awq 4-bit lm_head transplant
#16
by dexgeh - opened
Hello, if anyone is interested I successfully transplanted (with the help of an llm) the lm_head of this model over cyankiwi/gemma-4-31B-it-AWQ-4bit.
The model keep the whole capability intact, but with the styletune applied.
For us gpu-poor this can be helpful.
The script I used: https://hastebin.com/share/ulubeseced.py
This benchmark results are close (not formal I know, just a quick run), so the quality is preserved without damaging the results:
https://llm-bench.io/benchmarks/cmvk17elv001201p5yiq1nqwx
https://llm-bench.io/benchmarks/cmvmvqu6o00kj01p5fz3wfarp