[CUDA] Fix beam search of num_beams > 32 (#23599)

### Description * Pass topk_scores to beam scorer in slow topk path. * Add an env variable `ORT_BEAM_SEARCH_USE_FAST_TOPK` to enable/disable fast topk. * Add a test case for slow topk path. ### Motivation and Context This bug was introduced in #16272 Beam search uses fast cuda kernel when number of beams <= 32. When beam size is larger than that threshold, we use another code path (slower cuda kernel) to get topk. In such `slow topk path`, topk_scores shall be passed to beam scorer but it is not. This bug will cause incorrect result when num_beams > 32. It was not found previously since such large beam size is rarely used.
microsoft · Feb 7, 2025 · 09e5724 · 09e5724
1 parent 82840f6
commit 09e5724
Show file tree

Hide file tree

Showing 4 changed files with 35 additions and 15 deletions.
diff --git a/onnxruntime/contrib_ops/cpu/transformers/beam_search_parameters.cc b/onnxruntime/contrib_ops/cpu/transformers/beam_search_parameters.cc
@@ -2,6 +2,7 @@
 // Licensed under the MIT License.
 
 #include "contrib_ops/cpu/transformers/beam_search_parameters.h"
+#include "core/platform/env_var_utils.h"
 
 namespace onnxruntime {
 namespace contrib {
@@ -136,7 +137,11 @@ void BeamSearchParameters::ParseFromInputs(OpKernelContext* context) {
       temperature = 1.0f;
     }
   }
+
+  // The following parameter is read from environment variable for testing purpose.
+  use_fast_topk = ParseEnvironmentVariableWithDefault<bool>(kBeamSearchUseFastTopK, true);
 }
+
 void BeamSearchParameters::SetSubgraphParameters(int vocabulary_size, int heads, int hidden_size_per_head, int layers) {
   // Override vocab_size using the inferred shape from the decoder subgraph ONLY IF
   // the vocab_size hasn't been explicitly specified by the user (as an attribute of BeamSearch)

diff --git a/onnxruntime/contrib_ops/cpu/transformers/generation_shared.h b/onnxruntime/contrib_ops/cpu/transformers/generation_shared.h
@@ -199,8 +199,14 @@ struct IGenerationParameters {
   int extra_decoding_ids_input_id = -1;
   int cross_qk_output_id = -1;
   int no_speech_probs_output_id = -1;
+
+  // Parameter for testing slow topk path. It can be updated by the below environment variable.
+  bool use_fast_topk = true;
 };
 
+// Environment variable to enable/disable fast topk kernel on GPU. Default is 1 (enabled).
+constexpr const char* kBeamSearchUseFastTopK = "ORT_BEAM_SEARCH_USE_FAST_TOPK";
+
 }  // namespace transformers
 }  // namespace contrib
 }  // namespace onnxruntime
diff --git a/onnxruntime/contrib_ops/cuda/transformers/generation_device_helper.cc b/onnxruntime/contrib_ops/cuda/transformers/generation_device_helper.cc
@@ -524,7 +524,8 @@ Status ProcessLogits(const OrtValue& logits,                                 //
     beam_state->remaining_scores = beam_state->remaining_scores.subspan(next_token_scores.size());
   }
 
-  if (num_beams <= 32) {
+  gsl::span<float> scores_to_process = beam_state->next_scores;
+  if (parameters->use_fast_topk && num_beams <= 32) {
     constexpr size_t max_parts_of_vocab = 128;
     size_t candidate_count = SafeInt<size_t>(batch_beam_size) * 2 * num_beams;
     float* topk_tmp_buffer = beam_state->topk_buffer.data();
@@ -546,13 +547,6 @@ Status ProcessLogits(const OrtValue& logits,                                 //
                          beam_state->next_tokens.data(),
                          beam_state->next_indices.data(),
                          cuda_stream);
-
-    // Select [batch_size, 2 * num_beams] from [batch_size * num_beams, 2 * num_beams]
-#ifdef DEBUG_GENERATION
-    dumper->Print("next_tokens before scorer", beam_state->next_tokens.data(), batch_size, 2 * num_beams);
-    dumper->Print("next_indices before scorer", beam_state->next_indices.data(), batch_size, 2 * num_beams);
-    dumper->Print("next_scores before scorer", beam_state->next_scores.data(), batch_size, 2 * num_beams);
-#endif
   } else {
     // Apply top-k selection like the following:
     //   next_token_scores = next_token_scores.view(batch_size, num_beams * vocab_size)
@@ -588,18 +582,20 @@ Status ProcessLogits(const OrtValue& logits,                                 //
     cuda::LaunchNextTokenKernel(next_token_indices, beam_state->next_indices.data(), beam_state->next_tokens.data(),
                                 batch_size, top_k, vocab_size, cuda_stream);
 
