Fix text prefiller start pos being updated twice

Differential Revision: D86013259

Pull Request resolved: pytorch#15509

Author: larryliu0820

examples/models/llama/main.cpp

@@ -40,7 +40,12 @@ DEFINE_double(
 DEFINE_int32(
     seq_len,
     128,
-    "Total number of tokens to generate (prompt + output). Defaults to max_seq_len. If the number of input tokens + seq_len > max_seq_len, the output will be truncated to max_seq_len tokens.");
+    "DEPRECATED: Please use max_seq_len instead. Total number of tokens to generate (prompt + output). Defaults to max_seq_len. If the number of input tokens + seq_len > max_seq_len, the output will be truncated to max_seq_len tokens.");
+
+DEFINE_int32(
+    max_new_tokens,
+    -1,
+    "Total number of tokens to generate, excluding the prompt, will be capped by max_seq_len - # prompt tokens.");
 
 DEFINE_int32(
     cpu_threads,
@@ -122,20 +127,33 @@ int32_t main(int32_t argc, char** argv) {
   }
 
   if (warmup) {
-    auto error = runner->warmup(prompt, /*max_new_tokens=*/seq_len);
+    int32_t warmup_max_new_tokens =
+        FLAGS_max_new_tokens != -1 ? FLAGS_max_new_tokens : seq_len;
+    auto error =
+        runner->warmup(prompt, /*max_new_tokens=*/warmup_max_new_tokens);
     if (error != executorch::runtime::Error::Ok) {
       ET_LOG(Error, "Failed to warmup llama runner");
       return 1;
     }
-    // reset kv cache pos to 0
-    runner->reset();
   }
   // generate
   executorch::extension::llm::GenerationConfig config{
-      .seq_len = seq_len, .temperature = temperature};
+      .temperature = temperature};
+
+  if (FLAGS_max_new_tokens != -1) {
+    config.max_new_tokens = FLAGS_max_new_tokens;
+  } else {
+    ET_LOG(
+        Info,
+        "max_new_tokens not provided, falling back to seq_len=%d. "
+        "Consider using --max_new_tokens instead of --seq_len for specifying generation length.",
+        seq_len);
+    config.seq_len = seq_len;
+  }
+
   auto error = runner->generate(prompt, config);
   if (error != executorch::runtime::Error::Ok) {
-    ET_LOG(Error, "Failed to warmup llama runner");
+    ET_LOG(Error, "Failed to run llama runner");
     return 1;
   }
 

extension/llm/runner/test/test_text_prefiller.cpp

@@ -138,113 +138,68 @@ TEST_F(TextPrefillerTest, PrefillCallsPrefillChunkOnceWhenPromptFits) {
 TEST_F(
     TextPrefillerTest,
     PrefillCallsPrefillChunkMultipleTimesWhenPromptExceedsMaxLen) {
-  // Create a spy TextPrefiller with max_seq_len = 3
+  // Create a real TextPrefiller with max_seq_len = 3 and parallel prefill
   const int64_t max_seq_len = 3;
-  auto prefiller = createMockTextPrefiller(max_seq_len);
+  auto prefiller = createTextPrefiller(max_seq_len, true, true);
 
   // Create prompt tokens with size > max_seq_len
   std::vector<uint64_t> prompt_tokens = {1, 2, 3, 4, 5, 6, 7, 8};
   int64_t start_pos = 0;
 
