ubergarm

/

Ling-1T-GGUF

like 11

Did you receive any official response from HF or the people tagged into the discussion?

If I delete enough to get below the 12TB, can I start uploading again automatically?
Or are our accounts stuck in the blocked state until they unlock it manually?

I created a post under models-and-benchmarks in the Aider discord called "ubergarm2/Ling-1T-GGUF" will hopefully be posting full aider polyglot benchmarks report for the IQ2_KS in the next day or so /neo

edit 2: nah it failed the next 4 test_cases completeley again.. hmm maybe its not just performing so good for aider polyglot. the aider benchmark they provided was for the older benchmark not the newer polyglot so we don't have a reference score either

another old message: edit: i updated Aider to latest and it nailed the first test_case. so I think we are back on track now! :).It also works in Roo Code!

old message: its failing every test and sub test within the each test without creating any errors i.e no malformed responses or error outputs other then the solutions being completely wrong. its weird. the responses are coherent. and it works in open-webui. i've only done 18/225 test cases so its possible although unlikely its very bad luck in the start. i don't think so tough since its failing every step within each test case as well.

@Fernanda24

i updated Aider to latest and it nailed the first test_case. so I think we are back on track now! :).It also works in Roo Code!

Wahoo! Great to hear, hopefully it keeps going lol... There may also be things like running with --jinja but its hot off the presses so I haven't fiddled with it much.

I have a few more quants landing now and have the initial perplexity vs size graph available if the relative "quality" of the quants is of interest to you. Of course the bigger ones will be slower too. The iq2_k is interesting in that it uses full q8_0 for attn.*/first 4 dense/shexp which does make it a bit slower for the size given the proportion of active tokens...