python 3.12.2 + torch2.9 + cu130
MultiGPU repo pulled 26.3.
After upgrading ComfyUI to 0.18.1 having 2x CLIPLoaderMultiGPU (different clips, different models with 2 stage samplers ) in single workflow both loaded into cuda:1 used for CLIPs ( cuda:0 only for models) there is garbled output from 2nd Sampler after 2nd or 3rd run. First sampler always fine.
Second CLIP becomes corrupted and needs to unload models and cache to make it work again for another run. Situation repeats each second generation after cleaning cache/VRAM.
Models are loaded using native ComfyUI loader, both CLIPs are loaded using CLIPLoaderMultiGPU.
Switching to native Load CLIP fixes the issue. In ComfyUI 0.14 there was no problem.
Edit:
I switched both CLIP loaders to CLIPLoaderGGUFDisTorch2MultiGPU (1st stage) CLIPLoaderDisTorch2MultiGPU (2nd stage) and it's same issue which seems to be tied to 2nd stage qwen_3_8b_fp8mixed.safetensors of Klein 9B (type flux2) in my case.
First stage (CLIPLoaderGGUFDisTorch2MultiGPU) is CLIP type chroma without issue.
Edit2 (issue bypass by replacing node):
I replaced 2nd stage CLIPLoaderDisTorch2MultiGPU with CLIPLoaderGGUFDisTorch2MultiGPU and downloaded qwen_3_8b_Q8_0 GGUF static quants and the issue is not present.
So the issue is either with both CLIPLoaderMultiGPU and CLIPLoaderDisTorch2MultiGPU or with qwen_3_8b_fp8mixed.safetensors or it's combination with CLIPLoaderGGUFDisTorch2MultiGPU in the same WF.
I can see both GGUF remain loaded on cuda:1 which confirms that CLIPLoaderGGUFDisTorch2MultiGPU is working as expected.
python 3.12.2 + torch2.9 + cu130
MultiGPU repo pulled 26.3.
After upgrading ComfyUI to 0.18.1 having 2x CLIPLoaderMultiGPU (different clips, different models with 2 stage samplers ) in single workflow both loaded into cuda:1 used for CLIPs ( cuda:0 only for models) there is garbled output from 2nd Sampler after 2nd or 3rd run. First sampler always fine.
Second CLIP becomes corrupted and needs to unload models and cache to make it work again for another run. Situation repeats each second generation after cleaning cache/VRAM.
Models are loaded using native ComfyUI loader, both CLIPs are loaded using CLIPLoaderMultiGPU.
Switching to native Load CLIP fixes the issue. In ComfyUI 0.14 there was no problem.
Edit:
I switched both CLIP loaders to CLIPLoaderGGUFDisTorch2MultiGPU (1st stage) CLIPLoaderDisTorch2MultiGPU (2nd stage) and it's same issue which seems to be tied to 2nd stage qwen_3_8b_fp8mixed.safetensors of Klein 9B (type flux2) in my case.
First stage (CLIPLoaderGGUFDisTorch2MultiGPU) is CLIP type chroma without issue.
Edit2 (issue bypass by replacing node):
I replaced 2nd stage CLIPLoaderDisTorch2MultiGPU with CLIPLoaderGGUFDisTorch2MultiGPU and downloaded qwen_3_8b_Q8_0 GGUF static quants and the issue is not present.
So the issue is either with both CLIPLoaderMultiGPU and CLIPLoaderDisTorch2MultiGPU or with qwen_3_8b_fp8mixed.safetensors or it's combination with CLIPLoaderGGUFDisTorch2MultiGPU in the same WF.
I can see both GGUF remain loaded on cuda:1 which confirms that CLIPLoaderGGUFDisTorch2MultiGPU is working as expected.