model : add support for apple/DiffuCoder-7B-cpGRPO #14502
Conversation
| # Read merges.txt | ||
| merges_file = self.dir_model / "merges.txt" | ||
| merges = [] | ||
| with open(merges_file, 'r', encoding='utf-8') as f: | ||
| for i, line in enumerate(f): | ||
| line = line.strip() | ||
| if i == 0 and line.startswith("#version:"): | ||
| continue | ||
| if not line: | ||
| continue | ||
| merges.append(line) |
There was a problem hiding this comment.
Why is this needed? SpecialVocab should be able to handle this.
| layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert}, 0); | ||
| } | ||
| } break; | ||
| case LLM_ARCH_DREAM: |
There was a problem hiding this comment.
This looks (unsurprisingly) identical to Qwen2, I don't think it deserves the duplication.
|
Also really puzzled by the fact that it's just a finetuned Qwen2.5 model while at the same time being based on Dream which claims to be a diffusion model?! |
According to PDF they started from Qwen but it's not just a finetune |
|
Yeah I was up too late and skimmed through the DiffuCoder paper too fast. The HuggingFace model tree, and section 3 of the paper made it seem like a normal Qwen2.5 finetune. I'll close this request for now. |
3 participants
This pull request adds initial support for Apple's new DiffuCoder model.
I'll be uploading the F16 gguf to https://huggingface.co/gabriellarson/DiffuCoder-7B-cpGRPO-GGUF