Thanks to visit codestin.com
Credit goes to github.com

Commits

Commits on Feb 24, 2024

Adding flag to enable QK projection in one step.

lingvo-bot
authored and
copybara-github
committed

Commits on Feb 22, 2024

Goodbye lingvo/jax!

zhangqiaorjc
authored and
copybara-github
committed

Commits on Feb 8, 2024

Support GLU for MoE (https://arxiv.org/abs/2002.05202)

lingvo-bot
authored and
copybara-github
committed

Commits on Feb 7, 2024

Define Base Class TypeVar for type annotation.

lingvo-bot
authored and
copybara-github
committed

Commits on Jan 30, 2024

Add a flag to disable grouping. This is no-op as disable_grouping is false

lingvo-bot
authored and
copybara-github
committed

Commits on Jan 25, 2024

Quantization aware training with activation quantization, by using QEinsum and QMatmal instead of einsum and matmul.

rybakov
authored and
copybara-github
committed
Silence some pytype errors.

lingvo-bot
authored and
copybara-github
committed

Commits on Jan 23, 2024

Added pytype None checks to gshard_utils.py.

lingvo-bot
authored and
copybara-github
committed

Commits on Jan 19, 2024

Make expert_capacity_factor configurable

lingvo-bot
authored and
copybara-github
committed

Commits on Jan 3, 2024

Allow a scalar `tasks` tensor in `MultiTaskProjection`.

lingvo-bot
authored and
copybara-github
committed

Commits on Dec 28, 2023

Add some pytype annotations.

lingvo-bot
authored and
copybara-github
committed

Commits on Dec 26, 2023

`num_tasks` and `max_task_id` *is* the same. Don't pass `max_task_id`, infer the value from the `weights` shape.

lingvo-bot
authored and
copybara-github
committed

Commits on Dec 21, 2023

Minor cleanups in `MultiTaskProjection`

lingvo-bot
authored and
copybara-github
committed
Clarify the warning text when setting an inefficient `einsum_order`.

lingvo-bot
authored and
copybara-github
committed

Commits on Dec 19, 2023

Commits on Dec 18, 2023

Cleanup: remove duplicated line

lingvo-bot
authored and
copybara-github
committed

Commits on Dec 17, 2023

Avoid using inference-time arithmetic on inputs of reshapes, use `-1` instead.

lingvo-bot
authored and
copybara-github
committed

Commits on Dec 14, 2023

Fix pylint warnings.

lingvo-bot
authored and
copybara-github
committed

Commits on Dec 13, 2023

Fix some more pylint warnings.

lingvo-bot
authored and
copybara-github
committed
Fix some pylint warnings.

lingvo-bot
authored and
copybara-github
committed

Commits on Dec 12, 2023

Commits on Dec 9, 2023

Allow setting a device_mesh of list type.

lingvo-bot
authored and
copybara-github
committed
Make `layers` a `pytype_strict_library`.

lingvo-bot
authored and
copybara-github
committed

Commits on Dec 8, 2023

Change `layers_with_attention` to `pytype_strict_library`.

lingvo-bot
authored and
copybara-github
committed

Commits on Dec 5, 2023

Avoid using inference-time arithmetic on inputs of reshapes, use -1 instead.

lingvo-bot
authored and
copybara-github
committed

Commits on Dec 4, 2023

Move parameter 'time_alignment' to DepthwiseConv2DLayer.

rybakov
authored and
copybara-github
committed
Avoid using inference-time arithmetic on inputs of reshapes, use `-1` instead.

lingvo-bot
authored and
copybara-github
committed

Commits on Dec 1, 2023

`tf.matmul` supports batch multiplications, drop unnecessary reshape.

lingvo-bot
authored and
copybara-github
committed

Commits on Nov 28, 2023

Commits on Nov 27, 2023

Clean up build dep

protoget
authored and
copybara-github
committed

Commits on Nov 24, 2023

Make precise the example for ``CompositeOptimizer`` usage.

lingvo-bot
authored and
copybara-github
committed

Commits on Nov 23, 2023

Rename `source_contexts_reshaped` to `source_contexts_projected`.

lingvo-bot
authored and
copybara-github
committed