One full-stack MiniMax-H3 policy, adapted to two generations of NVIDIA data center GPUs. Four GPUs generate a 1344x768, 124-frame audio-video clip with no model offload and the original BF16 checkpoint.
MiniMax-H3 packs text, audio, conditioning inputs, and target video into a single sequence. At this release setting, every attention layer processes 38,247 tokens, making attention and repeated denoising the natural targets for end-to-end acceleration.
End-to-end inference time comes directly from the released SGLang runtime logging after one warmup request. Dense and full-opt use the same prompt, seed, model, resolution, and four-GPU topology on each platform.
Warm, batch 1, 1344x768, 124 frames, 50 steps, Ulysses-4, no offload. Peak full-opt memory is 49,154 MiB per H100 and 49,126 MiB per A100.
Full optimization is the complete Sol Engine stack. Every pair uses the same prompt and seed, and each clip carries its own generated audio track.
Cinematic push-in on a starship bridge: a captain in a high-collared navy tunic stands silhouetted at the observation window as a dreadnought armada charges its hyperdrives. A blinding flash, the bridge shudders, and the fleet is gone. [full prompt compacted]
A flamenco dancer spins fast on a wooden stage under a single warm spotlight, her red skirt flaring wide, heels striking the boards in sharp rapid bursts, a guitar driving underneath and hands clapping on the offbeat, camera circling with her.
Two mechanics in a garage at night lean over an open engine bay. The older one wipes his hands and says “It’s not the pump.” The younger one answers “Then what is it?” A radio murmurs in the corner, a wrench clatters onto the concrete, fluorescent tubes buzz overhead.
The model-level decisions are identical on both systems. Only the sparse-attention kernel is specialized for the target GPU generation.
First 10 steps and first two layers stay dense. Sol-Attn uses tau=1.0, no token reordering, and an exact 951-token multimodal prefix sink.
The SM90 path uses a CuTe sparse-attention backend while retaining the same online routing and exact prefix behavior.
The SM80 path uses a portable Triton backend, preserving the policy and distributed execution without requiring Hopper-only instructions.
| Measured property | 4× H100 | 4× A100 |
|---|---|---|
| Effective Sol-Attn route density | 18.04% | 18.04% |
| FirstBlockCache reuse rate | 67.35% | 67.35% |
| Full-opt peak memory per GPU | 49,154 MiB | 49,126 MiB |
| Distributed execution | Ulysses-4 | Ulysses-4 |
| Model offload | Disabled | Disabled |
From the Sol Engine repository root, choose the manifest matching the allocated GPU node.
python scripts/launch_candidate.py \ candidates/minimax_h3_h100_fullopt_exact.toml \ --mode sbatch --confirm-submit
python scripts/launch_candidate.py \ candidates/minimax_h3_a100_fullopt_exact.toml \ --mode sbatch --confirm-submit