qad+vsa 2026-07-22 19:55:19 -07:00
SolitaryThinker 17a5d8400f [fix]: capture shim must not recast fp8-quantized model to bf16 2026-07-22 19:48:25 -07:00
SolitaryThinker 574bdc44dc [fix]: fp8 device-placement check; capture shim disables fsdp/offload like the released example; fastwan anchor CLI 2026-07-22 19:45:52 -07:00
SolitaryThinker 3b134a41fb [feat]: FastWan-QAD-FP8 port — vendored fastvideo-main forward, fp8 layer, main-faithful DMD loop, card, capture shim 2026-07-22 19:41:34 -07:00
SolitaryThinker 23176a48b6 [misc]: installable package — console script, evidence as package data, editable-install convention (no PYTHONPATH) 2026-07-22 18:59:31 -07:00
SolitaryThinker 9d7daad178 [feat]: WanDMDLoop — FastWan few-step x0+renoise sampler with main's exact sigma-table semantics (wan.dmd.x0renoise/v1) 2026-07-22 18:47:28 -07:00
SolitaryThinker 89ffc4b68d [docs]: examples/ — canonical SDK text-to-video script 2026-07-22 18:18:41 -07:00
SolitaryThinker a3aa600bd9 [misc]: wan21/gates/ — separate the measurement side (goldens, anchor, capture, compare, probe) from logic code 2026-07-22 16:25:43 -07:00
SolitaryThinker d1e2d2c974 [misc]: backend-band evidence — SDPA-policy anchor records alongside bitwise AUTO records 2026-07-22 16:20:03 -07:00
SolitaryThinker de2743d443 [feat]: attention backend registry — FLASH_ATTN/SDPA, FASTVIDEO2_ATTENTION_BACKEND override, fail-closed selection, policy in env fingerprint 2026-07-22 16:16:42 -07:00
SolitaryThinker dcec5644c2 [misc]: model_id is the only load key — HF repo strings are card ingredients, never resolvable 2026-07-22 16:06:59 -07:00
SolitaryThinker ac11912063 [feat]: identity policy — model_id primary, HF repos as fail-closed aliases (weights + component sources) 2026-07-22 16:04:47 -07:00
SolitaryThinker 97d653456e [feat]: offline SDK — Model/Result (modality-neutral handle over the gated engine path) 2026-07-22 15:54:30 -07:00
SolitaryThinker ec1e32e070 [misc]: layers-extraction equivalence evidence (anchor bitwise 0.0 on unchanged baseline) 2026-07-22 15:44:32 -07:00
SolitaryThinker 3a709f78cc [feat]: shared layers/ (norms, MLP, embeddings, rotary, attention) extracted from the Wan model — key- and cast-compatible 2026-07-22 15:42:14 -07:00
SolitaryThinker 0c824beb7b [misc]: aligned evidence — bitwise DiT anchor (0.0 bf16+fp32), full ladder green, aligned 50-step sample 2026-07-22 15:34:45 -07:00
SolitaryThinker a37198afb7 [fix]: attention fallback keyed on autocast state, not q dtype — official casts fp32-promoted q/k to bf16 and runs flash 2026-07-22 15:28:31 -07:00
SolitaryThinker ee18433a6e [fix]: split storage vs compute dtype — DiT stores fp32 like official, computes bf16 via provenance.precision autocast 2026-07-22 15:25:50 -07:00
SolitaryThinker 2000e8a0ff [feat]: align the DiT with official — vendored WanModel (pinned 9737cba) + official weights + official autocast regime 2026-07-22 15:20:45 -07:00
SolitaryThinker 3333d959c0 [misc]: fp32 exact-math golden + final 3-way numerics report + decomposition ledger records 2026-07-22 15:03:17 -07:00
SolitaryThinker 7db862e6ff [misc]: dit_fp32 gate bound above the measured 2.6e-3 diffusers-vs-official fp32 delta (suspected RoPE; growth fails) 2026-07-22 15:01:38 -07:00
SolitaryThinker 787e17f500 [fix]: sdpa-fp32 exact-math golden (flash helper force-casts bf16); isolated anchor probes; robust report over record unions 2026-07-22 14:59:23 -07:00
SolitaryThinker 93f80b5c4b [feat]: fp32 exact-math anchor row (dit_fp32 golden + adapters); document load-bearing 512 zero-pad context 2026-07-22 14:55:43 -07:00
