Spaces:
Sleeping
No Merge?
I have been trying for weeks to merge my models but no matter what I do it doesn't work. This space, my last resort, does not work. Could somebody please help me? Here's my merge I want -
method: linear
base_model: Qwen/Qwen2.5-14B-Instruct-1M
model: Qwen/Qwen2.5-14B-Instruct-1M
alpha: 0.20model: YOYO-AI/ZYH-LLM-Qwen2.5-14B-V4
alpha: 0.20model: JungZoona/T3Q-qwen2.5-14b-v1.0-e3
alpha: 0.40model: prithivMLmods/Gauss-Opus-14B-R999
alpha: 0.20
tokenizer_source: base
I will leave this here, as it may be of help to some people.
https://colab.research.google.com/drive/1obulZ1ROXHjYLn6PPZJwRR6GzgQogxxb?usp=sharing
If you are using the free tier, your time will be limited.
Usually, you have about an hour of service per day (within about 24 hours), which might not be enough for most mergers. However, note that it is not the merge itself that will take long; it is downloading the files into the notebook that will take hours, depending on their size and service load.
If you use it only once every two to three days (yeah, 48 to 72 hours), you will have up to 3~ hours of service (the free time "stacks up" if unused, but only up to about 3 hours of service), which should be enough for most sensibly sized merges.
I recommend using the TPU runtime. It has some 300 gigs of RAM and 200 gigs of storage.
Do take some time to read about the notebook. Good luck.
It is not good, but it is "free" (it costs your time, and might force you to make concessions in the merge's design)
i have been using a modification of that notebook to make my merges (12B max) ever since this space died. So I know from experience that it works, despite being so constrained.
tried on colab with TPU but the network speed was so bad i couldnt finish. Then i did it on runpod and it was finished in 15min for just 0.12$. I regret wasting time searching for a free alternative.