disable some models for torchbench

benchmarks/experiment_runner.py

@@ -113,7 +113,8 @@ def generate_and_run_all_configs(self):
 
         # Skip unsupported config.
         if not self.model_loader.is_compatible(benchmark_model,
-                                               benchmark_experiment):
+                                               benchmark_experiment,
+                                               self._args.strict_compatible):
           logger.warning("SKIP incompatible model and experiment configs.")
           self._save_results(benchmark_experiment.to_dict(),
                              benchmark_model.to_dict(), {"error": "SKIP"})
@@ -841,7 +842,6 @@ def __str__(self):
   parser.add_argument(
       "--disable-tf32",
       action="store_true",
-      default=False,
       help="Whether to enable fast F32 multiplication in PyTorch.",
   )
   parser.add_argument(
@@ -864,21 +864,23 @@ def __str__(self):
   parser.add_argument(
       "--pure-wall-time",
       action="store_true",
-      default=False,
       help="Times wall time measurements with pure CUDA events. No kernel launch overhead.",
   )
   parser.add_argument(
       "--filter-by-single-graph",
       action="store_true",
-      default=False,
       help="Runs the experiment with hard-failing when it detects there will be multiple graphs out of a single compiled region.",
   )
   parser.add_argument(
       "--verify",
       action="store_true",
-      default=False,
       help="""If set, verifies the model output with PT Eager mode, and saves relative error to the output file."""
   )
+  parser.add_argument(
+      "--strict-compatible",
+      action="store_true",
+      help="Strictly skips some models including models without installation file or causing stackdump.",
+  )
   return parser.parse_args(args)
 
 

benchmarks/torchbench_model.py

@@ -56,13 +56,28 @@
     "timm_vovnet",
     "vgg16",
     "hf_T5",
-    # PyTorch/benchmark sets its optimizer as SGD.
-    # Otherwise, OOMs.
-    "llama_v2_7b_16h",
 }
 
 # Skip the experiment of a model if any of the experiment configs in the list is fully matched
 DENY_LIST = {
+    "cm3leon_generate": [
+        {
+            "test": "train",
+        },
+        {
+            "test": "eval",
+            "xla": "PJRT",
+        },
+    ],  # no install.py
+    "hf_T5_generate": [
+        {
+            "test": "train",
+        },
+        {
+            "test": "eval",
+            "xla": "PJRT",
+        },
+    ],  # no install.py
     "doctr_det_predictor": [{
         "test": "train"
     },],  # not implemented
@@ -127,6 +142,25 @@
     "vision_maskrcnn": [{}],
 }
 
+# This strict deny list denies tests that hold for too long and timeoout.
+STRICT_DENY_LIST = {
+    **{
+        "opacus_cifar10": [{
+            "accelerator": "tpu",
+        },],  # stackdump issue in TPU
+        "pytorch_stargan": [{
+            "accelerator": "tpu",
+        },],  # stackdump issue in TPU
+        "soft_actor_critic": [{
+            "accelerator": "tpu",
+        },],  # stackdump issue in TPU
+        "speech_transformer": [{
+            "accelerator": "tpu",
+        },],  # stackdump issue in TPU
+    },
+    **DENY_LIST
+}
+
 
 class TorchBenchModelLoader(ModelLoader):
 
@@ -179,9 +213,13 @@ def list_model_configs(self):
 
     return model_configs
 
-  def is_compatible(self, dummy_benchmark_model, benchmark_experiment):
-    if dummy_benchmark_model.model_name in DENY_LIST:
-      for deny_experiment_config in DENY_LIST[dummy_benchmark_model.model_name]:
+  def is_compatible(self,
+                    dummy_benchmark_model,
+                    benchmark_experiment,
+                    use_strict_deny=False):
+    deny_list = STRICT_DENY_LIST if use_strict_deny else DENY_LIST
+    if dummy_benchmark_model.model_name in deny_list:
+      for deny_experiment_config in deny_list[dummy_benchmark_model.model_name]:
         matched = True
         for k, v in deny_experiment_config.items():
           if getattr(benchmark_experiment, k) != v: