Why is t3_cfg.safetensors twice the size of t3_cfg.pt? 🤔
Why is t3_cfg.safetensors twice the size of t3_cfg.pt? To my knowledge, the safetensors format conversion doesn't inflate the checkpoint size this much. Are we looking at two completely different checkpoints? @ResembleAI Please clarify!
After using the safetensors files, the vram usage also seems to increase significantly. Now it seems to use about 9gb of vram, although I cannot remember what the exact vram usage was originally.
If someone from the team can clarify this within the next two weeks (starting 06/09/2025), please do. Many users are losing trust in adopting this open-source TTS model due to insufficient access to architectural documentation and lack of transparent communication from the authors. We recognize you've mentioned being a small team of three, though your HuggingFace page shows 15 members. Still, we request any details on the architecture and the clarification on inconsistencies in checkpoint weights. If this cannot be resolved officially, community members will consider reaching out to HuggingFace moderation for further action.
Hi there, let's clarify a couple of things
- there are 3 people in the generative research team - the rest of the employees are in dfd-research/prod/sales/marketing/biz/etc
- I converted the weights to safetensors. So in training we normally save our models in fp16. Now I've never actually used safetensors before (remember we were mainly closed source before this release) and admittedly vibe-coded my way through converting them since i'm super busy rn. So I guess the reason why the safetensors are twice the filesize is because the default setting for safetensor conversion is fp32.
Hi, thanks for the clarification. Is it possible for you to reupload in the original precision then?
Hi there, let's clarify a couple of things
- there are 3 people in the generative research team - the rest of the employees are in dfd-research/prod/sales/marketing/biz/etc
- I converted the weights to safetensors. So in training we normally save our models in fp16. Now I've never actually used safetensors before (remember we were mainly closed source before this release) and admittedly vibe-coded my way through converting them since i'm super busy rn. So I guess the reason why the safetensors are twice the filesize is because the default setting for safetensor conversion is fp32.
Appreciate the response @ollieollie . I figured the issue might be mismatched precision, but I wanted to double-check with you. Safetensors files usually store model weights in FP16 precision by default, it must have been some problem with conversion configuration. I’ll point others interested to this thread. It would be great if you could update and reupload the .safetensors weights so people don’t panic when they see this size mismatch.
Mingyi do you have github? I was wondering if you got anywhere with the safetensors files
Mingyi do you have github? I was wondering if you got anywhere with the safetensors files
Huh? I just waited for the PRs in the official github repo to be merged, before running the code again?