Repo-only preparation for running FastH3 video generation on a four-card NVIDIA CMP 170HX node (SM 8.0, 64 GiB per card, 180 W per card — generic, public hardware labeling only; no private infrastructure identifiers appear in this repository).
Internal status: WAITING_LICENSE + WAITING_RESOURCE
Current status: WAITING_LICENSE + WAITING_RESOURCE. This repository is documentation and planning only until both gates clear: written jurisdiction confirmation for the MiniMax H3 Community License territory closure, and node access. Pins re-verified against upstream sources on 2026-08-31 (see MANIFEST.md).
We want to know whether FastH3 (a 4-step distilled MiniMax-H3 video model with VSA attention) is usable on a CMP 170HX node. CMP 170HX is a GA100 part: compute capability 8.0 (SM 8.0), not Blackwell, so the checkpoint's native sm100a kernels cannot run and the documented non-Blackwell route must be used. Two candidate lanes are prepared; nothing has been executed.
| Lane A — BF16 baseline | Lane B — Kijai INT8 "convrot" | |
|---|---|---|
| Checkpoint | FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree @ b65818d41939b5085451074fe8ca8b799f8d4921 |
Kijai/MiniMax-H3-experimental @ f4cac997f880e93cf6940af61ee8d58ef31ff7f3 |
| Size | 147,870,258,925 B = 137.71 GiB total (verified from the revision's file listing) | 22,898,594,920 B (DiT) + 3,171,670,912 B (VAE), sha256 pinned in MANIFEST.md |
| Runtime | FastVideo @ commit 48a047c05ff4138f20cfa33351499c6ec5945f5d |
ComfyUI 0.31.0 + ComfyUI PR #15958 (draft, pinned head) + comfy-kitchen PR #117 (merged) |
| Non-default flags | --no-replicated-dit --vsa-kernel triton --no-fa4 (documented non-Blackwell path) |
pinned ComfyUI / PR heads |
| Static fit (arithmetic only) | ≈ 34.4 GiB weights/card across 4 cards vs 64 GiB VRAM | ≈ 24.28 GiB total → single-card possible; 2-card topology hypothesis |
| Untested on CMP 170HX? | Yes — untested | Yes — untested (sm_80+ support is community-reported only) |
Lane B's checkpoint appears to be an INT8 repack of the same
fastvideo_vsa_datafree baseline (inferred from the file name; to be
confirmed before any comparative claims are made).
The pinned model weights are licensed under the MiniMax H3 Community License, whose Applicable Territory excludes the EU, the UK, the Republic of Korea, and the United States. Use in any jurisdiction requires confirming the deployment location is inside the Applicable Territory before any download or run. The harness explicitly treats this as a gate it must not assume away — see docs/harness.md (license gate stop condition) and MANIFEST.md. This repository's code/docs are Apache-2.0 (see LICENSE); that does not change the weights' own license.
| File | Contents |
|---|---|
| MANIFEST.md | Pinned revisions, byte counts, sha256, license clause summary |
| docs/compatibility.md | SM 8.0 vs sm100a, triton VSA route, head-divisibility rule, ComfyUI lane status — with claim labels |
| docs/static-fit.md | Paper arithmetic for both lanes vs 64 GiB/card; exact vs estimated numbers marked |
| docs/harness.md | One-command harness spec: pinned env, exact flags, benchmark suite, storage plan, gates, stop conditions |
| docs/quality-eval.md | Blinded 1–5 rubric protocol, AI-generation identifier requirement, publication labeling |
| docs/publication-plan.md | Two-PR rule, sanitization checklist, release gate |
This scaffold is published for repo-only preparation. No weights have been downloaded, no GPU run has been performed, and no packages have been installed. Everything here is static analysis of externally reported facts. The next physical step (download + smoke run) is blocked on resource availability and on the license/fit gates in docs/harness.md.