support ET dump for llama3 runner (pytorch#7507)

Summary:

Support ET dump for llama3 runner to easy understand the regression performance issue

Reviewed By: Andriyluck, limintang

Differential Revision: D67656207

Author: billmguo

examples/qualcomm/oss_scripts/llama3_2/CMakeLists.txt

@@ -5,7 +5,7 @@
 # LICENSE file in the root directory of this source tree.
 
 # model sharding with custom op
-set(CUSTOM_OP_SRCS_FILE 
+set(CUSTOM_OP_SRCS_FILE
   "${EXECUTORCH_SOURCE_DIR}/extension/llm/custom_ops/op_fallback.cpp"
 )
 add_library(custom_ops ${CUSTOM_OP_SRCS_FILE})
@@ -45,7 +45,7 @@ list(
 # build qnn llama3.2 1b runner
 add_executable(qnn_llama3_2_runner ${_llama3_2_runner__srcs})
 target_include_directories(
-  qnn_llama3_2_runner PUBLIC ${_common_include_directories}
+  qnn_llama3_2_runner PUBLIC ${_common_include_directories} ${EXECUTORCH_SOURCE_DIR}/devtools/etdump
 )
 
 target_link_libraries(
@@ -58,6 +58,7 @@ target_link_libraries(
   gflags
   re2::re2
   custom_ops
+  etdump
 )
 target_compile_options(
   qnn_llama3_2_runner PUBLIC ${_common_compile_options}

examples/qualcomm/oss_scripts/llama3_2/qnn_llama3_2_runner.cpp

@@ -49,6 +49,11 @@ DEFINE_int32(
     0,
     "0: PromptProcessor(prefill) / 1: TokenGenerator(kv) / 2: HybridMode (prefill+kv)");
 
+DEFINE_bool(
+    gen_etdump,
+    false,
+    "false: Disable ET dump/ True: Enable ET dump (default: false)");
+
 int main(int argc, char** argv) {
   gflags::ParseCommandLineFlags(&argc, &argv, true);
 
@@ -57,7 +62,8 @@ int main(int argc, char** argv) {
       {FLAGS_model_path},
       FLAGS_tokenizer_path.c_str(),
       FLAGS_temperature,
-      FLAGS_eval_mode);
+      FLAGS_eval_mode,
+      FLAGS_gen_etdump);
   std::vector<char> buf;
   buf.reserve(5 * FLAGS_seq_len); // assume each token is around 5 char
   std::ofstream fout(FLAGS_output_path.c_str());

examples/qualcomm/oss_scripts/llama3_2/runner/runner.cpp

@@ -41,7 +41,8 @@ Runner::Runner(
     const std::vector<std::string>& models_path,
     const std::string& tokenizer_path,
     const float temperature,
-    const int eval_mode)
+    const int eval_mode,
+    const bool gen_etdump)
     : n_bos_(1),
       n_eos_(1),
       tokenizer_path_(tokenizer_path),
@@ -54,6 +55,28 @@ Runner::Runner(
   }
   ET_LOG(Info, "creating runner: tokenizer_path=%s", tokenizer_path_.c_str());
   ET_LOG(Info, "eval mode=%d", eval_mode);
+  if (gen_etdump) {
+    gen_etdump_ = true;
+    switch (eval_mode) {
+      case EvalMode::kPrefill:
+        prefill_dump_ = std::make_unique<torch::executor::ETDumpGen>();
+        break;
+      case EvalMode::kKVCached:
+        decode_dump_ = std::make_unique<torch::executor::ETDumpGen>();
+        break;
+      case EvalMode::kHybrid:
+        prefill_dump_ = std::make_unique<torch::executor::ETDumpGen>();
+        decode_dump_ = std::make_unique<torch::executor::ETDumpGen>();
+        break;
+      default:
+        ET_CHECK_MSG(false, "Unsupported eval mode");
+        break;
+    }
+    std::string etdump_dir =
+        models_path[0].substr(0, models_path[0].find_last_of("/\\") + 1);
+    prefill_etdump_path_ = etdump_dir + "prefill_etdump.etdp";
+    decode_etdump_path_ = etdump_dir + "decode_etdump.etdp";
+  }
 }
 
