fix: Align code with flash_api.cpp when num_splits > 1 (#22)

AlibabaPAI · Nov 21, 2024 · 0451484 · 0451484
1 parent ee9724f
commit 0451484
Show file tree

Hide file tree

Showing 2 changed files with 25 additions and 5 deletions.
diff --git 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
@@ -234,15 +234,31 @@ void custom_call_flash_attention_forward(cudaStream_t stream, void** buffers,
   // kernel. In any case we don't expect seqlen_q to be larger than 64 for
   // 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
+  launch_params.num_splits = params.num_splits;
+  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);
     }
   }
+  const int num_splits =
+      launch_params.num_splits == 0 ? 1 : launch_params.num_splits;
+  at::Tensor softmax_lse_accum =
+      torch::empty({num_splits, params.b, params.h, launch_params.seqlen_q},
+                   opts.dtype(at::kFloat));
+  at::Tensor out_accum =
+      torch::empty({num_splits, params.b, params.h, launch_params.seqlen_q,
+                    launch_params.d_rounded},
+                   opts.dtype(at::kFloat));
+  if (fabs(1 - params.p_dropout) <
+      1e-6) {  // SplitKV is not implemented for dropout
+    if (launch_params.num_splits > 1) {
+      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;
 
@@ -269,8 +285,12 @@ void custom_call_flash_attention_forward(cudaStream_t stream, void** buffers,
 
   FP16_SWITCH(!launch_params.is_bf16, [&] {
     HEADDIM_SWITCH(launch_params.d, [&] {
-      // TODO(wenting.swt): support split_kv
-      run_mha_fwd_<elem_type, kHeadDim>(launch_params, torch_stream);
+      if (launch_params.num_splits <= 1) {
+        run_mha_fwd_<elem_type, kHeadDim>(launch_params, torch_stream);
+      } else {
+        run_mha_fwd_splitkv_dispatch<elem_type, kHeadDim>(launch_params,
+                                                          torch_stream);
+      }
     });
   });