Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Refactor QModuleMixin and Calibration and fix stream-lining bug #249

Merged
merged 9 commits into from
Jul 19, 2024
Merged

Refactor QModuleMixin and Calibration and fix stream-lining bug #249

merged 9 commits into from
Jul 19, 2024

Commits on Jul 19, 2024

  1. test(quantize): fix bogus test

    @dacorvo
    dacorvo committed Jul 19, 2024
    Configuration menu
    Copy the full SHA
    d33a775 View commit details
    Browse the repository at this point in the history

  2. test(QConv2D): use float inputs

    @dacorvo
    dacorvo committed Jul 19, 2024
    Configuration menu
    Copy the full SHA
    7e42886 View commit details
    Browse the repository at this point in the history

  3. test(qlinear): avoid passing quantized inputs when activation is none 

    This reveals an overflow in qfloat8_e5m2 activations test, that is removed for now.
    @dacorvo
    dacorvo committed Jul 19, 2024
    Configuration menu
    Copy the full SHA
    3f47114 View commit details
    Browse the repository at this point in the history

  4. test(calibrate): use correct activation qtype

    @dacorvo
    dacorvo committed Jul 19, 2024
    Configuration menu
    Copy the full SHA
    4696e81 View commit details
    Browse the repository at this point in the history

  5. refactor(QModuleMixin): quantize inputs if needed 

    Only QLinear might request its inputs to be always quantized, as it
is the only layer for which an optimized kernel exists.
    @dacorvo
    dacorvo committed Jul 19, 2024
    Configuration menu
    Copy the full SHA
    ed3cc71 View commit details
    Browse the repository at this point in the history

  6. feat(QModuleMixin): sanitize inputs

    @dacorvo
    dacorvo committed Jul 19, 2024
    Configuration menu
    Copy the full SHA
    3a4acd0 View commit details
    Browse the repository at this point in the history

  7. refactor(QModuleMixin): avoid multiple forward calls in calibration 

    By putting the input/output quantization code inside module forward hooks,
it allows them to be called only after the calibration hooks.
This simplifies a lot the calibration code, in particular avoiding several
calls to forward during output calibration.
    @dacorvo
    dacorvo committed Jul 19, 2024
    Configuration menu
    Copy the full SHA
    abf97a9 View commit details
    Browse the repository at this point in the history

  8. fix(calibration): only disable output quantization when streamlining

    @dacorvo
    dacorvo committed Jul 19, 2024
    Configuration menu
    Copy the full SHA
    d783a3b View commit details
    Browse the repository at this point in the history

  9. refactor(QModuleMixin): remove forward indirection

    @dacorvo
    dacorvo committed Jul 19, 2024
    Configuration menu
    Copy the full SHA
    59b9983 View commit details
    Browse the repository at this point in the history