issues
search
google
/
aqt
Apache License 2.0
247
stars
25
forks
source link
issues
Newest
Newest
Most commented
Recently updated
Oldest
Least commented
Least recently updated
Reduce checkpoint size for Flax models
#720
liamclarkza
opened
8 hours ago
2
fix tracer leakage
#719
copybara-service[bot]
opened
18 hours ago
0
Rename `IntNumerics` to `IntSymmetric` in advance of adding asymmetric integer numerics.
#718
copybara-service[bot]
closed
49 minutes ago
0
Change loss value for v5e
#717
copybara-service[bot]
closed
22 hours ago
0
Remove TODO straggling from cl/588117622.
#716
copybara-service[bot]
closed
1 day ago
0
Rename `abs_val_mapped_to` to `get_quant_bound`.
#715
copybara-service[bot]
closed
1 hour ago
0
Add kwarg to disable rounding separately from `use_fwd_quant` in `aqt_dot_general_test.py`.
#714
copybara-service[bot]
closed
21 hours ago
0
Support per-channel scaling in AQTv2 `ConstantCalibration` and integrate into mojax.
#713
copybara-service[bot]
closed
4 days ago
0
Adding flax wrapper for quantized convolutions
#712
copybara-service[bot]
opened
6 days ago
0
require python 3.10
#711
copybara-service[bot]
closed
6 days ago
0
Fix bug of FP numeircs.
#710
copybara-service[bot]
closed
4 days ago
0
Add test for e1m2 and e0m3
#709
copybara-service[bot]
closed
1 week ago
0
Add numerics config for FP5 with subnormals
#708
copybara-service[bot]
closed
1 week ago
0
FP6 ocp format
#707
copybara-service[bot]
closed
1 week ago
0
Best practices to reduce memory usage in SERVE mode
#706
mar-muel
opened
1 week ago
0
Add fp8 numeric to dot_general and quantizer make functions.
#705
copybara-service[bot]
closed
1 week ago
0
Update fp8 cpu training loss target due to XLA changes.
#704
copybara-service[bot]
closed
2 weeks ago
0
Support Python versions <3.10
#703
mar-muel
opened
2 weeks ago
1
Replace tuple unpacking with jnp.s_ for variable indexing
#702
copybara-service[bot]
closed
2 weeks ago
0
Bump AQT version to 0.8.0
#701
copybara-service[bot]
closed
2 weeks ago
0
Refactor `Calibration` to calculate scales and biases instead of just scale bounds.
#700
copybara-service[bot]
closed
6 days ago
0
Roll back the input / output signature of the quantized conv operations
#699
copybara-service[bot]
closed
3 weeks ago
0
Add a new implementation of fp_round that improves readability.
#698
copybara-service[bot]
closed
3 weeks ago
0
fix the function signature of my_conv_general_dilated and make it identical as default the conv function.
#697
copybara-service[bot]
closed
4 weeks ago
0
Comments for fp numerics
#696
copybara-service[bot]
closed
1 month ago
0
Use a different method name other than the variable name.
#695
copybara-service[bot]
closed
1 month ago
0
make dummy gradient into qtensor configurable
#694
copybara-service[bot]
closed
1 month ago
0
Adjust test mnist training loss because of changes in XLA
#693
copybara-service[bot]
closed
1 month ago
0
QTensor refactoring: holds AqtTileMap instead of TilingState.
#692
copybara-service[bot]
opened
1 month ago
0
Gradient quantization support
#691
lixcli
opened
1 month ago
0
Add a setter to enable int2 weight dtype.
#690
copybara-service[bot]
closed
1 month ago
0
cleaner FP numerics with subnormals supported
#689
copybara-service[bot]
closed
3 weeks ago
0
subnormal implementation with both static and stochastic rounding
#688
copybara-service[bot]
closed
1 month ago
0
Add support for Delayed Per-Tensor Scaling
#687
copybara-service[bot]
closed
3 weeks ago
0
Support subnormal in FP simulation
#686
copybara-service[bot]
closed
4 weeks ago
0
Allow space-to-batch propagations on Reduce
#685
copybara-service[bot]
closed
1 month ago
0
Consolidate the calculation of scale tensors into a function.
#684
copybara-service[bot]
closed
3 weeks ago
0
Support negative index on `AxisTiling` and `tiled_dot_general.to_tiled_axes_transposed`.
#683
copybara-service[bot]
closed
1 month ago
0
Add simple quantization range validation tests for `int_numerics`.
#682
copybara-service[bot]
closed
1 month ago
0
Add support for passing a dataset to `flax_e2e_model.train_and_evaluate` and make statistics printing easier to read.
#681
copybara-service[bot]
closed
1 month ago
0
Add fake-quantization support for quantization biases to AQTv2.
#680
copybara-service[bot]
closed
1 month ago
0
Remove intermediate `new_snr_highs` tensor.
#679
copybara-service[bot]
closed
1 month ago
0
[xla:cpu] Update expected target_loss for xla:cpu thunks
#678
copybara-service[bot]
closed
1 month ago
0
Modifying conv ops to optionally accept quantized lhs and rhs and also return them
#677
copybara-service[bot]
closed
1 month ago
0
[pallas] Relands PR #22552: Simplify handling of BlockMapping and GridMapping
#676
copybara-service[bot]
closed
1 month ago
0
Add support asymmetric fake-quantization to AQTv2.
#675
copybara-service[bot]
opened
1 month ago
0
Support subchannel quantization on `quant` and `calibrate`.
#674
copybara-service[bot]
closed
1 month ago
0
Bugfix on e2e example: fix wrong activation quantization mode.
#673
copybara-service[bot]
closed
2 months ago
0
internal
#672
copybara-service[bot]
opened
2 months ago
0
Add mean of abs calibration
#671
copybara-service[bot]
closed
2 months ago
0
Next