sayzardIvan Fioravanti ᯅ (@ivanfioravanti)
Qwen3.5-9B 모델의 'LIBERATION' 작업이 완료되었고, 전체 작업 완료 시간은 9분 40초였습니다. 이는 Qwen3.5 계열의 중대형 모델을 로컬/비GPU 또는 에지 환경에서 준비·실행하는 과정에서의 시간 성능을 가늠할 수 있는 사례로, 모델 배포·변환 워크플로에 참고가 되는 정보입니다.
Ivan Fioravanti ᯅ (@ivanfioravanti)
Obliteratus가 Apple Silicon에서도 작동함을 확인했습니다. Qwen3-4B 모델의 'LIBERATION'(모델 해방/실행 준비) 작업을 3분 47초 만에 완료했고, 더 큰 모델도 시도해보겠다는 예고입니다. Apple Silicon에서 경량/중형 Qwen3 계열 모델을 빠르게 실행·변환할 수 있음을 시사하는 기술적 성과입니다.
Python Trending (@pythontrending)
Megatron-Bridge는 Megatron 기반 모델을 위한 학습 라이브러리로, Hugging Face와의 양방향 변환 기능을 제공해 Megatron 계열 모델과 Hugging Face 포맷 간의 상호 변환 및 학습 워크플로를 단순화한다는 발표. 모델 이식성과 훈련 파이프라인 통합에 유용함.
think independent (@ThinkIndepende2)
작성자가 PyTorch 코드를 JAX 또는 XLA 코드로 효율적으로 변환하는 방법을 묻고 있습니다. 여러 개발자(@antigravity 등)를 태그하며 자동 변환 기능이 'killer feature'가 될 것이라고 제안해, 프레임워크 간 코드 변환 도구나 워크플로 개선에 대한 관심을 드러냅니다.
Hacker NewsONNX Runtime and CoreML May Silently Convert Your Model to FP16
https://ym2132.github.io/ONNX_MLProgram_NN_exploration
#HackerNews #ONNX #Runtime #CoreML #FP16 #ModelConversion #AIdevelopment
Fahim FarookSince yesterday evening, when I finally perfected the process for converting existing #StableDiffusion models (in either CKPT or SAFETENSOR format) to CoreML format, I’ve converted six models. Of them, all except for one work fine 🙂
So, in case this helps anybody else, here are the important things to remember:
1. You *must* use Python 3.8. If you any other Python version, you will end up with errors. (Not that I’ve tested all Python versions, but I did have errors with Python 3.9 and have read reports of others …)
2. You should be on Ventura 13.1 or higher.
3. You need the models to be in Diffusers format to run the conversion, but the easiest way that has worked for me is to download a CKPT file, convert it to Diffusers and point the script at the local folder with the Diffusers format model.
4. HuggingFace folks have a bunch of conversion scripts here: https://github.com/huggingface/diffusers/tree/main/scripts
5. The above scripts don’t mention SAFETENSOR format in the file names but SAFETENSOR is just CKPT with some changes. The CKPT conversion file has an extra argument named “--from_safetensors” so you can use the same script for CKPT to convert SAFETENSOR files with that extra argument.
6. You can use the Apple conversion script to convert one element at at a time using the different arguments such as “--convert-unet”, “--convert-text-encoder” etc. You don’t have to run all of them together. In fact, it turned out when I ran them all together, sometimes a component might be left out — generally the text encoder.
7. Once you’ve converted all the components and have them in one folder, you have to run the Apple conversion script once more with the “--bundle-resources-for-swift-cli” argument (pointing at your output folder) to create the final compiled CoreML model files (.mlmodelc) from your .mlpackage files.
That’s it 🙂 If you do all of the above, it should be fairly straightforward to create new CoreML models from existing StableDiffusion models.
Feel free to hit me up should you run into issues. Since I’ve gone through all this, I’d be happy to help anybody else facing the same issues ….
#CoreML #StableDiffusion #MachineLearning #DeepLearning #ModelConversion
So, in case this helps anybody else, here are the important things to remember:
1. You *must* use Python 3.8. If you any other Python version, you will end up with errors. (Not that I’ve tested all Python versions, but I did have errors with Python 3.9 and have read reports of others …)
2. You should be on Ventura 13.1 or higher.
3. You need the models to be in Diffusers format to run the conversion, but the easiest way that has worked for me is to download a CKPT file, convert it to Diffusers and point the script at the local folder with the Diffusers format model.
4. HuggingFace folks have a bunch of conversion scripts here: https://github.com/huggingface/diffusers/tree/main/scripts
5. The above scripts don’t mention SAFETENSOR format in the file names but SAFETENSOR is just CKPT with some changes. The CKPT conversion file has an extra argument named “--from_safetensors” so you can use the same script for CKPT to convert SAFETENSOR files with that extra argument.
6. You can use the Apple conversion script to convert one element at at a time using the different arguments such as “--convert-unet”, “--convert-text-encoder” etc. You don’t have to run all of them together. In fact, it turned out when I ran them all together, sometimes a component might be left out — generally the text encoder.
7. Once you’ve converted all the components and have them in one folder, you have to run the Apple conversion script once more with the “--bundle-resources-for-swift-cli” argument (pointing at your output folder) to create the final compiled CoreML model files (.mlmodelc) from your .mlpackage files.
That’s it 🙂 If you do all of the above, it should be fairly straightforward to create new CoreML models from existing StableDiffusion models.
Feel free to hit me up should you run into issues. Since I’ve gone through all this, I’d be happy to help anybody else facing the same issues ….
#CoreML #StableDiffusion #MachineLearning #DeepLearning #ModelConversion