AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Reverse logic of context tracking enablement from #1927 (#1971)

Browse Source 
In this commit it disabled by default and can be enabled via
`RCCL_ENABLE_CONTEXT_TRACKING=1` for both (CDNA, RDNA)
Original PR https://github.com/ROCm/rccl/pull/1927
This commit is contained in:
Artem Kuzmitckii
2025-10-09 10:24:09 +02:00
committed by GitHub
parent cede6d0134
commit 00a42c80f3
4 changed files with 11 additions and 23 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/how-to/rccl-usage-tips.rst
+6 -4
View File
@@ -269,11 +269,13 @@ The default allreduce PyTorch benchmark peak bus bandwidth performance is
~170 GB/s on a single OAM with ROCm 6.2.4, while the optimized run for CPX on a
single OAM peaks at ~315 GB/s.
Context tracking on Radeon GPUs
Context tracking on GPUs
----------------------------------------
Context tracking is disabled on the AMD Instinct™ series of GPUs for better performance but is enabled for Radeon GPUs.
To disable context tracking for Radeon GPUs, set the following environment variable:
Context tracking is disabled by default for optimal performance. However, enabling of context tracking can significantly improve performance
in certain scenarios. To enable context tracking, set the following environment variable:
.. code-block:: shell
export RCCL_DISABLE_CONTEXT_TRACKING=1
export RCCL_ENABLE_CONTEXT_TRACKING=1