[REFACTOR] Cleanup Metrics (#2392)

This PR run another round of cleanup of metrics. - Remove less useful ones - Reorganize by labels in prometheus style

Author: tqchen

cpp/serve/engine_actions/action_commons.cc

@@ -94,8 +94,6 @@ void ProcessFinishedRequestStateEntries(std::vector<RequestStateEntry> finished_
  static_cast<double>((trequest_finish - rstate->tprefill_finish).count()) / 1e9;
  estate->metrics.sum_num_input_tokens += rsentry->request->num_input_tokens;
  estate->metrics.sum_num_prefill_tokens += root_rsentry->num_prefill_tokens;
- estate->metrics.sum_request_prefill_time += prefill_elapsed_time;
- estate->metrics.sum_request_decode_time += decode_elapsed_time;
  int64_t num_output_tokens = 0;
  for (const RequestStateEntry& entry : rstate->entries) {
  num_output_tokens += entry->mstates[0]->committed_tokens.size();
@@ -105,11 +103,11 @@ void ProcessFinishedRequestStateEntries(std::vector<RequestStateEntry> finished_
  num_output_tokens -= rsentry->request->generation_cfg->n;
  estate->metrics.sum_num_output_tokens += num_output_tokens;
 
- estate->metrics.num_last_finished_req_input_tokens.Set(rsentry->request->num_input_tokens);
- estate->metrics.num_last_finished_req_prefill_tokens.Set(root_rsentry->num_prefill_tokens);
- estate->metrics.num_last_finished_req_output_tokens.Set(num_output_tokens);
- estate->metrics.last_finished_req_prefill_time.Set(prefill_elapsed_time);
- estate->metrics.last_finished_req_decode_time.Set(decode_elapsed_time);
+ estate->metrics.last_finished_req_num_input_tokens = rsentry->request->num_input_tokens;
+ estate->metrics.last_finished_req_num_prefill_tokens = root_rsentry->num_prefill_tokens;
+ estate->metrics.last_finished_req_num_output_tokens = num_output_tokens;
+ estate->metrics.last_finished_req_prefill_time = prefill_elapsed_time;
+ estate->metrics.last_finished_req_decode_time = decode_elapsed_time;
  }
  }
 }

cpp/serve/engine_actions/batch_decode.cc

@@ -130,7 +130,7 @@ class BatchDecodeActionObj : public EngineActionObj {
  auto tend = std::chrono::high_resolution_clock::now();
  double elapsed_time = static_cast<double>((tend - tstart).count()) / 1e9;
  estate->metrics.sum_engine_decode_time += elapsed_time;
- estate->metrics.UpdateBatchDecodeTime(num_rsentries, elapsed_time);
+ estate->metrics.UpdateDecodeTimeByBatchSize(num_rsentries, elapsed_time);
 
  return estate->running_queue;
  }

cpp/serve/engine_actions/batch_draft.cc

@@ -134,11 +134,10 @@ class BatchDraftActionObj : public EngineActionObj {
  &model_workspaces_[0].draft_probs_storage);
  for (int i = 0; i < num_rsentries; ++i) {
  mstates[i]->AddDraftToken(sample_results[i], draft_token_slots_[i]);
- estate->metrics.sum_num_draft_tokens += 1;
  }
 
  auto tdraft_end = std::chrono::high_resolution_clock::now();
- estate->metrics.UpdateBatchDraftTime(
+ estate->metrics.UpdateDraftTimeByBatchSize(
  num_rsentries, static_cast<double>((tdraft_end - tdraft_start).count()) / 1e9);
  }
  }

cpp/serve/engine_actions/batch_verify.cc

@@ -147,9 +147,8 @@ class BatchVerifyActionObj : public EngineActionObj {
  rsentries[i]->mstates[verify_model_id_]->CommitToken(sample_result);
  rsentries[i]->mstates[draft_model_id_]->CommitToken(sample_result);
  }
- estate->metrics.sum_num_accepted_tokens += accept_length;
- estate->metrics.UpdateSpecDecodingStats(cum_verify_lengths[i + 1] - cum_verify_lengths[i],
- accept_length);
+ estate->metrics.spec_decode.Update(cum_verify_lengths[i + 1] - cum_verify_lengths[i],
+ accept_length);
  int rollback_length =
  std::max(cum_verify_lengths[i + 1] - cum_verify_lengths[i] - accept_length, 0);
  // rollback kv cache
@@ -210,7 +209,7 @@ class BatchVerifyActionObj : public EngineActionObj {
  auto tend = std::chrono::high_resolution_clock::now();
  double elapsed_time = static_cast<double>((tend - tstart).count()) / 1e9;
  estate->metrics.sum_engine_decode_time += elapsed_time;
- estate->metrics.UpdateBatchVerificationTime(total_verify_length, elapsed_time);
+ estate->metrics.UpdateVerifyTimeByBatchSize(total_verify_length, elapsed_time);
 
  return estate->running_queue;
  }

cpp/serve/engine_actions/eagle_batch_draft.cc

@@ -151,11 +151,10 @@ class EagleBatchDraftActionObj : public EngineActionObj {
  // No need to save hidden states as they are not used by subsequent engine actions
  for (int i = 0; i < num_rsentries; ++i) {
  mstates[i]->AddDraftToken(sample_results[i], draft_token_slots_[i]);
- estate->metrics.sum_num_draft_tokens += 1;
  }
 
  auto tdraft_end = std::chrono::high_resolution_clock::now();
- estate->metrics.UpdateBatchDraftTime(
+ estate->metrics.UpdateDraftTimeByBatchSize(
  num_rsentries, static_cast<double>((tdraft_end - tdraft_start).count()) / 1e9);
  }
  }

cpp/serve/engine_actions/eagle_batch_verify.cc

@@ -153,9 +153,8 @@ class EagleBatchVerifyActionObj : public EngineActionObj {
  rsentries[i]->mstates[verify_model_id_]->CommitToken(sample_result);
  rsentries[i]->mstates[draft_model_id_]->CommitToken(sample_result);
  }
- estate->metrics.UpdateSpecDecodingStats(cum_verify_lengths[i + 1] - cum_verify_lengths[i],
- accept_length);
- estate->metrics.sum_num_accepted_tokens += accept_length - 1;
+ estate->metrics.spec_decode.Update(cum_verify_lengths[i + 1] - cum_verify_lengths[i],
+ accept_length);
  // - Minus one because the last draft token has no kv cache entry
  // - Take max with 0 in case of all accepted.
  int rollback_length =
@@ -304,7 +303,7 @@ class EagleBatchVerifyActionObj : public EngineActionObj {
  auto tend = std::chrono::high_resolution_clock::now();
  double elapsed_time = static_cast<double>((tend - tstart).count()) / 1e9;
  estate->metrics.sum_engine_decode_time += elapsed_time;
- estate->metrics.UpdateBatchVerificationTime(cum_verify_lengths.back(), elapsed_time);
+ estate->metrics.UpdateVerifyTimeByBatchSize(cum_verify_lengths.back(), elapsed_time);
 
  return estate->running_queue;
  }
@@ -380,7 +379,6 @@ class EagleBatchVerifyActionObj : public EngineActionObj {
  }
  for (int i = 0; i < static_cast<int>(mstates.size()); ++i) {
  mstates[i]->AddDraftToken(sample_results[i], draft_token_slots_[i]);
- estate->metrics.sum_num_draft_tokens += 1;
  }
  }
  /*!

cpp/serve/engine_actions/eagle_new_request_prefill.cc

@@ -340,7 +340,6 @@ class EagleNewRequestPrefillActionObj : public BatchPrefillBaseActionObj {
  for (int i = 0; i < static_cast<int>(rsentries_for_sample.size()); ++i) {
  rsentries_for_sample[i]->mstates[model_id]->AddDraftToken(sample_results[i],
  draft_token_slots_[i]);
- estate->metrics.sum_num_draft_tokens += 1;
  }
  }
 

cpp/serve/metrics.cc

@@ -5,134 +5,134 @@
  */
 #include "metrics.h"
 
