Clarification needed: Are the FP16 weights original or upconverted from 4-bit?
Hi everyone,
I noticed that while the official release came in 4-bit quantized format, @unsloth has published FP16 weights for this model. I'm hoping to get some clarity on the provenance of these FP16 weights.
Question: Are these FP16 weights:
- The original full-precision weights from OpenAI's training, or
- Upconverted/dequantized from the 4-bit release?
@unsloth - if you could share some details about how these FP16 weights were obtained, that would be really helpful for the community to understand what we're working with.
Thanks in advance for any clarification!
Curious as well
These are up converted for now - I'm unsure if OpenAI will release the original BF16 checkpoint, since they did post-train it to FP4, so maybe the BF16 anyways is different
I think upcast is a reasonable approach and it's also mentioned in OpenAI's official implementation here:
In this implementation, we upcast all weights to BF16 and run the model in BF16.
I appreciate you pointing this out!
How to serve this BF16 weight?