-  // Set up expectations for prefill_chunk calls
-  {
-    InSequence seq; // Ensure calls happen in the expected order
-
-    // First chunk: tokens [1, 2, 3]
-    EXPECT_CALL(*prefiller, prefill_chunk(_, _))
-        .WillOnce([&](std::vector<uint64_t>& tokens, int64_t& pos) {
-          EXPECT_EQ(tokens.size(), 3);
-          EXPECT_EQ(tokens[0], 1);
-          EXPECT_EQ(tokens[1], 2);
-          EXPECT_EQ(tokens[2], 3);
-          EXPECT_EQ(pos, 0);
-          return Result<uint64_t>(10);
-        });
-
-    // Second chunk: tokens [4, 5, 6]
-    EXPECT_CALL(*prefiller, prefill_chunk(_, _))
-        .WillOnce([&](std::vector<uint64_t>& tokens, int64_t& pos) {
-          EXPECT_EQ(tokens.size(), 3);
-          EXPECT_EQ(tokens[0], 4);
-          EXPECT_EQ(tokens[1], 5);
-          EXPECT_EQ(tokens[2], 6);
-          EXPECT_EQ(pos, 3);
-          return Result<uint64_t>(20);
-        });
-
-    // Third chunk: tokens [7, 8]
-    EXPECT_CALL(*prefiller, prefill_chunk(_, _))
-        .WillOnce([&](std::vector<uint64_t>& tokens, int64_t& pos) {
-          EXPECT_EQ(tokens.size(), 2);
-          EXPECT_EQ(tokens[0], 7);
-          EXPECT_EQ(tokens[1], 8);
-          EXPECT_EQ(pos, 6);
-          return Result<uint64_t>(30);
-        });
-  }
+  // Track all tokens and positions passed to text_decoder_runner step
+  struct StepCall {
+    std::vector<uint64_t> tokens;
+    int64_t pos;
+  };
+  std::vector<StepCall> step_calls;
+
+  // Set up expectations for text_decoder_runner step calls
+  EXPECT_CALL(text_decoder_runner_, step(_, _))
+      .Times(3) // Should be called 3 times for 3 chunks
+      .WillRepeatedly(
+          [&](executorch::extension::TensorPtr& tokens, int64_t pos) {
+            // Extract token values from tensor
+            std::vector<uint64_t> token_values;
+            int64_t num_tokens = tokens->size(1);
+            auto* token_data = tokens->const_data_ptr<int64_t>();
+            for (int64_t i = 0; i < num_tokens; i++) {
+              token_values.push_back(static_cast<uint64_t>(token_data[i]));
+            }
+            step_calls.push_back({token_values, pos});
+            return Result<executorch::aten::Tensor>(tensor);
+          });
 
   // Call prefill
   auto result = prefiller->prefill(prompt_tokens, start_pos);
 
   // Verify the result
   EXPECT_EQ(result.error(), Error::Ok);
-  EXPECT_EQ(result.get(), 30); // Should return the token from the last chunk
-
-  // Verify that start_pos has been updated correctly
-  EXPECT_EQ(start_pos, prompt_tokens.size());
-}
-
-// Test that prefill() handles edge cases correctly
-TEST_F(TextPrefillerTest, PrefillHandlesEdgeCasesCorrectly) {
-  // Create a spy TextPrefiller with max_seq_len = 1
-  const int64_t max_seq_len = 1;
-  auto prefiller = createMockTextPrefiller(max_seq_len);
-
-  // Create prompt tokens with size > max_seq_len
-  std::vector<uint64_t> prompt_tokens = {1, 2, 3};
-  int64_t start_pos = 5; // Non-zero starting position
-
-  // Set up expectations for prefill_chunk calls
-  {
-    InSequence seq;
-
-    // First chunk: token [1]
-    EXPECT_CALL(*prefiller, prefill_chunk(_, _))
-        .WillOnce([&](std::vector<uint64_t>& tokens, int64_t& pos) {
-          EXPECT_EQ(tokens.size(), 1);
-          EXPECT_EQ(tokens[0], 1);
-          EXPECT_EQ(pos, 5);
-          return Result<uint64_t>(10);
-        });
-
-    // Second chunk: token [2]
-    EXPECT_CALL(*prefiller, prefill_chunk(_, _))
-        .WillOnce([&](std::vector<uint64_t>& tokens, int64_t& pos) {
-          EXPECT_EQ(tokens.size(), 1);
-          EXPECT_EQ(tokens[0], 2);
-          EXPECT_EQ(pos, 6);
-          return Result<uint64_t>(20);
-        });
-
-    // Third chunk: token [3]
-    EXPECT_CALL(*prefiller, prefill_chunk(_, _))
-        .WillOnce([&](std::vector<uint64_t>& tokens, int64_t& pos) {
-          EXPECT_EQ(tokens.size(), 1);
-          EXPECT_EQ(tokens[0], 3);
-          EXPECT_EQ(pos, 7);
-          return Result<uint64_t>(30);
-        });
-  }
-
-  // Call prefill
-  auto result = prefiller->prefill(prompt_tokens, start_pos);
 
