8000 Commits · andportnoy/praxis · GitHub

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

Commits

Commits on Aug 23, 2023

Support prefix vectorization with tree_map_params

ishark
authored and
pax authors
committed
< 8000 button data-component="IconButton" type="button" class="prc-Button-ButtonBase-c50BI prc-Button-IconButton-szpyj" data-loading="false" data-no-visuals="true" data-size="small" data-variant="invisible" aria-describedby=":R3prnt8t5rab:-loading-announcement" aria-labelledby=":R9rnt8t5rab:">

Commits on Aug 22, 2023

Update pytype to Py3.10 way (Part 3)

laurentes
authored and
pax authors
committed
Update pytype to Py3.10 way (Part 2)

laurentes
authored and
pax authors
committed
Add trees.copy
The praxis Authors
committed
Switch from jnp.ndarray to jax.Array in pytypes.
The praxis Authors
committed
Add token accuracy to FullSoftmax.

m-orsini
authored and
pax authors
committed

Commits on Aug 21, 2023

Update test to not have one of the input matrix dimension == sparsity group size.
The praxis Authors
committed
Refactor ShadedAdafactor for easier extension.

Zarjagen
authored and
pax authors
committed

Commits on Aug 18, 2023

Allow turning off bias in softmax
The praxis Authors
committed
deleted "step" and "start_step" from beam search output
The praxis Authors
committed

Commits on Aug 17, 2023

Add the QK norm option to the dot product attention.
The praxis Authors
committed
Workaround for decoding when bld and blnh are sharded differently on b.

ukoxyz
authored and
pax authors
committed
Implement cross attention for multi query attention.
The praxis Authors
committed
include non-trainable in the variables for extend_step vmap.
The praxis Authors
committed
Allow fakequant activations with bitwidths other than 8
The praxis Authors
committed
Prevent crashes when the local attention slice is longer than the sequence.

m-orsini
authored and
pax authors
committed

Commits on Aug 16, 2023

Place DPO objective formula in Sphinx math block
The praxis Authors
committed
pax/praxis style updates pt.1
The praxis Authors
committed
Nits: typos and typing updates in ShardedAdafactor
The praxis Authors
committed
internal change

saberkun
authored and
pax authors
committed

Commits on Aug 15, 2023

Add basic support to quantize in native S4 format.

Yun Ni
authored and
pax authors
committed
DatasetInputSpecsProvider: instantiate the input pipeline only once and cache

aaroey
authored and
pax authors
committed
fix sparsity apply condition.
The praxis Authors
committed
Update pytype to Py3.10 way (Part 1)

laurentes
authored and
pax authors
committed

Commits on Aug 14, 2023

Add support for passing the extra input kwargs in the decode data.

Hyperparticle
authored and
pax authors
committed
Enable per-channel quantization during inference time of quantizer

lenscloth
authored and
pax authors
committed

Commits on Aug 12, 2023

Add support for chunked attention in MQA.

laurentes
authored and
pax authors
committed

Commits on Aug 11, 2023

Simplify @jaxtyped to @typed
The praxis Authors
committed
Add Ngrammer Stub which only computes group norm

vlad17
authored and
pax authors
committed
MultiQueryAttention layer to support extend n steps.

sirakiin
authored and
pax authors
committed

Commits on Aug 10, 2023

Fix conflicting naming in Transformer attributes.
The praxis Authors
committed
(2/3) Improves the import time of `paxml.ParameterizedExperiment` from ~32s to ~17s, by making the imports of some packages lazy, and postponing the evaluation of some typing annotations (by either…

wangpengmit
authored and
pax authors
committed
Rename the shape argument in maybe_slice_uneven_sharding.

hsidahmed865
authored and
pax authors
committed
Speed up EagerTensor to Numpy conversion in `LingvoInputAdaptor`.
The praxis Authors
committed
Allow ability to scale the update by the norm of the variance.
The praxis Authors
committed

0