OSSForks/tensorflow
1
0
Fork 0
mirror of https://github.com/zebrajr/tensorflow.git synced 2025-12-06 12:20:11 +01:00
Code Issues Actions 5 Packages Projects Releases Wiki Activity

Increase the maximum HLO op chain length for profiling from 8192 to 16384.

Browse Source 
This prevents max chains of trivial ops (e.g. add.fp32) to run faster than copying the data, which results in 'too fast to measure' error.

PiperOrigin-RevId: 826079017
This commit is contained in:
Christian Sigg 2025-10-30 10:01:34 -07:00 committed by TensorFlower Gardener
parent 71e640f242
commit 3943b53326
1 changed files with 3 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
third_party/xla/xla/service/gpu/model/hlo_op_profiler.cc vendored
View File

@ -397,7 +397,9 @@ absl::StatusOr<HloInstructionProfile> HloOpProfiler::MeasureClockCyclesPerOp(
// Longer chains are too slow to compile.
constexpr int kMinOpChainLength = 16;
constexpr int kMaxOpChainLength = 8192;
// If you get "too fast to measure" errors on faster GPUs, try increasing
// kMaxOpChainLength.
constexpr int kMaxOpChainLength = 16 * 1024;
absl::Duration duration = absl::ZeroDuration();
int chain_length = kMinOpChainLength;