diff --git a/MaxText/configs/a3/llama_2_7b/16vm.sh b/MaxText/configs/a3/llama_2_7b/16vm.sh index fa07a470e..819407578 100644 --- a/MaxText/configs/a3/llama_2_7b/16vm.sh +++ b/MaxText/configs/a3/llama_2_7b/16vm.sh @@ -26,7 +26,7 @@ export NVTE_FUSED_ATTN=1 export NCCL_DEBUG=VERSION export XLA_FLAGS="--xla_dump_to=$OUTPUT_PATH/$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_simplify_all_fp_conversions --xla_gpu_graph_level=0 +--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_all_reduce_combine_threshold_bytes=1073741824 --xla_gpu_all_gather_combine_threshold_bytes=134217728 --xla_gpu_reduce_scatter_combine_threshold_bytes=134217728 --xla_gpu_enable_pipelined_all_gather=true diff --git a/MaxText/configs/a3/llama_2_7b/1vm.sh b/MaxText/configs/a3/llama_2_7b/1vm.sh index 9120b9311..fb6fac5fe 100644 --- a/MaxText/configs/a3/llama_2_7b/1vm.sh +++ b/MaxText/configs/a3/llama_2_7b/1vm.sh @@ -26,7 +26,7 @@ export NVTE_FUSED_ATTN=1 export NCCL_DEBUG=VERSION export XLA_FLAGS="--xla_dump_to=$OUTPUT_PATH/$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_simplify_all_fp_conversions +--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_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 diff --git a/MaxText/configs/a3/llama_2_7b/2vm.sh b/MaxText/configs/a3/llama_2_7b/2vm.sh index d1ee728ac..3723eb522 100644 --- a/MaxText/configs/a3/llama_2_7b/2vm.sh +++ b/MaxText/configs/a3/llama_2_7b/2vm.sh @@ -26,7 +26,7 @@ export NVTE_FUSED_ATTN=1 export NCCL_DEBUG=VERSION export XLA_FLAGS="--xla_dump_to=$OUTPUT_PATH/$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_simplify_all_fp_conversions --xla_gpu_graph_level=0 +--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_all_reduce_combine_threshold_bytes=67108864 --xla_gpu_all_gather_combine_threshold_bytes=134217728 --xla_gpu_reduce_scatter_combine_threshold_bytes=67108864 --xla_gpu_enable_pipelined_all_gather=true diff --git a/MaxText/configs/a3/llama_2_7b/4vm.sh b/MaxText/configs/a3/llama_2_7b/4vm.sh index 7712d036a..931c3a9bf 100644 --- a/MaxText/configs/a3/llama_2_7b/4vm.sh +++ b/MaxText/configs/a3/llama_2_7b/4vm.sh @@ -26,7 +26,7 @@ export NVTE_FUSED_ATTN=1 export NCCL_DEBUG=VERSION export XLA_FLAGS="--xla_dump_to=$OUTPUT_PATH/$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_simplify_all_fp_conversions +--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_all_reduce_combine_threshold_bytes=536870912 --xla_gpu_all_gather_combine_threshold_bytes=134217728 --xla_gpu_reduce_scatter_combine_threshold_bytes=67108864 --xla_gpu_enable_pipelined_all_gather=true diff --git a/MaxText/configs/a3/llama_2_7b/8vm.sh b/MaxText/configs/a3/llama_2_7b/8vm.sh index bb003f2f7..057c5f0e9 100644 --- a/MaxText/configs/a3/llama_2_7b/8vm.sh +++ b/MaxText/configs/a3/llama_2_7b/8vm.sh @@ -26,7 +26,7 @@ export NVTE_FUSED_ATTN=1 export NCCL_DEBUG=VERSION export XLA_FLAGS="--xla_dump_to=$OUTPUT_PATH/$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_simplify_all_fp_conversions --xla_gpu_graph_level=0 +--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_all_reduce_combine_threshold_bytes=1073741824 --xla_gpu_all_gather_combine_threshold_bytes=134217728 --xla_gpu_reduce_scatter_combine_threshold_bytes=67108864 --xla_gpu_enable_pipelined_all_gather=true diff --git a/end_to_end/gpu/a3/test_llama2_7b.sh b/end_to_end/gpu/a3/test_llama2_7b.sh index 7a6a9e0c3..213f03c82 100644 --- a/end_to_end/gpu/a3/test_llama2_7b.sh +++ b/end_to_end/gpu/a3/test_llama2_7b.sh @@ -54,7 +54,7 @@ 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_simplify_all_fp_conversions +--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_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