Add ttfb to available metrics

Summary: Add ttfb to available metrics

Reviewed By: dshi7

Differential Revision: D50225322

fbshipit-source-id: e1814224332bcd2c9345653dceb325fca3571646

run.py

@@ -103,6 +103,8 @@ def printResultSummaryTime(result_summary, model, metrics_needed=[], flops_model
             flops = model.get_flops()
             tflops = flops / (cpu_walltime / 1.0e3) / 1.0e12
         print('{:<20} {:>20}'.format("GPU FLOPS:", "%.4f TFLOPs per second" % tflops, sep=''))
+    if 'ttfb' in metrics_needed:
+        print('{:<20} {:>20}'.format("Time to first batch:", "%.4f ms" % model.ttfb, sep=''))
     if model_flops is not None:
         tflops = model_flops / (cpu_walltime / 1.0e3) / 1.0e12
         print('{:<20} {:>20}'.format("Model Flops:", "%.4f TFLOPs per second" % tflops, sep=''))
@@ -356,8 +358,8 @@ def _validate_profile_options(profile_options: str):
     parser.add_argument(
         "--metrics",
         type=str,
-        default="cpu_peak_mem,gpu_peak_mem",
-        help="Specify metrics [cpu_peak_mem,gpu_peak_mem,flops,model_flops]to be collected. You can also set `none` to disable all metrics. The metrics are separated by comma such as cpu_peak_mem,gpu_peak_mem.",
+        default="cpu_peak_mem,gpu_peak_mem,ttfb",
+        help="Specify metrics [cpu_peak_mem,gpu_peak_mem,ttfb,flops,model_flops]to be collected. You can also set `none` to disable all metrics. The metrics are separated by comma such as cpu_peak_mem,gpu_peak_mem.",
     )
     parser.add_argument(
         "--metrics-gpu-backend",

torchbenchmark/util/experiment/metrics.py

@@ -20,6 +20,7 @@ class TorchBenchModelMetrics:
     throughputs: List[float]
     cpu_peak_mem: Optional[float]
     gpu_peak_mem: Optional[float]
+    ttfb: Optional[float] # time-to-first-batch
     pt2_compilation_time: Optional[float]
     pt2_graph_breaks: Optional[float]
     model_flops: Optional[float]
@@ -112,6 +113,7 @@ def get_model_test_metrics(model: Union[BenchmarkModel, ModelTask], metrics=[],
     throughputs = None
     cpu_peak_mem = None
     gpu_peak_mem = None
+    ttfb = None
     pt2_compilation_time = None
     pt2_graph_breaks = None
     model_flops = None
@@ -133,7 +135,10 @@ def get_model_test_metrics(model: Union[BenchmarkModel, ModelTask], metrics=[],
                            if isinstance(model, ModelTask) else model.pt2_graph_breaks
     if 'model_flops' in metrics:
         model_flops = get_model_flops(model)
-    return TorchBenchModelMetrics(latencies, throughputs, cpu_peak_mem, gpu_peak_mem, pt2_compilation_time, pt2_graph_breaks, model_flops)
+    if 'ttfb' in metrics:
+        ttfb = model.get_model_attribute('ttfb') \
+                if isinstance(model, ModelTask) else model.ttfb
+    return TorchBenchModelMetrics(latencies, throughputs, cpu_peak_mem, gpu_peak_mem, ttfb, pt2_compilation_time, pt2_graph_breaks, model_flops)
 
 def get_model_accuracy(model_config: TorchBenchModelConfig, isolated: bool=True) -> str:
     import copy

torchbenchmark/util/model.py

@@ -3,6 +3,7 @@
 from contextlib import contextmanager, ExitStack
 import warnings
 import yaml
+import time
 from pathlib import Path
 from typing import ContextManager, Optional, List, Tuple, Generator
 from torchbenchmark import REPO_PATH
@@ -79,6 +80,7 @@ class BenchmarkModel(metaclass=PostInitProcessor):
     See [Adding Models](#../models/ADDING_MODELS.md)
     """
     def __init__(self, test: str, device: str, batch_size: Optional[int]=None, extra_args: List[str]=[]):
+        self._start_init_time = time.time_ns()
         self.metadata = self._load_metadata()
         self.test = test
         # sanity checks of the options
@@ -149,6 +151,7 @@ def __post__init__(self):
         # Need to clean up the cache because we run deep copy within correceness check
         if self.device == "cuda":
             torch.cuda.empty_cache()
+        self._end_init_time = time.time_ns()
 
     def _skip_by_device_name(self):
         if not self.device == "cuda":
@@ -390,3 +393,8 @@ def pt2_graph_breaks(self):
         from torch._dynamo.utils import counters
         num_graph_breaks = len(counters["graph_break"].keys())
         return num_graph_breaks
+
+    @property
+    def ttfb(self):
+        """Return the time taken to the first batch in ms."""
+        return (self._end_init_time - self._start_init_time) / 1_000_000