-
e4b14e124d
[feat] v2 Wan loop+adapter: optional i2v conditioning hooks (cond concat + CLIP context); T2V unchanged
Will Lin
2026-06-18 11:42:35 +00:00
-
60ca94796c
[feat] v2 torch backend: CLIP image-encoder adapter + image_encoder component kind (i2v groundwork)
Will Lin
2026-06-18 11:35:18 +00:00
-
060bdd85b6
[revert] v2: drop FastWan/VSA registry entries — generic Wan loader can't map their gated-attn params
Will Lin
2026-06-18 11:31:31 +00:00
-
0a18b6a031
[feat] v2: alias FastVideo/Wan2.1-VSA-T2V-14B-720P to the Wan-14B card (bucket B)
Will Lin
2026-06-18 11:19:17 +00:00
-
2e5bad1d6c
[feat] v2: port Wan2.1-T2V-14B (bucket B) + document the all-models backlog
Will Lin
2026-06-18 11:10:05 +00:00
-
705f580542
[refactor] v2 torch backend: TorchComponent base + one generic builder + v2.* facade (Phase 1b/1c)
Will Lin
2026-06-18 11:01:45 +00:00
-
cbe67337c3
[feat] v2: per-model sampling defaults on ModelCard (Phase 1a)
Will Lin
2026-06-18 10:47:28 +00:00
-
113478fd74
[refactor] v2: rename models/ (recipe layer) -> recipes/; move toy backend -> platform/backends/toy.py
Will Lin
2026-06-18 10:35:41 +00:00
-
9b2007c058
[refactor] v2: use absolute imports (v2.*) everywhere instead of relative
SolitaryThinker
2026-06-18 09:28:52 +00:00
-
1542b7a03d
[feat] v2 VideoGenerator: A/V convenience path (generate_video -> T2VS -> mp4 + 24kHz wav)
SolitaryThinker
2026-06-18 08:59:43 +00:00
-
e4b7261b66
[feat] v2 LTX-2.3 T2VS GPU-verified: audio VAE/vocoder wiring + dual-connector audio fix
SolitaryThinker
2026-06-18 08:44:07 +00:00
-
b39b68c6f5
[feat] v2 LTX-2.3 T2VS: single-stage joint audio+video card/loop/program (CPU-verified)
SolitaryThinker
2026-06-18 08:36:59 +00:00
-
5f698f03f5
[feat] v2 LTX-2 adapters: A/V foundation (joint DiT forward + dual text connector + audio decode)
SolitaryThinker
2026-06-18 08:23:01 +00:00
-
799a9b69a2
[refactor] v2: shared model registry (HF-id primary + arch fallback) for all entrypoints
SolitaryThinker
2026-06-18 07:57:55 +00:00
-
27b320afd8
[fix] v2: name LTX-2 cards by architecture (2-stage vs single-stage) + Wan2.1 is T2V-only
SolitaryThinker
2026-06-18 07:46:09 +00:00
-
c74e9cb4fe
[docs] v2: LTX-2 base/2.3 GPU-verified on rebuilt x86 stack + remaining-port mechanisms
SolitaryThinker
2026-06-18 07:13:28 +00:00
-
0a2c985b56
[docs] v2: LTX-2.3 example + roadmap (A14B offload working, base/2.3 ported, env status)
SolitaryThinker
2026-06-18 06:27:16 +00:00
-
8b343db604
[feat] v2: Wan2.2-A14B MoE CPU offload (fits 1 GPU) + LTX-2 base/2.3 single-stage port
SolitaryThinker
2026-06-18 06:24:15 +00:00
-
24b76d372c
[feat] v2: architecture-driven dispatch + real LTX-2 upsampler + Wan2.2-A14B MoE card
SolitaryThinker
2026-06-18 05:41:42 +00:00
-
f7256ab197
[feat] v2 port: Wan2.2-TI2V-5B (T2V) — 4th GPU-verified model
SolitaryThinker
2026-06-18 04:55:29 +00:00
-
9de0e2bcd1
[feat] v2 VideoGenerator: convenience API (from_pretrained/generate_video) + porting roadmap
SolitaryThinker
2026-06-18 04:33:56 +00:00
-
5f33f80d6c
[feat] v2 VideoGenerator: typed fastvideo.api entrypoint over the v2 engine
SolitaryThinker
2026-06-18 03:56:17 +00:00
-
e7d4aaeef4
