it's works for me 3090ti, INT8, 32ram

@Felldude 谢谢回复，我增加了一个4步加速lora后解决了问题～

Key Differences: Model Loading and Peak Attention Strategies

The fact that other models run successfully implies that your 12GB VRAM—utilizing VRAM-to-RAM swapping strategies—should theoretically handle the 0.8M/15S workload. However, this merged model fails, likely because it does not employ an optimal low-VRAM strategy.

Differences in Model Loading Strategies

Running H3 on a 12GB card relies on a "sequential offload" strategy: only the layer currently being computed remains in VRAM, while the rest are offloaded to system RAM, keeping VRAM usage around 2–3GB. Since other models run normally, their loaders are evidently executing this strategy correctly. In contrast, logs for this merged model show 10.12 GiB of weights resident in VRAM; it likely failed to offload them promptly, effectively "pinning" a large volume of weights in VRAM and leaving negligible free space.

Differences in Peak Attention Mechanism Usage

With long sequences like 0.8M/15S, peak VRAM usage for the attention mechanism spikes dramatically. Tests show that in long-sequence scenarios, specific attention implementations can reduce peak VRAM from ~11.3GB to ~7.1GB. While other models may default to memory-efficient attention paths (such as chunked attention), this merged model appears forced onto a high-peak path, immediately hitting the 12GB limit.

⚙️ The "Last Straw" Triggering the OOM Error

The error in the logs occurred during an int8_linear computation within comfy_kitchen, specifically failing when attempting to allocate a 2.33 GiB temporary tensor.

This indicates a "heavy" step in the merged model's INT8 quantization computation path: it requires the allocation of a massive intermediate activation tensor all at once during linear layer calculations. Other models likely employ smarter chunked quantization or streaming processing to avoid such large, instantaneous allocations. Because this model's internal structure has been modified, it inadvertently triggered a path requiring high temporary VRAM. 💡 Recommendations

Since you have verified that the 0.6M/10S setting works fine while 0.8M/15S consistently crashes, the most straightforward approach is to use the 0.6M/10S or 0.7M configuration for this model.

If you are determined to run 0.8M/15S with this model, you could try the following:

Force-enable a more VRAM-efficient attention mechanism: Ensure that a node like "MiniMax H3 Memory Efficient Sage Attention" (or a similar low-VRAM attention patch) is connected in your workflow.

Add the --lowvram launch argument: Include --lowvram in your launcher settings to force ComfyUI to offload weights to system RAM more aggressively.

If these measures do not resolve the issue, the conclusion is clear: this specific merged model cannot stably handle 0.8M/15S tasks on 12GB of VRAM until its INT8 computation path and loading strategy are optimized. Sticking to 0.6M or 0.7M, or reverting to a model you have already verified, would be the more hassle-free choice.

Thank you very much for releasing this model. I have several H3 models on hand and wanted to try the one you published, but I ran into an issue. I felt it was important to provide timely feedback—this isn't a complaint, and I appreciate your reply!

Environment:

GPU: RTX 4070 SUPER 12GB

RAM: 32GB

ComfyUI: 0.36.0

PyTorch: 2.13.0+cu130

Model: baseMinimaxh38Step_comfyFI2vaPruned_int8.safetensors

Task:

0.8M / 15S, 8 steps, euler + simple

Error Details:

text

torch.OutOfMemoryError: Allocation on device 0 would exceed allowed memory.

Currently allocated: 10.12 GiB

Requested: 2.33 GiB

Device limit: 11.99 GiB

Occurred during the sampling phase: comfy_kitchen int8_linear → quantize_and_rotate_rowwise (before reaching VAE decoding).

"You mentioned changing the transformation size from 256 to 64; where exactly can this parameter be modified? Is it convrot_groupsize? I didn't see this option in the loader or the quantization configuration. If the default is 256, could you provide a quantized version with a size of 64 or 128, or expose this parameter in the node? Other users with 12GB cards might encounter the same issue."

@Felldude Understood. Thank you for your efforts—thank you!