+#include <tvm/runtime/logging.h>
+
+#include <sstream>
+
 namespace mlc {
 namespace llm {
 namespace serve {
 
-picojson::value EngineMetrics::AsJSON() const {
- picojson::object metrics;
- metrics["sum_request_prefill_time"] = sum_request_prefill_time.AsJSON();
- metrics["sum_request_decode_time"] = sum_request_decode_time.AsJSON();
- metrics["sum_engine_prefill_time"] = sum_engine_prefill_time.AsJSON();
- metrics["sum_engine_decode_time"] = sum_engine_decode_time.AsJSON();
- metrics["sum_num_input_tokens"] = sum_num_input_tokens.AsJSON();
- metrics["sum_num_prefill_tokens"] = sum_num_prefill_tokens.AsJSON();
- metrics["sum_num_output_tokens"] = sum_num_output_tokens.AsJSON();
- metrics["sum_num_accepted_tokens"] = sum_num_accepted_tokens.AsJSON();
- metrics["sum_num_draft_tokens"] = sum_num_draft_tokens.AsJSON();
-
- metrics["last_finished_req_prefill_time"] = last_finished_req_prefill_time.AsJSON();
- metrics["last_finished_req_decode_time"] = last_finished_req_decode_time.AsJSON();
- metrics["num_last_finished_req_input_tokens"] = num_last_finished_req_input_tokens.AsJSON();
- metrics["num_last_finished_req_prefill_tokens"] = num_last_finished_req_prefill_tokens.AsJSON();
- metrics["num_last_finished_req_output_tokens"] = num_last_finished_req_output_tokens.AsJSON();
-
- picojson::array batch_decode_time_obj;
- picojson::array batch_draft_time_obj;
- picojson::array batch_verification_time_obj;
- batch_decode_time_obj.reserve(batch_decode_time_list.size());
- batch_draft_time_obj.reserve(batch_draft_time_list.size());
- batch_verification_time_obj.reserve(batch_verification_time_list.size());
- for (const Metric& batch_decode_time : batch_decode_time_list) {
- if (batch_decode_time.label.empty()) {
- continue;
- }
- batch_decode_time_obj.push_back(batch_decode_time.AsJSON());
+picojson::value TimeCost::AsJSON() const {
+ picojson::object config;
+ config["count"] = picojson::value(count);
+ if (count != 0) {
+ config["mean"] = picojson::value(sum / count);
  }
- for (const Metric& batch_draft_time : batch_draft_time_list) {
- if (batch_draft_time.label.empty()) {
- continue;
- }
- batch_draft_time_obj.push_back(batch_draft_time.AsJSON());
- }
- for (const Metric& batch_verification_time : batch_verification_time_list) {
- if (batch_verification_time.label.empty()) {
- continue;
- }
- batch_verification_time_obj.push_back(batch_verification_time.AsJSON());
- }
- metrics["batch_decode_time_per_batch_size"] = picojson::value(batch_decode_time_obj);
- metrics["batch_draft_time_per_batch_size"] = picojson::value(batch_draft_time_obj);
- metrics["batch_verification_time_per_batch_size"] = picojson::value(batch_verification_time_obj);
+ return picojson::value(config);
+}
 
+picojson::value SpecDecodeMetrics::AsJSON() const {
+ picojson::object metrics;
  auto f_vector_to_array = [](const std::vector<int64_t>& vec) {
  picojson::array arr;
  for (int64_t v : vec) {
  arr.push_back(picojson::value(v));
  }
  return picojson::value(arr);
  };
- metrics["accept_count"] = f_vector_to_array(accept_count);
  metrics["draft_count"] = f_vector_to_array(draft_count);
- return picojson::value(metrics);
-}
+ metrics["accept_count"] = f_vector_to_array(accept_count);
 