[feat] v2 ltx2: two-stage distilled GPU bring-up (LTX2Transformer3DModel 18.88B)
SolitaryThinker
2026-06-18 02:51:39 +00:00
-
2a085a5c39
[feat] v2 wan_causal: causal DiT (CausalWanTransformer3DModel) GPU bring-up
SolitaryThinker
2026-06-18 02:34:41 +00:00
-
261d3007a9
[fix] v2 tests: make no-torch-import guards GPU-aware (skipif torch installed)
SolitaryThinker
2026-06-17 23:30:17 +00:00
-
b2da3b2b4d
[feat] v2 wan21: backend-aware latent geometry + checkpoint stamping
SolitaryThinker
2026-06-17 23:30:17 +00:00
-
7ec4d01afc
[feat] v2 cuda backend: real fastvideo construction + risk A-E fixes (Wan2.1 verified on H100)
SolitaryThinker
2026-06-17 23:30:17 +00:00
-
8b0e2bbb4b
[docs] add v2/HANDOFF.md for GPU-side bring-up of the torch backend
SolitaryThinker
2026-06-17 14:24:12 -07:00
-
7fd58f26ea
[fix] correct GPU adapters against real fastvideo API (cross-check findings)
SolitaryThinker
2026-06-17 14:16:23 -07:00
-
15b6278e9f
[feat] real torch/CUDA backend (written-not-run) behind the cuda cells
SolitaryThinker
2026-06-17 13:52:28 -07:00
-
cd0fbbd3a0
[feat] static-buffer capture form for the cudagraph step body (Path A)
SolitaryThinker
2026-06-17 13:28:56 -07:00
-
02151584cd
[feat] piecewise CUDA-graph capture/replay at the step boundary (Path A)
SolitaryThinker
2026-06-17 13:14:40 -07:00
-
51bb5e349a
[feat] route all diffusion loops + RL recompute through the kernel table
SolitaryThinker
2026-06-17 12:55:52 -07:00
-
1bdbf1d5b3
[feat] multi-backend dispatch substrate (device/arch/kernel registries)
SolitaryThinker
2026-06-17 12:26:11 -07:00
-
de295403c7
[feat] Adapter plane, non-linear workflows, RL→distill flywheel
SolitaryThinker
2026-06-16 23:51:32 -07:00
-
34834eb4f4
[feat] #8b speculative (draft-verify) decoding — exact + lower-latency AR
SolitaryThinker
2026-06-16 23:32:01 -07:00
-
bd36d69680
[feat] Five more stress tests: LTX-2 A/V, weight-sync, served reward, cache-dit, nested workflows
SolitaryThinker
2026-06-16 23:20:23 -07:00
-
94b84b03b9
[examples] Add v2_examples/{training,omni,workflows}/ — runnable examples
SolitaryThinker
2026-06-16 22:13:43 -07:00
-
a16ce1a07f
[examples] Add v2_examples/inference/ — runnable Wan2.1 inference examples
SolitaryThinker
2026-06-16 22:04:56 -07:00
-
37544c68a5
[docs] Update v2/README to designv4 + current scope (127 tests)
SolitaryThinker
2026-06-16 21:24:25 -07:00
-
88e4a4355c
[feat] Three stress tests: interactive sessions, workflow RL, heterogeneous co-scheduling
SolitaryThinker
2026-06-16 19:44:40 -07:00
-
f1dead1361
[feat] Register cross-model workflows as first-class named servables
SolitaryThinker
2026-06-16 19:13:36 -07:00
-
cfc2ad7487
[feat] Cross-model T2I→I2V workflow + N-way joint RL over arbitrary experts
SolitaryThinker
2026-06-16 18:38:10 -07:00
-
97fedaf432
[feat] Add Qwen-Omni thinker→talker→vocoder model (3 experts, 3 loops)
SolitaryThinker
2026-06-16 15:00:12 -07:00
-
3ec511419f
[feat] UniRL/PromptRL joint LM+generator RL stress test + designv4
SolitaryThinker
2026-06-16 13:37:18 -07:00
-
145082cf74
[refactor] rename package mini_fastvideo → v2
SolitaryThinker
2026-06-16 10:57:01 -07:00
-
71caa58ba9
[fix] mini-fastvideo serving: address adversarial-review findings (capacity/credit leaks, robustness)
