feat: fastwan22_5b family — FastWan2.2-TI2V-5B-FullAttn (recommended recipe) - #27
Merged
Merged
Conversation
…recipe) New flat videogen family cloned from fastwan/ (QAD), serving FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers per FastVideo's recommended recipe: bf16, 3-step DMD CFG-free, full Wan2.2 VAE decode, 1280x704x121@24fps (+ portrait transpose), no quantization, no TAEHV, no torch.compile (each a possible later, separately quality-gated delta). Key difference vs fastwan/: the factory PINS pipeline_config to FastWan2_2_TI2V_5B_Config. At the anonymized /data/default serving mount, fastvideo's path-based preset resolution falls through to detectors, and this model's model_index _class_name (WanDMDPipeline) matches the FastWan2.1 detector -> silently selects the 480p Wan2.1 config. Pinning + a boot-time check turns that failure loud. Menu hash e0c0ed3c (2 shapes). Bench basis (H100, warm): full-VAE 16.9s/clip; denoise 3x1.36s/it. VRAM for the K=2 pool on one 80GB H100 is flagged for bake-time validation in shapes.json. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Johan-de-R
temporarily deployed
to
external_collaborator
August 1, 2026 00:54 — with 
GitHub Actions
Inactive
GitHub Actions
Inactive
…ockerfile Fresh builds from the plain Dockerfile install the XOR-op fork, which is stale against cutlass-dsl 4.5+ (cute.core.ThrMma AttributeError at import) and kills any worker that selects FLASH_ATTN -- caught by the fastwan22_5b bake validation (both shapes died before READY). Ports the exact fix Dockerfile.dreamverse has carried since the LTX-2.3 FA4 work: pin Dao-AILab/flash-attention@82d6441e (FastVideo#1564's own resolution) and apply the [:2] wrapper patch for its 4-tuple returns. Older images (QAD a25aab8) predate the cutlass drift and are unaffected. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Johan-de-R
had a problem deploying
to
external_collaborator
August 3, 2026 17:01 — with
GitHub Actions
Failure
GitHub Actions
Failure
The pin declares nvidia-cutlass-dsl==4.6.0.dev0 (pre-release), which uv refuses to resolve; the venv's 4.6.0 release already satisfies it. Verified in-container: install + flash_attn.cute.interface import clean. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Johan-de-R
had a problem deploying
to
external_collaborator
August 3, 2026 17:04 — with
GitHub Actions
Failure
GitHub Actions
Failure
…mpile
Johan quality-approved the tiny decoder via seed-locked split-screen A/B
vs full-VAE ('equally good to the eye'); ship target moves to 1xB200
(~3.0s/clip vs 9.8s full-VAE, measured). Factory: output_type=latent +
CPU-offloaded VAE + taew2_2_super decode wrapper (width/height/frames
plumbed into SamplingParam — the QAD portrait lesson) + compile on.
Config pin unchanged and empirically verified at /data/default.
Dockerfile: gated pinned-commit fetch of taehv.py + taew2_2_super.pth
(HF mirror is private; GitHub raw at a pinned ref).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Johan-de-R
had a problem deploying
to
external_collaborator
August 3, 2026 20:33 — with
GitHub Actions
Failure
GitHub Actions
Failure
(480p|720p) x (landscape|portrait), all 121f@24fps/3-step. 480p is off the distillation distribution: launch inclusion is gated on quality review of the bake-session probe clips + the measured 4-shape pool VRAM fit (expected ~115GB/180GB B200). Falls back to the 2-shape 720p menu (hash e0c0ed3c) if either gate fails. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Johan-de-R
had a problem deploying
to
external_collaborator
August 3, 2026 21:21 — with
GitHub Actions
Failure
GitHub Actions
Failure
Johan-de-R
marked this pull request as ready for review
August 3, 2026 21:46
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
New videogen family serving FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers (720p+480p T2V, 5s@24fps, 3-step DMD CFG-free), plus two Dockerfile fixes that affect all future video-image builds.
Final shipped config (evolved during review-ready validation; commit trail tells the story):
output_type=latent, VAE CPU-offloaded) + torch.compile — Johan quality-approved via seed-locked split-screen A/B vs full-VAE; ~3.1s/clip 720p, ~1.4s 480p on 1xB200 through the full serving path4fbb8f99; measured pool peak 46.6GB/180GBpipeline_config=FastWan2_2_TI2V_5B_Config: at the anonymized/data/defaultmount, fastvideo's path-based preset detection falls through to the WanDMDPipeline detector and silently selects the FastWan2.1 480p config. Pin + boot-time check make it loud. Empirically verified in the ship image:RESOLVED_CONFIG=FastWan2_2_TI2V_5B_Config flow_shift=5.0 dmd=[1000,757,522]. (Same latent trap likely affects the deployedfastwanQAD family — follow-up audit tracked separately.)cute.core.ThrMmaAttributeError kills any FLASH_ATTN worker; this is FastVideo#1564's own fix, previously only in Dockerfile.dreamverse) + gated pinned-commit fetch oftaew2_2_super.pth(HF mirror went private; GitHub raw at pinned ref).Deploy state: the branch tip is byte-identical to the serving image
fastvideo-runtime:2.1.5-fastwan22-5b-b200-4fbb8f99(private model, live on 1xB200) — merging changes nothing in prod; it makes the repo match what's deployed. If review requires code changes, the image gets rebuilt from the merged result.Backend counterpart: deepinfra/backend#3940. Full evidence trail: backend
claude_plans/2026-07-28-fastwan22-5b-ship.md.🤖 Generated with Claude Code