-void EngineMetrics::Reset() {
- sum_request_prefill_time.Reset(/*warmed_up=*/true);
- sum_request_decode_time.Reset(/*warmed_up=*/true);
- sum_engine_prefill_time.Reset(/*warmed_up=*/true);
- sum_engine_decode_time.Reset(/*warmed_up=*/true);
- sum_num_input_tokens.Reset(/*warmed_up=*/true);
- sum_num_prefill_tokens.Reset(/*warmed_up=*/true);
- sum_num_output_tokens.Reset(/*warmed_up=*/true);
- sum_num_accepted_tokens.Reset(/*warmed_up=*/true);
- sum_num_draft_tokens.Reset(/*warmed_up=*/true);
- last_finished_req_prefill_time.Reset(/*warmed_up=*/true);
- last_finished_req_decode_time.Reset(/*warmed_up=*/true);
- num_last_finished_req_input_tokens.Reset(/*warmed_up=*/true);
- num_last_finished_req_prefill_tokens.Reset(/*warmed_up=*/true);
- num_last_finished_req_output_tokens.Reset(/*warmed_up=*/true);
- batch_decode_time_list.clear();
- batch_draft_time_list.clear();
- batch_verification_time_list.clear();
- batch_decode_time_list.resize(kMaxEffectiveBatchSize);
- batch_draft_time_list.resize(kMaxEffectiveBatchSize);
- batch_verification_time_list.resize(kMaxEffectiveBatchSize);
- accept_count.clear();
- draft_count.clear();
-}
+ ICHECK_EQ(draft_count.size(), accept_count.size());
+ // NOTE: label follows prometheus with full context
+ // so it can be flattened and used in metrics reoorting end point
+ picojson::object accept_prob_metrics;
+ picojson::object accept_rate_metrics;
+ picojson::object accept_len_metrics;
 
-void EngineMetrics::UpdateBatchDecodeTime(int batch_size, double time) {
- if (batch_size > kMaxEffectiveBatchSize) {
- return;
- }
- if (batch_decode_time_list[batch_size].label.empty()) {
- batch_decode_time_list[batch_size].label = std::to_string(batch_size);
- }
- batch_decode_time_list[batch_size].Update(time);
-}
+ double accept_len_value = 0;
 
-void EngineMetrics::UpdateBatchDraftTime(int batch_size, double time) {
- if (batch_size > kMaxEffectiveBatchSize) {
- return;
- }
- if (batch_draft_time_list[batch_size].label.empty()) {
- batch_draft_time_list[batch_size].label = std::to_string(batch_size);
- }
- batch_draft_time_list[batch_size].Update(time);
-}
+ for (size_t i = 0; draft_count.size(); ++i) {
+ std::ostringstream accept_prob_label;
+ accept_prob_label << "accept_prob{step=" << i << "}";
+ double accept_prob_value =
+ (static_cast<double>(accept_count[i]) / static_cast<double>(draft_count[i]));
+ accept_prob_metrics[accept_prob_label.str()] = picojson::value(accept_prob_value);
+ accept_len_value += accept_prob_value;
 
-void EngineMetrics::UpdateBatchVerificationTime(int batch_size, double time) {
- if (batch_size > kMaxEffectiveBatchSize) {
- return;
- }
- if (batch_verification_time_list[batch_size].label.empty()) {
- batch_verification_time_list[batch_size].label = std::to_string(batch_size);
+ std::ostringstream accept_len_label;
+ accept_len_label << "accept_len{step=" << i << "}";
+ accept_len_metrics[accept_len_label.str()] = picojson::value(accept_len_value);
+
+ if (i != 0) {
+ std::ostringstream accept_rate_label;
+ accept_rate_label << "accept_rate{step=" << i << "}";
+ double accept_rate_value =
+ (static_cast<double>(accept_count[i]) / static_cast<double>(accept_count[i - 1]));
+ accept_rate_metrics[accept_rate_label.str()] = picojson::value(accept_rate_value);
+ }
  }
- batch_verification_time_list[batch_size].Update(time);
+ metrics["accept_prob"] = picojson::value(accept_prob_metrics);
+ metrics["accept_rate"] = picojson::value(accept_rate_metrics);
+ metrics["accept_len"] = picojson::value(accept_len_metrics);
+
+ return picojson::value(metrics);
 }
 
