PyTorch/XLA 2.4

Researchers and practitioners can use the open-source PyTorch machine learning (ML) library and XLA ML compiler for flexible, powerful model

[{"selector":"#anim-017b0989-43f5-4e9a-91de-ece42b00e436 [data-leaf-element=\"true\"]","keyframes":{"transform":["translate3d(-34.4642856200156%, 0, 0) translate(-25%, 0%) scale(1.5)","translate3d(0%, 0, 0) translate(0%, 0%) scale(1)"]},"delay":0,"duration":2000,"fill":"forwards"}] [{"selector":"#anim-863c4c12-0571-4bb3-9897-78be3a0d2269","keyframes":{"opacity":[0,1]},"delay":0,"duration":2000,"easing":"cubic-bezier(.3,0,.55,1)","fill":"both"}] [{"selector":"#anim-f3b58b53-8337-4647-bbd7-cdbbf1fbd486","keyframes":{"transform":["scale(0.3333333333333333)","scale(1)"]},"delay":0,"duration":2000,"easing":"cubic-bezier(.3,0,.55,1)","fill":"forwards"}]

The PyTorch/XLA team is pleased to release 2.4 today. This version improves on the previous release and addresses developer concerns

[{"selector":"#anim-611fe3cb-fd2e-4ef9-9e9f-b458d73aaeef [data-leaf-element=\"true\"]","keyframes":{"transform":["translate(0%, 0%) scale(1.5)","translate(0%, 0%) scale(1)"]},"delay":0,"duration":2000,"easing":"cubic-bezier(.3,0,.55,1)","fill":"forwards"}] [{"selector":"#anim-f0825cd7-d563-4afe-94d7-170df9cb2b59","keyframes":{"transform":["rotate(-540deg) scale(0.1)","none"],"opacity":[0,1]},"delay":0,"duration":1000,"fill":"both","iterations":1}]

Although the XLA compiler can optimise your models, custom kernel code can give model writers with better performance

[{"selector":"#anim-0e327745-733e-4518-8f5b-17d546ee646b [data-leaf-element=\"true\"]","keyframes":{"transform":["translate3d(-21.81318664215033%, 0, 0) translate(-25%, 0%) scale(1.5)","translate3d(0%, 0, 0) translate(0%, 0%) scale(1)"]},"delay":0,"duration":2000,"fill":"forwards"}] [{"selector":"#anim-a615373d-3325-44ff-8f1f-5394a3ae7672","keyframes":{"transform":["rotate(-540deg) scale(0.1)","none"],"opacity":[0,1]},"delay":0,"duration":1000,"fill":"both","iterations":1}]

Pallas, a bespoke kernel language that supports TPUs and GPUs, lets you write faster Python code closer to the hardware than C++

[{"selector":"#anim-19700265-29c8-46d3-8249-62e7834ee4f5 [data-leaf-element=\"true\"]","keyframes":{"transform":["translate(0%, 0%) scale(1.5)","translate(0%, 0%) scale(1)"]},"delay":0,"duration":2000,"easing":"cubic-bezier(.3,0,.55,1)","fill":"forwards"}] [{"selector":"#anim-3fa44d15-d397-479e-a6c2-6d58d07d4e02","keyframes":{"transform":["rotate(-540deg) scale(0.1)","none"],"opacity":[0,1]},"delay":0,"duration":1000,"fill":"both","iterations":1}]

This means PyTorch/XLA 2.4 generates the compute graph of operation before sending models to the XLA device target hardware

[{"selector":"#anim-5de5138f-5c91-43a5-80aa-a357f6d41180","keyframes":{"transform":["translate3d(118.66668%, 0px, 0)","translate3d(0px, 0px, 0)"]},"delay":0,"duration":1000,"easing":"cubic-bezier(.2, 0, .8, 1)","fill":"both"}] [{"selector":"#anim-02de53d3-c7ae-495d-b58e-8f3493dce6be","keyframes":{"transform":["rotateZ(180deg)","rotateZ(0deg)"]},"delay":0,"duration":1000,"easing":"cubic-bezier(.2, 0, .5, 1)","fill":"forwards"}] [{"selector":"#anim-9bf2982a-cc82-487a-b7c1-b7bb85c45c96 [data-leaf-element=\"true\"]","keyframes":{"transform":["translate3d(35.937499914669296%, 0, 0)","translate3d(0%, 0, 0)"]},"delay":0,"duration":2000,"easing":"cubic-bezier(.3,0,.55,1)","fill":"both"}]

