smollm2-135m-instruct-hidden — hidden-state ONNX export
Prefill-only ONNX export that adds a last_hidden_state output (post-final-norm,
the input to lm_head) alongside logits, for training a LoRA adapter on the
output head in the browser (the choochoo tool).
- outputs:
logits [batch, seq, 49152],last_hidden_state [batch, seq, 576] lm_head(last_hidden_state) == logits- dtype:
q8(onnx/model_quantized.onnx)
See onnx_hidden.py.
- Downloads last month
- 2