Accelerate MiniMax H3 inference on DGX Spark

[LynnColeArt]

Summary

• rebase the cumulative optimization work onto current main (7bd148f)
• retain cuDNN + regional compile + CUTLASS, without cache, as the measured full-compute default
• add request-level mask reuse, fused SwiGLU, static prompt/RoPE conditioning, and direct modality scatters
• retain corrected deterministic Humming as an explicit eager/no-cache alternative with fail-closed profile validation
• persist linear, Triton, and TorchInductor kernel caches across container recreation
• gate readiness on a valid two-step 768x448x56 FL2VA text-to-video warmup
• add a two-render fixed-seed decoded video/audio acceptance command

Current default-profile validation

The rebased review candidate was built and cold-started on one DGX Spark with the documented defaults:

• model load: 89.1659 GiB in 482.705 seconds
• production-shape startup warmup: 33.738 seconds
• application readiness: about 9 minutes 2 seconds after container start
• fixed-seed full HTTP renders: 115.204 and 113.295 seconds (114.250-second mean)
• both outputs: 56 H.264 frames at 768x448/24 fps plus AAC-LC stereo
• every decoded video frame was byte-identical; both audio channels compared at infinite PSNR
• no cuDNN or linear-kernel fallback, OOM kill, or restart

The optional Humming measurements are documented separately as historical eager/no-cache evidence. Humming is not the default; the launcher rejects it with regional compile or Cache-DiT.

Validation

• 18 pinned-image pytest cases passed
• launcher attention, execution, cache, linear, and startup-warmup rejection paths passed
• network security tests passed
• public repository audit passed after acceptance artifacts were generated
• Compose configuration, shell syntax, and branch-wide whitespace checks passed
• current-default cold startup, health/model identity, two full renders, full FFmpeg decode, fixed-seed frame/audio comparison, and no-fallback checks passed

Operator controls

• attention fallback: H3_DIFFUSION_ATTENTION_BACKEND=TORCH_SDPA
• eager execution: H3_EXECUTION_MODE=eager
• optional deterministic Humming (eager/no-cache only): H3_LINEAR_BACKEND=HUMMING
• diagnostic lazy first request: H3_STARTUP_WARMUP=false

Generated media is not included because MiniMax H3 outputs are license-sensitive.