THUDM

/

glm-4-voice-9b

like 75

Follow

Knowledge Engineering Group (KEG) & Data Mining at Tsinghua University 1.44k

nnsight seems to produce nan on every logits what so ever .. chat doesnt do that ??

can anyone from THUDM actually explain what was changed on the model expect for the heads and the embeddings for the audio / special tokens ? as this is the weirdest thing ive seen in a long time