Update torchbench commit pin, add sam_fast benchmark (#121420)

After this, the sam_fast benchmark can now be run in the pytorch repo:
```
SEGMENT_ANYTHING_FAST_USE_FLASH_4=0 benchmarks/dynamo/torchbench.py --inference --amp --performance --backend=inductor --explain --only sam_fast
```

sam_fast is designed for inference only, with cuda and amp on. The code adds these restrictions to the benchmark.

Pull Request resolved: https://github.com/pytorch/pytorch/pull/121420
Approved by: https://github.com/oulgen, https://github.com/msaroufim

.github/ci_commit_pins/torchbench.txt

@@ -1 +1 @@
-1ef0a39e13872e4cf810c430ae4c87e46938f2ba
+d6015d42d9a1834bc7595c4bd6852562fb80b30b

benchmarks/dynamo/ci_expected_accuracy/aot_eager_torchbench_inference.csv

@@ -310,6 +310,10 @@ sam,pass,0
 
 
 
+sam_fast,pass,0
+
+
+
 shufflenet_v2_x1_0,pass,0
 
 

benchmarks/dynamo/ci_expected_accuracy/aot_inductor_torchbench_inference.csv

@@ -282,6 +282,10 @@ sam,fail_to_run,0
 
 
 
+sam_fast,fail_to_run,0
+
+
+
 shufflenet_v2_x1_0,pass,0
 
 

benchmarks/dynamo/ci_expected_accuracy/dynamo_eager_torchbench_inference.csv

@@ -310,6 +310,10 @@ sam,pass,0
 
 
 
+sam_fast,pass,0
+
+
+
 shufflenet_v2_x1_0,pass,0
 
 

benchmarks/dynamo/ci_expected_accuracy/inductor_torchbench_inference.csv

@@ -310,6 +310,10 @@ sam,pass,0
 
 
 
+sam_fast,pass,0
+
+
+
 shufflenet_v2_x1_0,pass,0
 
 

benchmarks/dynamo/common.py

@@ -3483,6 +3483,7 @@ def run(runner, args, original_dir=None):
             "Wav2Vec2ForCTC",
             "Wav2Vec2ForPreTraining",
             "sam",
+            "sam_fast",
             "resnet50_quantized_qat",
             "mobilenet_v2_quantized_qat",
         }:

benchmarks/dynamo/torchbench.py

@@ -229,6 +229,11 @@ class TorchBenchmarkRunner(BenchmarkRunner):
         if part:
             extra_args += ["--part", part]
 
+        # sam_fast only runs with amp
+        if model_name == "sam_fast":
+            self.args.amp = True
+            self.setup_amp()
+
         if model_name == "vision_maskrcnn" and is_training:
             # Output of vision_maskrcnn model is a list of bounding boxes,
             # sorted on the basis of their scores. This makes accuracy
@@ -259,7 +264,6 @@ class TorchBenchmarkRunner(BenchmarkRunner):
                 extra_args=extra_args,
             )
         model, example_inputs = benchmark.get_module()
-
         # Models that must be in train mode while training
         if is_training and (
             not use_eval_mode or model_name in self._config["only_training"]
@@ -269,7 +273,6 @@ class TorchBenchmarkRunner(BenchmarkRunner):
             model.eval()
         gc.collect()
         batch_size = benchmark.batch_size
-
         # Torchbench has quite different setup for yolov3, so directly passing
         # the right example_inputs
         if model_name == "yolov3":

benchmarks/dynamo/torchbench.yaml

@@ -94,6 +94,7 @@ slow:
 non_deterministic:
   # https://github.com/pytorch/pytorch/issues/98355
   - mobilenet_v3_large
+  - sam_fast
 
 
 dtype:
@@ -173,6 +174,8 @@ skip:
       # timeout
       - sam
       # model is CUDA only
+      - sam_fast
+      # model is CUDA only
       - llama_v2_7b_16h
       # flaky
       - stable_diffusion
@@ -197,6 +200,7 @@ skip:
       - llama
       - llama_v2_7b_16h
       - simple_gpt
+      - sam_fast
       # Model's DEFAULT_TRAIN_BSIZE is not implemented
       - cm3leon_generate
       - hf_T5_generate