Publish CUDA 13 inference profiles and ComfyUI proof

.gitattributes

@@ -50,3 +50,6 @@ comfyui/examples/hevc/comfyui-t2va-608x480.mp4 filter=lfs diff=lfs merge=lfs -te
 comfyui/examples/timelines/comfyui-ref2va-608x480.png filter=lfs diff=lfs merge=lfs -text
 comfyui/examples/timelines/comfyui-t2va-608x480.png filter=lfs diff=lfs merge=lfs -text
 comfyui/workflow-export.png filter=lfs diff=lfs merge=lfs -text
+comfyui/examples/**/*.mp4 filter=lfs diff=lfs merge=lfs -text
+comfyui/examples/**/*.png filter=lfs diff=lfs merge=lfs -text
+comfyui/examples/**/*.jpg filter=lfs diff=lfs merge=lfs -text

README.md

@@ -16,172 +16,209 @@ tags:
   - audio-video-generation
 ---
 
-# MiniMax H3 — native OrbitQuant W4A4, source FP32 VAEs
+# MiniMax H3 — OrbitQuant W4A4 with source FP32 VAEs
 
-OrbitQuant conversion of [MiniMaxAI/MiniMax-H3](https://huggingface.co/MiniMaxAI/MiniMax-H3), pinned to source revision [`73372e6cf53e414edd3ab03e357717fb0602e758`](https://huggingface.co/MiniMaxAI/MiniMax-H3/tree/73372e6cf53e414edd3ab03e357717fb0602e758).
+OrbitQuant conversion of
+[MiniMaxAI/MiniMax-H3](https://huggingface.co/MiniMaxAI/MiniMax-H3), pinned to
+source revision
+[`73372e6cf53e414edd3ab03e357717fb0602e758`](https://huggingface.co/MiniMaxAI/MiniMax-H3/tree/73372e6cf53e414edd3ab03e357717fb0602e758).
 
-Eligible linear weights in `transformer`, `transformer_ref`, and the Qwen3-VL `text_encoder` are stored and executed through OrbitQuant's native packed W4A4 path. BF16 is the compute dtype for non-quantized modules and W4 activations; it is not a full-weight reconstruction. The visual `vae` and `audio_vae` are byte-for-byte FP32 source copies and are never quantized.
+Eligible linear weights in `transformer`, `transformer_ref`, and the Qwen3-VL
+`text_encoder` are stored and executed through OrbitQuant's native packed W4A4
+path. Non-quantized boundaries use BF16 compute. The visual `vae` and
+`audio_vae` are byte-for-byte FP32 source copies and are never quantized.
 
-## ComfyUI workflow
-
-Download the ready-to-import [MiniMax H3 OrbitQuant T2VA workflow](comfyui/workflows/MiniMax-H3-OrbitQuant-T2VA.json). It is derived from Comfy-Org's bundled [`video_minimax_h3_t2v.json`](https://github.com/Comfy-Org/workflow_templates/blob/7653f1cdef1d92394b6ef9946018c0a8aa4136b8/templates/video_minimax_h3_t2v.json), preserving its human-readable composition: model and usage notes on the left, generation in the center, and the standard ComfyUI `SaveVideo` output on the right.
-
-![MiniMax H3 OrbitQuant public ComfyUI workflow](comfyui/workflow-export.png)
-
-The 6120×2620 PNG above comes from ComfyUI's `Workflow Image → Export → png`
-action rather than a browser screenshot. It embeds the loadable workflow in its
-`tEXtworkflow` metadata.
+## Final 608×480 example
 
-Install [ComfyUI-OrbitQuant](https://github.com/iamwavecut/ComfyUI-OrbitQuant/tree/feature/minimax-h3-comfyui) in `ComfyUI/custom_nodes`, restart ComfyUI, import the workflow, and set `OrbitQuant Release Loader.model_path` to this downloaded model directory. The tested graph uses the two generic public nodes `OrbitQuant Release Loader` and `OrbitQuant Generate Video`; there are no MiniMax-specific public node classes. Model-specific component selection and offload rules come from [`comfyui_orbitquant.json`](comfyui_orbitquant.json) and an internal allowlisted adapter.
-
-The exact imported graph was executed through ComfyUI `9a9fdb10ed144ce760d9682cb247526ea23cc525` on an RTX PRO 6000 at 608×480, 124 frames, seed 42, and 50 sigma points. Its live-exported [API prompt](comfyui/workflows/MiniMax-H3-OrbitQuant-T2VA-api.json), [machine-readable report](comfyui/report.json), and reviewed T2VA/Ref2VA media are included.
-
-### ComfyUI T2VA — generic public workflow
-
-<video controls playsinline preload="metadata" width="608" aria-label="MiniMax H3 OrbitQuant ComfyUI T2VA proof">
-  <source src="https://huggingface.co/WaveCut/MiniMax-H3-OrbitQuant-W4A4/resolve/main/comfyui/examples/hevc/comfyui-t2va-608x480.mp4" type="video/mp4; codecs=&quot;hvc1&quot;">
-  <source src="https://huggingface.co/WaveCut/MiniMax-H3-OrbitQuant-W4A4/resolve/main/comfyui/examples/h264/comfyui-t2va-608x480.mp4" type="video/mp4">
-  <a href="https://huggingface.co/WaveCut/MiniMax-H3-OrbitQuant-W4A4/resolve/main/comfyui/examples/h264/comfyui-t2va-608x480.mp4">MiniMax H3 OrbitQuant ComfyUI T2VA proof MP4</a>
+<video controls playsinline preload="metadata" width="608" aria-label="MiniMax H3 OrbitQuant balanced ComfyUI proof">
+  <source src="https://huggingface.co/WaveCut/MiniMax-H3-OrbitQuant-W4A4/resolve/main/comfyui/examples/hevc/comfyui-balanced-t2va-608x480-crf10.mp4" type="video/mp4; codecs=&quot;hvc1&quot;">
+  <source src="https://huggingface.co/WaveCut/MiniMax-H3-OrbitQuant-W4A4/resolve/main/comfyui/examples/h264/comfyui-balanced-t2va-608x480.mp4" type="video/mp4">
+  <a href="https://huggingface.co/WaveCut/MiniMax-H3-OrbitQuant-W4A4/resolve/main/comfyui/examples/h264/comfyui-balanced-t2va-608x480.mp4">MiniMax H3 OrbitQuant proof MP4</a>
 </video>
 
-[H.265/HEVC](comfyui/examples/hevc/comfyui-t2va-608x480.mp4) · [H.264 fallback](comfyui/examples/h264/comfyui-t2va-608x480.mp4) · [five-frame timeline](comfyui/examples/timelines/comfyui-t2va-608x480.png)
+[H.265 10-bit CRF 10](comfyui/examples/hevc/comfyui-balanced-t2va-608x480-crf10.mp4)
+· [H.264 fallback](comfyui/examples/h264/comfyui-balanced-t2va-608x480.mp4)
+· [CRF 1 yuv444p master](comfyui/examples/masters/comfyui-balanced-t2va-608x480-crf1.mp4)
+· [16-frame overview](comfyui/examples/timelines/comfyui-balanced-t2va-608x480.jpg)
+· [adjacent-frame review](comfyui/examples/review/comfyui-balanced-adjacent-triplets.jpg)
+· [audio spectrum](comfyui/examples/review/comfyui-balanced-audio-spectrum.png)
 
-### ComfyUI Ref2VA — generic public workflow
-
-<video controls playsinline preload="metadata" width="608" aria-label="MiniMax H3 OrbitQuant ComfyUI Ref2VA proof">
-  <source src="https://huggingface.co/WaveCut/MiniMax-H3-OrbitQuant-W4A4/resolve/main/comfyui/examples/hevc/comfyui-ref2va-608x480.mp4" type="video/mp4; codecs=&quot;hvc1&quot;">
-  <source src="https://huggingface.co/WaveCut/MiniMax-H3-OrbitQuant-W4A4/resolve/main/comfyui/examples/h264/comfyui-ref2va-608x480.mp4" type="video/mp4">
-  <a href="https://huggingface.co/WaveCut/MiniMax-H3-OrbitQuant-W4A4/resolve/main/comfyui/examples/h264/comfyui-ref2va-608x480.mp4">MiniMax H3 OrbitQuant ComfyUI Ref2VA proof MP4</a>
-</video>
+This live ComfyUI run uses 608×480, 124 frames at 24 FPS, seed 42, and 24
+sigma points / 23 denoiser forwards. All 300 eligible denoiser linears use the
+native packed W4A4 path with no exact INT8 weight cache. Source FP32 tiled visual
+decode produced the retained CRF 1 master; the HEVC card copy was derived from
+that master at CRF 10. The output contains AAC stereo at 32 kHz.
 
-[H.265/HEVC](comfyui/examples/hevc/comfyui-ref2va-608x480.mp4) · [H.264 fallback](comfyui/examples/h264/comfyui-ref2va-608x480.mp4) · [five-frame timeline](comfyui/examples/timelines/comfyui-ref2va-608x480.png)
+Full-resolution frames and adjacent triplets were reviewed for face geometry,
+eyes, lips, grid artifacts, ghosting, texture breakup, and abrupt section
+redraw. The macro-to-face shot remains coherent. The audio spectrum is broadband
+without a persistent narrow electronic whistle.
 
-## Working recipe
-
-1. Move the W4 text encoder onto the GPU only for prompt/reference conditioning.
-2. Offload the text encoder to RAM before denoising.
-3. Move the selected W4 denoiser onto the GPU and run native packed W4A4 with BF16 activations.
-4. Use the official default schedule: **50 sigma-grid points / 49 denoiser forwards**.
-5. Persist every scheduler-step checkpoint atomically.
-6. After denoising exits and releases the GPU, decode the saved latents with the original FP32 visual and audio VAEs.
-
-The earlier eight-point shortcut was the principal quality failure: even the untouched BF16 source produced broken faces, lost temporal detail, and damaged audio with that schedule. VAE loading was separately falsified as the cause by decoding identical bad latents with the exact source VAE.
-
-The apparent native-W4 slowdown had a separate cause: a prequantized Diffusers model loaded without an explicit compute dtype left its non-quantized modules in FP32. All 300 OrbitQuant linears then rejected the BF16/FP16 packed W4 kernel and fell back to the generic FP32 path. OrbitQuant 0.9.1 makes BF16 the default compute dtype for prequantized Diffusers artifacts while preserving an explicitly requested dtype. On the exact 50-block, 11,264-token transformer profile this changed one forward from **114.12 s** to **2.87 s** (39.8×); the package-default load measured **2.87 s** with all 300 linears on `triton_cuda_packed_w4`.
-
-![Source, isolated text component, and final T2VA full-timeline comparisons](examples/quality-timeline-matrix.webp)
-
-Rows from top to bottom: source BF16 control, W4A4 text-only isolation control, final native W4A4 text + transformer. Columns sample frames 0, 31, 62, 92, and 123.
+## ComfyUI workflow
 
-## Embedded examples
+Download the ready-to-import
+[MiniMax H3 OrbitQuant T2VA workflow](comfyui/workflows/MiniMax-H3-OrbitQuant-T2VA.json).
+It is based on Comfy-Org's bundled
+[`video_minimax_h3_t2v.json`](https://github.com/Comfy-Org/workflow_templates/blob/7653f1cdef1d92394b6ef9946018c0a8aa4136b8/templates/video_minimax_h3_t2v.json)
+and preserves the official preset's readable composition.
 
-### Source BF16 control — dynamic zoom and dialogue
+![MiniMax H3 OrbitQuant public ComfyUI workflow](comfyui/workflow-export.png)
 
-<video controls playsinline preload="metadata" width="608" aria-label="Source BF16 control — dynamic zoom and dialogue">
-  <source src="https://huggingface.co/WaveCut/MiniMax-H3-OrbitQuant-W4A4/resolve/main/examples/hevc/source-bf16-control.mp4" type="video/mp4; codecs=&quot;hvc1&quot;">
-  <source src="https://huggingface.co/WaveCut/MiniMax-H3-OrbitQuant-W4A4/resolve/main/examples/h264/source-bf16-control.mp4" type="video/mp4">
-  <a href="https://huggingface.co/WaveCut/MiniMax-H3-OrbitQuant-W4A4/resolve/main/examples/h264/source-bf16-control.mp4">Source BF16 control — dynamic zoom and dialogue MP4</a>
-</video>
+The PNG above is a 3060×1310 ComfyUI Workflow Image Export, not a browser
+screenshot. Its `tEXt` `workflow` chunk contains the same six-node graph with
+`balanced`, 608×480, 124 frames, 24 steps, and the detailed example prompt.
 
-[H.265/HEVC](examples/hevc/source-bf16-control.mp4) · [H.264 fallback](examples/h264/source-bf16-control.mp4) · [five-frame timeline](examples/timelines/source-bf16-control.png)
-### OrbitQuant W4A4 text encoder, BF16 denoiser
+Install [ComfyUI-OrbitQuant](https://github.com/iamwavecut/ComfyUI-OrbitQuant)
+into `ComfyUI/custom_nodes`, restart ComfyUI, import the workflow, and set
+`OrbitQuant Release Loader.model_path` to this downloaded model directory. The
+graph uses only the generic public nodes `OrbitQuant Release Loader` and
+`OrbitQuant Generate Video`; there are no MiniMax-specific public node classes.
 
