[ExecuTorch][Llama] Change runner to enable chunked prefill (#9805)

This diff adds code to chunk prompt longer than max_seq_len to enable prefill of larger context

Differential Revision: [D71833061](https://our.internmc.facebook.com/intern/diff/D71833061/)

Author: pytorchbot

examples/models/llama/runner/runner.cpp

@@ -11,6 +11,7 @@
 
 #include <executorch/examples/models/llama/runner/runner.h>
 
+#include <algorithm>
 #include <ctime>
 
 #include <executorch/extension/llm/runner/util.h>
@@ -140,7 +141,8 @@ Error Runner::load() {
   text_prefiller_ = std::make_unique<llm::TextPrefiller>(
       text_decoder_runner_.get(),
       metadata_.at(kUseKVCache),
-      metadata_.at(kEnableDynamicShape));
+      metadata_.at(kEnableDynamicShape),
+      metadata_.at(kMaxSeqLen));
 
   text_token_generator_ = std::make_unique<llm::TextTokenGenerator>(
       tokenizer_.get(),
@@ -221,11 +223,11 @@ Error Runner::generate(
 
   ET_CHECK_MSG(num_prompt_tokens >= 1, "Expected at least 1 prompt token");
   ET_CHECK_MSG(
-      num_prompt_tokens < metadata_.at(kMaxSeqLen),
+      num_prompt_tokens < metadata_.at(kMaxContextLen),
       "num_prompt_tokens %d >= max_seq_len_ %" PRId64
       ", Max seq length exceeded - please increase max seq len value in .../llama2/model.py",
       num_prompt_tokens,
-      metadata_.at(kMaxSeqLen));
+      metadata_.at(kMaxContextLen));
   ET_CHECK_MSG(
       num_prompt_tokens < seq_len,
       "num_prompt_tokens %d >= seq_len %d, Sequence length exceeded - please increase the seq_len value passed to generate()",
@@ -242,10 +244,10 @@ Error Runner::generate(
   }
   int64_t pos = 0;
   auto prefill_res = text_prefiller_->prefill(prompt_tokens, pos);
-  stats_.first_token_ms = llm::time_in_ms();
-  stats_.prompt_eval_end_ms = llm::time_in_ms();
   ET_CHECK_OK_OR_RETURN_ERROR(prefill_res.error());
   uint64_t cur_token = prefill_res.get();
+  stats_.first_token_ms = llm::time_in_ms();
+  stats_.prompt_eval_end_ms = llm::time_in_ms();
 
   // print the first token from prefill. No prev_token so use cur_token for it.
   wrapped_callback(

examples/models/llava/runner/llava_runner.cpp

@@ -55,7 +55,8 @@ Error LlavaRunner::load() {
   text_prefiller_ = std::make_unique<llm::TextPrefiller>(
       text_decoder_runner_.get(),
       /*use_kv_cache=*/true,
-      /*enable_parallel_prefill=*/true);
+      /*enable_parallel_prefill=*/true,
+      /*max_seq_len=*/128);
 
   // Load the image prefiller
   image_prefiller_ = std::make_unique<LlavaImagePrefiller>(module_.get());

extension/llm/runner/text_prefiller.cpp

@@ -10,6 +10,7 @@
 // LLM.
 
 #include <executorch/extension/llm/runner/text_prefiller.h>
+#include <algorithm>
 
 namespace executorch {
 namespace extension {
@@ -18,10 +19,13 @@ namespace llm {
 TextPrefiller::TextPrefiller(
     TextDecoderRunner* text_decoder_runner,
     bool use_kv_cache,
-    bool enable_parallel_prefill)
+    bool enable_parallel_prefill,
+    int64_t max_seq_len)
     : text_decoder_runner_(text_decoder_runner),
       use_kv_cache_(use_kv_cache),
-      enable_parallel_prefill_(enable_parallel_prefill) {}
+      enable_parallel_prefill_(enable_parallel_prefill),
+      max_seq_len_(max_seq_len > 0 ? max_seq_len - 1 : 127) {
+} // -1 because for some reason tracing results in this upperbound
 
 ::executorch::runtime::Result<uint64_t> TextPrefiller::prefill(
     std::vector<uint64_t>& prompt_tokens,
@@ -30,6 +34,45 @@ ::executorch::runtime::Result<uint64_t> TextPrefiller::prefill(
   if (!text_decoder_runner_->is_method_loaded()) {
     ET_CHECK_OK_OR_RETURN_ERROR(text_decoder_runner_->load());
   }
+
+  // Check if we need to chunk the prompt tokens
+  int32_t num_prompt_tokens = prompt_tokens.size();
+
+  // If prompt tokens exceed max_seq_len_, we need to chunk them
+  if (num_prompt_tokens > max_seq_len_) {
+    uint64_t cur_token = 0;
+    int num_tokens_to_process = 0;
+
+    while (num_tokens_to_process < num_prompt_tokens) {
+      auto num_tokens_to_prefill_with = std::min<int>(
+          num_prompt_tokens - num_tokens_to_process, max_seq_len_);
+
+      std::vector<uint64_t> prompt_tokens_to_process(
+          num_tokens_to_prefill_with);
+      std::copy(
+          prompt_tokens.begin() + num_tokens_to_process,
+          prompt_tokens.begin() + num_tokens_to_process +
+              num_tokens_to_prefill_with,
+          prompt_tokens_to_process.begin());
+
+      // Process this chunk
+      auto chunk_result = prefillChunk(prompt_tokens_to_process, start_pos);
+      ET_CHECK_OK_OR_RETURN_ERROR(chunk_result.error());
+      cur_token = chunk_result.get();
+
+      num_tokens_to_process += num_tokens_to_prefill_with;
+    }
+
+    return cur_token;
+  } else {
+    // If prompt tokens don't exceed max_seq_len_, process them directly
+    return prefillChunk(prompt_tokens, start_pos);
+  }
+}
+
+::executorch::runtime::Result<uint64_t> TextPrefiller::prefillChunk(
+    std::vector<uint64_t>& prompt_tokens,
+    int64_t& start_pos) {
   // enable_parallel_prefill_ maybe set even when not using kv cache
   // When kv cache is not used, start pos is ignored
   int32_t num_prompt_tokens = prompt_tokens.size();

extension/llm/runner/text_prefiller.h

@@ -22,7 +22,8 @@ class ET_EXPERIMENTAL TextPrefiller {
   TextPrefiller(
       TextDecoderRunner* text_decoder_runner,
       bool use_kv_cache_,
-      bool enable_parallel_prefill);
+      bool enable_parallel_prefill,
+      int64_t max_seq_len = 128);
   /**
    * Prefill an LLM Module with the given text input.
    * @param prompt_tokens The text prompt tokens to the LLM Module. Encoded by
@@ -35,10 +36,22 @@ class ET_EXPERIMENTAL TextPrefiller {
       std::vector<uint64_t>& prompt_tokens,
       int64_t& start_pos);
 
+  /**
+   * Helper method to prefill a chunk of tokens.
+   * @param prompt_tokens The chunk of text prompt tokens to process.
+   * @param start_pos The starting position in KV cache of the input in the LLM
+   * Module.
+   * @return The next token of the LLM Module after prefilling this chunk.
+   */
+  ::executorch::runtime::Result<uint64_t> prefillChunk(
+      std::vector<uint64_t>& prompt_tokens,
+      int64_t& start_pos);
+
  private:
   TextDecoderRunner* text_decoder_runner_;
   bool use_kv_cache_;
   bool enable_parallel_prefill_;
+  int64_t max_seq_len_;
 };
 
 } // namespace llm