8000 Commits · hx89/praxis · GitHub

More Web Proxy on the site http://driver.im/

Commits

Commits on May 6, 2024

Support start step in LinearRampupCosineDecay
The praxis Authors
committed

Commits on May 10000 1, 2024

Add eps for weight quantization.

jianlijianli
authored and
pax authors
committed

Commits on Apr 30, 2024

Temporarily disable channel-wise pruning in praxis to align its implementation with the sparsity/jax channelwise pruning implementation.
The praxis Authors
committed

Commits on Apr 29, 2024

Skip mask recalculation in praxis sparsity library.
The praxis Authors
committed
Move pax sparsity variable constants out of jax library and into pax specific sparsifier.py
The praxis Authors
committed

Commits on Apr 22, 2024

Fix Nvidia PR.

zhangqiaorjc
authored and
pax authors
committed
Merge pull request #58 from abhinavgoel95:patch-3
pax authors
committed

Commits on Apr 18, 2024

Fix in DotProductAttention to also support relative position bias when segment positions are None.
The praxis Authors
committed
Moves the suffix to sparsity so that we used a shared values with layers when creating variables.

shivaniag
authored and
pax authors
committed

Commits on Apr 17, 2024

Create grok.py
abhinavgoel95
authored
Open source pipeline_test.py

laurentes
authored and
pax authors
committed
Add sparsity core library which will eventually replace sparsity_core_depr.

laurentes
authored and
pax authors
committed

Commits on Apr 15, 2024

Quantized training, design general custom einsum.

rybakov
authored and
pax authors
committed

Commits on Apr 11, 2024

Add sparsity core library which will eventually replace sparsity_core_depr.
The praxis Authors
committed

Commits on Apr 10, 2024

Add MobileNetV2 layer.
The praxis Authors
committed

Commits on Apr 9, 2024

Fix rank of partition specs for `count` variable with OptaxOptimizer.

laurentes
authored and
pax authors
committed

Commits on Apr 8, 2024

Merge pull request #55 from abhinavgoel95:abhinavgoel95-patch-1
pax authors
committed
Prepare Praxis for 1.4.0 release
The praxis Authors
committed
Update `requirements.txt`
The praxis Authors
committed

Commits on Apr 5, 2024

Adding support for classifier-free guidance for sequence models.
The praxis Authors
committed
Adding support for expert parallelism
abhinavgoel95
authored

Commits on Apr 4, 2024

tests for quantized training

rybakov
authored and
pax authors
committed
Quantized training: extend custom einsum with control of forward and backward quantization bits.

rybakov
authored and
pax authors
committed

Commits on Apr 2, 2024

Exposing early_exit parameter for sampling.
The praxis Authors
committed

Commits on Mar 30, 2024

Cleans up a trivial unnecessary logic.
The praxis Authors
committed
Apply pytype to more unit tests.

laurentes
authored and
pax authors
committed

Commits on Mar 27, 2024

Pass arguments that will become keyword only soon as named arguments.

mtthss
authored and
pax authors
committed
Do not flatten the whole matrices while computing top_k, but only reshape the dimension we want to sparsity.

shivaniag
authored and
pax authors
committed

Commits on Mar 26, 2024

Use jnp.minimum to compare tensor dimensions, to cover inputs with dynamic length.
The praxis Authors
committed
Replace deprecated `jax.tree_map` with `jax.tree.map` to remove warnings from logs.

galenmandrew
authored and
pax authors
committed

Commits on Mar 21, 2024

Adding mixed precision quantization decorator.

dryman
authored and
pax authors
committed

Commits on Mar 20, 2024

Access thread_resources via jax.interpreters.pxla instead of jax.experimental.maps

superbobry
authored and
pax authors
committed

Commits on Mar 19, 2024

Skip the key state update if consolidate_rope_key_state is set.

changlan
authored and
pax authors
committed
Use None for replicated axis.

jianlijianli
authored and
pax authors
committed

Commits on Mar 18, 2024

Split `by` to `bsc`.

jianlijianli
authored and
pax authors
committed

0