Google Cloud adds a “mark step” call to each PyTorch action on TPUs to force compilation and execution

[{"selector":"#anim-1411ecdd-7504-42a1-b9cf-d60126a29d51 [data-leaf-element=\"true\"]","keyframes":{"transform":["translate3d(-34.179687404002955%, 0, 0) translate(-25%, 0%) scale(1.5)","translate3d(0%, 0, 0) translate(0%, 0%) scale(1)"]},"delay":0,"duration":2000,"fill":"forwards"}] [{"selector":"#anim-0b93bc26-1fbe-4ad9-9d8f-3d6a654e791a","keyframes":{"opacity":[1,1]},"delay":0,"duration":2000,"easing":"cubic-bezier(.3,0,.55,1)","fill":"both"}] [{"selector":"#anim-6feafd38-a5b2-4021-a2e0-1b1505c9b7c0","keyframes":{"transform":["scale(3)","scale(1)"]},"delay":0,"duration":2000,"easing":"cubic-bezier(.3,0,.55,1)","fill":"forwards"}]

Good thing your code works with PyTorch/XLA 2.4 despite API changes. New API methods will facilitate future development

[{"selector":"#anim-de5edab3-b1e5-4ce9-8791-27dd4bad2d5a [data-leaf-element=\"true\"]","keyframes":{"transform":["translate3d(36.15771803681094%, 0, 0)","translate3d(0%, 0, 0)"]},"delay":0,"duration":2000,"easing":"cubic-bezier(.3,0,.55,1)","fill":"both"}] [{"selector":"#anim-80d1a2c7-1449-4c50-9b8b-71bd989950e9","keyframes":{"transform":["translate3d(108.54431%, 0px, 0)","translate3d(0px, 0px, 0)"]},"delay":0,"duration":600,"easing":"cubic-bezier(0.4, 0.4, 0.0, 1)","fill":"both"}] [{"selector":"#anim-1817eb95-15ac-420d-8680-3f4b2aadf5ae","keyframes":{"opacity":[0,1]},"delay":0,"duration":600,"easing":"cubic-bezier(0.4, 0.4, 0.0, 1)","fill":"both"}] [{"selector":"#anim-c58d7932-ad81-4ca5-895a-553fd7ae85c7","keyframes":{"transform":["scale(0.15)","scale(1)"]},"delay":0,"duration":600,"easing":"cubic-bezier(0.4, 0.4, 0.0, 1)","fill":"forwards"}]

Accelerated Linear Algebra, or XLA, is an open-source machine learning compiler

[{"selector":"#anim-f0368297-1d2e-43b4-bed4-7e363d317c1c [data-leaf-element=\"true\"]","keyframes":{"transform":["translate3d(-34.179687404002955%, 0, 0) translate(-25%, 0%) scale(1.5)","translate3d(0%, 0, 0) translate(0%, 0%) scale(1)"]},"delay":0,"duration":2000,"fill":"forwards"}] [{"selector":"#anim-0f0d65ac-802a-482c-ab73-e2714cdce652","keyframes":{"opacity":[1,1]},"delay":0,"duration":2000,"easing":"cubic-bezier(.3,0,.55,1)","fill":"both"}] [{"selector":"#anim-940c49ec-3cd1-4e65-a4b4-f76611b92937","keyframes":{"transform":["scale(3)","scale(1)"]},"delay":0,"duration":2000,"easing":"cubic-bezier(.3,0,.55,1)","fill":"forwards"}]

The XLA compiler optimises PyTorch, TensorFlow, and JAX models for high-performance execution on GPUs, CPUs, and ML accelerators For more details Govindhtech.com

[{"selector":"#anim-a1ab1076-c4ba-41d3-b746-e0da68a401cc","keyframes":{"opacity":[1,1]},"delay":0,"duration":2000,"easing":"cubic-bezier(.3,0,.55,1)","fill":"both"}] [{"selector":"#anim-71f36370-b543-4c7f-b42a-89d6d0825683","keyframes":{"transform":["scale(3)","scale(1)"]},"delay":0,"duration":2000,"easing":"cubic-bezier(.3,0,.55,1)","fill":"forwards"}] [{"selector":"#anim-31a5f3c7-e542-40e4-93e1-efc5e464e0ae [data-leaf-element=\"true\"]","keyframes":{"transform":["translate3d(-21.81318664215033%, 0, 0) translate(-25%, 0%) scale(1.5)","translate3d(0%, 0, 0) translate(0%, 0%) scale(1)"]},"delay":0,"duration":2000,"fill":"forwards"}]