True CUDA text-to-video with live UMT5 encode, dependency-minimal (no cuBLAS)

1. Contribution

Claimed. A self-contained CUDA text-to-video pipeline (Wan2.1) that performs a live UMT5 text encode — with nothing injected from the reference (no pre-computed prompt embeddings) — then denoises with classifier-free guidance over the Wan DiT and decodes through the 3D-VAE. Against the diffusers reference with matched seed and schedule, the early-step (step 0-3) latents and the output frames reach cosine 1.0 at frame PSNR ~89.8 dB. End-to-end wall time is ~92.5 s on a consumer-class NVIDIA GPU, and the build is dependency-minimal: a linkage check shows the CUDA runtime only, no cuBLAS. The altitude is correctness validation plus an engineering port, not a novel algorithm: it mirrors an existing AMD/ROCm standalone pipeline onto CUDA and pins numeric parity against diffusers.

NOT claimed. This does not establish generation quality (cosine/PSNR vs a reference are numeric-parity metrics, not perceptual or human-preference quality). It does not claim throughput or speed competitiveness as a benchmark — the ~92.5 s wall is a single end-to-end run, not a profiled, multi-trial throughput figure. It does not claim parity at higher resolution, longer clips, other Wan variants, other quantization, or on other GPU architectures. The speed contrast with an AMD/ROCm path (~92.5 s vs ~1130 s) is across different hardware and is reported as context, not as a controlled speedup.

2. Methods

3. Results

4. Threats to validity / Limitations

• Sample size. Parity is established on what the source describes as a single end-to-end generation (early-step latents + frames). This is a correctness smoke test, not a benchmark-scale evaluation across prompts, seeds, resolutions, or clip lengths.
• Single run, no variance. The ~92.5 s wall is one run; no repeated trials, variance, or profiling are recorded. It is a wall-clock observation, not a throughput benchmark.
• Parity is not quality. Cosine 1.0 and PSNR ~89.8 dB measure numeric agreement with diffusers, not perceptual quality, temporal coherence, or human preference. A pipeline can match a reference numerically and still produce poor video if the reference is poor.
• Oracle/version pins missing. The diffusers version, precision (fp16/fp32), the parity gate threshold, the prompt, the seed value, and the schedule/step count are not recorded in the source notes. Reproducibility and the strength of the cosine 1.0 / PSNR claim depend on these and cannot be fully audited here.
• Cross-hardware contrast. The ~92.5 s vs ~1130 s comparison spans an NVIDIA/CUDA target vs an AMD/ROCm target — different silicon and backends — so it is not a controlled speedup measurement.
• Scope. Bounded to a self-contained Wan2.1 text-to-video pipeline on the NVIDIA target; no claim for other Wan variants, quantizations, or other GPU architectures.
• Evidence level. The source describes parity against the diffusers oracle, but the absence of recorded version/seed/gate pins (above) means the result should be read as oracle-parity-validated rather than independently re-run from a fully pinned recipe.

5. Reproducibility / Provenance

• Build: CUDA toolchain with optimization enabled. Run command: not recorded in source notes.
• Dependency footprint (linkage check): CUDA runtime only (no cuBLAS/cuDNN/cuTLASS).
• Version pins (CUDA, diffusers, UMT5/Wan weights): not recorded in source notes.
• Oracle caveat: validate the diffusers reference (seed, schedule, precision) before trusting the cosine 1.0 / PSNR ~89.8 dB parity; these pins are not captured in the source.
• Relation: CUDA counterpart to an AMD/ROCm standalone Wan2.1 text-to-video pipeline; correctness-pinned to the same diffusers reference.