SolitaryThinker
2026-06-16 09:53:34 -07:00
-
8240970ea6
[feat] mini-fastvideo serving + fleet (our own version, Dynamo-optional)
SolitaryThinker
2026-06-16 09:25:12 -07:00
-
44d10b5c79
[feat] mini-fastvideo phase 2: omni/MoT — Cosmos3 + canonical vllm-omni (BAGEL/lance)
SolitaryThinker
2026-06-16 04:40:10 -07:00
-
9d874aa609
[fix] mini-fastvideo: address adversarial-review findings (admission liveness + §7.1 cache key)
SolitaryThinker
2026-06-16 04:33:53 -07:00
-
8e828680c6
[feat] mini-fastvideo: model-native runtime per design_v3 (Wan2.1/LTX2 + 4 training methods)
SolitaryThinker
2026-06-16 04:21:10 -07:00
-
735ee6aa56
update
SolitaryThinker
2026-06-15 16:09:35 -07:00
-
068a532b31
design
SolitaryThinker
2026-06-13 15:39:21 -07:00
-
-
-
e73521d04e
[misc] record InterleaveThinker validation handoff
Mac Lee
2026-06-18 02:47:52 +00:00
-
0c56a4aa53
[feat] add interleave trace runner
Mac Lee
2026-06-18 02:44:01 +00:00
-
eb33639cbd
[feat] add InterleaveThinker compatibility service
Mac Lee
2026-06-18 02:40:54 +00:00
-
7e70f8598a
[misc] track InterleaveThinker integration plan
Mac Lee
2026-06-18 02:28:24 +00:00
-
87f98c9b8b
[kernel] Add varlen support for block-sparse attention (#1319)
sumyyyyy
2026-06-18 09:06:00 +08:00
-
e60601df7f
[feat] QAD 5090: QAT training recipe — finetune + DMD distillation (12/12) (#1462)
2026-06-17 15:18:52 -07:00
-
eed9c4bfbf
[kernel] QAD 5090: Add Attn-QAT training Triton kernels (11/12) (#1460)
2026-06-17 13:35:12 -07:00
-
1dee77f4a4
[feat] QAD 5090: Wire the Attn-QAT training attention backend (10/12) (#1459)
2026-06-17 13:32:56 -07:00
-
b80148819c
[ci] Add performance dashboard visualisation scripts (#1469)
Satyam Srivastava
2026-06-16 23:29:29 -07:00
-
c3b971488e
[docs] QAD 5090: Add NVFP4 + Attn-QAT inference example and how-to (9/12) (#1458)
2026-06-16 14:36:50 -07:00
-
88e753f281
[feat] QAD 5090: Wire the Attn-QAT inference attention backend (8/12) (#1457)
2026-06-16 13:12:33 -07:00
-
77832059cc
[kernel] QAD 5090: Add modified SageAttention3 FP4 inference kernels (7/12) (#1455)
2026-06-16 12:02:30 -07:00
-
-
5671ebc718
[perf] LTX-2.3 NVFP4 bench: VAE-tiling / latent-output knobs + long-token docs
alex/opt-ltx23-nvfp4
alexzms
2026-06-15 16:55:13 -07:00
-
5a01741ea0
update
will/design
SolitaryThinker
2026-06-15 16:09:35 -07:00
-
57a2fa9ec1
[perf] LTX-2.3 distilled NVFP4 + FA4-FP4 + CUDA-graphs benchmark on GB300
alexzms
2026-06-14 11:57:10 -07:00
-
-
-
de4d430d87
design
SolitaryThinker
2026-06-13 15:39:21 -07:00
-
-
-
ee607bd0d0
feat(api): typed omni request plane (OmniRequest/Output/Event)
will/f_0
SolitaryThinker
2026-06-13 01:36:24 -07:00
-
-
-
633d393568
[ci] layer-0 grad-norm regression for per-method training tests (5a-ii) (#1396)
alexzmsandmergify[bot]
2026-06-11 21:45:07 -07:00
-
5854aec2ce
[feat] Add Wan RL DiffusionNFT training (#1450)
Junda SuandPeiyuan Zhang
2026-06-11 21:18:59 -07:00
-
30e45c2411
[bugfix] Classify new config/sampling fields in schema parity inventory (#1446)
reason
Mook
2026-06-10 13:38:36 -07:00
-
2a4fe697a6
[docs] LTX-2.3 distilled i2v: typed-API example (from_config + generate) (#1448)
alexzms
2026-06-10 10:58:14 -07:00
-