 bool Runner::is_loaded() const {
@@ -91,9 +114,17 @@ Error Runner::load() {
 
   for (std::shared_ptr<Module>& module : modules_) {
     if (!prefill_forward_name_.empty()) {
+      if (gen_etdump_) {
+        ET_CHECK_OK_OR_RETURN_ERROR(
+            module->load_method(prefill_forward_name_, prefill_dump_.get()));
+      }
       ET_CHECK_OK_OR_RETURN_ERROR(module->load_method(prefill_forward_name_));
     }
     if (!kv_forward_name_.empty()) {
+      if (gen_etdump_) {
+        ET_CHECK_OK_OR_RETURN_ERROR(
+            module->load_method(kv_forward_name_, decode_dump_.get()));
+      }
       ET_CHECK_OK_OR_RETURN_ERROR(module->load_method(kv_forward_name_));
     }
   }
@@ -395,6 +426,8 @@ Error Runner::generate(
 
   stats_.num_prompt_tokens = num_prompt_tokens;
   stats_.num_generated_tokens = pos - num_prompt_tokens;
+  if (gen_etdump_)
+    gen_etdump_data();
   printReport(stats_);
   if (stats_callback) {
     stats_callback(stats_);
@@ -403,6 +436,22 @@ Error Runner::generate(
   return Error::Ok;
 }
 
+void Runner::gen_etdump_data() {
+  // dump the prefill and decode etdump data
+  if (prefill_dump_.get() != nullptr) {
+    torch::executor::etdump_result result = prefill_dump_->get_etdump_data();
+    FILE* ptr = fopen(prefill_etdump_path_.c_str(), "w+");
+    fwrite(result.buf, 1, result.size, ptr);
+    fclose(ptr);
+  }
+  if (decode_dump_.get() != nullptr) {
+    torch::executor::etdump_result result = decode_dump_->get_etdump_data();
+    FILE* ptr = fopen(decode_etdump_path_.c_str(), "w+");
+    fwrite(result.buf, 1, result.size, ptr);
+    fclose(ptr);
+  }
+}
+
 namespace {
 void printReport(const Runner::Stats& stats) {
   printf("PyTorchObserver %s\n", statsToJsonString(stats).c_str());

examples/qualcomm/oss_scripts/llama3_2/runner/runner.h

@@ -17,6 +17,7 @@
 #include <string>
 #include <unordered_map>
 
+#include <executorch/devtools/etdump/etdump_flatcc.h>
 #include <executorch/examples/qualcomm/oss_scripts/llama3_2/runner/io_memory.h>
 #include <executorch/extension/llm/sampler/sampler.h>
 #include <executorch/extension/llm/tokenizer/tokenizer.h>
@@ -30,7 +31,8 @@ class Runner {
       const std::vector<std::string>& models_path,
       const std::string& tokenizer_path,
       const float temperature,
-      const int eval_mode);
+      const int eval_mode,
+      const bool gen_etdump);
 
   struct Stats {
     // Scaling factor for timestamps - in this case, we use ms.
@@ -69,6 +71,7 @@ class Runner {
   void stop();
   std::vector<executorch::runtime::Result<executorch::runtime::MethodMeta>>
   get_methods_meta(std::string& method_name);
+  void gen_etdump_data();
 
  private:
   template <typename T>
@@ -93,6 +96,11 @@ class Runner {
   float temperature_;
   std::unique_ptr<executorch::extension::llm::Tokenizer> tokenizer_;
   std::unique_ptr<executorch::extension::llm::Sampler> sampler_;
+  std::unique_ptr<torch::executor::ETDumpGen> prefill_dump_;
+  std::unique_ptr<torch::executor::ETDumpGen> decode_dump_;
+  bool gen_etdump_ = false;
+  std::string prefill_etdump_path_;
+  std::string decode_etdump_path_;
   Stats stats_;
   std::unique_ptr<Memory> io_mem_;
   EvalMode eval_mode_;

examples/qualcomm/oss_scripts/llama3_2/targets.bzl

@@ -28,6 +28,7 @@ def define_common_targets():
             "//executorch/extension/llm/tokenizer:bpe_tokenizer",
             "//executorch/extension/evalue_util:print_evalue",
             "//executorch/backends/qualcomm/runtime:runtime",
+            "//executorch/devtools/etdump:etdump_flatcc",
         ],
         external_deps = [
             "gflags",