From 2f3198c0b0194122e24e43476345ffc146c0747a Mon Sep 17 00:00:00 2001 From: Seventeen17 <17aloha@gmail.com> Date: Tue, 19 Nov 2024 14:38:44 +0800 Subject: [PATCH] align code with flash_api.cpp when num_splits > 1 --- test/test_flash_attention_forward.py | 2 +- torch_xla/csrc/ops/flash_attention_forward.cpp | 14 +++++++++++++- 2 files changed, 14 insertions(+), 2 deletions(-) diff --git a/test/test_flash_attention_forward.py b/test/test_flash_attention_forward.py index 9ccb91481efe..9e019ed29325 100644 --- a/test/test_flash_attention_forward.py +++ b/test/test_flash_attention_forward.py @@ -27,7 +27,7 @@ def setup_env(): @pytest.mark.parametrize("alibi", [False, True]) @pytest.mark.parametrize("local", [False, True]) @pytest.mark.parametrize("causal", [False, True]) -@pytest.mark.parametrize("d", [8]) +@pytest.mark.parametrize("d", [256]) @pytest.mark.parametrize("softmax_scale", [0.25]) @pytest.mark.parametrize( "seqlen_q,seqlen_k", diff --git a/torch_xla/csrc/ops/flash_attention_forward.cpp b/torch_xla/csrc/ops/flash_attention_forward.cpp index 0bf9342395e5..7c9386fcfcc9 100644 --- a/torch_xla/csrc/ops/flash_attention_forward.cpp +++ b/torch_xla/csrc/ops/flash_attention_forward.cpp @@ -235,13 +235,25 @@ void custom_call_flash_attention_forward(cudaStream_t stream, void** buffers, // inference. const int num_m_blocks = (params.seqlen_q + 64 - 1) / 64; launch_params.num_splits = 0; - if (1 - params.p_dropout == 0.0f) { // SplitKV is not implemented for dropout + if (params.p_dropout == 1.0f) { // SplitKV is not implemented for dropout if (launch_params.num_splits < 1) { auto dprops = at::cuda::getCurrentDeviceProperties(); launch_params.num_splits = num_splits_heuristic(params.b * params.h * num_m_blocks, dprops->multiProcessorCount, num_n_blocks, 128); } + if (launch_params.num_splits > 1) { + at::Tensor softmax_lse_accum = + torch::empty({launch_params.num_splits, params.b, params.h, + launch_params.seqlen_q}, + opts.dtype(at::kFloat)); + at::Tensor out_accum = + torch::empty({params.num_splits, params.b, params.h, + launch_params.seqlen_q, launch_params.d_rounded}, + opts.dtype(at::kFloat)); + launch_params.softmax_lseaccum_ptr = softmax_lse_accum.data_ptr(); + launch_params.oaccum_ptr = out_accum.data_ptr(); + } } int64_t counter_offset = params.b * params.h * 32;