Hi, thanks for the great work on IndexCache!
I'd like to reproduce and deploy the method, but I noticed the repository only provides the SGLang/vLLM patches, and the release list doesn't include the 30B DSA model used for the main experiments (Tables 1–3 in the paper).
In Section 4.1 of the paper, the 30B DSA model is described as initialized from zai-org/GLM-4.7-Flash (30B-A3B MoE, MLA, 47 layers) and trained with a 1,000-step dense warm-up + 4,000-step sparse training phase on SFT data at 200K context. This DSA-finetuned checkpoint is currently not available on Hugging Face.
Thanks in advance!
Hi, thanks for the great work on IndexCache!
I'd like to reproduce and deploy the method, but I noticed the repository only provides the SGLang/vLLM patches, and the release list doesn't include the 30B DSA model used for the main experiments (Tables 1–3 in the paper).
In Section 4.1 of the paper, the 30B DSA model is described as initialized from
zai-org/GLM-4.7-Flash(30B-A3B MoE, MLA, 47 layers) and trained with a 1,000-step dense warm-up + 4,000-step sparse training phase on SFT data at 200K context. This DSA-finetuned checkpoint is currently not available on Hugging Face.Thanks in advance!