PR openxla#12255: [GPU] Fix FMHA hangs by moving compilation to thunk…

… initialization. Imported from GitHub PR openxla#12255 Applies the same fix as in openxla#12228 to FMHA. Copybara import of the project: -- 70a4282 by Ilia Sergachev <[email protected]>: [GPU] Fix FMHA hangs by moving compilation to thunk initialization. Merging this change closes openxla#12255 COPYBARA_INTEGRATE_REVIEW=openxla#12255 from openxla:fix_fmha_hang 70a4282 PiperOrigin-RevId: 633217196
ROCm · May 13, 2024 · c7a2865 · c7a2865
1 parent 9f9d7f6
commit c7a2865
Show file tree

Hide file tree

Showing 4 changed files with 70 additions and 43 deletions.
diff --git a/xla/service/gpu/gpu_fused_mha_runner.cc b/xla/service/gpu/gpu_fused_mha_runner.cc
@@ -68,29 +68,13 @@ absl::Status RunFusedMHA(GpufMHAParams params, se::Stream *stream,
     dropout_rate = *params.config->dropout_rate;
   }
 
-  double scale = 1.0;
-  if (params.config->fmha_scale) {
-    scale = *params.config->fmha_scale;
-  }
-
   std::optional<int64_t> seed;
   if (params.config->seed) {
     seed = *params.config->seed;
   }
-  TF_ASSIGN_OR_RETURN(
-      se::dnn::FMHAMaskKind mask_type,
-      GetDNNFmhaMaskKindFromCudnnFmhaMaskKind(params.config->mask_type));
-  se::dnn::FusedMHAOp::Config config{scale,
-                                     params.config->lhs_bmm1,
-                                     params.config->rhs_bmm1,
-                                     params.config->rhs_bmm2,
-                                     params.config->intermediate_lhs_bmm2,
-                                     params.config->output,
-                                     params.config->bias,
-                                     params.config->activation,
-                                     dropout_rate,
-                                     seed,
-                                     mask_type};
+
+  TF_ASSIGN_OR_RETURN(se::dnn::FusedMHAOp::Config config,
+                      params.config->AsDnnFusedMHAOpConfig());
   TF_ASSIGN_OR_RETURN(auto *runner,
                       lazy_runner->GetOrCreateRunner(config, stream));
   return (*runner)(stream, options.profile_result, scratch_memory,
@@ -183,35 +167,13 @@ absl::Status RunFusedMHABackward(
     dropout_rate = *params.config->dropout_rate;
   }
 
-  double scale = 1.0;
-  if (params.config->fmha_scale) {
-    scale = *params.config->fmha_scale;
-  }
-
   std::optional<int64_t> seed;
   if (params.config->seed) {
     seed = *params.config->seed;
   }
 
-  TF_ASSIGN_OR_RETURN(
-      se::dnn::FMHAMaskKind mask_type,
-      GetDNNFmhaMaskKindFromCudnnFmhaMaskKind(params.config->mask_type));
-  se::dnn::FusedMHABackwardOp::Config config{scale,
-                                             params.config->bmm1_grad_gemm1_rhs,
-                                             params.config->bmm1_grad_gemm2_rhs,
-                                             params.config->bmm2_grad_gemm1_lhs,
-                                             params.config->bmm2_grad_gemm2_rhs,
-                                             params.config->d_output,
-                                             params.config->d_bmm1_lhs,
-                                             params.config->d_bmm1_rhs,
-                                             params.config->d_bmm2_rhs,
-                                             params.config->d_s,
-                                             params.config->d_bias,
-                                             params.config->fwd_output,
-                                             params.config->bias,
-                                             dropout_rate,
-                                             seed,
-                                             mask_type};
+  TF_ASSIGN_OR_RETURN(se::dnn::FusedMHABackwardOp::Config config,
+                      params.config->AsDnnFusedMHABackwardOpConfig());
   TF_ASSIGN_OR_RETURN(auto *runner,
                       lazy_runner->GetOrCreateRunner(config, stream));
   // TODO: pass in real softmax_sum, dQ_accum, fwd_output
@@ -404,6 +366,21 @@ absl::Status RunGpuFMHABackwardImpl(const GpufMHABackwardParams &params,
   return config;
 }
 
+absl::StatusOr<se::dnn::FusedMHAOp::Config>
+GpufMHAConfig::AsDnnFusedMHAOpConfig() const {
+  double scale = 1.0;
+  if (fmha_scale.has_value()) {
+    scale = *fmha_scale;
+  }
+  TF_ASSIGN_OR_RETURN(se::dnn::FMHAMaskKind mask_type,
+                      GetDNNFmhaMaskKindFromCudnnFmhaMaskKind(mask_type));
+
+  return se::dnn::FusedMHAOp::Config{
+      scale,    lhs_bmm1, rhs_bmm1,   rhs_bmm2,     intermediate_lhs_bmm2,
+      output,   bias,     activation, dropout_rate, seed,
+      mask_type};
+}
+
 /*static*/ absl::StatusOr<GpufMHABackwardConfig> GpufMHABackwardConfig::For(
     const GpufMHABackwardDescriptor &desc) {
   // Get shapes from desc.
@@ -546,6 +523,32 @@ absl::Status RunGpuFMHABackwardImpl(const GpufMHABackwardParams &params,
   return config;
 }
 