SolitaryThinker b57f3ad47d [fix]: probe sdpa shim k_lens device 2026-07-22 14:52:52 -07:00
SolitaryThinker a31f19ab5a [misc]: DiT delta decomposition probe (MATH/PAD/CAST/ATTN) + fp32 exact-math golden capture 2026-07-22 14:49:15 -07:00
SolitaryThinker 194d27220f [misc]: official Wan2.1 goldens (commit 9737cba) + 3-way numerics report + anchor ledger records 2026-07-22 14:39:11 -07:00
SolitaryThinker b28b01e405 [fix]: official text cleaning (ftfy width-fold retokenizes CJK; anchor caught rel 0.42 on negative prompt); fastvideo_main VAE device; resilient compare 2026-07-22 14:35:54 -07:00
SolitaryThinker 8e17148bd1 [misc]: pin py3.12/torch2.12; goldens captured in the fastvideo env by policy (env_policy in manifest) 2026-07-22 14:25:34 -07:00
SolitaryThinker 1387a9c5e7 [feat]: official-goldens anchor — capture shim, per-impl adapters, verify --anchor, 3-way numerics comparison 2026-07-22 14:18:57 -07:00
SolitaryThinker ad1dd41c99 [docs]: tighten card docstring to standard terms (serializable, content-addressed) 2026-07-22 13:54:56 -07:00
SolitaryThinker ceb5a24a30 [docs]: spell out what JSON round-trip and the content digest guarantee (and don't) on cards 2026-07-22 13:51:01 -07:00
SolitaryThinker ae99a82ccf [misc]: GB200 evidence — blessed fingerprints, T0-T3 ledger records, 50-step sample (md5-identical to reference) 2026-07-22 07:45:26 -07:00
SolitaryThinker 89a55be5f5 [fix]: card-declared compute dtypes — preserve loader fp32 islands, never introspect mixed-dtype modules 2026-07-22 07:39:09 -07:00
SolitaryThinker 964769efb7 [fix]: force card-declared dtypes after loading (torch_dtype renamed upstream; loader default must not win) 2026-07-22 07:34:39 -07:00
SolitaryThinker e10d45d8be [feat]: fastvideo2 MVP — data cards, driven loop, enforced pipelines, wan2.1 reference oracle, tiered verifier + evidence ledger 2026-07-22 07:24:24 -07:00
SolitaryThinker 7cd5d02104 [misc]: strip repo for the v2.1 clean-slate MVP branch 2026-07-22 07:24:14 -07:00
pkisfaludi-nvandClaude Opus 4.8 9fb74b9732 Make LTX-2 RMSNorm out-of-place so torch_tensorrt + Ulysses SP compiles (#1623)
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-07-22 04:32:04 -07:00
Adriel FungandSolitaryThinker 521dee0e82 [perf]: enable per-block torch.compile for LTX2 with persistent cache (#1602)
Co-authored-by: SolitaryThinker <wlsaidhi@gmail.com>
2026-07-20 15:59:13 -07:00
William Lin 229419208e [ci]: opt in to fork-PR head checkout after actions/checkout guard change (#1625) 2026-07-20 13:12:22 -07:00
William Lin 65f3b946b9 [feat] Add LTX-2 and LTX-2.3 fine-tuning to the modular trainer (#1624) 2026-07-20 11:58:39 -07:00
Zhang Peiyuan 191fcbf46c [feat] add qat docs (#1621) 2026-07-19 18:37:47 -07:00
Junda Su 755a4e4470 [new-model] Add LingBot-Video Dense and MoE/refiner T2V inference (#1595) 2026-07-18 20:18:16 -07:00
9709b7513b [feat] Port NVFP4 QAT/QAD to modular train framework (#1619)
Co-authored-by: Peiyuan Zhang <email>

Co-authored-by: Peiyuan <a>
2026-07-18 15:01:09 -07:00
William Lin 32cd603515 Revert docs trusted-branch-only workflow (#1618) 2026-07-17 00:22:56 -07:00
Junda Su d4bdd3621a [new-model] Port LingBot-World-v2 (#1579) 2026-07-16 19:51:53 -07:00
William Lin e2f8322842 [ci]: skip unused Buildkite submodule checkout (#1614) 2026-07-16 18:37:57 -07:00
6966f9e0bc [fix] Z-Image (#1236) draft port: rebase + strict-load contract + bf16 encoder parity + PORT_STATUS (#1339)
Co-authored-by: Mrinaal Dogra <mdogra@ucsd.edu>
Co-authored-by: SolitaryThinker <wlsaidhi@gmail.com>
2026-07-16 18:18:25 -07:00
Mac Lee 743f4ed5f9 [ci]: harden Modal repository checkout (#1590) 2026-07-16 18:05:50 -07:00