The correct eos_token_id value for Kimi-K2-Instruct
#11
by
anikifoss
- opened
Thanks for uploding the converted and patched BF16 model for custom quantization needs!
I noticed that BF16 safetensors version of Kimi-K2-Instruct was updated with eos_token_id = 163586
, but the GGUFs I checked (Q4_K_M and Q8_0) still have the eos_token_id = 163585
.
I'm not sure which value to use. Could you share what is the correct value of eos_token_id
for this model?
anikifoss
changed discussion title from
The correct value eos_token_id for Kimi-K2-Instruct
to The correct eos_token_id value for Kimi-K2-Instruct
I fixed it sorry! They all should be 163586
Thanks for the update. And thanks for doing the heavy lifting with finding and fixing all the issues in the upstream models!