-void EngineMetrics::UpdateSpecDecodingStats(int draft_length, int accept_length) {
- if (accept_count.size() < draft_length) {
- this->accept_count.resize(draft_length, 0);
- this->draft_count.resize(draft_length, 0);
- }
- for (int j = 0; j < draft_length; ++j) {
- if (j < accept_length) {
- this->accept_count[j]++;
+picojson::value EngineMetrics::AsJSON() const {
+ picojson::object metrics;
+ metrics["sum_engine_prefill_time"] = picojson::value(sum_engine_prefill_time);
+ metrics["sum_engine_decode_time"] = picojson::value(sum_engine_decode_time);
+ metrics["sum_num_input_tokens"] = picojson::value(sum_num_input_tokens);
+ metrics["sum_num_prefill_tokens"] = picojson::value(sum_num_prefill_tokens);
+ metrics["sum_num_output_tokens"] = picojson::value(sum_num_output_tokens);
+
+ metrics["last_finished_req_prefill_time"] = picojson::value(last_finished_req_prefill_time);
+ metrics["last_finished_req_decode_time"] = picojson::value(last_finished_req_decode_time);
+ metrics["last_finished_req_num_input_tokens"] =
+ picojson::value(last_finished_req_num_input_tokens);
+ metrics["last_finished_req_num_prefill_tokens"] =
+ picojson::value(last_finished_req_num_prefill_tokens);
+ metrics["last_finished_req_num_output_tokens"] =
+ picojson::value(last_finished_req_num_output_tokens);
+
+ metrics["spec_decode"] = spec_decode.AsJSON();
+
+ auto f_create_time_list = [](const std::string& label_name,
+ const std::vector<TimeCost>& time_list) {
+ picojson::object result;
+ for (size_t i = 1; i < time_list.size(); ++i) {
+ const TimeCost& item = time_list[i];
+ if (item.count == 0) continue;
+ std::ostringstream label_mean;
+ label_mean << "mean_" << label_name << "{batch_size=" << i << "}";
+ double mean = item.sum / item.count;
+ result[label_mean.str()] = picojson::value(mean);
+ std::ostringstream label_count;
+ label_count << "count_" << label_name << "{batch_size=" << i << "}";
+ result[label_count.str()] = picojson::value(item.count);
  }
- this->draft_count[j]++;
- }
+ return picojson::value(result);
+ };
+
+ metrics["decode_time_by_batch_size"] =
+ f_create_time_list("decode_time", decode_time_by_batch_size);
+ metrics["draft_time_by_batch_size"] = f_create_time_list("draft_time", draft_time_by_batch_size);
+ metrics["verify_time_by_batch_size"] =
+ f_create_time_list("verify_time", verify_time_by_batch_size);
+
+ return picojson::value(metrics);
+}
+
+void EngineMetrics::Reset() {
+ sum_engine_prefill_time = 0.0;
+ sum_engine_decode_time = 0.0;
+ sum_num_input_tokens = 0;
+ sum_num_prefill_tokens = 0;
+ sum_num_output_tokens = 0;
+ last_finished_req_prefill_time = 0.0;
+ last_finished_req_decode_time = 0.0;
+ last_finished_req_num_input_tokens = 0.0;
+ last_finished_req_num_prefill_tokens = 0.0;
+ last_finished_req_num_output_tokens = 0.0;
+ spec_decode.Reset();
+ decode_time_by_batch_size.clear();
+ draft_time_by_batch_size.clear();
+ verify_time_by_batch_size.clear();
+ decode_time_by_batch_size.resize(kEndFineGrainedTrackingBatchSize);
+ draft_time_by_batch_size.resize(kEndFineGrainedTrackingBatchSize);
+ verify_time_by_batch_size.resize(kEndFineGrainedTrackingBatchSize);
 }
 
 } // namespace serve