JAX key features and enhancements (#5708)

Co-authored-by: Pratik Basyal <prbasyal@amd.com>
This commit is contained in:
Istvan Kiss
2025-12-01 19:52:07 +01:00
committed by Istvan Kiss
parent e8fdc34b71
commit f37007bf42
2 changed files with 32 additions and 0 deletions

View File

@@ -36,6 +36,7 @@ Andrej
Arb
Autocast
autograd
Backported
BARs
BatchNorm
BLAS
@@ -202,9 +203,11 @@ GenAI
GenZ
GitHub
Gitpod
hardcoded
HBM
HCA
HGX
HLO
HIPCC
hipDataType
HIPExtension
@@ -329,6 +332,7 @@ MoEs
Mooncake
Mpops
Multicore
multihost
Multithreaded
mx
MXFP
@@ -1020,6 +1024,7 @@ uncacheable
uncorrectable
underoptimized
unhandled
unfused
uninstallation
unmapped
unsqueeze

View File

@@ -269,6 +269,33 @@ For a complete and up-to-date list of JAX public modules (for example, ``jax.num
JAX API modules are maintained by the JAX project and is subject to change.
Refer to the official Jax documentation for the most up-to-date information.
Key features and enhancements for ROCm 7.1
===============================================================================
- Enabled compilation of multihost HLO runner Python bindings.
- Backported multihost HLO runner bindings and some related changes to
:code:`FunctionalHloRunner`.
- Added :code:`requirements_lock_3_12` to enable building for Python 3.12.
- Removed hardcoded NHWC convolution layout for ``fp16`` precision to address the performance drops for ``fp16`` precision on gfx12xx GPUs.
- ROCprofiler-SDK integration:
- Integrated ROCprofiler-SDK (v3) to XLA to improve profiling of GPU events,
support both time-based and step-based profiling.
- Added unit tests for :code:`rocm_collector` and :code:`rocm_tracer`.
- Added Triton unsupported conversion from ``f8E4M3FNUZ`` to ``fp16`` with
rounding mode.
- Introduced :code:`CudnnFusedConvDecomposer` to revert fused convolutions
when :code:`ConvAlgorithmPicker` fails to find a fused algorithm, and removed
unfused fallback paths from :code:`RocmFusedConvRunner`.
Key features and enhancements for ROCm 7.0
===============================================================================