-<video controls playsinline preload="metadata" width="608" aria-label="OrbitQuant W4A4 text encoder, BF16 denoiser">
-  <source src="https://huggingface.co/WaveCut/MiniMax-H3-OrbitQuant-W4A4/resolve/main/examples/hevc/w4-text-control.mp4" type="video/mp4; codecs=&quot;hvc1&quot;">
-  <source src="https://huggingface.co/WaveCut/MiniMax-H3-OrbitQuant-W4A4/resolve/main/examples/h264/w4-text-control.mp4" type="video/mp4">
-  <a href="https://huggingface.co/WaveCut/MiniMax-H3-OrbitQuant-W4A4/resolve/main/examples/h264/w4-text-control.mp4">OrbitQuant W4A4 text encoder, BF16 denoiser MP4</a>
-</video>
+On the RunPod ComfyUI image, launch ComfyUI with:
 
-[H.265/HEVC](examples/hevc/w4-text-control.mp4) · [H.264 fallback](examples/h264/w4-text-control.mp4) · [five-frame timeline](examples/timelines/w4-text-control.png)
-### Final native W4A4 text + W4A4 T2VA transformer
+```bash
+python main.py --listen 0.0.0.0 --port 8188 \
+  --disable-cuda-malloc \
+  --disable-dynamic-vram \
+  --disable-async-offload
+```
 
-<video controls playsinline preload="metadata" width="608" aria-label="Final native W4A4 text + W4A4 T2VA transformer">
-  <source src="https://huggingface.co/WaveCut/MiniMax-H3-OrbitQuant-W4A4/resolve/main/examples/hevc/w4a4-native-t2va.mp4" type="video/mp4; codecs=&quot;hvc1&quot;">
-  <source src="https://huggingface.co/WaveCut/MiniMax-H3-OrbitQuant-W4A4/resolve/main/examples/h264/w4a4-native-t2va.mp4" type="video/mp4">
-  <a href="https://huggingface.co/WaveCut/MiniMax-H3-OrbitQuant-W4A4/resolve/main/examples/h264/w4a4-native-t2va.mp4">Final native W4A4 text + W4A4 T2VA transformer MP4</a>
-</video>
+These supported flags let the OrbitQuant subprocess enforce its own allocator
+cap instead of competing with ComfyUI's global DynamicVRAM and async-offload
+layers.
 
-[H.265/HEVC](examples/hevc/w4a4-native-t2va.mp4) · [H.264 fallback](examples/h264/w4a4-native-t2va.mp4) · [five-frame timeline](examples/timelines/w4a4-native-t2va.png)
+## Inference profiles
 
-Each player prefers H.265/HEVC (`hvc1`, CRF 24) and retains H.264 as a browser fallback. Both files carry AAC stereo audio at 32 kHz.
+All numbers use CUDA 13, 608×480, 124 frames, 24 sigma points / 23 forwards,
+native-auto Torch Flash SDPA, no weight cache, sequential CUDA text conditioning,
+and source FP32 VAEs.
 
-## Component precision and size
+| Profile | GPU | Task | Placement | Process peak | Denoise | Generation |
+| --- | --- | --- | --- | ---: | ---: | ---: |
+| `balanced` (default) | RTX PRO 6000 | T2VA | streamed leaf offload, 12 GiB cap | 6.36 GiB child; 6.90 GiB incl. idle ComfyUI | 46.68 s | — |
+| `speed` | RTX PRO 6000 | T2VA | resident transformer | 21.14 GiB | 46.84 s | 51.10 s |
+| `minimum_vram` | RTX 4090 | T2VA | low-CPU-memory streamed leaf offload, 8 GiB cap | 4.07 GiB | 154.25 s | 188.70 s |
+| `speed` | RTX PRO 6000 | Ref2VA | resident `transformer_ref` | 24.06 GiB | 118.48 s | 155.42 s |
 
-| component | stored mode | artifact GiB | eligible linear coverage | OrbitQuant modules | AdaLN INT4 |
-|---|---|---:|---:|---:|---:|
-| `transformer` | `w4a4` | 17.03 | 97.45% | 300 | 50 |
-| `transformer_ref` | `w4a4` | 17.03 | 97.45% | 300 | 50 |
-| `text_encoder` | `w4a4` | 18.55 | 95.80% | 448 | 0 |
-| `vae` | `source_copy` | 9.70 | exact source copy | 0 | 0 |
-| `audio_vae` | `source_copy` | 0.56 | exact source copy | 0 | 0 |
+`balanced` is the recommended Pareto recipe. On the tested PRO 6000, streamed
+weight movement overlaps denoising closely enough to match the resident path
+while cutting the child process's physical CUDA peak by about 70%.
+`minimum_vram` is the verified absolute-minimum endpoint. `speed` removes
+transformer transfers when VRAM is available.
 
-Input/output projections, time/context/refiner boundaries, embeddings, norms, and language-head boundaries that are excluded by the pinned H3/Qwen policy remain in source precision. “Four bit” describes the eligible packed linear weights, not every tensor in the architecture.
+SageAttention2's available CUDA 13 binary did not include SM120 code for this
+PRO 6000, and forced cuDNN attention was slower. Native-auto Torch Flash SDPA is
+therefore the shipped supported attention path.
 
-## Measurements
+## Install
 
-Hardware: one RunPod NVIDIA RTX PRO 6000 Blackwell Server Edition (94.97 GiB), PyTorch 2.8.0+cu128. Every sample is exactly **608×480**, **124 frames**, **24 FPS**, and **50 sigma points / 49 forwards**.
+```bash
+pip install "orbitquant[hf,kernels]>=0.9.2,<0.10"
+pip install "diffusers @ git+https://github.com/huggingface/diffusers.git@abc5e9bf71fd38f53cd471bc3acaa84bc5ecbfdc"
+pip install "transformers>=5.13,<6" accelerate av soundfile
+```
 
-| example | task | generation seconds | generation peak VRAM GiB | FP32 VAE decode seconds | decode peak VRAM GiB |
-|---|---|---:|---:|---:|---:|
-| `source-bf16-control` | `t2va` | 451.58 | 63.76 | 19.56 | 11.31 |
-| `w4-text-control` | `t2va` | 381.57 | 63.76 | 13.15 | 11.31 |
-| `w4a4-native-t2va` | `t2va` | 114.67 | 43.18 | 4.71 | 11.31 |
-| `comfyui-t2va` | `t2va` | 137.14 | 18.91 | 5.48 | 11.29 |
-| `comfyui-ref2va` | `ref2va` | 318.86 | 30.46 | 5.50 | 11.29 |
+Or install all pinned runtime requirements from this repository:
 
-The final native-W4A4 generation through durable latents completed in **114.67 s** at **43.18 GiB** peak VRAM. That is **18.7×** faster than the previously published W4A16 compatibility run (2145.74 s) and **3.94×** faster than the measured source-BF16 control (451.58 s). The first-step native-vs-W4A16 latent drift was 0.0416% video NRMSE and 0.0734% audio NRMSE.
+```bash
+pip install -r runtime-requirements.txt
+```
 
-Raw measurements, review decisions, media probes, and early checkpoint comparisons are in [`benchmarks/`](benchmarks/).
+## Direct runner
 
-## Reproduction
+The runner writes each scheduler checkpoint atomically and saves the latent
+bundle before decode. The examples below keep the prompt in a file to avoid
+shell quoting a multi-kilobyte description.
 
-Install the pinned runtimes:
+Balanced T2VA:
 
 ```bash
-pip install "orbitquant[hf,kernels]>=0.9.1,<0.10"
-pip install "diffusers @ git+https://github.com/huggingface/diffusers.git@abc5e9bf71fd38f53cd471bc3acaa84bc5ecbfdc"
-pip install "transformers>=5.13,<6" accelerate av soundfile
+python scripts/run_quantized_example.py \
+  --release . \
+  --output balanced.mp4 \
+  --save-latents balanced.latents.pt \
+  --prompt "$(cat prompt.txt)" \
+  --seed 42 --width 608 --height 480 --num-frames 124 --steps 24 \
+  --manual-stage-offload \
+  --text-encoder-sequential-offload \
+  --transformer-group-offload-type leaf_level \
+  --group-offload-use-stream \
+  --cuda-memory-cap-gib 12 \
+  --transformer-runtime-mode auto_fused \
+  --checkpoint-dir checkpoints/balanced
 ```
 
-Generate and atomically persist latents:
+Maximum-speed T2VA: remove the group-offload and allocator-cap options while
+keeping `--manual-stage-offload --text-encoder-sequential-offload`.
+
+Minimum-VRAM T2VA: use the balanced command with
+`--group-offload-low-cpu-mem-usage --cuda-memory-cap-gib 8`.
+
+Ref2VA speed:
 
 ```bash
 python scripts/run_quantized_example.py \
   --release . \
-  --output dynamic-zoom.mp4 \
-  --save-latents dynamic-zoom.latents.pt \
-  --prompt "Single continuous live-action shot with fast running, a smooth push-in, stable face identity, Russian dialogue, and rain ambience." \
-  --seed 42 --height 480 --width 608 --num-frames 124 --steps 50 \
+  --output ref2va.mp4 \
+  --save-latents ref2va.latents.pt \
+  --prompt "$(cat prompt.txt)" \
+  --task ref2va --reference reference.png \
+  --seed 42 --width 608 --height 480 --num-frames 124 --steps 24 \
   --manual-stage-offload \
+  --text-encoder-sequential-offload \
+  --reference-vae-sequential-offload --reference-vae-tile-size 128 \
   --transformer-runtime-mode auto_fused \
-  --checkpoint-dir checkpoints/dynamic-zoom
+  --checkpoint-dir checkpoints/ref2va
 ```
 
-Only after that process exits, decode with the untouched source VAEs:
+Decode only after the latent-producing process exits:
 
 ```bash
 python scripts/decode_h3_latents.py \
-  --latents dynamic-zoom.latents.pt \
-  --vae vae --vae-dtype fp32 \
-  --audio-vae audio_vae --audio-vae-dtype fp32 \
-  --output dynamic-zoom.mp4
+  --latents balanced.latents.pt \
+  --vae vae \
+  --audio-vae audio_vae \
+  --output balanced.master-crf1.mp4 \
+  --preview-output balanced.mp4
 ```
 
-For Ref2VA replace manual stage offload with `--task ref2va --reference examples/reference/astronaut.png --offload-reserve-margin 64GB`. The runner selects `transformer_ref` automatically and lets H3's ComponentsManager place the untouched source FP32 VAE for reference encoding. The verified full ComfyUI generation completed in **318.86 s** at **30.46 GiB** peak VRAM, dispatched 300/300 eligible linears to `triton_cuda_packed_w4`, and returned the text encoder to CPU after conditioning.
+The decoder always loads the release's untouched source FP32 visual and audio
+VAEs. The visual VAE is tiled and sequentially offloaded; the audio VAE enters
+GPU only for the audio stage.
 
-## Validation
+## Component precision and size
 
-- The BF16 source control establishes that the official 50-point schedule is temporally coherent.
-- The W4-text-only control was reviewed in full resolution and judged ideal, including audio.
-- The final T2VA example uses native W4A4 conditioning and native W4A4 `transformer` execution before source-FP32 VAE decode.
-- The exact 50-block transformer profile verifies 300/300 eligible linears dispatch to the packed W4 Triton kernel under the package default.
-- The full ComfyUI Ref2VA run completed 49/49 forwards with 300/300 packed-W4 dispatch before source-FP32 visual/audio decode; its reviewed H.264 and HEVC outputs are embedded above.
-- The runner rejects an FP32 native-W4 compute model before the first denoiser forward, preventing a silent generic-kernel fallback from consuming a full run.
-- Every retained review explicitly checks face integrity, motion ghosting, texture breakup, repeating tiles, and at least four full-timeline positions.
-- Every H.264 and HEVC file is independently probed for dimensions, readable frame count, codec, and audio stream.
-- Visual/audio VAE weight-file SHA256 values match the pinned source revision; see [`validation/source_component_copy_audit.json`](validation/source_component_copy_audit.json).
+| Component | Stored mode | Artifact GiB | Eligible linear coverage | OrbitQuant modules | AdaLN INT4 |
+| --- | --- | ---: | ---: | ---: | ---: |
+| `transformer` | W4A4 | 17.03 | 97.45% | 300 | 50 |
+| `transformer_ref` | W4A4 | 17.03 | 97.45% | 300 | 50 |
+| `text_encoder` | W4A4 | 18.55 | 95.80% | 448 | 0 |
+| `vae` | source FP32 copy | 9.70 | exact source copy | 0 | 0 |
+| `audio_vae` | source FP32 copy | 0.56 | exact source copy | 0 | 0 |
+
+Input/output projections, time/context/refiner boundaries, embeddings, norms,
+and language-head boundaries excluded by the pinned H3/Qwen policy remain in
+source precision. “Four bit” describes eligible packed linear weights, not
+every tensor in the architecture.
+
+## Validation and provenance
+
+- OrbitQuant 0.9.2 revision `cd58b4ecf77f22b8c4116b3d0b7d4af258e16ba3`.
+- Diffusers revision `abc5e9bf71fd38f53cd471bc3acaa84bc5ecbfdc`.
+- The live ComfyUI workflow reached terminal `pass` through `/prompt` and
+  produced the standard `VIDEO` output.
+- All 23 denoiser checkpoints and the final latent were persisted before
+  source-FP32 decode.
+- Visual/audio VAE weight SHA256 values match the pinned source revision; see
+  [`validation/source_component_copy_audit.json`](validation/source_component_copy_audit.json).
 - Exact artifact hashes are in [`SHA256SUMS`](SHA256SUMS).
+- Full machine-readable profile and media evidence is in
+  [`comfyui/report.json`](comfyui/report.json).
 
 ## License and modifications
 
-The original [MiniMax H3 Community License Agreement](https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/73372e6cf53e414edd3ab03e357717fb0602e758/LICENSE) is copied as [`LICENSE`](LICENSE). See [`NOTICE`](NOTICE), [`MODIFICATIONS.md`](MODIFICATIONS.md), and the upstream [`QA-about-License`](docs/QA-about-License.md).
+The original
+[MiniMax H3 Community License Agreement](https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/73372e6cf53e414edd3ab03e357717fb0602e758/LICENSE)
+is copied as [`LICENSE`](LICENSE). See [`NOTICE`](NOTICE),
+[`MODIFICATIONS.md`](MODIFICATIONS.md), and the upstream
+[`QA-about-License`](docs/QA-about-License.md).

SHA256SUMS

@@ -1,112 +1,119 @@
-8116657dfaca81919460b4473c57a69b31a9a6ec97a3b1747df55e59500557e9  .gitattributes
-59b99642b95ea21630e311198ddbfffbfe05aadba0c2f5d884cbdf4efcc90f44  LICENSE
-e6685eda9517bad6751f538dcc0ed967a1a08cacd000a20fe12a3fdd01023b83  MODIFICATIONS.md
-154fe7fbdf198395d53c57d6786b2fbf7e63a083ff70843c5ef78c0c7303d91a  NOTICE
-34fc76ba3797f5d4dbc5e956dde7d1d0e84cfd01c6922094ca1202b05a8bd319  PERFORMANCE.md
-7b140b9f115e729213add98ce65e61c039e11ceb900228d84e109d21275ba046  README.md
-804780f7133477067bd6bbfbc02dc8b3cf9feeb400f97c08f5b1d5f6cbab3840  audio_scheduler/scheduler_config.json
-9a3c645ff892b376c6f5f4c8685964cd75474731af594ff058492a0000caabb6  audio_vae/config.json
-52c59e67ba8de5477c81bfbced0327aabf500f1bfdeefd5ee754529241cb26cb  audio_vae/diffusion_pytorch_model.safetensors
-c088e81b20500cdfcc77da8589f7b1408d080ebabf2051acae8c2f3426ffe1e7  benchmarks/astronaut-w4both-w4a16-early-gate.json
-415a42c8f0c2146ec7cf4bd70c582e5420d87517bef33135cebdd11593d62c9e  benchmarks/astronaut-w4both-w4a16-step001-health.json
-3d3a387b791332d68260f01dcc7c47f2944b21b55a5cbbcc7d966fc095dafa19  benchmarks/early-gate/native-dynamic-zoom-early.metrics.json
-32b73981afb8f5a3deacebf42631b673365ef423f7a6bea2ab41edb42cdf4c73  benchmarks/early-gate/native-vs-w4a16-step001.json
-6422124c7a9fff01207f0f3a0d31c48ee17670def930a27705d98a212c728bc6  benchmarks/early-gate/w4a16-dynamic-zoom-early.metrics.json
-064de721be4f0401f6e852fc4f426375c8d595ab6b30d232fdf270ba0e75747a  benchmarks/example_metrics.csv
-eeff4d8dda3228614ec753d184f908ff9425fd90ca97c6c9e2edcba9b406751d  benchmarks/example_metrics.json
-281bdb716d3e4db82e2e90e74c6bf74f05c483702502ce7df30dd448012dad02  benchmarks/media_probe.json
-47a26334e37b41f5d20068f7131409c0b2fa508ea358a18902e90140837d873c  benchmarks/profiles/native-full-forward-bf16-load.json
-b7a97b2c0d19bcc6bb6e052ec5a552c725083354be375ad12ee6afee38dc4130  benchmarks/profiles/native-full-forward-fp32-load.json
-72c7369bb28fffa42ca6a017c16d63aa86908be2b8f8ff9bc146c4473d394a60  benchmarks/profiles/native-full-forward-package-default.json
-22513428f1923473bae58d43de4b2a9f4bad05e062fa7911b035a6808a49d09e  benchmarks/profiles/native-full-forward-transformer-ref-cold.json
-05e2ed4075c91e75d9834a16dcf0a81397be49639372efe8b2360c73206bb2a9  benchmarks/profiles/native-full-forward-transformer-ref-hot.json
-02233da18b3c01761889df2206181e0fc7b9f120165b37a0fe4c9f7bbdda0096  comfyui/examples/h264/comfyui-ref2va-608x480.mp4
-4cceecedd28dfc593cf5e61145836405d18475bd8109b2ef7a984d96d91a13e6  comfyui/examples/h264/comfyui-t2va-608x480.mp4
-0c3040f81aefdb14b30047a5c972347d012d72dbfa026b8a9b2e245e555b7996  comfyui/examples/hevc/comfyui-ref2va-608x480.mp4
-bc31c456f9fdfae29c4e89d063baa953bbd73141c05a237ed43006336661eaab  comfyui/examples/hevc/comfyui-t2va-608x480.mp4
-78d579a40bf159ed3039900d6503d992843b4ec67662eb777751be908e9237d4  comfyui/examples/timelines/comfyui-ref2va-608x480.png
-7adf6cc9259ae6eb367f8217912bf76ec9cd88ef358150d80b4742d57b50bf01  comfyui/examples/timelines/comfyui-t2va-608x480.png
-f8bbd080f6676741ce5bd4d1c5dfda874d7077a3c3cc566a6e6c2e5fd0ea1a32  comfyui/report.json
-df49d477b234e71c23e0a7d457d66315408333db62433885db4b20f3606328f0  comfyui/workflow-export.png
-295ac493e659067aa1b39b9bf70143922c4459d1897636d896327011f167e1c1  comfyui/workflows/MiniMax-H3-OrbitQuant-Ref2VA-api.json
-5ed9b7600b6869bd887459d4d3e8f740201f6b4bd8ca40d8fd786bb6f8721f29  comfyui/workflows/MiniMax-H3-OrbitQuant-T2VA-api.json
-7c1a82427b80203e4252726a0ab3a336a6859fe2ff43302a07de00bce5722acb  comfyui/workflows/MiniMax-H3-OrbitQuant-T2VA.json
-b248e45ce21157a41c298a491f95ea2bc25468c3fe3cb7eaf939818056853250  comfyui_orbitquant.json
-26900a2868636f886d241efe94002ad11d858b3c3eb005e58c0c2ad60f0de7ae  docs/QA-about-License.md
-87044e59a72a2b8ed3df285de044a097d7366f3004dcfc26ffd11fc71ed32d17  examples/h264/source-bf16-control.mp4
-513f09b20cb4e59a47a2f4ff19e544cb29d41870ee2c4759ad0a2ff6cb16fe0c  examples/h264/w4-text-control.mp4
-679c0b690fce47ed2dd19ba7fa185d0cea69ee3a3b308a84faa909ce0762d24b  examples/h264/w4a4-native-t2va.mp4
-eb4eb5d19d231887c6dfba5be70ac3d9107fe076a5f7c69e6fa77ecc0ad5ced9  examples/hevc/source-bf16-control.mp4
-ea89a3cc2a37c4294da4142f22d9d3778f1d086efb6125d73e1d5ccf392c6cbb  examples/hevc/w4-text-control.mp4
-0bef87dbecbd9963e6febb43fb44cc2caf142192e59f7deee13950e55789b235  examples/hevc/w4a4-native-t2va.mp4
-cc97132c8132fa431d375c1e0b0dc8a85b536a91bc13bd5494e789e592126426  examples/quality-timeline-matrix.webp
-b1257de01d8440baea06865ea35ce557d885b6dda700281ba25202193c58285c  examples/reference/astronaut.png
-1c33ff3ec4f637b15496b01130cd83c7393f742fce45b5986d158e6970cfdd27  examples/reviews/source-bf16-control.json
-fad45c95f068f9abfc5125e1e0993004fe036b793803eebc3e28255827200a25  examples/reviews/w4-text-control.json
-9740fde8102c96aaa2bb9f41ecfbf3eaea651d2393944a21107889789d21f6e7  examples/reviews/w4a4-native-t2va.json
-efa068559ff6c2b7d0f6fbf2c835ba4d1b654f4cd016e7be737e27b50c0c380e  examples/timelines/source-bf16-control.png
-367855aa5fd90e4a6b6baece0894a3f4343a6a6ffc6b16750711d14a05bb82fe  examples/timelines/w4-text-control.png
-a69f89026fde4cacf6a52047ccdf5909a9ba218658d13b9b514a9fd11ece4482  examples/timelines/w4a4-native-t2va.png
-a6c5f6fca6c2c92726ddb03407ff11d9c1287bbfa164e489e58fb84376073faa  modular_model_index.json
-5c72a170d2a4a1a3bc5adad2e689ae28138a9700e5b8c96c0266331e86c0acce  processor/chat_template.json
-599bab54075088774b1733fde865d5bd747cbcc7a547c5bc12610e874e26f5e3  processor/merges.txt
-27225450ac9c6529872ee1924fcb0962ff5634834f817040f444118116f4e516  processor/preprocessor_config.json
-a5d85b6dcc535e6b93115a9ef287e6132fdbf30270da6218194ba742261173c7  processor/tokenizer.json
-a07e942ac874baa13758de8d1fbdb186683cc03416b5589e1b6671c6b3057c68  processor/tokenizer_config.json
-7768af27c1fafa9cc9011c1dc20067e03f8915e03b63504550e11d5066986d13  processor/video_preprocessor_config.json
-ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910  processor/vocab.json
-432518f56fa7e8fc1b6f48f98d1a93606eb1d3ba3aabb028343bcb7f960a1b2a  prompts.json
-95676df192c32b659a46cc8ad1c5b1c7556d8258730eea761399c60abdb94894  quantization_manifest.json
-894c94241bf053846755f74c639e86cceb03335c40bf36745d0c8a3780c61821  runtime-requirements.txt
-8fa6c3aa70dc9e691e1a6df899fd1b6f75f70481a27cee6e18a303817075c304  scheduler/scheduler_config.json
-075a3208f5fba960a4eae3003a5b7fcc57ad3d6a0ba9e8d91391f0d7b64cae04  scripts/audit_source_component_copies.py
-5f970b7f378349add8381ad5fa809d711ef315966d0c3a6f4aa36ae7b72665f9  scripts/checkpoint_io.py
-b1bb1643636c0a83b31eed65996ab4a8c8bc9c2f458bae5a5dccde02158ce432  scripts/decode_h3_latents.py
-5f58a74b4408abf18e0617dd35a999ce0f8b8d238fd69f45e163ae0171421145  scripts/decode_precision_policy.py
-e9ea546b28f1fc4ec01f1270072280be684d6d92cea3546e268b9b5d3e5b1c2d  scripts/latent_io.py
-ea79fe8d056bbfe08e5f9ddfadd83f74d11539641221679e07a2a82dc2cf5086  scripts/manual_stage_offload.py
-5d1fb5bc4f429447d68f7647a566e90320d55f7ac6b630a320be39642e1f202f  scripts/media_packaging.py
-655098f3ec10bdbb41bb79b2d38a326d5d7aad115beec90edbfd46611446676a  scripts/offload_policy.py
-64b2b294d46ce22b3fe9a3e7642e3b1344a190f4a83dd9bb44a3a48204b541d5  scripts/orbitquant_h3_compat.py
-8eb302a23053b6d96ecea86974cb55cc5193eb295fc41c314a34c452779f9a07  scripts/quality_gate.py
-295f3c21780891df90f9876c822fb46425dffefbfd05e16e370585ed04af62d9  scripts/run_quantized_example.py
-841b69c59e4eff19737414b9547599c138581dded744467ab8c6b86d9005a009  scripts/runtime_cache_policy.py
-496abd4859f92b354ae9b34d2028864ea1a3c13d92f7067164aea5465d4ba61d  scripts/verification_policy.py
-de100ce1d13e83fe378b1a8543b2e906ce2cecc4e7153d126864975c5378a7ed  scripts/verify_components.py
-a2b186303b94eeb97750e4cc498103217c1888794e160c7c6b79dd39220267a9  text_encoder/config.json
-19c0e38d660f3366f146fba332780419eab679529fb6381783b1cacc7b863d53  text_encoder/generation_config.json
-7461c06c1ee7f5e606bd014e1825183407fcbc18893ef3ec7099398b8274a20f  text_encoder/model-00001-of-00005.safetensors
-c510d3757606e1cf375e983bf54ff1c2c0ba470c83b74679224bcf8d51745a5f  text_encoder/model-00002-of-00005.safetensors
-988e2362f4fbe353b761afddff94c69a64eb821409fe226fe835d252edb056ec  text_encoder/model-00003-of-00005.safetensors
-bfb4cb2a0c7e64ee2960596a5049a797e6a45776c96913e06ee8c1964d11125d  text_encoder/model-00004-of-00005.safetensors
-3cf949a827326855d2dc847557c2f71bb0d464c9a9cafb444b6c2e0aac86b742  text_encoder/model-00005-of-00005.safetensors
-36b2d7912dffc98f6f5e0bd7b11234c451f6c99f351f5f28a87ce30979f73271  text_encoder/model.safetensors.index.json
-599bab54075088774b1733fde865d5bd747cbcc7a547c5bc12610e874e26f5e3  tokenizer/merges.txt
-a5d85b6dcc535e6b93115a9ef287e6132fdbf30270da6218194ba742261173c7  tokenizer/tokenizer.json
-a07e942ac874baa13758de8d1fbdb186683cc03416b5589e1b6671c6b3057c68  tokenizer/tokenizer_config.json
-ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910  tokenizer/vocab.json
-eac2b97122aeab6cb7a324fd8b451e4ad021777d6137d285e2098b089bafb926  transformer/config.json
-fa30399f163a85ab89d7e23a93237d8188ea1bf09a6bbd489e0f82102796daa7  transformer/diffusion_pytorch_model-00001-of-00005.safetensors
-6d929ad9edbedd0eacde8ecac20f227ce543b1f63edaf7b283340a485e319bdc  transformer/diffusion_pytorch_model-00002-of-00005.safetensors
-65a6880aad54e1dac2aa64bbe29fbf1199f179a5f6d5b16dd0e9e3448322811f  transformer/diffusion_pytorch_model-00003-of-00005.safetensors
-8cc4dd8cd6e5dbe8d7610a361ef6fcbbfd7a2a0e4a61fffc73b116f88aad13a7  transformer/diffusion_pytorch_model-00004-of-00005.safetensors
-def316191099ee999ee21c7a34b6954c74ffe18d4b7faf92847457afab494cc9  transformer/diffusion_pytorch_model-00005-of-00005.safetensors
-127932208cc61680a73ed835cdbf971a8190c7c275668864a855a15c614cb765  transformer/diffusion_pytorch_model.safetensors.index.json
-eac2b97122aeab6cb7a324fd8b451e4ad021777d6137d285e2098b089bafb926  transformer_ref/config.json
-357baa234bc04ed4843eefe19ce98a23995e24f012bcfc5117b464c9fc649815  transformer_ref/diffusion_pytorch_model-00001-of-00005.safetensors
-60b0663b551b9cf1ccc62c14571689c3cf54de8efef0971b1687cbbed9ba2149  transformer_ref/diffusion_pytorch_model-00002-of-00005.safetensors
-ee2c6a0690df2c2d13e21c82cd6c25ef499a5e317d71944ce70cb83f1747b1b0  transformer_ref/diffusion_pytorch_model-00003-of-00005.safetensors
-a353924509eff7dd7be99f3efc8c57eecf4d87e24976ab2b02e2367141891651  transformer_ref/diffusion_pytorch_model-00004-of-00005.safetensors
-006e19769149a3f3d96a5df7341bca346cfb0113d6e2b1b984167acd3daf5e12  transformer_ref/diffusion_pytorch_model-00005-of-00005.safetensors
-127932208cc61680a73ed835cdbf971a8190c7c275668864a855a15c614cb765  transformer_ref/diffusion_pytorch_model.safetensors.index.json
-78f67deec3d63aae807f2bfe7154bc1e26f6372cb20b63265fcbae1b62bb5745  vae/config.json
-72f4c6be84ac0674f27398cde991dd9d719762f3952c4921aa66b2ce542f6374  vae/diffusion_pytorch_model-00001-of-00003.safetensors
-2e05e8bc23fa4071043e17fd242be8acd0685e781a43987432b2eae925be4198  vae/diffusion_pytorch_model-00002-of-00003.safetensors
-c05d6ac4b1a33de372799d708531da6320f6a3ce6d1ce6d895e770988e004a39  vae/diffusion_pytorch_model-00003-of-00003.safetensors
-15f6d44553c3c616b0dc999920aa784f92ecee7e4201f1f99ac405cfbf3061ca  vae/diffusion_pytorch_model.safetensors.index.json
-d79d95cb98ff188056ff93668935f434cde9374f64ac7f683141d868e23a7ffe  validation/audio_vae.json
-bb84e8c4b30042de4306035ac8c7b401bddc6fcadb33336bd9933993d9c8bb3a  validation/source_component_copy_audit.json
-43e8558e13737d94de0e5739adf10c4e371b00149d5c5965eb8557042749a6fc  validation/text_encoder.json
-34b7c710e3a1a13ae078f497064703f898b10d782a67d3861efb29c7a0a14c02  validation/transformer.json
-8ac5758c586d3b2ac170e6766f0fc5c191ad6e3e7d8de2a372961df29308eda6  validation/transformer_ref.json
-18e908a81caf2a84520fd1813b40948867bcc9d1a74f6029f800779cf1417a77  validation/vae.json
+24c2039a0ccb86c392e9fefe8a26b135b4cf46da58a0a7d14729c62196a75ddc  ./.gitattributes
+59b99642b95ea21630e311198ddbfffbfe05aadba0c2f5d884cbdf4efcc90f44  ./LICENSE
+e6685eda9517bad6751f538dcc0ed967a1a08cacd000a20fe12a3fdd01023b83  ./MODIFICATIONS.md
+154fe7fbdf198395d53c57d6786b2fbf7e63a083ff70843c5ef78c0c7303d91a  ./NOTICE
+34fc76ba3797f5d4dbc5e956dde7d1d0e84cfd01c6922094ca1202b05a8bd319  ./PERFORMANCE.md
+eaaff934eda5b84c4c5bb1c007e53817fe401231e42b958152810be4798dd1d4  ./README.md
+804780f7133477067bd6bbfbc02dc8b3cf9feeb400f97c08f5b1d5f6cbab3840  ./audio_scheduler/scheduler_config.json