-  // Verify the result
-  EXPECT_EQ(result.error(), Error::Ok);
-  EXPECT_EQ(result.get(), 30);
+  // Verify that step was called 3 times with correct tokens and positions
+  ASSERT_EQ(step_calls.size(), 3);
+
+  // First chunk: tokens [1, 2, 3] at position 0
+  EXPECT_EQ(step_calls[0].tokens.size(), 3);
+  EXPECT_EQ(step_calls[0].tokens[0], 1);
+  EXPECT_EQ(step_calls[0].tokens[1], 2);
+  EXPECT_EQ(step_calls[0].tokens[2], 3);
+  EXPECT_EQ(step_calls[0].pos, 0);
+
+  // Second chunk: tokens [4, 5, 6] at position 3
+  EXPECT_EQ(step_calls[1].tokens.size(), 3);
+  EXPECT_EQ(step_calls[1].tokens[0], 4);
+  EXPECT_EQ(step_calls[1].tokens[1], 5);
+  EXPECT_EQ(step_calls[1].tokens[2], 6);
+  EXPECT_EQ(step_calls[1].pos, 3);
+
+  // Third chunk: tokens [7, 8] at position 6
+  EXPECT_EQ(step_calls[2].tokens.size(), 2);
+  EXPECT_EQ(step_calls[2].tokens[0], 7);
+  EXPECT_EQ(step_calls[2].tokens[1], 8);
+  EXPECT_EQ(step_calls[2].pos, 6);
 
   // Verify that start_pos has been updated correctly
-  EXPECT_EQ(start_pos, 8); // 5 (initial) + 3 (tokens)
+  EXPECT_EQ(start_pos, prompt_tokens.size());
 }
 