-#ifdef DEBUG_GENERATION
-    dumper->Print("next_scores before scorer", topk_scores->Data<float>(), batch_size, top_k);
-    dumper->Print("next_tokens before scorer", beam_state->next_tokens.data(), batch_size, top_k);
-    dumper->Print("next_indices before scorer", beam_state->next_indices.data(), batch_size, top_k);
-#endif
+    scores_to_process = gsl::span<float>(topk_scores->MutableData<float>(), batch_size * top_k);
   }
 
   // gsl::span doesn't convert from non const to const, so all we're doing here is making each const.
-  gsl::span<const float> next_scores(beam_state->next_scores.data(), beam_state->next_scores.size());
+  gsl::span<const float> next_scores(scores_to_process.data(), scores_to_process.size());
   gsl::span<const int32_t> next_tokens(beam_state->next_tokens.data(), beam_state->next_tokens.size());
   gsl::span<const int32_t> next_indices(beam_state->next_indices.data(), beam_state->next_indices.size());
 
+#ifdef DEBUG_GENERATION
+  dumper->Print("next_scores before scorer", next_scores.data(), batch_size, 2 * num_beams);
+  dumper->Print("next_tokens before scorer", next_tokens.data(), batch_size, 2 * num_beams);
+  dumper->Print("next_indices before scorer", next_indices.data(), batch_size, 2 * num_beams);
+#endif
+
   beam_scorer->Process(
       *sequences,
       next_scores,
@@ -735,6 +731,7 @@ void CudaBeamSearchScorer::Process(transformers::ISequences& sequences,
                                        next_tokens,
                                        next_indices,
                                        stream_);
+
   CUDA_CALL_THROW(cudaEventRecord(event_process_complete_.Get(), stream_));
 
   cuda::LaunchBeamSearchScorer_AppendNextTokenToSequences(*state_cpu_,

diff --git a/onnxruntime/test/contrib_ops/beam_search_test.cc b/onnxruntime/test/contrib_ops/beam_search_test.cc
@@ -9,6 +9,8 @@
 #include "test/common/cuda_op_test_utils.h"
 #include "test/providers/model_tester.h"
 #include "test/util/include/current_test_name.h"
+#include "test/util/include/scoped_env_vars.h"
+#include "contrib_ops/cpu/transformers/generation_shared.h"
 
 #ifdef USE_CUDA
 #include "core/providers/cuda/cuda_provider_options.h"
@@ -19,7 +21,7 @@ extern std::unique_ptr<Ort::Env> ort_env;
 namespace onnxruntime {
 namespace test {
 
-TEST(BeamSearchTest, GptBeamSearchFp32) {
+void RunGptBeamSearchFp32() {
   std::vector<int64_t> input_ids_shape{3, 12};
   std::vector<int32_t> input_ids{
       0, 0, 0, 0, 0, 52, 195, 731, 321, 301, 734, 620,
@@ -107,6 +109,16 @@ TEST(BeamSearchTest, GptBeamSearchFp32) {
   ASSERT_TRUE(std::equal(expected_output.cbegin(), expected_output.cend(), result_span.begin(), result_span.end()));
 }
 
+TEST(BeamSearchTest, GptBeamSearchFp32) {
+  RunGptBeamSearchFp32();
+}
+
+TEST(BeamSearchTest, GptBeamSearchFp32_DisableFastTopK) {
+  ScopedEnvironmentVariables scoped_env_vars{
+      EnvVarMap{{onnxruntime::contrib::transformers::kBeamSearchUseFastTopK, "0"}}};
+  RunGptBeamSearchFp32();
+}
+
 TEST(BeamSearchTest, GptBeamSearchFp16) {
   std::vector<int64_t> input_ids_shape{3, 12};
   std::vector<int32_t> input_ids{