+9a3c645ff892b376c6f5f4c8685964cd75474731af594ff058492a0000caabb6  ./audio_vae/config.json
+52c59e67ba8de5477c81bfbced0327aabf500f1bfdeefd5ee754529241cb26cb  ./audio_vae/diffusion_pytorch_model.safetensors
+c088e81b20500cdfcc77da8589f7b1408d080ebabf2051acae8c2f3426ffe1e7  ./benchmarks/astronaut-w4both-w4a16-early-gate.json
+415a42c8f0c2146ec7cf4bd70c582e5420d87517bef33135cebdd11593d62c9e  ./benchmarks/astronaut-w4both-w4a16-step001-health.json
+3d3a387b791332d68260f01dcc7c47f2944b21b55a5cbbcc7d966fc095dafa19  ./benchmarks/early-gate/native-dynamic-zoom-early.metrics.json
+32b73981afb8f5a3deacebf42631b673365ef423f7a6bea2ab41edb42cdf4c73  ./benchmarks/early-gate/native-vs-w4a16-step001.json
+6422124c7a9fff01207f0f3a0d31c48ee17670def930a27705d98a212c728bc6  ./benchmarks/early-gate/w4a16-dynamic-zoom-early.metrics.json
+064de721be4f0401f6e852fc4f426375c8d595ab6b30d232fdf270ba0e75747a  ./benchmarks/example_metrics.csv
+eeff4d8dda3228614ec753d184f908ff9425fd90ca97c6c9e2edcba9b406751d  ./benchmarks/example_metrics.json
+281bdb716d3e4db82e2e90e74c6bf74f05c483702502ce7df30dd448012dad02  ./benchmarks/media_probe.json
+47a26334e37b41f5d20068f7131409c0b2fa508ea358a18902e90140837d873c  ./benchmarks/profiles/native-full-forward-bf16-load.json
+b7a97b2c0d19bcc6bb6e052ec5a552c725083354be375ad12ee6afee38dc4130  ./benchmarks/profiles/native-full-forward-fp32-load.json
+72c7369bb28fffa42ca6a017c16d63aa86908be2b8f8ff9bc146c4473d394a60  ./benchmarks/profiles/native-full-forward-package-default.json
+22513428f1923473bae58d43de4b2a9f4bad05e062fa7911b035a6808a49d09e  ./benchmarks/profiles/native-full-forward-transformer-ref-cold.json
+05e2ed4075c91e75d9834a16dcf0a81397be49639372efe8b2360c73206bb2a9  ./benchmarks/profiles/native-full-forward-transformer-ref-hot.json
+6c17fa38c5fee6a396a58d7f55981d1cdb644fae5cbbac87a2eb8d4194ad4d2f  ./comfyui/examples/h264/comfyui-balanced-t2va-608x480.mp4
+02233da18b3c01761889df2206181e0fc7b9f120165b37a0fe4c9f7bbdda0096  ./comfyui/examples/h264/comfyui-ref2va-608x480.mp4
+4cceecedd28dfc593cf5e61145836405d18475bd8109b2ef7a984d96d91a13e6  ./comfyui/examples/h264/comfyui-t2va-608x480.mp4
+1bbc05a09f3f49d1564afb1b0c0f49a5f3c3a05a34a28e4f8b28f1d3295f5335  ./comfyui/examples/hevc/comfyui-balanced-t2va-608x480-crf10.mp4
+0c3040f81aefdb14b30047a5c972347d012d72dbfa026b8a9b2e245e555b7996  ./comfyui/examples/hevc/comfyui-ref2va-608x480.mp4
+bc31c456f9fdfae29c4e89d063baa953bbd73141c05a237ed43006336661eaab  ./comfyui/examples/hevc/comfyui-t2va-608x480.mp4
+32711ef0f3a3a341340ab88e31537b26eefd486909154b0ba03ada49896f7f21  ./comfyui/examples/masters/comfyui-balanced-t2va-608x480-crf1.mp4
+a1114d1ba23da74c0ee1d911599cdd50cab3deef5e2de9e9ff9b9ff9167f97db  ./comfyui/examples/review/comfyui-balanced-adjacent-triplets.jpg
+36b4c549e3bf39716325a81ea14417be811ac8b5dd7a71bfe596aef8288ab385  ./comfyui/examples/review/comfyui-balanced-audio-spectrum.png
+5a5139474ed5aed90706a002bef1d574b2e737b24164e2367e008167bde34f48  ./comfyui/examples/review/comfyui-balanced-frame-120.png
+f081bb49aa12d73c108f8fef963c51c387662590cc41c6b069a0c7a47e7552e5  ./comfyui/examples/timelines/comfyui-balanced-t2va-608x480.jpg
+78d579a40bf159ed3039900d6503d992843b4ec67662eb777751be908e9237d4  ./comfyui/examples/timelines/comfyui-ref2va-608x480.png
+7adf6cc9259ae6eb367f8217912bf76ec9cd88ef358150d80b4742d57b50bf01  ./comfyui/examples/timelines/comfyui-t2va-608x480.png
+c4eec7582b072e752e0491b7c79c3e76a1fc207a6cf3f1e86d95be13155b7782  ./comfyui/report.json
+164ac2cd2dd51a1aa7a99c213e6777585b48b9efdbe6ad742e39cad8a5143caf  ./comfyui/workflow-export.png
+497dede417244161108ed46cfe0be2c183cba7851863f492ae67f6f92e66871c  ./comfyui/workflows/MiniMax-H3-OrbitQuant-Ref2VA-api.json
+83fc4ee22497d42cac1dbc49d68c51260c795c5b12a21a41fff5ac296b866c6b  ./comfyui/workflows/MiniMax-H3-OrbitQuant-T2VA-api.json
+e519ae3cd3763b9b832cee341eb4f982681a834ae0d3f80237b6d30a05350f92  ./comfyui/workflows/MiniMax-H3-OrbitQuant-T2VA.json
+ec9ad075deab2f51b96d96f90f747db410440c65fe16ced65bf8d910deb58f5b  ./comfyui_orbitquant.json
+26900a2868636f886d241efe94002ad11d858b3c3eb005e58c0c2ad60f0de7ae  ./docs/QA-about-License.md
+87044e59a72a2b8ed3df285de044a097d7366f3004dcfc26ffd11fc71ed32d17  ./examples/h264/source-bf16-control.mp4
+513f09b20cb4e59a47a2f4ff19e544cb29d41870ee2c4759ad0a2ff6cb16fe0c  ./examples/h264/w4-text-control.mp4
+679c0b690fce47ed2dd19ba7fa185d0cea69ee3a3b308a84faa909ce0762d24b  ./examples/h264/w4a4-native-t2va.mp4
+eb4eb5d19d231887c6dfba5be70ac3d9107fe076a5f7c69e6fa77ecc0ad5ced9  ./examples/hevc/source-bf16-control.mp4
+ea89a3cc2a37c4294da4142f22d9d3778f1d086efb6125d73e1d5ccf392c6cbb  ./examples/hevc/w4-text-control.mp4
+0bef87dbecbd9963e6febb43fb44cc2caf142192e59f7deee13950e55789b235  ./examples/hevc/w4a4-native-t2va.mp4
+cc97132c8132fa431d375c1e0b0dc8a85b536a91bc13bd5494e789e592126426  ./examples/quality-timeline-matrix.webp
+b1257de01d8440baea06865ea35ce557d885b6dda700281ba25202193c58285c  ./examples/reference/astronaut.png
+1c33ff3ec4f637b15496b01130cd83c7393f742fce45b5986d158e6970cfdd27  ./examples/reviews/source-bf16-control.json
+fad45c95f068f9abfc5125e1e0993004fe036b793803eebc3e28255827200a25  ./examples/reviews/w4-text-control.json
+9740fde8102c96aaa2bb9f41ecfbf3eaea651d2393944a21107889789d21f6e7  ./examples/reviews/w4a4-native-t2va.json
+efa068559ff6c2b7d0f6fbf2c835ba4d1b654f4cd016e7be737e27b50c0c380e  ./examples/timelines/source-bf16-control.png
+367855aa5fd90e4a6b6baece0894a3f4343a6a6ffc6b16750711d14a05bb82fe  ./examples/timelines/w4-text-control.png
+a69f89026fde4cacf6a52047ccdf5909a9ba218658d13b9b514a9fd11ece4482  ./examples/timelines/w4a4-native-t2va.png
+a6c5f6fca6c2c92726ddb03407ff11d9c1287bbfa164e489e58fb84376073faa  ./modular_model_index.json
+5c72a170d2a4a1a3bc5adad2e689ae28138a9700e5b8c96c0266331e86c0acce  ./processor/chat_template.json
+599bab54075088774b1733fde865d5bd747cbcc7a547c5bc12610e874e26f5e3  ./processor/merges.txt
+27225450ac9c6529872ee1924fcb0962ff5634834f817040f444118116f4e516  ./processor/preprocessor_config.json
+a5d85b6dcc535e6b93115a9ef287e6132fdbf30270da6218194ba742261173c7  ./processor/tokenizer.json
+a07e942ac874baa13758de8d1fbdb186683cc03416b5589e1b6671c6b3057c68  ./processor/tokenizer_config.json
+7768af27c1fafa9cc9011c1dc20067e03f8915e03b63504550e11d5066986d13  ./processor/video_preprocessor_config.json
+ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910  ./processor/vocab.json
+432518f56fa7e8fc1b6f48f98d1a93606eb1d3ba3aabb028343bcb7f960a1b2a  ./prompts.json
+576410ac684295f68af5f2d64cb236452afc0d7dfab7989d786f2a0e1126ba14  ./quantization_manifest.json
+1637243cc87dc53123f28988cc72c9fb1ac986af4995ba47d07ddc4b4f161af7  ./runtime-requirements.txt
+8fa6c3aa70dc9e691e1a6df899fd1b6f75f70481a27cee6e18a303817075c304  ./scheduler/scheduler_config.json
+075a3208f5fba960a4eae3003a5b7fcc57ad3d6a0ba9e8d91391f0d7b64cae04  ./scripts/audit_source_component_copies.py
+5f970b7f378349add8381ad5fa809d711ef315966d0c3a6f4aa36ae7b72665f9  ./scripts/checkpoint_io.py
+2a742dc0234370ad1ebfd909e00e4a9bad18e0b554b26a34c71aff6802f9c68e  ./scripts/decode_h3_latents.py
+5f58a74b4408abf18e0617dd35a999ce0f8b8d238fd69f45e163ae0171421145  ./scripts/decode_precision_policy.py
+e9ea546b28f1fc4ec01f1270072280be684d6d92cea3546e268b9b5d3e5b1c2d  ./scripts/latent_io.py
+fcd00e50a8230c185f63c33fa09d5b111507df800011b9a6d33b6acf033a1d88  ./scripts/manual_stage_offload.py
+5d1fb5bc4f429447d68f7647a566e90320d55f7ac6b630a320be39642e1f202f  ./scripts/media_packaging.py
+655098f3ec10bdbb41bb79b2d38a326d5d7aad115beec90edbfd46611446676a  ./scripts/offload_policy.py
+64b2b294d46ce22b3fe9a3e7642e3b1344a190f4a83dd9bb44a3a48204b541d5  ./scripts/orbitquant_h3_compat.py
+8eb302a23053b6d96ecea86974cb55cc5193eb295fc41c314a34c452779f9a07  ./scripts/quality_gate.py
+cfa82c50292f1b62cd95cb201b67a191c65e972f55a9b026b71feaded49dba3e  ./scripts/run_quantized_example.py
+841b69c59e4eff19737414b9547599c138581dded744467ab8c6b86d9005a009  ./scripts/runtime_cache_policy.py
+496abd4859f92b354ae9b34d2028864ea1a3c13d92f7067164aea5465d4ba61d  ./scripts/verification_policy.py
+de100ce1d13e83fe378b1a8543b2e906ce2cecc4e7153d126864975c5378a7ed  ./scripts/verify_components.py
+a2b186303b94eeb97750e4cc498103217c1888794e160c7c6b79dd39220267a9  ./text_encoder/config.json
+19c0e38d660f3366f146fba332780419eab679529fb6381783b1cacc7b863d53  ./text_encoder/generation_config.json
+7461c06c1ee7f5e606bd014e1825183407fcbc18893ef3ec7099398b8274a20f  ./text_encoder/model-00001-of-00005.safetensors
+c510d3757606e1cf375e983bf54ff1c2c0ba470c83b74679224bcf8d51745a5f  ./text_encoder/model-00002-of-00005.safetensors
+988e2362f4fbe353b761afddff94c69a64eb821409fe226fe835d252edb056ec  ./text_encoder/model-00003-of-00005.safetensors
+bfb4cb2a0c7e64ee2960596a5049a797e6a45776c96913e06ee8c1964d11125d  ./text_encoder/model-00004-of-00005.safetensors
+3cf949a827326855d2dc847557c2f71bb0d464c9a9cafb444b6c2e0aac86b742  ./text_encoder/model-00005-of-00005.safetensors
+36b2d7912dffc98f6f5e0bd7b11234c451f6c99f351f5f28a87ce30979f73271  ./text_encoder/model.safetensors.index.json
+599bab54075088774b1733fde865d5bd747cbcc7a547c5bc12610e874e26f5e3  ./tokenizer/merges.txt
+a5d85b6dcc535e6b93115a9ef287e6132fdbf30270da6218194ba742261173c7  ./tokenizer/tokenizer.json
+a07e942ac874baa13758de8d1fbdb186683cc03416b5589e1b6671c6b3057c68  ./tokenizer/tokenizer_config.json
+ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910  ./tokenizer/vocab.json
+eac2b97122aeab6cb7a324fd8b451e4ad021777d6137d285e2098b089bafb926  ./transformer/config.json
+fa30399f163a85ab89d7e23a93237d8188ea1bf09a6bbd489e0f82102796daa7  ./transformer/diffusion_pytorch_model-00001-of-00005.safetensors
+6d929ad9edbedd0eacde8ecac20f227ce543b1f63edaf7b283340a485e319bdc  ./transformer/diffusion_pytorch_model-00002-of-00005.safetensors
+65a6880aad54e1dac2aa64bbe29fbf1199f179a5f6d5b16dd0e9e3448322811f  ./transformer/diffusion_pytorch_model-00003-of-00005.safetensors
+8cc4dd8cd6e5dbe8d7610a361ef6fcbbfd7a2a0e4a61fffc73b116f88aad13a7  ./transformer/diffusion_pytorch_model-00004-of-00005.safetensors
+def316191099ee999ee21c7a34b6954c74ffe18d4b7faf92847457afab494cc9  ./transformer/diffusion_pytorch_model-00005-of-00005.safetensors
+127932208cc61680a73ed835cdbf971a8190c7c275668864a855a15c614cb765  ./transformer/diffusion_pytorch_model.safetensors.index.json
+eac2b97122aeab6cb7a324fd8b451e4ad021777d6137d285e2098b089bafb926  ./transformer_ref/config.json
+357baa234bc04ed4843eefe19ce98a23995e24f012bcfc5117b464c9fc649815  ./transformer_ref/diffusion_pytorch_model-00001-of-00005.safetensors
+60b0663b551b9cf1ccc62c14571689c3cf54de8efef0971b1687cbbed9ba2149  ./transformer_ref/diffusion_pytorch_model-00002-of-00005.safetensors
+ee2c6a0690df2c2d13e21c82cd6c25ef499a5e317d71944ce70cb83f1747b1b0  ./transformer_ref/diffusion_pytorch_model-00003-of-00005.safetensors
+a353924509eff7dd7be99f3efc8c57eecf4d87e24976ab2b02e2367141891651  ./transformer_ref/diffusion_pytorch_model-00004-of-00005.safetensors
+006e19769149a3f3d96a5df7341bca346cfb0113d6e2b1b984167acd3daf5e12  ./transformer_ref/diffusion_pytorch_model-00005-of-00005.safetensors
+127932208cc61680a73ed835cdbf971a8190c7c275668864a855a15c614cb765  ./transformer_ref/diffusion_pytorch_model.safetensors.index.json
+78f67deec3d63aae807f2bfe7154bc1e26f6372cb20b63265fcbae1b62bb5745  ./vae/config.json
+72f4c6be84ac0674f27398cde991dd9d719762f3952c4921aa66b2ce542f6374  ./vae/diffusion_pytorch_model-00001-of-00003.safetensors
+2e05e8bc23fa4071043e17fd242be8acd0685e781a43987432b2eae925be4198  ./vae/diffusion_pytorch_model-00002-of-00003.safetensors
+c05d6ac4b1a33de372799d708531da6320f6a3ce6d1ce6d895e770988e004a39  ./vae/diffusion_pytorch_model-00003-of-00003.safetensors
+15f6d44553c3c616b0dc999920aa784f92ecee7e4201f1f99ac405cfbf3061ca  ./vae/diffusion_pytorch_model.safetensors.index.json
+d79d95cb98ff188056ff93668935f434cde9374f64ac7f683141d868e23a7ffe  ./validation/audio_vae.json
+bb84e8c4b30042de4306035ac8c7b401bddc6fcadb33336bd9933993d9c8bb3a  ./validation/source_component_copy_audit.json
+43e8558e13737d94de0e5739adf10c4e371b00149d5c5965eb8557042749a6fc  ./validation/text_encoder.json
+34b7c710e3a1a13ae078f497064703f898b10d782a67d3861efb29c7a0a14c02  ./validation/transformer.json
+8ac5758c586d3b2ac170e6766f0fc5c191ad6e3e7d8de2a372961df29308eda6  ./validation/transformer_ref.json
+18e908a81caf2a84520fd1813b40948867bcc9d1a74f6029f800779cf1417a77  ./validation/vae.json

comfyui/examples/h264/comfyui-balanced-t2va-608x480.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6c17fa38c5fee6a396a58d7f55981d1cdb644fae5cbbac87a2eb8d4194ad4d2f
+size 946049

comfyui/examples/hevc/comfyui-balanced-t2va-608x480-crf10.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1bbc05a09f3f49d1564afb1b0c0f49a5f3c3a05a34a28e4f8b28f1d3295f5335
+size 4102196

comfyui/examples/masters/comfyui-balanced-t2va-608x480-crf1.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:32711ef0f3a3a341340ab88e31537b26eefd486909154b0ba03ada49896f7f21
+size 14540221

comfyui/report.json

@@ -3,18 +3,20 @@
   "date": "2026-08-04",
   "hardware": {
     "provider": "RunPod",
-    "gpu": "NVIDIA RTX PRO 6000 Blackwell Server Edition",
-    "gpu_memory_mib": 97887,
-    "template": "runpod-torch-v280",
-    "image": "runpod/pytorch:1.0.2-cu1281-torch280-ubuntu2404"
+    "speed_and_comfy_gpu": "NVIDIA RTX PRO 6000 Blackwell Workstation Edition",
+    "speed_and_comfy_gpu_memory_mib": 97887,
+    "minimum_vram_gpu": "NVIDIA GeForce RTX 4090",
+    "cuda": "13.0",
+    "template": "runpod-torch-v280"
   },
   "revisions": {
     "comfyui": "9a9fdb10ed144ce760d9682cb247526ea23cc525",
     "comfyui_frontend": "1.47.12",
-    "comfyui_orbitquant_candidate": "51d5473bc2bd4f14572724d362770c8ffacb10ed",
+    "comfyui_orbitquant_branch": "main",
     "model_release_tested": "e434bbea523349576e7c3d2f6090744aa4597123",
     "source_model": "73372e6cf53e414edd3ab03e357717fb0602e758",
-    "orbitquant": "0.9.1",
+    "orbitquant": "0.9.2",
+    "orbitquant_revision": "cd58b4ecf77f22b8c4116b3d0b7d4af258e16ba3",
     "diffusers": "abc5e9bf71fd38f53cd471bc3acaa84bc5ecbfdc",
     "official_workflow_templates": "7653f1cdef1d92394b6ef9946018c0a8aa4136b8"
   },
@@ -24,73 +26,116 @@
     "release_type": "ORBITQUANT_RELEASE",
     "output_type": "VIDEO",
     "model_specific_public_nodes": false,
-    "adapter": "minimax_h3"
+    "adapter": "minimax_h3",
+    "inference_profiles": [
+      "balanced",
+      "speed",
+      "minimum_vram"
+    ]
   },
-  "workflow": {
-    "source": "Comfy-Org/workflow_templates/templates/video_minimax_h3_t2v.json",
-    "ui_file": "workflows/minimax-h3-t2va-ui.json",
-    "api_file_from_live_graph": "workflows/minimax-h3-public-ui-api.json",
-    "workflow_image_export": "workflow-export.png",
-    "live_import": {
-      "node_count": 6,
-      "missing_nodes": [],
-      "links": 2,
-      "seed": 42,
-      "width": 608,
-      "height": 480,
-      "num_frames": 124,
-      "steps": 50
-    },
-    "terminal_prompt_id": "823b3407-ae85-43fd-b51a-129ea7a11f4d",
-    "save_video_output": {
-      "codec": "h264",
-      "width": 608,
-      "height": 480,
-      "frames": 124,
-      "fps": 24,
-      "audio": "aac stereo 32000 Hz",
-      "duration_seconds": 5.175
-    }
+  "common_recipe": {
+    "width": 608,
+    "height": 480,
+    "num_frames": 124,
+    "fps": 24,
+    "sigma_points": 24,
+    "denoiser_forwards": 23,
+    "native_w4_modules": "300/300",
+    "attention": "native_auto_torch_flash_sdpa",
+    "int8_weight_cache_modules": 0,
+    "conditioner": "sequential_cuda_layers_then_cpu",
+    "vae": "untouched_source_fp32_tiled_decode",
+    "audio_vae": "untouched_source_fp32_cuda_only_during_audio_stage"
   },
-  "runs": {
-    "t2va": {
+  "profiles": {
+    "balanced": {
+      "status": "pass",
+      "task": "t2va",
+      "gpu": "NVIDIA RTX PRO 6000 Blackwell Workstation Edition",
+      "offload": "streamed_leaf_group_offload",
+      "cuda_memory_cap_gib": 12.0,
+      "live_comfyui_terminal_prompt_id": "b23d25f6-3b6c-4be5-89c7-b3f2604159d0",
+      "api_harness_elapsed_seconds": 105.09174448787235,
+      "load_seconds": 10.83373383898288,
+      "denoise_seconds": 46.68325518886559,
+      "generation_torch_peak_gib": 1.9883627891540527,
+      "full_run_sampled_child_process_peak_gib_approx": 6.04,
+      "four_forward_delivery_gate_child_peak_gib": 6.365,
+      "four_forward_delivery_gate_total_device_peak_gib": 6.902344,
+      "comfyui_idle_context_gib": 0.537,
+      "rss_peak_gib": 42.852699279785156,
+      "video_decode_seconds": 31.002969308989123,
+      "video_decode_torch_peak_gib": 1.0246772766113281,
+      "audio_decode_seconds": 0.1771434131078422,
+      "audio_decode_torch_peak_gib": 0.9356474876403809,
+      "audio_std": 0.17499302327632904,
+      "semantic_review": "pass: coherent macro-to-face shot; aligned eyes and lips; no grid, ghosting, texture collapse, section redraw, narrow whistle, NaN, Inf, or clipping failure"
+    },
+    "speed_t2va": {
       "status": "pass",
-      "prompt_id": "32d20b0b-f8b9-4a8f-aa25-98db4ccc9644",
-      "component": "transformer",
-      "generation_seconds": 137.13920322200283,
-      "generation_peak_vram_gib": 18.90608207695186,
-      "decode_seconds": 5.4824893129989505,
-      "decode_peak_vram_gib": 11.286694642156363,
-      "native_w4_modules": "300/300",
-      "offload": "manual_stage_offload",
-      "text_encoder_after_generation": "cpu",
-      "vae_dtype": "fp32",
-      "audio_vae_dtype": "fp32",
-      "semantic_review": "pass: stable face, eyes, lips, wet-hair and jacket texture; no grid, ghosting, or temporal collapse"
+      "gpu": "NVIDIA RTX PRO 6000 Blackwell Workstation Edition",
+      "offload": "resident_transformer",
+      "process_peak_gib": 21.136719,
+      "torch_peak_gib": 19.136168,
+      "rss_peak_gib": 21.008953,
+      "median_warm_forward_seconds": 2.036486,
+      "denoise_seconds": 46.839179,
+      "generation_seconds": 51.09528,
+      "harness_seconds": 57.535302,
+      "semantic_review": "pass: source-FP32 decode retains final face, wet hair, skin texture, hard-surface rings, and broadband storm audio without a persistent squeal"
     },
-    "ref2va": {
+    "minimum_vram_t2va": {
       "status": "pass",
-      "prompt_id": "4aa7325a-98e2-44ea-9f24-46ec404213b1",
-      "component": "transformer_ref",
-      "generation_seconds": 318.855212665163,
-      "generation_peak_vram_gib": 30.459273928776383,
-      "decode_seconds": 5.50284016598016,
-      "decode_peak_vram_gib": 11.286694642156363,
-      "native_w4_modules": "300/300",
-      "offload": "components_manager_auto_cpu_offload",
-      "offload_reserve_margin": "64GB",
-      "text_encoder_after_generation": "cpu",
-      "vae_dtype": "fp32",
-      "audio_vae_dtype": "fp32",
-      "semantic_review": "pass: face geometry, eyes, lips, visor reflection and suit texture remain coherent through the push-in"
+      "gpu": "NVIDIA GeForce RTX 4090",
+      "offload": "low_cpu_memory_streamed_leaf_group_offload",
+      "cuda_memory_cap_gib": 8.0,
+      "process_peak_gib": 4.066406,
+      "torch_peak_gib": 2.108679,
+      "rss_peak_gib": 35.763912,
+      "median_warm_forward_seconds": 6.706353,
+      "denoise_seconds": 154.246112,
+      "generation_seconds": 188.701789,
+      "harness_seconds": 206.821594,
+      "semantic_review": "pass: independently decoded streamed-order latent retains the dynamic sequence and final face without a grid, collapse, or persistent narrow squeal"
+    },
+    "speed_ref2va": {
+      "status": "pass",
+      "gpu": "NVIDIA RTX PRO 6000 Blackwell Workstation Edition",
+      "offload": "resident_transformer_ref",
+      "process_peak_gib": 24.058594,
+      "torch_peak_gib": 21.026865,
+      "rss_peak_gib": 22.914082,
+      "median_warm_forward_seconds": 5.151465,
+      "denoise_seconds": 118.483702,
+      "generation_seconds": 155.421968,
+      "harness_seconds": 161.40482,
+      "semantic_review": "pass: source-FP32 reference and output VAE stages; all-frame face, rigid-edge, boundary, playback, and audio gates passed"
     }
   },
+  "workflow": {
+    "source": "Comfy-Org/workflow_templates/templates/video_minimax_h3_t2v.json",
+    "ui_file": "workflows/minimax-h3-t2va-ui.json",
+    "api_file_from_live_graph": "workflows/minimax-h3-public-ui-api.json",
+    "workflow_image_export": "workflow-export.png",
+    "export_method": "ComfyUI Workflow Image Export action",
+    "export_width": 3060,
+    "export_height": 1310,
+    "embedded_workflow": true,
+    "embedded_node_count": 6,
+    "embedded_link_count": 2,
+    "embedded_profile": "balanced",
+    "embedded_steps": 24,
+    "embedded_prompt_characters": 1909,
+    "missing_nodes_after_png_import": []
+  },
   "media": {
-    "h264_t2va_sha256": "4cceecedd28dfc593cf5e61145836405d18475bd8109b2ef7a984d96d91a13e6",
-    "hevc_t2va_sha256": "bc31c456f9fdfae29c4e89d063baa953bbd73141c05a237ed43006336661eaab",
-    "h264_ref2va_sha256": "02233da18b3c01761889df2206181e0fc7b9f120165b37a0fe4c9f7bbdda0096",
-    "hevc_ref2va_sha256": "0c3040f81aefdb14b30047a5c972347d012d72dbfa026b8a9b2e245e555b7996",
-    "workflow_sha256": "7c1a82427b80203e4252726a0ab3a336a6859fe2ff43302a07de00bce5722acb",
-    "workflow_image_export_sha256": "df49d477b234e71c23e0a7d457d66315408333db62433885db4b20f3606328f0"
+    "balanced_master_crf1_sha256": "32711ef0f3a3a341340ab88e31537b26eefd486909154b0ba03ada49896f7f21",
+    "balanced_hevc_crf10_sha256": "1bbc05a09f3f49d1564afb1b0c0f49a5f3c3a05a34a28e4f8b28f1d3295f5335",
+    "balanced_h264_preview_sha256": "6c17fa38c5fee6a396a58d7f55981d1cdb644fae5cbbac87a2eb8d4194ad4d2f",
+    "balanced_timeline_sha256": "f081bb49aa12d73c108f8fef963c51c387662590cc41c6b069a0c7a47e7552e5",
+    "balanced_adjacent_triplets_sha256": "a1114d1ba23da74c0ee1d911599cdd50cab3deef5e2de9e9ff9b9ff9167f97db",
+    "balanced_frame_120_sha256": "5a5139474ed5aed90706a002bef1d574b2e737b24164e2367e008167bde34f48",
+    "balanced_audio_spectrum_sha256": "36b4c549e3bf39716325a81ea14417be811ac8b5dd7a71bfe596aef8288ab385",
+    "workflow_image_export_sha256": "164ac2cd2dd51a1aa7a99c213e6777585b48b9efdbe6ad742e39cad8a5143caf"
   }
 }

comfyui/workflows/MiniMax-H3-OrbitQuant-Ref2VA-api.json

@@ -17,12 +17,13 @@
       ],
       "prompt": "Single continuous cinematic shot of the referenced astronaut. The astronaut turns toward the camera and raises one gloved hand as the camera makes a fast smooth push-in toward the helmet visor. The face, eyes, visor reflections, suit patches, and fabric texture remain coherent during motion. No cuts.",
