vLLM-Omni H3-Image Support: 4K Generation Performance on NVIDIA Vera Rubin
Overview
vLLM has introduced Day-0 support for the MiniMax H3-Image model, enabling high-performance text-to-image and image-editing workflows on diverse hardware, including NVIDIA Vera Rubin. By leveraging vLLM-Omni optimizations, the system achieves a median latency of 0.87 seconds for native 4K (4096×4096) image generation on four GPUs. Beyond latency, the implementation supports high-throughput configurations capable of generating over 6,000 images per hour.
Technical Performance and Optimization
The H3-Image model consists of a Qwen3-VL encoder, a 33.1B dense Diffusion Transformer (DiT), and an image VAE. The vLLM-Omni implementation optimizes this architecture through several key techniques:
- MXFP8 GEMMs: On Rubin and Blackwell architectures, MXFP8 precision is used for eligible DiT linears to accelerate computation and reduce weight footprints.
- Sequence Parallelism (SP4): Sequence work is distributed across four GPUs, while DiT weights remain replicated to maximize efficiency.
- VAE Decoding Overlap: Decoding is optimized by distributing image tiles across GPUs and combining CUDA Graph replay with multiple CUDA streams, which reduced VAE decoding latency by approximately 2.5×.
- AdaLN Caching: The system reuses conditioning projections when inputs and weights match, avoiding redundant calculations across the six denoising steps.
Hardware Support and Deployment
Beyond NVIDIA Vera Rubin, vLLM-Omni provides recipes and support for a wide range of data-center and consumer hardware:
- NVIDIA: Support spans multiple generations, including Hopper and Ampere, with specific optimizations for memory-constrained environments using layerwise CPU offloading.
- AMD Instinct: Recipes for MI325X and MI350X support BF16 and Flash Attention, utilizing Ulysses sequence parallelism.
- Intel Arc Pro: Documentation is provided for one-, two-, and four-GPU deployments, including tensor parallelism configurations for 32 GB cards.
- Huawei Ascend: Support for Atlas 800I A2, 800T A3, and 350 (A5) NPUs using BF16 and eager execution.
- Moore Threads: MUSA implementation enables VAE graph decoding and tensor parallelism on S5000 cards.
Economic Projections
Based on MiniMax's standard 4K API pricing of USD 0.06879 per image, the high-throughput configuration (6,115 images/hour) yields significant revenue potential. At full billable utilization, the system projects to USD 105.16 per GPU/hour in gross revenue. At 75% utilization, this figure adjusts to USD 78.87 per GPU/hour. These figures serve as projections based on measured throughput and do not account for operational costs.