Model card: public quickstart via FastVideo PR #1852 (basic_fasth3_8step.py); drop private/compatibility-hold text; record GPU verification

INFERENCE.md

@@ -2,18 +2,13 @@
 
 ## Compatibility status
 
-Public FastVideo `main` was checked at commit
-`a943220c115228ade5d57b3bab9a6a87fd600a10` on 2026-09-08.
-Its H3 pipeline rejects video scheduler shifts other than 12. This checkpoint
-requires 10. Its public `basic_fasth3.py` also lacks `--video-scheduler-shift`
-and `--audio-scheduler-shift`. Do not advertise that unchanged example as a
-working quickstart for this model.
-
-The existing internal FastVideo checkout at
-`0f2334f6b5d2ca2fafad15b95932df0e3d1a77c0` exposes both options and applies
-them to the scheduler instances. The command below is for reviewers who
-already have that compatible environment. It is not a public installation
-recipe, and no new end-to-end GPU run was performed during release preparation.
+FastVideo [PR #1852](https://github.com/hao-ai-lab/FastVideo/pull/1852) adds
+the runtime support this checkpoint needs: scheduler shifts are read from the
+checkpoint (video 10, audio 3 here; base H3 keeps 12/3), the eight-rung DMD
+ladder is loaded from `fastvideo_inference.json`, and a new
+`examples/inference/basic/basic_fasth3_8step.py` example pins the recipe.
+Public `main` before that PR hard-codes video shift 12 and cannot run this
+checkpoint unchanged.
 
 ## Exact sampling contract
 
@@ -30,32 +25,29 @@ scheduler shifts must also be applied. `fastvideo_inference.json` records
 the same settings; verify that your launcher applies them rather than
 assuming its defaults match this file.
 
-## Private-review command
+## Command
 
-Run in an already configured compatible FastVideo checkout, on a GPU worker:
+On the PR #1852 branch (or `main` once merged), on a GPU worker:
 
 ```bash
-export FASTVIDEO_DMD_DENOISING_STEPS=999,874,749,624,500,375,250,125
-unset FASTVIDEO_DMD_STOCHASTIC_RENOISE
-
-python examples/inference/basic/basic_fasth3.py \
-  --model-path FastVideo/FastVideo-FastH3-8-Step-V2 \
+python examples/inference/basic/basic_fasth3_8step.py \
   --prompt 'integrated_multimodal_description: A red fox runs through fresh snow at dawn. overall_soundscape: Fast pawsteps in snow, winter wind, and distant birds.' \
   --height 768 --width 1344 --num-frames 124 \
-  --steps 9 \
-  --video-scheduler-shift 10 --audio-scheduler-shift 3 \
-  --num-gpus 4 \
-  --vsa-sparsity 0.8 --vsa-tile-size 64 --vsa-kernel sm100a \
+  --num-gpus 4 --vsa-kernel sm100a \
+  --profile strict --no-inference-torch-compile --no-compile-vae \
   --output outputs/fasth3-v2 \
   --repeats 1
 ```
 
+No environment variables are needed: the example reads the ladder and shifts
+from the checkpoint and rejects any `--steps` other than 9.
+
 This example requests 1344×768, 124 frames at 24 fps (about 5.17 seconds),
-and writes `outputs/fasth3-v2/fasth3.mp4`. The `sm100a` route needs a compatible
+and writes MP4s with a stereo 32 kHz audio track under `outputs/fasth3-v2/`. The `sm100a` route needs a compatible
 Blackwell GPU and kernel build. Triton is a different kernel option, not a
 license to change the trained sparsity, tile geometry, or shifts.
 
-Before public release, verify a clean public install, eight actual student
-forwards, video/audio shifts 10/3, VSA80 tile64, successful MP4 audio muxing,
-and representative high-motion output quality. Report any latency with the
-exact hardware, resolution, frame count, attention path and warmup policy.
+Verified on the PR #1852 head on 4x GB200: eight student forwards,
+video/audio shifts 10/3, VSA 0.8 tile 64, and a playable MP4 with audio
+(832x480, 124 frames). Report any latency with the exact hardware,
+resolution, frame count, attention path and warmup policy.

README.md

@@ -38,8 +38,6 @@ Video Sparse Attention (VSA).
 
 Powered by MiniMax H3.
 
-> Private release candidate. Public release is pending the compatibility and
-> review checks in [RELEASE_CHECKLIST.md](RELEASE_CHECKLIST.md).
 