       "task": "ref2va",
+      "inference_profile": "speed",
       "reference_path": "/workspace/evidence/astronaut.png",
       "seed": 314159,
       "width": 608,
       "height": 480,
       "num_frames": 124,
-      "steps": 50,
+      "steps": 24,
       "filename_prefix": "orbitquant/minimax-h3-ref2va"
     },
     "_meta": {

comfyui/workflows/MiniMax-H3-OrbitQuant-T2VA-api.json

@@ -16,14 +16,15 @@
   },
   "105": {
     "inputs": {
-      "prompt": "Single continuous live-action shot of a young woman running toward the camera in rain. The camera rapidly dollies backward, then smoothly pushes into a tight close-up while she turns her head and says in Russian: «Я здесь». Stable face identity, natural eyes and lips, coherent wet hair and jacket texture, strong motion, no cuts.",
+      "prompt": "Create one uninterrupted five-second photorealistic cinematic shot inside a vast abandoned Victorian astronomical conservatory during a midnight thunderstorm. Begin centimetres above a flooded mosaic floor on an extreme macro view of a scratched silver compass as a raindrop creates concentric ripples through reflected amber lamps. In the background, the same young female astronomer with dark hazel eyes, a small scar through her right eyebrow, rain-darkened curls and a burgundy wool coat runs toward an antique brass orrery. The camera performs one fast but smooth backward-and-upward crane with a clockwise half-orbit, passes cleanly between rotating copper rings, racks focus from the compass to her face, then transitions into a controlled push-in. Preserve strong physical parallax, straight architecture, attached compass parts, stable orrery topology, readable glass planets, plausible splashes, coherent wet fabric, fixed facial identity, aligned eyes, natural lips, five stable fingers and consistent hairline throughout every frame. A blue-white lightning flash crosses the dome and illuminates the woman, water, brass and glass coherently, followed by physically delayed thunder. In the final close-up she catches her breath, looks into the lens, grips a small brass lever and says clearly in Russian: «Не бойся грома — смотри, как небо запоминает нас». Synchronize her lips and preserve a single natural voice. Build continuous stereo sound from rain, footsteps, brass gears, wind through cracked glass, lightning and thunder; no electronic squeal, duplicated speech or clipping. Use believable anamorphic photography, restrained halation, realistic skin, natural motion blur, coherent reflections and subtle sensor grain. No cuts, temporal resets, face morphing, crossed eyes, melting mouth, texture crawling, checkerboard grids, duplicated limbs, subtitles, logos, watermarks or black frames.",
       "task": "t2va",
+      "inference_profile": "balanced",
       "reference_path": "",
       "seed": 42,
       "width": 608,
       "height": 480,
       "num_frames": 124,
-      "steps": 50,
+      "steps": 24,
       "filename_prefix": "orbitquant/minimax-h3-t2va",
       "release": [
         "119",

comfyui/workflows/MiniMax-H3-OrbitQuant-T2VA.json

@@ -32,7 +32,7 @@
       "outputs": [],
       "title": "Note: MiniMax H3 OrbitQuant",
       "properties": {},
-      "widgets_values": ["MINIMAX H3 — ORBITQUANT W4A4\n\nThis workflow follows ComfyUI's bundled\nMiniMax H3 Text to Video preset.\n\nLayout:\n- model and usage notes on the left\n- compact generation in the center\n- core Save Video on the right\n\nOfficial preset:\nComfy-Org/workflow_templates\nvideo_minimax_h3_t2v.json\n\nKEY INPUTS\n\nprompt\nDescribe the shot, camera motion, dialogue,\nsound effects, and music in one block.\n\ntask\nt2va for text-to-video. Use ref2va plus a\nlocal reference_path for reference-to-video.\n\nwidth / height\nVerified at 608 × 480.\n\nnum_frames\n124 frames: the shortest verified\nfive-second H3 packing sequence.\n\nsteps\n50 sigma points / 49 denoiser forwards.\n\nOFFLOAD AND PRECISION\n\nThe text encoder enters GPU for conditioning\nand returns to CPU afterward.\n\nT2VA uses manual stage offload. Ref2VA uses\nH3 ComponentsManager for source FP32 VAE\nreference encoding. Visual and audio decode\nuse untouched source FP32 VAEs."],
+      "widgets_values": ["MINIMAX H3 — ORBITQUANT W4A4\n\nThis workflow follows ComfyUI's bundled\nMiniMax H3 Text to Video preset.\n\nLayout:\n- model and usage notes on the left\n- compact generation in the center\n- core Save Video on the right\n\nOfficial preset:\nComfy-Org/workflow_templates\nvideo_minimax_h3_t2v.json\n\nKEY INPUTS\n\nprompt\nDescribe the shot, camera motion, dialogue,\nsound effects, and music in one block.\n\ntask\nt2va for text-to-video. Use ref2va plus a\nlocal reference_path for reference-to-video.\n\ninference_profile\nbalanced is the safe default; speed keeps the\ntransformer resident; minimum_vram streams\nweights with low-CPU-memory loading.\n\nwidth / height\nVerified at 608 × 480.\n\nnum_frames\n124 frames: the shortest verified\nfive-second H3 packing sequence.\n\nsteps\n24 sigma points / 23 denoiser forwards.\n\nOFFLOAD AND PRECISION\n\nThe text encoder executes sequentially on GPU\nfor conditioning and returns to CPU afterward.\nRef2VA reference encoding uses tiled sequential\nGPU VAE execution. Visual and audio decode use\nuntouched source FP32 VAEs."],
       "color": "#222",
       "bgcolor": "#000"
     },
@@ -65,6 +65,7 @@
         {"name": "release", "type": "ORBITQUANT_RELEASE", "link": 1},
         {"name": "prompt", "type": "STRING", "widget": {"name": "prompt"}, "link": null},
         {"name": "task", "type": "COMBO", "widget": {"name": "task"}, "link": null},
+        {"name": "inference_profile", "type": "COMBO", "widget": {"name": "inference_profile"}, "link": null},
         {"name": "reference_path", "type": "STRING", "widget": {"name": "reference_path"}, "link": null},
         {"name": "seed", "type": "INT", "widget": {"name": "seed"}, "link": null},
         {"name": "width", "type": "INT", "widget": {"name": "width"}, "link": null},
@@ -80,14 +81,15 @@
       "title": "Generate Video (Verified 608 × 480)",
       "properties": {"Node name for S&R": "OrbitQuantGenerateVideo", "previewExposures": []},
       "widgets_values": [
-        "Single continuous live-action shot of a young woman running toward the camera in rain. The camera rapidly dollies backward, then smoothly pushes into a tight close-up while she turns her head and says in Russian: «Я здесь». Stable face identity, natural eyes and lips, coherent wet hair and jacket texture, strong motion, no cuts.",
+        "Create one uninterrupted five-second photorealistic cinematic shot inside a vast abandoned Victorian astronomical conservatory during a midnight thunderstorm. Begin centimetres above a flooded mosaic floor on an extreme macro view of a scratched silver compass as a raindrop creates concentric ripples through reflected amber lamps. In the background, the same young female astronomer with dark hazel eyes, a small scar through her right eyebrow, rain-darkened curls and a burgundy wool coat runs toward an antique brass orrery. The camera performs one fast but smooth backward-and-upward crane with a clockwise half-orbit, passes cleanly between rotating copper rings, racks focus from the compass to her face, then transitions into a controlled push-in. Preserve strong physical parallax, straight architecture, attached compass parts, stable orrery topology, readable glass planets, plausible splashes, coherent wet fabric, fixed facial identity, aligned eyes, natural lips, five stable fingers and consistent hairline throughout every frame. A blue-white lightning flash crosses the dome and illuminates the woman, water, brass and glass coherently, followed by physically delayed thunder. In the final close-up she catches her breath, looks into the lens, grips a small brass lever and says clearly in Russian: «Не бойся грома — смотри, как небо запоминает нас». Synchronize her lips and preserve a single natural voice. Build continuous stereo sound from rain, footsteps, brass gears, wind through cracked glass, lightning and thunder; no electronic squeal, duplicated speech or clipping. Use believable anamorphic photography, restrained halation, realistic skin, natural motion blur, coherent reflections and subtle sensor grain. No cuts, temporal resets, face morphing, crossed eyes, melting mouth, texture crawling, checkerboard grids, duplicated limbs, subtitles, logos, watermarks or black frames.",
         "t2va",
+        "balanced",
         "",
         42,
         608,
         480,
         124,
-        50,
+        24,
         "orbitquant/minimax-h3-t2va"
       ]
     },
@@ -120,7 +122,7 @@
       "outputs": [],
       "title": "Note: Verified Profile",
       "properties": {},
-      "widgets_values": ["VERIFIED 480P PROFILE\n\nwidth                 608\nheight                480\nframes                124\nfps                    24\nsigma points           50\ndenoiser forwards      49\nvisual VAE     source FP32\naudio VAE      source FP32\n\nH3 accepts 5–15 seconds. The runner rejects\nshorter frame counts before allocating the\ndenoiser, preventing expensive invalid jobs.\n\nScheduler checkpoints, the latent bundle,\nmetrics, and stage logs are persisted as soon\nas each artifact exists."],
+      "widgets_values": ["VERIFIED 480P PROFILE\n\nwidth                 608\nheight                480\nframes                124\nfps                    24\nsigma points           24\ndenoiser forwards      23\ndefault profile  balanced\nvisual VAE     source FP32\naudio VAE      source FP32\n\nH3 accepts 5–15 seconds. The runner rejects\nshorter frame counts before allocating the\ndenoiser, preventing expensive invalid jobs.\n\nScheduler checkpoints, the latent bundle,\nmetrics, and stage logs are persisted as soon\nas each artifact exists."],
       "color": "#222",
       "bgcolor": "#000"
     }

