-
Notifications
You must be signed in to change notification settings - Fork 28
Pull requests: google/aqt
Author
Label
Projects
Milestones
Reviews
Assignee
Sort
Pull requests list
Avoid jax.promote_dtype(float8_e4m3fn, x) failure when using aqt_einsum with fp8_e4m3fn dtype
#759
opened Mar 4, 2025 by
copybara-service
bot
Add rounding bias that controls the rounding threshold
#750
opened Nov 20, 2024 by
copybara-service
bot
Update the accumulation dtype if FP8 precision is used for AQT.
#745
opened Nov 5, 2024 by
copybara-service
bot
[sharding_in_types][Take 2] Add
out_type
argument to einsum
and dot_general
to allow specifying for the output type. Right now, it only accept a NamedSharding
but in the future we can allow a polymorphic type of: jax.ShapeDtypeStruct | Sharding | Layout
.
#743
opened Oct 22, 2024 by
copybara-service
bot
Allows setting array of bounds in set_calibration_config.
#729
opened Sep 24, 2024 by
copybara-service
bot
QTensor refactoring: holds AqtTileMap instead of TilingState.
#692
opened Aug 9, 2024 by
copybara-service
bot
Set config.set_bits params' default values as None.
#659
opened Jun 25, 2024 by
copybara-service
bot
Change a way to split tiling axes without configuration
#644
opened Jun 7, 2024 by
copybara-service
bot
Support pl.BlockSpec that holds block_shape with None.
#630
opened May 24, 2024 by
copybara-service
bot
Add
from_qtensor
which transform QTensor
into PallasQTensor
.
#629
opened May 24, 2024 by
copybara-service
bot
cast scale to dequant_type when restoring scale to original shape in pallas.
#627
opened May 24, 2024 by
copybara-service
bot
Put numerics and numerics-related logic into QTensor
#623
opened May 22, 2024 by
copybara-service
bot
[Pallas support] Add
quant_blockwisely
function, which quantizes tensor according to the BlockSpec of the inputs.
#596
opened Apr 24, 2024 by
copybara-service
bot
Previous Next
ProTip!
Follow long discussions with comments:>50.