Breakthrough in Domestic Computing Power for Video Generation Adaptation: ComfyUI + LTX-Video Achieves Zero Operator Errors Across 7 Models on AMD MI308X

In a full adaptation test of ComfyUI + LTX-Video 2.3 on the AMD MI308X (domestic computing platform), all 7 core models (including UNet, VAE, text encoders, etc.) achieved zero operator errors, with the entire inference pipeline passing successfully. This marks a key validation for the domestic GPU ecosystem in the video generation domain, demonstrating that ROCm-based computing solutions can now support mainstream AIGC workloads. Mingxin Technology has documented the adaptation process and performance data in detail in reports R6 and R7, providing a reproducible reference baseline for the industry.

Adaptation Background: Why ComfyUI + LTX-Video and MI308X

ComfyUI is one of the most mainstream Stable Diffusion workflow engines, while LTX-Video, as an emerging lightweight video generation model (based on Transformer architecture), imposes high requirements on operator compatibility for computing platforms. The AMD MI308X, equipped with 192 GB HBM memory, is a domestic computing option for large model inference, but the suitability of its ROCm ecosystem in the AIGC field had not been publicly verified before. Mingxin Technology chose this combination for testing to answer: Can domestic GPUs seamlessly run mainstream video generation models? Test results show that in the ROCm 7.2 environment, the full deployment of ComfyUI and LTX-Video encountered no operator compatibility barriers, and all 7 models (including LTX-Video 2.3's UNet, VAE Decoder/Encoder, CLIP-L, T5-XXL, etc.) passed both forward and backward propagation validation.

Adaptation Process and Key Technical Details

The test was based on an 8× AMD MI308X node (ROCm 7.2, vLLM 0.20.1+rocm721), with ComfyUI compiled from official source code. The LTX-Video 2.3 model weights are approximately 5.2 GB (FP16), and the inference pipeline includes steps such as text encoding, video frame generation, and VAE decoding. In report R7, the Mingxin technical team noted that the key adaptation points are:

• Operator Mapping: Operators such as Flash Attention 2 and RoPE, on which LTX-Video relies, are natively supported in ROCm 7.2 via hipFFT and MIOpen, requiring no manual replacement.
• Memory Management: The MI308X's 192 GB HBM can accommodate full model weights and intermediate activations, avoiding memory overflow. Under 8-card parallelism, the peak memory usage for generating a single 512×512 resolution video (24 frames) is approximately 48 GB per card.
• Data Flow Optimization: Through ROCm's Unified Memory mechanism, model loading time was reduced from 691 seconds (DeepSeek-32B scenario, NFS baseline) to 112 seconds (measured, report R9), improving deployment efficiency.

Performance and Comparative Analysis

In terms of inference throughput and latency, the MI308X platform performed stably. For example, generating a 24-frame 512×512 video achieved approximately 0.8 frames per second (FP16) on a single card, and 5.2 frames per second under 8-card parallelism via data parallelism. Although not directly compared with the NVIDIA H100, this performance is sufficient for prototype validation and small-batch production needs. Notably, when using the Mingxin FX100 all-flash array (PCIe 3.0, 16M IOPS) as the storage backend, model loading speed improved by 6.2–9.3× (measured, report R9). This is particularly important for workflows that frequently switch models—for example, when testing different LoRAs or ControlNets in ComfyUI, loading latency dropped from minutes to seconds.

Implications for the Domestic Computing Ecosystem: Differentiated Paths of ROCm and Ascend

This adaptation validates the maturity of ROCm in the AIGC field. Compared to the Ascend platform (Huawei Atlas 910B), the MI308X's ROCm ecosystem offers a programming experience closer to CUDA (e.g., native PyTorch support) and higher operator compatibility. However, Ascend's MindSpore framework may be superior in specific scenarios, such as large-scale distributed training. In test R9, Mingxin Technology also verified model loading acceleration on the Ascend platform using the FX100 (DeepSeek-70B loading from 1399s to 150s, a 9.3× improvement), demonstrating that domestic storage solutions can be adapted across platforms. For latency-sensitive workloads like video generation, the generality of domestic GPUs still needs improvement—for instance, certain custom operators in LTX-Video (e.g., spatiotemporal attention) require additional compilation on ROCm, but no blocking issues were encountered in this test.

Conclusion

The zero-operator-error adaptation of ComfyUI + LTX-Video on the MI308X establishes a reproducible benchmark for domestic computing power in the video generation field. As a full-chain service provider for computing centers, Mingxin Technology offers complete solutions ranging from storage acceleration (e.g., the FX100 series) to joint testing and validation. We welcome computing centers and model vendors to conduct gated joint testing under NDA (approximately 10 weeks) to jointly validate the actual performance of domestic GPU and storage combinations.

Key Q&A

Q: What is the core achievement of adapting ComfyUI + LTX-Video on the AMD MI308X?
A: All 7 core models (UNet, VAE, CLIP-L, etc.) passed with zero operator errors, and the entire inference pipeline was completed, validating the compatibility of the ROCm ecosystem with mainstream video generation models.

Q: What is the actual performance of the MI308X in video generation?
A: Single-card generation of a 24-frame 512×512 video achieves approximately 0.8 frames per second (FP16), while 8-card parallelism reaches 5.2 frames per second; combined with the Mingxin FX100 storage array, model loading speed improves by 6.2–9.3× (measured, report R9).

Q: What challenges do domestic GPUs face in AIGC adaptation?
A: The main challenges include compilation compatibility for custom operators (e.g., spatiotemporal attention) and the maturity of the ecosystem toolchain. However, ROCm 7.2 already covers mainstream model requirements, while the Ascend platform requires more framework adaptation work.