921db7479d
[perf] LTX-2.3 distilled i2v: drop max-autotune from compile kwargs (#1445)
alexzms
2026-06-10 10:06:41 -07:00
-
7f539424cb
[feat]: add Lucy Edit inference scaffold (#1363)
Aryan KumarandAryan Kumar
2026-06-09 15:21:43 -07:00
-
19a838f54f
[bugfix]: release VSA tile cache during training (#1434)
2026-06-09 15:01:22 -07:00
-
d922ab2cbc
[model] Flux2 Klein Port (#1349)
2026-06-09 14:55:55 -07:00
-
9ea77d37f3
[bugfix] EMA shadow on resume and EMA under MoE path (#1441)
Kaiqin Kongandmergify[bot]
2026-06-08 17:55:43 -07:00
-
2e35b0c6bd
[refactor]: linear/mlp FP4 path additions for Wan-2.1 (Attn-QAT 6/12) (#1390)
2026-06-08 14:51:57 -07:00
-
a520eceed7
document vendored reward runtimes in agent memory
py/add_rl
Davids048
2026-06-08 20:22:13 +00:00
-
b2c75534ac
fix mypy issues for vendored reward runtimes
Davids048
2026-06-08 20:07:46 +00:00
-
0b9c325984
fix mypy issues in RL integration files
Davids048
2026-06-08 19:54:35 +00:00
-
2afa837d96
handle yapf and ruff issues
Davids048
2026-06-08 19:22:11 +00:00
-
5b1c779041
[misc]: simplify cosmos3 action (fold action split into _split_flat_latent)
feat/cosmos3-action
SolitaryThinker
2026-06-07 16:42:15 -07:00
-
c6f15de8c7
[feat] cosmos3 PR3: action (domain-aware) DiT pathway + packing + CFG glue
SolitaryThinker
2026-06-07 13:12:23 -07:00
-
245515f653
[misc]: simplify cosmos3 audio (fold sound split into _split_flat_latent)
feat/cosmos3-audio
SolitaryThinker
2026-06-07 16:37:55 -07:00
-
b24ee58f8d
[docs] cosmos3 PR2 (audio/t2vs) complete; bit-exact across components
SolitaryThinker
2026-06-07 12:56:24 -07:00
-
62a0a0bbfb
[feat] cosmos3 PR2: t2vs pipeline (joint denoise + AVAE decode + AV mux)
SolitaryThinker
2026-06-07 12:55:13 -07:00
-
0245218e34
[feat] cosmos3 PR2: DiT sound pathway + sound packing (bit-exact parity)
SolitaryThinker
2026-06-07 12:29:17 -07:00
-
14147c57c2
[docs] cosmos3 PR2: AVAE sound decoder done; remaining audio components
SolitaryThinker
2026-06-07 09:51:51 -07:00
-
4fda151b7a
[feat] cosmos3 PR2: native AVAE sound decoder + bit-exact parity
SolitaryThinker
2026-06-07 09:50:51 -07:00
-
3d74fe1896
[docs] cosmos3 PR2: audio port plan (AVAE + DiT sound pathway + t2vs)
SolitaryThinker
2026-06-07 09:36:30 -07:00
-
06321bb364
[docs] cosmos3: record T2I verification + resolution-based flow_shift
feat/cosmos3-i2v
SolitaryThinker
2026-06-07 09:33:02 -07:00
-
0a5d9ecbc7
[feat] cosmos3: resolution-based flow_shift + T2I check & example
SolitaryThinker
2026-06-07 09:32:09 -07:00
-
744ca86ced
[docs] cosmos3: record I2V real-weights verification (feat/cosmos3-i2v)
SolitaryThinker
2026-06-07 04:46:06 -07:00
-
a5be058cf1
[feat] cosmos3: I2V image conditioning (static-repeat) + parity + example
SolitaryThinker
2026-06-07 04:45:32 -07:00
-
1a7e5b1665
[misc]: simplify cosmos3 port (reuse embed, dedup prompt, hoist)
feat/cosmos3-tier-a-port
SolitaryThinker
2026-06-07 16:33:45 -07:00
-
77488851b4
[skill] add-model: fold in Cosmos3 omni-port lessons
feat/add-model-skill-revision
SolitaryThinker
2026-06-07 14:00:02 -07:00
-
-
-
255721c1c2
[docs] cosmos3: record real-weights E2E acceptance + scheduler fix (I003)
SolitaryThinker
2026-06-07 01:33:28 -07:00