 // Test that prefill() handles errors from prefill_chunk correctly
@@ -305,4 +260,119 @@ TEST_F(TextPrefillerTest, PrefillChunkWorksWithParallelPrefill) {
   // Verify that start_pos has been updated correctly
   EXPECT_EQ(start_pos, prompt_tokens.size());
 }
+// Test that prefill_chunk updates start_pos correctly with parallel prefill
+TEST_F(TextPrefillerTest, PrefillChunkUpdatesStartPosCorrectlyParallel) {
+  // Create a TextPrefiller with parallel prefill enabled
+  auto prefiller = createTextPrefiller(10, true, true);
+
+  // Set up expectations for the text decoder runner
+  int64_t captured_pos = -1;
+  EXPECT_CALL(text_decoder_runner_, step(_, _))
+      .WillOnce([&](executorch::extension::TensorPtr& tokens, int64_t pos) {
+        captured_pos = pos;
+        // Verify tokens shape is [1, num_tokens]
+        EXPECT_EQ(tokens->dim(), 2);
+        EXPECT_EQ(tokens->size(0), 1);
+        EXPECT_EQ(tokens->size(1), 3);
+        return Result<executorch::aten::Tensor>(tensor);
+      });
+
+  // Create prompt tokens
+  std::vector<uint64_t> prompt_tokens = {1, 2, 3};
+  int64_t start_pos = 5; // Non-zero starting position
+
+  // Call prefill_chunk directly
+  auto result = prefiller->prefill_chunk(prompt_tokens, start_pos);
+
+  // Verify the result
+  EXPECT_EQ(result.error(), Error::Ok);
+
+  // Verify that step was called with the original start_pos
+  EXPECT_EQ(captured_pos, 5);
+
+  // Verify that start_pos has been updated by the number of tokens
+  // This is the key test: start_pos should be updated exactly once
+  EXPECT_EQ(start_pos, 8); // 5 + 3 tokens
+}
+
+// Test that prefill_chunk updates start_pos correctly with sequential prefill
+TEST_F(TextPrefillerTest, PrefillChunkUpdatesStartPosCorrectlySequential) {
+  // Create a TextPrefiller with sequential prefill (parallel disabled)
+  auto prefiller = createTextPrefiller(10, true, false);
+
+  // Track all positions passed to step
+  std::vector<int64_t> captured_positions;
+  EXPECT_CALL(text_decoder_runner_, step(_, _))
+      .Times(3)
+      .WillRepeatedly(
+          [&](executorch::extension::TensorPtr& tokens, int64_t pos) {
+            captured_positions.push_back(pos);
+            // Verify tokens shape is [1, 1] for sequential prefill
+            EXPECT_EQ(tokens->dim(), 2);
+            EXPECT_EQ(tokens->size(0), 1);
+            EXPECT_EQ(tokens->size(1), 1);
+            return Result<executorch::aten::Tensor>(tensor);
+          });
+
+  // Create prompt tokens
+  std::vector<uint64_t> prompt_tokens = {1, 2, 3};
+  int64_t start_pos = 10; // Non-zero starting position
+
+  // Call prefill_chunk directly
+  auto result = prefiller->prefill_chunk(prompt_tokens, start_pos);
+
+  // Verify the result
+  EXPECT_EQ(result.error(), Error::Ok);
+
+  // Verify that step was called with incrementing positions
+  ASSERT_EQ(captured_positions.size(), 3);
+  EXPECT_EQ(captured_positions[0], 10); // First token at initial start_pos
+  EXPECT_EQ(captured_positions[1], 11); // Second token at start_pos + 1
+  EXPECT_EQ(captured_positions[2], 12); // Third token at start_pos + 2
+
+  // Verify that start_pos has been updated by the number of tokens
+  // This is the key test: start_pos should be updated exactly once per token
+  EXPECT_EQ(start_pos, 13); // 10 + 3 tokens
+}
+
+// Test that prefill with chunking updates start_pos correctly across chunks.
+// This test would have caught the bug where start_pos was being updated twice.
+TEST_F(
+    TextPrefillerTest,
+    PrefillWithChunkingUpdatesStartPosCorrectlyAcrossChunks) {
+  // Create a TextPrefiller with max_seq_len = 3 and parallel prefill
+  auto prefiller = createTextPrefiller(3, true, true);
+
+  // Track all positions passed to step
+  std::vector<int64_t> captured_positions;
+  EXPECT_CALL(text_decoder_runner_, step(_, _))
+      .Times(3) // Should be called 3 times: [1,2,3], [4,5,6], [7,8]
+      .WillRepeatedly(
+          [&](executorch::extension::TensorPtr& tokens, int64_t pos) {
+            captured_positions.push_back(pos);
+            return Result<executorch::aten::Tensor>(tensor);
+          });
+
+  // Create prompt tokens that exceed max_seq_len
+  std::vector<uint64_t> prompt_tokens = {1, 2, 3, 4, 5, 6, 7, 8};
+  int64_t start_pos = 100; // Non-zero starting position
+
+  // Call prefill (which will chunk internally)
+  auto result = prefiller->prefill(prompt_tokens, start_pos);
+
+  // Verify the result
+  EXPECT_EQ(result.error(), Error::Ok);
+
+  // Verify that step was called with correct positions for each chunk
+  // If start_pos were updated twice (the bug), these would be wrong
+  ASSERT_EQ(captured_positions.size(), 3);
+  EXPECT_EQ(captured_positions[0], 100); // Chunk 1: tokens [1,2,3]
+  EXPECT_EQ(captured_positions[1], 103); // Chunk 2: tokens [4,5,6]
+  EXPECT_EQ(captured_positions[2], 106); // Chunk 3: tokens [7,8]
+
+  // Verify that final start_pos is correct
+  // This is the key test for the bug: start_pos should be exactly
+  // initial_pos + num_tokens, not double-incremented
+  EXPECT_EQ(start_pos, 108); // 100 + 8 tokens
+}
 } // namespace

extension/llm/runner/text_llm_runner.cpp

@@ -247,7 +247,7 @@ Error TextLLMRunner::warmup(const std::string& prompt, int32_t max_new_tokens) {
   Error err = generate(prompt, config);
 
   // Reset stats after warmup, not resetting the std::unique_ptr!
-  stats_->reset();
+  reset();
   return err;
 }
 

extension/llm/runner/text_prefiller.cpp

@@ -59,7 +59,6 @@ ::executorch::runtime::Result<uint64_t> TextPrefiller::prefill(
       ET_CHECK_OK_OR_RETURN_ERROR(chunk_result.error());
       cur_token = chunk_result.get();
 
-      start_pos += num_tokens_to_prefill_with;
       num_tokens_to_process += num_tokens_to_prefill_with;
     }