comfyui_orbitquant.json

@@ -10,6 +10,7 @@
     "width": 608,
     "height": 480,
     "num_frames": 124,
-    "steps": 50
+    "steps": 24,
+    "inference_profile": "balanced"
   }
 }

quantization_manifest.json

@@ -4,7 +4,7 @@
   "source_revision": "73372e6cf53e414edd3ab03e357717fb0602e758",
   "source_license": "minimax-h3-community-license-agreement",
   "repo_id": "WaveCut/MiniMax-H3-OrbitQuant-W4A4",
-  "visibility": "private",
+  "visibility": "public",
   "quant_method": "orbitquant",
   "recipe": "native-w4a4-text-and-denoisers-source-fp32-vaes",
   "weight_bits": 4,
@@ -19,8 +19,8 @@
     "audio_vae"
   ],
   "calibration_data": null,
-  "orbitquant_version": "0.9.0",
-  "orbitquant_revision": "888d9bd02cc718e9e50e478e61fe390654c1da5f",
+  "orbitquant_version": "0.9.2",
+  "orbitquant_revision": "cd58b4ecf77f22b8c4116b3d0b7d4af258e16ba3",
   "diffusers_revision": "abc5e9bf71fd38f53cd471bc3acaa84bc5ecbfdc",
   "components": [
     {
@@ -1554,13 +1554,21 @@
     "transformer_ref": "native_packed_w4a4_bf16_compute",
     "vae": "source_fp32_separate_decode",
     "audio_vae": "source_fp32_separate_decode",
-    "num_inference_steps": 50,
-    "model_evaluations": 49
+    "num_inference_steps": 24,
+    "model_evaluations": 23,
+    "default_inference_profile": "balanced",
+    "inference_profiles": {
+      "balanced": "streamed_leaf_group_offload_with_12_gib_allocator_cap",
+      "speed": "resident_selected_transformer",
+      "minimum_vram": "low_cpu_memory_streamed_leaf_group_offload_with_8_gib_allocator_cap"
+    },
+    "exact_int8_weight_cache": false,
+    "attention_backend": "native_auto_torch_flash_sdpa"
   },
   "validated_runtime_provenance": {
-    "orbitquant_version": "0.9.1",
-    "orbitquant_revision": "73bb2be9add3c86ad60430ec8e2d0a28d7126fc7",
-    "orbitquant_requirement": "orbitquant[hf,kernels]>=0.9.1,<0.10",
+    "orbitquant_version": "0.9.2",
+    "orbitquant_revision": "cd58b4ecf77f22b8c4116b3d0b7d4af258e16ba3",
+    "orbitquant_requirement": "orbitquant[hf,kernels]>=0.9.2,<0.10",
     "diffusers_revision": "abc5e9bf71fd38f53cd471bc3acaa84bc5ecbfdc"
   }
 }

runtime-requirements.txt

@@ -1,4 +1,4 @@
-orbitquant[hf,kernels]>=0.9.1,<0.10
+orbitquant[hf,kernels]>=0.9.2,<0.10
 diffusers @ git+https://github.com/huggingface/diffusers.git@abc5e9bf71fd38f53cd471bc3acaa84bc5ecbfdc
 transformers>=5.13,<6
 accelerate

scripts/decode_h3_latents.py

@@ -2,93 +2,169 @@
 from __future__ import annotations
 
 import argparse
+import gc
 import json
 import resource
 import time
 from pathlib import Path
 
+import av
 import torch
-
-import orbitquant  # noqa: F401 - register OrbitQuant Diffusers loaders
+from accelerate import cpu_offload
 from diffusers import AutoencoderKLMiniMaxH3, AutoencoderKLMiniMaxH3Audio
-from diffusers.modular_pipelines.minimax_h3.packing import MINIMAX_H3_PIXEL_MEAN, MINIMAX_H3_PIXEL_STD
-from diffusers.utils.export_utils import encode_video
+from diffusers.modular_pipelines.minimax_h3.packing import (
+    MINIMAX_H3_PIXEL_MEAN,
+    MINIMAX_H3_PIXEL_STD,
+)
+from diffusers.utils.export_utils import _prepare_audio_stream, _write_audio, encode_video
 from diffusers.video_processor import VideoProcessor
 
 from checkpoint_io import atomic_json_write
-from decode_precision_policy import torch_dtype_name
 from latent_io import load_latent_bundle, prepare_inference_model
 from media_packaging import atomic_media_output
 
 
+def encode_master_crf1(
+    frames: list,
+    *,
+    fps: int,
+    output_path: Path,
+    audio: torch.Tensor,
+    audio_sample_rate: int,
+) -> None:
+    """Write the pre-delivery master before any HEVC recompression."""
+    container = av.open(str(output_path), mode="w", format="mp4")
+    stream = container.add_stream(
+        "libx264",
+        rate=fps,
+        options={"crf": "1", "preset": "slow"},
+    )
+    stream.width = frames[0].width
+    stream.height = frames[0].height
+    stream.pix_fmt = "yuv444p"
+    audio_stream = _prepare_audio_stream(container, audio_sample_rate)
+    try:
+        for image in frames:
+            frame = av.VideoFrame.from_image(image)
+            for packet in stream.encode(frame):
+                container.mux(packet)
+        for packet in stream.encode():
+            container.mux(packet)
+        _write_audio(container, audio_stream, audio, audio_sample_rate, av)
+    finally:
+        container.close()
+
+
 def main() -> int:
     parser = argparse.ArgumentParser()
     parser.add_argument("--latents", type=Path, required=True)
     parser.add_argument("--vae", type=Path, required=True)
     parser.add_argument("--audio-vae", type=Path, required=True)
-    parser.add_argument("--vae-dtype", choices=("orbitquant", "fp32", "bf16"), required=True)
-    parser.add_argument(
-        "--audio-vae-dtype", choices=("orbitquant", "fp32", "bf16"), required=True
-    )
     parser.add_argument("--output", type=Path, required=True)
+    parser.add_argument("--preview-output", type=Path, required=True)
+    parser.add_argument("--disable-spatial-tiling", action="store_true")
     args = parser.parse_args()
 
     report_path = args.output.with_suffix(".decode.metrics.json")
-    report_path.parent.mkdir(parents=True, exist_ok=True)
     report: dict[str, object] = {
         "status": "running",
         "latent_bundle": str(args.latents),
         "vae": str(args.vae),
-        "vae_dtype": args.vae_dtype,
+        "vae_dtype": "fp32_source",
         "audio_vae": str(args.audio_vae),
-        "audio_vae_dtype": args.audio_vae_dtype,
+        "audio_vae_dtype": "fp32_source",
+        "offload": {
+            "video_vae": "sequential_cpu_offload",
+            "audio_vae": "resident_cuda_during_audio_stage",
+        },
+        "video_vae_tiling": not args.disable_spatial_tiling,
+        "master_video": {
+            "codec": "libx264",
+            "crf": 1,
+            "preset": "slow",
+            "pixel_format": "yuv444p",
+        },
         "gpu": torch.cuda.get_device_name(),
     }
+    report_path.parent.mkdir(parents=True, exist_ok=True)
+    atomic_json_write(report, report_path)
     try:
         bundle = load_latent_bundle(args.latents)
         metadata = dict(bundle["metadata"])
         report["metadata"] = metadata
 
-        load_started = time.perf_counter()
-        vae_kwargs = {"low_cpu_mem_usage": True}
-        vae_dtype_name = torch_dtype_name(args.vae_dtype)
-        if vae_dtype_name is not None:
-            vae_kwargs["dtype"] = getattr(torch, vae_dtype_name)
+        video_load_started = time.perf_counter()
         vae = prepare_inference_model(
-            AutoencoderKLMiniMaxH3.from_pretrained(args.vae, **vae_kwargs)
-        ).to("cuda")
-
-        audio_vae_kwargs = {"low_cpu_mem_usage": True}
-        audio_vae_dtype_name = torch_dtype_name(args.audio_vae_dtype)
-        if audio_vae_dtype_name is not None:
-            audio_vae_kwargs["dtype"] = getattr(torch, audio_vae_dtype_name)
-        audio_vae = prepare_inference_model(
-            AutoencoderKLMiniMaxH3Audio.from_pretrained(args.audio_vae, **audio_vae_kwargs)
-        ).to("cuda")
-        report["load_seconds"] = time.perf_counter() - load_started
+            AutoencoderKLMiniMaxH3.from_pretrained(
+                args.vae,
+                low_cpu_mem_usage=True,
+                dtype=torch.float32,
+            )
+        )
+        if args.disable_spatial_tiling:
+            vae.disable_tiling()
+        else:
+            vae.enable_tiling()
+        cpu_offload(
+            vae,
+            execution_device=torch.device("cuda"),
+            offload_buffers=True,
+        )
+        report["video_vae_load_seconds"] = time.perf_counter() - video_load_started
 
         torch.cuda.reset_peak_memory_stats()
-        decode_started = time.perf_counter()
+        video_decode_started = time.perf_counter()
         video_latents = bundle["video_latents"].to("cuda")
         with torch.inference_mode(), torch.autocast(device_type="cuda", dtype=torch.float16):
             video = vae.decode(video_latents, return_dict=False)[0]
-        pixel_mean = torch.tensor(MINIMAX_H3_PIXEL_MEAN, device="cuda").view(1, -1, 1, 1, 1)
-        pixel_std = torch.tensor(MINIMAX_H3_PIXEL_STD, device="cuda").view(1, -1, 1, 1, 1)
+        pixel_mean = torch.tensor(MINIMAX_H3_PIXEL_MEAN, device="cuda").view(
+            1, -1, 1, 1, 1
+        )
+        pixel_std = torch.tensor(MINIMAX_H3_PIXEL_STD, device="cuda").view(
+            1, -1, 1, 1, 1
+        )
         video = (video.float() * pixel_std + pixel_mean).clamp(0, 1)
         videos = VideoProcessor(vae_scale_factor=16, do_normalize=False).postprocess_video(
-            video, output_type="pil"
+            video,
+            output_type="pil",
         )
+        torch.cuda.synchronize()
+        report["video_decode_seconds"] = time.perf_counter() - video_decode_started
+        report["video_decode_torch_peak_bytes"] = torch.cuda.max_memory_allocated()
+        report["video_frames"] = len(videos[0])
+
+        del vae, video_latents, video, pixel_mean, pixel_std
+        gc.collect()
+        torch.cuda.empty_cache()
+
+        audio_load_started = time.perf_counter()
+        audio_vae = prepare_inference_model(
+            AutoencoderKLMiniMaxH3Audio.from_pretrained(
+                args.audio_vae,
+                low_cpu_mem_usage=True,
+                dtype=torch.float32,
+            )
+        ).to("cuda")
+        report["audio_vae_offload"] = "resident_cuda_during_audio_stage"
+        report["audio_vae_load_seconds"] = time.perf_counter() - audio_load_started
 
+        torch.cuda.reset_peak_memory_stats()
+        audio_decode_started = time.perf_counter()
         audio_latents = bundle["audio_latents"].to("cuda")
         with torch.inference_mode():
             audio = audio_vae.decode(audio_latents, return_dict=False)[0]
         audio = audio.float().permute(1, 0, 2)
+        if not torch.isfinite(audio).all() or torch.count_nonzero(audio) == 0:
+            raise RuntimeError("source audio VAE produced invalid or silent audio")
         torch.cuda.synchronize()
-        report["decode_seconds"] = time.perf_counter() - decode_started
-        report["cuda_decode_peak_bytes"] = torch.cuda.max_memory_allocated()
+        report["audio_decode_seconds"] = time.perf_counter() - audio_decode_started
+        report["audio_decode_torch_peak_bytes"] = torch.cuda.max_memory_allocated()
+        report["audio_min"] = float(audio.min())
+        report["audio_max"] = float(audio.max())
+        report["audio_std"] = float(audio.std())
 
