diff --git a/run.py b/run.py
index 827bf2e496..4911835d77 100644
--- a/run.py
+++ b/run.py
@@ -101,9 +101,9 @@ def printResultSummaryTime(result_summary, metrics_needed=[], model=None, flops_
         if flops_model_analyzer.metrics_backend_mapping['flops'] == 'dcgm':
             tflops_device_id, tflops = flops_model_analyzer.calculate_flops()
         else:
-            flops, batch_size = model.get_flops()
-            tflops = flops * batch_size / (cpu_walltime / 1.0e3) / 1.0e12
-        print('{:<20} {:>20}'.format("GPU %d FLOPS:" % tflops_device_id, "%.4f TFLOPs per second" % tflops, sep=''))
+            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 gpu_peak_mem is not None:
         print('{:<20} {:>20}'.format("GPU %d Peak Memory:" % mem_device_id, "%.4f GB" % gpu_peak_mem, sep=''))
     if cpu_peak_mem is not None:
diff --git a/torchbenchmark/util/backends/flops.py b/torchbenchmark/util/backends/flops.py
deleted file mode 100644
index fa0a965b51..0000000000
--- a/torchbenchmark/util/backends/flops.py
+++ /dev/null
@@ -1,10 +0,0 @@
-
-# By default, FlopCountAnalysis count one fused-mult-add (FMA) as one flop.
-# However, in our context, we count 1 FMA as 2 flops instead of 1.
-# https://github.com/facebookresearch/fvcore/blob/7a0ef0c0839fa0f5e24d2ef7f5d48712f36e7cd7/fvcore/nn/flop_count.py
-def enable_fvcore_flops(model: 'torchbenchmark.util.model.BenchmarkModel', flops_fma=2.0):
-    assert hasattr(model, 'TORCHVISION_MODEL') and model.TORCHVISION_MODEL, "fvcore flops is only available on torchvision models!"
-    assert model.test == "eval", "fvcore flops is only available on inference tests, as it doesn't measure backward pass."
-    from fvcore.nn import FlopCountAnalysis
-    model.flops = FlopCountAnalysis(model.model, tuple(model.example_inputs)).total()
-    model.flops = model.flops / model.batch_size * flops_fma
diff --git a/torchbenchmark/util/extra_args.py b/torchbenchmark/util/extra_args.py
index f23bcfe647..d27ecfb1c0 100644
--- a/torchbenchmark/util/extra_args.py
+++ b/torchbenchmark/util/extra_args.py
@@ -2,9 +2,7 @@
 import enum
 from typing import List, Optional, Tuple
 from torchbenchmark.util.backends import list_backends, BACKENDS
-
-from torchbenchmark.util.backends.flops import enable_fvcore_flops
-from torchbenchmark.util.env_check import is_torchvision_model, is_staged_train_test
+from torchbenchmark.util.env_check import is_staged_train_test
 
 TEST_STAGE = enum.Enum('TEST_STAGE', ['FORWARD', 'BACKWARD', 'OPTIMIZER', 'ALL'])
 AVAILABLE_PRECISIONS = ["fp32", "tf32", "fp16", "amp", "fx_int8", "bf16","amp_fp16", "amp_bf16"]
@@ -127,7 +125,6 @@ def apply_decoration_args(model: 'torchbenchmark.util.model.BenchmarkModel', dar
 def parse_opt_args(model: 'torchbenchmark.util.model.BenchmarkModel', opt_args: List[str]) -> argparse.Namespace:
     parser = argparse.ArgumentParser()
     parser.add_argument("--backend", choices=list_backends(), help="enable backends")
-    parser.add_argument("--flops", choices=["fvcore", "dcgm"], help="Return the flops result")
     args, extra_args = parser.parse_known_args(opt_args)
     if model.jit:
         args.backend = "torchscript"
@@ -137,7 +134,5 @@ def parse_opt_args(model: 'torchbenchmark.util.model.BenchmarkModel', opt_args:
     return args, extra_args
 
 def apply_opt_args(model: 'torchbenchmark.util.model.BenchmarkModel', args: argparse.Namespace):
-    if args.flops == "fvcore":
-        enable_fvcore_flops(model)
     if args.backend:
         model._enable_backend()
diff --git a/torchbenchmark/util/framework/vision/model_factory.py b/torchbenchmark/util/framework/vision/model_factory.py
index 97f3220f79..25e5dfadf0 100644
--- a/torchbenchmark/util/framework/vision/model_factory.py
+++ b/torchbenchmark/util/framework/vision/model_factory.py
@@ -48,7 +48,15 @@ def __init__(self, model_name, test, device, jit=False, batch_size=None, weights
             self.real_output = ( torch.rand_like(self.example_outputs), )
 
     def get_flops(self):
-        return self.flops, self.batch_size
+        # By default, FlopCountAnalysis count one fused-mult-add (FMA) as one flop.
+        # However, in our context, we count 1 FMA as 2 flops instead of 1.
+        # https://github.com/facebookresearch/fvcore/blob/7a0ef0c0839fa0f5e24d2ef7f5d48712f36e7cd7/fvcore/nn/flop_count.py
+        assert self.test == "eval", "fvcore flops is only available on inference tests, as it doesn't measure backward pass."
+        from fvcore.nn import FlopCountAnalysis
+        FLOPS_FMA = 2.0
+        self.flops = FlopCountAnalysis(self.model, tuple(self.example_inputs)).total()
+        self.flops = self.flops * FLOPS_FMA
+        return self.flops
 
     def gen_inputs(self, num_batches:int=1) -> Tuple[Generator, Optional[int]]:
         def _gen_inputs():
@@ -96,4 +104,3 @@ def cudagraph_eval(self):
             self.g.replay()
             break
         return (self.example_outputs, )
-