True standalone text-to-video on gfx1030, sovereign (no vendor BLAS)

1. Contribution

Claimed. A true standalone sovereign text-to-video pipeline on a single AMD RDNA2 GPU (gfx1030, 16 GB), where the entire chain runs in-process from a prompt string to decoded frames with nothing injected from a reference implementation: in-process UMT5-XXL encode (host-resident streamed weights, run first then freed) producing zero-padded prompt embeddings; a seeded random init latent; an in-process FlowMatch-Euler schedule (matched to diffusers, shift=3); a 20-step CFG denoise over a Wan video transformer (WanTransformer3DModel-equivalent); and a Wan 3D-VAE frame-parametric decode. Every linear (text encoder and transformer) is a hand-written tiled GEMM, and the binary links only the system HIP runtime and standard system libraries — no vendor BLAS. Against a diffusers fp32 oracle with matched seed and schedule, prompt embeddings, negative embeddings, early-step latents, and decoded frames all match at cosine 1.0 (verdict PASS).

NOT claimed. This is a correctness-validation and systems-integration result, not a novel algorithm and not a performance result. It does not claim throughput competitiveness (the hand-written GEMM is slower than vendor BLAS). It does not claim generation or perceptual quality — cosine 1.0 against a single reference is numeric parity of the computation, not an evaluation of video quality. It does not claim coverage of longer clips, higher resolutions, other Wan variants, or other backends. The altitude is honest: this supersedes an earlier bounded harness that injected the reference implementation's prompt embeddings, init latent, and sigmas — the contribution here is closing that gap so the pipeline is end-to-end self-contained, plus the sovereignty (no vendor BLAS) property.

2. Methods

The pipeline chain, in order: (1) UMT5-XXL encode via hand-written GEMM, host-resident streamed weights, run first then memory freed, producing zero-padded prompt embeddings; (2) a seeded random init latent; (3) an in-process FlowMatch-Euler schedule matched to diffusers, shift=3; (4) a 20-step CFG denoise over the Wan video transformer; (5) denormalize and decode via the Wan 3D-VAE.

3. Results

4. Threats to validity / Limitations

• Parity != quality. Cosine 1.0 establishes that the sovereign computation reproduces the diffusers fp32 computation; it says nothing about perceptual video quality. There is no human or metric-based quality evaluation.
• Single reference, single configuration. Parity is measured against one oracle (diffusers fp32) at one seed, one schedule (FlowMatch-Euler shift=3, 20 steps), one resolution (240x416), one clip length (13 frames), one Wan variant. This is a correctness smoke test pinning layout and orientation of the chain — not a benchmark-scale or multi-prompt evaluation.
• Single-run performance, no variance. The wall time, per-step transformer cost, and resident memory are single-run figures with no repetition, no variance or error bars, and no isolation of warm versus cold effects.
• Performance is explicitly non-competitive. The hand-written GEMM is slower than vendor BLAS; the wall time is the cost of the sovereignty- and correctness-first choice. No throughput claim is supported.
• Oracle version not pinned in source. The diffusers version used as the fp32 oracle is not recorded in these source notes; a skeptical reader cannot fully reproduce the exact oracle without it. Validate the oracle independently before trusting parity.
• Scope bounds. Results do not transfer to longer clips, higher resolution, other Wan variants, other GPUs or backends, or other quantization or precision regimes.
• Evidence level. The source notes describe a parity run with verdict PASS but do not, within these notes, document an independent second-party reproduction of the full standalone chain.

5. Reproducibility / Provenance

• Run. A reference-parity mode exercises the standalone string-to-frames chain and compares against diffusers fp32 with matched seed and schedule.
• Config pins. Fixed seed; FlowMatch-Euler shift=3; 20 denoise steps; small video latent; 13-frame output at 240x416.
• Sovereignty check. Inspecting the linked libraries of the binary should show only the system HIP runtime and standard system libraries — no vendor BLAS, DNN, or MKL.
• Validate the oracle first. The diffusers oracle version is not recorded in source notes; pin and validate it before relying on the cosine-1.0 parity.
• Build. Compiled ahead-of-time for the gfx1030 architecture against the system HIP runtime.