+absl::StatusOr<se::dnn::FusedMHABackwardOp::Config>
+GpufMHABackwardConfig::AsDnnFusedMHABackwardOpConfig() const {
+  double scale = 1.0;
+  if (fmha_scale.has_value()) {
+    scale = *fmha_scale;
+  }
+  TF_ASSIGN_OR_RETURN(se::dnn::FMHAMaskKind mask_type,
+                      GetDNNFmhaMaskKindFromCudnnFmhaMaskKind(mask_type));
+  return se::dnn::FusedMHABackwardOp::Config{scale,
+                                             bmm1_grad_gemm1_rhs,
+                                             bmm1_grad_gemm2_rhs,
+                                             bmm2_grad_gemm1_lhs,
+                                             bmm2_grad_gemm2_rhs,
+                                             d_output,
+                                             d_bmm1_lhs,
+                                             d_bmm1_rhs,
+                                             d_bmm2_rhs,
+                                             d_s,
+                                             d_bias,
+                                             fwd_output,
+                                             bias,
+                                             dropout_rate,
+                                             seed,
+                                             mask_type};
+}
+
 /*static*/ absl::StatusOr<GpufMHAParams> GpufMHAParams::For(
     const GpufMHAConfig &config, se::DeviceMemoryBase lhs_bmm1_buffer,
     se::DeviceMemoryBase rhs_bmm1_buffer, se::DeviceMemoryBase rhs_bmm2_buffer,

diff --git a/xla/service/gpu/gpu_fused_mha_runner.h b/xla/service/gpu/gpu_fused_mha_runner.h
@@ -101,10 +101,14 @@ struct GpufMHABackwardDescriptor {
   std::optional<Shape> d_bias_shape;
   std::optional<Shape> bias_shape;
 };
+
 // Structure to describe static properties of a GPU fused Multi-Headed
 // Attention.
 struct GpufMHAConfig {
   static absl::StatusOr<GpufMHAConfig> For(const GpufMHADescriptor& fmha_desc);
+
+  absl::StatusOr<se::dnn::FusedMHAOp::Config> AsDnnFusedMHAOpConfig() const;
+
   PrimitiveType
       input_type;  // Capture the primitive type of one of the inputs of BMM1
   PrimitiveType output_type;
@@ -133,6 +137,10 @@ struct GpufMHAConfig {
 struct GpufMHABackwardConfig {
   static absl::StatusOr<GpufMHABackwardConfig> For(
       const GpufMHABackwardDescriptor& fmha_desc);
+
+  absl::StatusOr<se::dnn::FusedMHABackwardOp::Config>
+  AsDnnFusedMHABackwardOpConfig() const;
+
   PrimitiveType
       input_type;  // Capture the primitive type of one of the inputs of BMM1
   PrimitiveType output_type;

diff --git a/xla/service/gpu/runtime/fused_mha_thunk.cc b/xla/service/gpu/runtime/fused_mha_thunk.cc
@@ -65,6 +65,13 @@ std::optional<se::DeviceMemoryBase> AssignBufferIfNotNull(
              : std::nullopt;
 }
 
+absl::Status FusedMHAThunk::Initialize(const InitializeParams& params) {
+  se::dnn::LazyOpRunner<se::dnn::FusedMHAOp>* lazy_runner =
+      GetOrCreateRunner(params.stream).AsFusedMHARunner();
+  TF_ASSIGN_OR_RETURN(auto config, config_.AsDnnFusedMHAOpConfig());
+  return lazy_runner->GetOrCreateRunner(config, params.stream).status();
+}
+
 absl::Status FusedMHAThunk::ExecuteOnStream(const ExecuteParams& params) {
   const auto& buffer_allocations = *params.buffer_allocations;
   se::DeviceMemoryBase lhs_bmm1_buffer =
@@ -143,6 +150,13 @@ FusedMHABackwardThunk::GetOrCreateRunner(
   return *it->second;
 }
 
+absl::Status FusedMHABackwardThunk::Initialize(const InitializeParams& params) {
+  se::dnn::LazyOpRunner<se::dnn::FusedMHABackwardOp>* lazy_runner =
+      GetOrCreateRunner(params.stream).AsFusedMHABackwardRunner();
+  TF_ASSIGN_OR_RETURN(auto config, config_.AsDnnFusedMHABackwardOpConfig());
+  return lazy_runner->GetOrCreateRunner(config, params.stream).status();
+}
+
 absl::Status FusedMHABackwardThunk::ExecuteOnStream(
     const ExecuteParams& params) {
   const auto& buffer_allocations = *params.buffer_allocations;

diff --git a/xla/service/gpu/runtime/fused_mha_thunk.h b/xla/service/gpu/runtime/fused_mha_thunk.h
@@ -53,6 +53,7 @@ class FusedMHAThunk : public Thunk {
   FusedMHAThunk(const FusedMHAThunk&) = delete;
   FusedMHAThunk& operator=(const FusedMHAThunk&) = delete;
 
+  absl::Status Initialize(const InitializeParams& params) override;
   absl::Status ExecuteOnStream(const ExecuteParams& params) override;
 
  private:
@@ -101,6 +102,7 @@ class FusedMHABackwardThunk : public Thunk {
   FusedMHABackwardThunk(const FusedMHABackwardThunk&) = delete;
   FusedMHABackwardThunk& operator=(const FusedMHABackwardThunk&) = delete;
 
+  absl::Status Initialize(const InitializeParams& params) override;
   absl::Status ExecuteOnStream(const ExecuteParams& params) override;
 
  private: