issues
search
google
/
aqt
Apache License 2.0
262
stars
27
forks
source link
issues
Newest
Newest
Most commented
Recently updated
Oldest
Least commented
Least recently updated
This CL adds ridge regression as a method for post-processing the scales, reducing quantization error.
#546
copybara-service[bot]
opened
7 months ago
0
Fix the flakiness of //third_party/py/aqt/jax/v2/numerics:fp8_numerics_test_cpu
#545
copybara-service[bot]
closed
7 months ago
0
Internal
#544
copybara-service[bot]
closed
7 months ago
0
Add an option to choose between flax.nn.custom_vjp and jax.custom_vjp.
#543
copybara-service[bot]
closed
7 months ago
0
Adds a tutorial for how to make a basic AQT quantizer
#542
copybara-service[bot]
closed
7 months ago
0
For AQT `intercept_method` API, Adds skeleton code and prototype for nn.Dense.
#541
copybara-service[bot]
closed
7 months ago
0
Replace jax.core.symbolic_equal_dim with ==
#540
copybara-service[bot]
closed
7 months ago
0
Freezer implementation saving the whole QTensor, instead of saving its members separately
#539
copybara-service[bot]
closed
7 months ago
0
Set the QTensor's dequant_dtype during SERVE mode to scale_t's dtype.
#538
copybara-service[bot]
closed
7 months ago
0
Add vlog to print out tensor dimensions before and after tiling.
#537
copybara-service[bot]
closed
7 months ago
0
Fix Presubmit failure after JAX update
#536
copybara-service[bot]
closed
7 months ago
0
Extract tiled dg input shapes and tiled dimension numbers from jaxpr in order to correctly initialize the Freezer when tiled dg is enabled. The Freezer shape initialization of the injected aqt_dg will then match the tensor shape stored inside.
#535
copybara-service[bot]
closed
7 months ago
0
improve assertion message in tiled_dg
#534
copybara-service[bot]
closed
7 months ago
0
Use sub-channel quantization for PAX AqtEinsum.
#533
copybara-service[bot]
closed
7 months ago
0
bump version number
#532
copybara-service[bot]
closed
7 months ago
0
Internal.
#531
copybara-service[bot]
closed
8 months ago
0
Enable static range quantization. Focusing only for statistics collection.
#530
copybara-service[bot]
closed
7 months ago
0
Add command to flaky test, and sync the version number to the published one.
#529
copybara-service[bot]
closed
8 months ago
0
Fixed import error for `type.Self` for Python < 3.11
#528
copybara-service[bot]
closed
8 months ago
0
Added pytype None checks
#527
copybara-service[bot]
closed
8 months ago
0
bump up version number
#526
copybara-service[bot]
closed
8 months ago
0
Refactor aqt_dot_general to run quantizer for both arguments at once
#525
copybara-service[bot]
closed
7 months ago
0
Separate QuantMode as a stand-alone module.
#524
copybara-service[bot]
closed
8 months ago
0
Internal.
#523
copybara-service[bot]
closed
8 months ago
0
Adds a `allow_dummy_gradient_into_qtensor` toggle for turning off `quant_grad` sanity check under serving mode.
#522
copybara-service[bot]
closed
8 months ago
0
Enable qvalue to be None, and divide a function to calculate the scale and a function to calculate the actual qvalue inside quantizer.
#521
copybara-service[bot]
closed
8 months ago
0
Make quantizer produce values already cast to low-precision.
#520
copybara-service[bot]
closed
8 months ago
0
Use partial() to further simplify the integration of tiled_dot_general. This will be very similar to the tiling integration into pax ops.
#519
copybara-service[bot]
closed
8 months ago
0
Fix flaky fp8 test on CPU
#518
copybara-service[bot]
closed
8 months ago
0
Remove unused import.
#517
copybara-service[bot]
closed
8 months ago
0
simplify the logic of tiled dot general in AqtDotGeneral. Results should be equivalent.
#516
copybara-service[bot]
closed
8 months ago
0
Fix cases of prng key reuse in aqt
#515
copybara-service[bot]
closed
8 months ago
0
Can AQT be used to calculate qk score?
#514
Lisennlp
opened
8 months ago
2
test tiling dot general in flax e2e model
#513
copybara-service[bot]
closed
8 months ago
0
Allow tiling configs to be more user friendly. Create a function that fills in the missing tiling specs and tiling axes in the config.
#512
copybara-service[bot]
closed
8 months ago
0
Partial integration of tiling into AqtDotGeneral class.
#511
copybara-service[bot]
closed
8 months ago
0
reduce the number of tests in tiled dot general test to fix flaky time out.
#510
copybara-service[bot]
closed
8 months ago
0
factor tiling process into a function
#509
copybara-service[bot]
closed
8 months ago
0
Move auxiliary functions and classes top-level. They will be needed for composing the tile function, which will be used for integration into aqt_dot_general.
#508
copybara-service[bot]
closed
8 months ago
0
remove the output assertion in tiled dot general since they appear in the test as well.
#507
copybara-service[bot]
closed
8 months ago
0
Add tests for tiled dot general.
#506
copybara-service[bot]
closed
8 months ago
0
Add a setter to set scale inside AbsMaxCalibration. Update IntNumerics.clip_gradient=True when scale < 1.0.
#505
copybara-service[bot]
closed
8 months ago
0
Let user set preserve_zero in IntNumerics inside a DotGeneral config.
#504
copybara-service[bot]
closed
8 months ago
0
Changes number of updates for structure decay to 8. 16 run into OOM for full model.
#503
copybara-service[bot]
closed
8 months ago
0
TypeError: dataclass() got an unexpected keyword argument 'frozen'
#502
SetoKaiba
closed
8 months ago
2
First stab at tiled_dot_general which will form a base of new LocalAQT implementation.
#501
copybara-service[bot]
closed
8 months ago
0
Run int4 model with bitcast in Freezer.
#500
copybara-service[bot]
closed
8 months ago
0
Add scale to AbsMax calibration
#499
copybara-service[bot]
closed
8 months ago
0
AqtEinsum 'not enough values to unpack'
#498
brandnewchoppa
closed
4 months ago
0
Move 3 function to be methods of DotGeneral class: dg_core, __call__, assert_config_validity
#497
copybara-service[bot]
closed
8 months ago
0
Previous
Next