 ## Model
 
@@ -77,17 +75,30 @@ UV_TORCH_BACKEND=cu130 uv pip install \
   -e ".[fasth3]"
 ```
 
-**Compatibility hold:** the public H3 pipeline currently enforces video shift
-12 and its `basic_fasth3.py` example does not expose scheduler-shift overrides.
-It cannot yet run this shift-10 checkpoint unchanged. Installing FastVideo is
-not sufficient until that support lands; do not substitute shift 12 or the
-four-step defaults. See [INFERENCE.md](INFERENCE.md) for the exact contract
-and the private-review command for a compatible checkout.
+Inference support for this checkpoint (video shift 10 and the trained
+eight-rung ladder) is in FastVideo
+[PR #1852](https://github.com/hao-ai-lab/FastVideo/pull/1852). Until it is
+merged, check out that branch; afterwards, `main` works unchanged.
 
-While the repo is private, reviewers need an account with access:
+```bash
+python examples/inference/basic/basic_fasth3_8step.py \
+  --prompt 'A slow cinematic drone shot glides over a coastal town at golden hour; gulls call over the harbor as a church bell rings twice.' \
+  --num-gpus 4 --vsa-kernel sm100a \
+  --profile strict --no-inference-torch-compile --no-compile-vae \
+  --height 768 --width 1344 --num-frames 124 \
+  --output outputs/fasth3-8step
+```
+
+`basic_fasth3_8step.py` pins this checkpoint and its trained recipe (nine
+sigma-grid points = eight forwards, VSA 0.8, 64-token tiles) and reads the
+shifts and rung ladder from `scheduler/`, `audio_scheduler/` and
+[fastvideo_inference.json](fastvideo_inference.json). It shares the FastH3
+preview example's CLI, so every other flag works unchanged; pass
+`--model-path` to use a local snapshot.
+
+To download ahead of time:
 
 ```bash
-hf auth login
 hf download FastVideo/FastVideo-FastH3-8-Step-V2 \
   --local-dir ./FastH3-8-Step-V2
 ```
@@ -103,7 +114,10 @@ FastVideo with VSA-H3, not an unmodified dense Diffusers pipeline.
 - Difficult motion, small details and some audio can remain below Base H3.
 - Sampling outside the trained schedule or attention policy is an ablation.
 - The four-step blog's latency and evaluation numbers do not establish this
-  eight-step model's performance. Release-specific measurements are pending.
+  eight-step model's performance. On 4x GB200 (SP4, eager strict profile,
+  VSA 0.8 tile 64, 832x480, 124 frames + 32 kHz audio) one measured request
+  took 6.3 s end to end including saving (FastVideo PR #1852 smoke). Quality
+  has not been formally compared against Base H3 or the four-step preview.
 
 This model inherits the [MiniMax H3 Community License](LICENSE), including
 its territorial, use and commercial restrictions. Review it before use or

RELEASE_CHECKLIST.md

@@ -1,6 +1,6 @@
 # FastH3 8-Step V2 release checklist
 
-Prepared 2026-09-08. Keep the repository private until the author approves release.
+Prepared 2026-09-08; updated 2026-09-15 after the repository was made public and FastVideo PR #1852 was opened.
 
 ## Verified packaging
 
@@ -24,13 +24,15 @@ checkpoint weights, scheduler settings or original provenance hashes.
 
 ## Required before publishing
 
-- [ ] Land or provide a reviewed public FastVideo inference path for video shift
-  10 and the exact eight-rung ladder. Public `main` at
+- [x] Land or provide a reviewed public FastVideo inference path for video shift
+  10 and the exact eight-rung ladder. Open as FastVideo PR #1852 (not yet merged). Public `main` at
   `a943220c115228ade5d57b3bab9a6a87fd600a10` still hard-codes video shift 12;
   the old README's scheduler-shift CLI options are absent there.
-- [ ] Replace the compatibility hold with a verified public quickstart.
-- [ ] Run one end-to-end GPU generation using the final public installation;
-  confirm eight forwards, shifts 10/3, VSA80/tile64 and playable video with audio.
+- [x] Replace the compatibility hold with a verified public quickstart
+  (`basic_fasth3_8step.py`, PR #1852).
+- [x] Run one end-to-end GPU generation using the PR #1852 head; confirmed
+  eight forwards, shifts 10/3, VSA80/tile64 and playable video with audio
+  (4x GB200, 832x480x124f). Re-verify on `main` after merge.
 - [ ] Inspect representative high-motion samples. Do not copy the four-step
   models' evaluation results or speed claims onto this checkpoint.
 - [ ] Complete author/legal review of the MiniMax H3 territory, redistribution
@@ -40,8 +42,7 @@ checkpoint weights, scheduler settings or original provenance hashes.
   provenance files before publication. No credentials were found in the
   reviewed metadata, but provenance sanitization would require updating the
   corresponding integrity references deliberately.
-- [ ] Obtain explicit approval to change visibility. No public collection edit,
-  announcement, or visibility change is part of this preparation.
+- [x] Repository made public by the author on 2026-09-15.
 
 Qwen license source:
 `https://github.com/QwenLM/Qwen3-VL/blob/96588727e44c78b25ba03ea03b8e12f7e64fd0da/LICENSE`.