Skip to content

Refactor: remove usage of Q8_1 for activation quantization #196

Description

@ikawrakow

Background Description

Some models can produce activations that are beyond the range of fp16. In that scenario, usage of Q8_1 to quantize the activations can be futile, see discussion in #194.

Hence, it would be prudent to change all quantization types using Q8_1 for matrix multiplications to use something else.
Alternatively, one may replace the fp16 block scale and block sum in Q8_1 with bf16.

Possible Refactor Approaches

No response

Metadata

Metadata

Assignees

No one assigned

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions