Merge pull request #652 from google:yooh/xla-flags-fix

PiperOrigin-RevId: 634572059
google · May 17, 2024 · 8930215 · 8930215
2 parents fa92a21 + b98ad6c
commit 8930215
Showing 1 changed file with 2 additions and 3 deletions.
diff --git a/end_to_end/gpu/a3/test_llama2_7b.sh b/end_to_end/gpu/a3/test_llama2_7b.sh
@@ -53,9 +53,8 @@ export NVTE_FUSED_ATTN=1
 export NCCL_DEBUG=VERSION
 
 export XLA_FLAGS="--xla_dump_to=$BASE_OUTPUT_PATH/$RUN_NAME/HLO_dumps/
---xla_gpu_enable_latency_hiding_scheduler=true --xla_gpu_enable_async_all_gather=true
---xla_gpu_enable_async_reduce_scatter=true --xla_gpu_enable_triton_gemm=false
- --xla_gpu_graph_level=0 --xla_gpu_enable_async_all_reduce=true --xla_gpu_enable_highest_priority_async_stream=true
+--xla_gpu_enable_latency_hiding_scheduler=true --xla_gpu_enable_triton_gemm=false
+ --xla_gpu_graph_level=0 --xla_gpu_enable_highest_priority_async_stream=true
  --xla_gpu_all_reduce_combine_threshold_bytes=134217728 --xla_gpu_all_gather_combine_threshold_bytes=134217728
  --xla_gpu_reduce_scatter_combine_threshold_bytes=67108864 --xla_gpu_enable_pipelined_all_gather=true
  --xla_gpu_enable_pipelined_reduce_scatter=true --xla_gpu_enable_pipelined_all_reduce=true