-        encode_started = time.perf_counter()
-        with atomic_media_output(args.output) as partial:
+        preview_started = time.perf_counter()
+        with atomic_media_output(args.preview_output) as partial:
             encode_video(
                 videos[0],
                 fps=int(metadata["fps"]),
@@ -96,8 +172,21 @@ def main() -> int:
                 audio=audio[0],
                 audio_sample_rate=int(metadata["sampling_rate"]),
             )
+        report["preview_encode_seconds"] = time.perf_counter() - preview_started
+        report["preview_output"] = str(args.preview_output)
+        report["preview_output_bytes"] = args.preview_output.stat().st_size
+        atomic_json_write(report, report_path)
+
+        encode_started = time.perf_counter()
+        with atomic_media_output(args.output) as partial:
+            encode_master_crf1(
+                videos[0],
+                fps=int(metadata["fps"]),
+                output_path=partial,
+                audio=audio[0],
+                audio_sample_rate=int(metadata["sampling_rate"]),
+            )
         report["encode_seconds"] = time.perf_counter() - encode_started
-        report["video_frames"] = len(videos[0])
         report["output_bytes"] = args.output.stat().st_size
         report["rss_peak_bytes"] = resource.getrusage(resource.RUSAGE_SELF).ru_maxrss * 1024
         report["status"] = "pass"

scripts/manual_stage_offload.py

@@ -3,6 +3,8 @@ from __future__ import annotations
 
 from collections.abc import Callable
 
+from accelerate import cpu_offload
+
 
 class _ExecutionDeviceHint:
     execution_device = "cuda"
@@ -40,6 +42,8 @@ def install_manual_h3_stage_offload(
     text_encoder,
     transformer,
     empty_cuda_cache: Callable[[], None],
+    place_transformer: bool = True,
+    sequential_text_encoder: bool = False,
 ) -> None:
     """Run conditioning on CUDA, then free it before placing the denoiser."""
     install_h3_rope_runtime_alignment(transformer)
@@ -63,13 +67,18 @@ def install_manual_h3_stage_offload(
 
     def call_then_place_denoiser(step, components, state):
         result = original_call(step, components, state)
-        text_encoder.to("cpu")
-        text_encoder._hf_hook = _ExecutionDeviceHint()
+        if not sequential_text_encoder:
+            text_encoder.to("cpu")
+            text_encoder._hf_hook = _ExecutionDeviceHint()
         empty_cuda_cache()
-        transformer.to("cuda")
-        move_registered_buffers(transformer, "cuda")
+        if place_transformer:
+            transformer.to("cuda")
+            move_registered_buffers(transformer, "cuda")
         return result
 
     encoder_step_cls.encode_prompt = staticmethod(encode_prompt_on_cuda)
     encoder_step_cls.__call__ = call_then_place_denoiser
-    text_encoder.to("cuda")
+    if sequential_text_encoder:
+        cpu_offload(text_encoder, execution_device="cuda", offload_buffers=True)
+    else:
+        text_encoder.to("cuda")

scripts/run_quantized_example.py

@@ -56,7 +56,7 @@ BF16_COMPUTE_COMPONENTS = frozenset({"transformer", "text_encoder"})
 
 def component_load_request(component: str, spec: dict[str, object]) -> tuple[object, dict]:
     kwargs: dict[str, object] = {"low_cpu_mem_usage": True}
-    if component in BF16_COMPUTE_COMPONENTS or spec["source"] == "bf16":
+    if component in BF16_COMPUTE_COMPONENTS or spec["source"] in {"bf16", "bf16_local"}:
         kwargs["dtype"] = torch.bfloat16
     if spec["source"] == "bf16":
         kwargs["revision"] = spec["revision"]
@@ -84,6 +84,14 @@ def main() -> int:
         default=[],
         help="Load one learned component from the pinned BF16 source for a controlled ablation.",
     )
+    parser.add_argument(
+        "--source-bf16-root",
+        type=Path,
+        help=(
+            "Use a completely downloaded local source-model root for BF16 "
+            "ablation components instead of fetching them from the Hub."
+        ),
+    )
     parser.add_argument(
         "--transformer-path",
         type=Path,
@@ -112,6 +120,24 @@ def main() -> int:
         default="64GB",
         help="ComponentsManager CUDA memory reserve margin, for example 12GB.",
     )
+    parser.add_argument(
+        "--transformer-group-offload-blocks",
+        type=int,
+        help="Enable block-level transformer group offload with this many blocks per group.",
+    )
+    parser.add_argument(
+        "--transformer-group-offload-type",
+        choices=("block_level", "leaf_level"),
+        default="block_level",
+    )
+    parser.add_argument("--group-offload-use-stream", action="store_true")
+    parser.add_argument("--group-offload-no-record-stream", action="store_true")
+    parser.add_argument("--group-offload-low-cpu-mem-usage", action="store_true")
+    parser.add_argument("--cuda-memory-cap-gib", type=float)
+    parser.add_argument("--text-encoder-sequential-offload", action="store_true")
+    parser.add_argument("--reference-vae-sequential-offload", action="store_true")
+    parser.add_argument("--reference-vae-tile-size", type=int, default=256)
+    parser.add_argument("--attention-backend")
     parser.add_argument(
         "--checkpoint-dir",
         type=Path,
@@ -150,12 +176,38 @@ def main() -> int:
     args = parser.parse_args()
     if args.task == "ref2va" and not args.reference:
         parser.error("--reference is required for ref2va")
+    if args.reference_vae_sequential_offload and args.task != "ref2va":
+        parser.error("--reference-vae-sequential-offload requires ref2va")
+    if args.reference_vae_tile_size < 64 or args.reference_vae_tile_size % 16:
+        parser.error("--reference-vae-tile-size must be a multiple of 16 and at least 64")
     if args.no_cpu_offload and args.manual_stage_offload:
         parser.error("--no-cpu-offload and --manual-stage-offload are mutually exclusive")
     if args.stop_after_denoise_steps is not None and not (
         1 <= args.stop_after_denoise_steps < args.steps
     ):
         parser.error("--stop-after-denoise-steps must be between 1 and steps - 1")
+    if args.transformer_group_offload_blocks is not None and args.transformer_group_offload_blocks < 1:
+        parser.error("--transformer-group-offload-blocks must be positive")
+    if args.cuda_memory_cap_gib is not None and args.cuda_memory_cap_gib <= 0:
+        parser.error("--cuda-memory-cap-gib must be positive")
+    if (
+        args.transformer_group_offload_blocks is not None
+        or args.transformer_group_offload_type == "leaf_level"
+    ) and not args.manual_stage_offload:
+        parser.error("transformer group offload requires --manual-stage-offload")
+    if (
+        args.group_offload_use_stream
+        and args.transformer_group_offload_type == "block_level"
+        and args.transformer_group_offload_blocks != 1
+    ):
+        parser.error("Diffusers stream group offload requires exactly one block per group")
+
+    cuda_memory_fraction = None
+    if args.cuda_memory_cap_gib is not None:
+        total_cuda_bytes = torch.cuda.get_device_properties(torch.cuda.current_device()).total_memory
+        requested_cuda_bytes = int(args.cuda_memory_cap_gib * 1024**3)
+        cuda_memory_fraction = min(1.0, requested_cuda_bytes / total_cuda_bytes)
+        torch.cuda.set_per_process_memory_fraction(cuda_memory_fraction)
 
     metrics_path = args.output.with_suffix(".metrics.json")
     metrics_path.parent.mkdir(parents=True, exist_ok=True)
@@ -168,6 +220,8 @@ def main() -> int:
         / f"{args.output.stem}-pid-{os.getpid()}"
     ).resolve()
     bf16_components = set(args.bf16_component)
+    if args.source_bf16_root is not None and "transformer" not in bf16_components:
+        parser.error("--source-bf16-root requires --bf16-component transformer")
     component_paths = (
         {"transformer": args.transformer_path.resolve()}
         if args.transformer_path is not None
@@ -179,6 +233,15 @@ def main() -> int:
         bf16_components=bf16_components,
         component_paths=component_paths,
     )
+    if args.source_bf16_root is not None:
+        transformer_subfolder = "transformer_ref" if args.task == "ref2va" else "transformer"
+        transformer_path = args.source_bf16_root.resolve() / transformer_subfolder
+        if not transformer_path.is_dir():
+            parser.error(f"local BF16 transformer directory does not exist: {transformer_path}")
+        component_plan["transformer"] = {
+            "source": "bf16_local",
+            "path": str(transformer_path),
+        }
     if args.transformer_runtime_mode == "dequant_bf16":
         variant = "orbitquant_w4a4_text_w4a16_transformer"
     elif not bf16_components:
@@ -204,9 +267,14 @@ def main() -> int:
         "checkpoint_dir": str(checkpoint_dir),
         "checkpoint_policy": "atomic_block_state_after_each_scheduler_step",
         "gpu": torch.cuda.get_device_name(),
+        "cuda_memory_cap_gib": args.cuda_memory_cap_gib,
+        "cuda_memory_fraction": cuda_memory_fraction,
         "pid": os.getpid(),
     }
     try:
+        if args.group_offload_use_stream:
+            report["stream_safe_orbitquant_buffer_identity"] = "package"
+
         def load_component(component: str, cls):
             spec = component_plan[component]
             path, kwargs = component_load_request(component, spec)
@@ -250,11 +318,49 @@ def main() -> int:
             else 0
         )
         report["native_w4_preflight"] = validate_native_w4_compute_dtype(transformer)
+        if args.attention_backend:
+            if args.attention_backend == "sage_hub":
+                from diffusers.models.attention_dispatch import (
+                    AttentionBackendName,
+                    _HUB_KERNELS_REGISTRY,
+                )
+
+                sage_config = _HUB_KERNELS_REGISTRY[AttentionBackendName.SAGE_HUB]
+                sage_config.revision = None
+                sage_config.version = 2
+                sage_config.kernel_fn = None
+                report["sage_hub_kernel_version"] = 2
+            transformer.set_attention_backend(args.attention_backend)
+        report["attention_backend"] = args.attention_backend or "native_auto"
         report["w4a4_int8_weight_cache_modules"] = (
             enable_w4a4_int8_weight_cache(transformer)
             if args.w4a4_int8_weight_cache
             else 0
         )
+        group_offload_enabled = (
+            args.transformer_group_offload_blocks is not None
+            or args.transformer_group_offload_type == "leaf_level"
+        )
+        if group_offload_enabled:
+            group_offload_kwargs = {
+                "onload_device": torch.device("cuda"),
+                "offload_device": torch.device("cpu"),
+                "offload_type": args.transformer_group_offload_type,
+                "non_blocking": args.group_offload_use_stream,
+                "use_stream": args.group_offload_use_stream,
+                "record_stream": (
+                    args.group_offload_use_stream
+                    and not args.group_offload_no_record_stream
+                ),
+                "low_cpu_mem_usage": args.group_offload_low_cpu_mem_usage,
+            }
+            if args.transformer_group_offload_type == "block_level":
+                group_offload_kwargs["num_blocks_per_group"] = (
+                    args.transformer_group_offload_blocks
+                )
+            transformer.enable_group_offload(
+                **group_offload_kwargs,
+            )
         report["load_seconds"] = time.perf_counter() - load_started
 
         move_started = time.perf_counter()
@@ -272,11 +378,35 @@ def main() -> int:
                 text_encoder=component_updates["text_encoder"],
                 transformer=transformer,
                 empty_cuda_cache=torch.cuda.empty_cache,
+                place_transformer=not group_offload_enabled,
+                sequential_text_encoder=args.text_encoder_sequential_offload,
             )
             report["offload"] = {
-                "mode": "manual_stage_offload",
-                "conditioner": "cuda_then_cpu",
-                "transformer": "cpu_then_cuda",
+                "mode": (
+                    "manual_stage_plus_transformer_group_offload"
+                    if group_offload_enabled
+                    else "manual_stage_offload"
+                ),
+                "conditioner": (
+                    "sequential_cuda_layers_then_cpu"
+                    if args.text_encoder_sequential_offload
+                    else "cuda_then_cpu"
+                ),
+                "transformer": (
+                    (
+                        f"block_level_{args.transformer_group_offload_blocks}"
+                        if args.transformer_group_offload_type == "block_level"
+                        else "leaf_level"
+                    )
+                    if group_offload_enabled
+                    else "cpu_then_cuda"
+                ),
+                "group_offload_use_stream": args.group_offload_use_stream,
+                "group_offload_record_stream": (
+                    args.group_offload_use_stream
+                    and not args.group_offload_no_record_stream
+                ),
+                "group_offload_low_cpu_mem_usage": args.group_offload_low_cpu_mem_usage,
                 "vae": "cpu",
                 "audio_vae": "cpu",
             }
@@ -285,6 +415,24 @@ def main() -> int:
                 components_manager,
                 memory_reserve_margin=args.offload_reserve_margin,
             )
+        if args.reference_vae_sequential_offload:
+            from accelerate import cpu_offload
+
+            reference_vae = component_updates["vae"]
+            reference_vae.enable_tiling(
+                tile_sample_min_height=args.reference_vae_tile_size,
+                tile_sample_min_width=args.reference_vae_tile_size,
+                tile_sample_min_overlap_height=args.reference_vae_tile_size // 4,
+                tile_sample_min_overlap_width=args.reference_vae_tile_size // 4,
+            )
+            cpu_offload(
+                reference_vae,
+                execution_device=torch.device("cuda"),
+                offload_buffers=True,
+            )
+            report["offload"]["vae"] = "sequential_cuda_layers_for_reference_then_cpu"
+            report["offload"]["reference_vae_tiling"] = True
+            report["offload"]["reference_vae_tile_size"] = args.reference_vae_tile_size
         report["transformer_dequant_cache_disabled_modules"] = (
             disable_dequantized_weight_cache(transformer, execution_device="cuda")
             if args.disable_transformer_dequant_cache