Skip slow tests for aarch64-inductor-benchmarks (#158842)

This PR suggests adding some models to `cpu_skip_list` which are currently being run in TIMM and Torchbench.
The suggested models takes a long time which leads to the benchmark runs being `timeout`.  [benchmark runs for aarch64](https://github.com/pytorch/pytorch/actions/workflows/inductor-perf-test-nightly-aarch64.yml)

•	The issue stems from unoptimized groupwise convolution (BF16 /F16 dtype) kernels for aarch64 platforms  , which significantly slow down execution leading to the timeout.
**Action:**
•	An optimized BF16 groupwise convolution kernel is currently being developed in oneDNN, targeted for release in Q4 2025.

To maintain dashboard consistency and signal clarity, I’ve skipped the affected tests in:
      * timm benchmarks
      * torchbench benchmarks

 As suggested, skip is applied at the CPU - arch level, explicitly branching for aarch64 and adding models which needs to be skipped. This keeps the logic clean, but:
•	An alternative considered was increasing shard counts for aarch64 runners, but given the known performance bottleneck, skipping avoids wasted compute cycles. Suggestions around this will be appreciated.

Benchmark does not timeout after the suggested change: https://github.com/pytorch/pytorch/actions/runs/16447200138

Pull Request resolved: https://github.com/pytorch/pytorch/pull/158842
Approved by: https://github.com/malfet

benchmarks/dynamo/common.py

@@ -14,6 +14,7 @@ import itertools
 import json
 import logging
 import os
+import platform
 import random
 import shutil
 import signal
@@ -1763,6 +1764,10 @@ class BenchmarkRunner:
     def skip_models_for_cpu(self):
         return set()
 
+    @property
+    def skip_models_for_cpu_aarch64(self):
+        return set()
+
     @property
     def skip_models_for_freezing_cpu(self):
         return set()
@@ -3715,7 +3720,10 @@ def run(runner, args, original_dir=None):
         runner.skip_models.update(runner.slow_models)
 
     if args.devices == ["cpu"]:
+        arch = platform.machine()
         runner.skip_models.update(runner.skip_models_for_cpu)
+        if arch == "aarch64":
+            runner.skip_models.update(runner.skip_models_for_cpu_aarch64)
     elif args.devices == ["cuda"]:
         runner.skip_models.update(runner.skip_models_for_cuda)
 

benchmarks/dynamo/runner.py

@@ -32,6 +32,7 @@ import io
 import itertools
 import logging
 import os
+import platform
 import re
 import shutil
 import subprocess
@@ -374,6 +375,7 @@ def get_skip_tests(suite, device, is_training: bool):
     original_dir = abspath(os.getcwd())
     module = importlib.import_module(suite)
     os.chdir(original_dir)
+    arch = platform.machine()
 
     if suite == "torchbench":
         skip_tests.update(module.TorchBenchmarkRunner().skip_models)
@@ -383,6 +385,10 @@ def get_skip_tests(suite, device, is_training: bool):
             )
         if device == "cpu":
             skip_tests.update(module.TorchBenchmarkRunner().skip_models_for_cpu)
+            if arch == "aarch64":
+                skip_tests.update(
+                    module.TorchBenchmarkRunner().skip_models_for_cpu_aarch64
+                )
         elif device == "cuda":
             skip_tests.update(module.TorchBenchmarkRunner().skip_models_for_cuda)
 

benchmarks/dynamo/timm_models.py

@@ -230,6 +230,14 @@ class TimmRunner(BenchmarkRunner):
     def _skip(self):
         return self._config["skip"]
 
+    @property
+    def skip_models_for_cpu(self):
+        return self._skip["device"]["cpu"]
+
+    @property
+    def skip_models_for_cpu_aarch64(self):
+        return self._skip["device"]["cpu_aarch64"]
+
     @property
     def skip_models(self):
         return self._skip["all"]

benchmarks/dynamo/timm_models.yaml

@@ -2,3 +2,14 @@
 skip:
   all:
     - ~
+  device:
+    cpu:
+      - ~
+    # Skip these additional models when running on aarch64
+    cpu_aarch64:
+      # timeout on aarch64
+      - dm_nfnet_f0
+      - nfnet_l0
+      - resnest101e
+      - swsl_resnext101_32x16d
+      - visformer_small

benchmarks/dynamo/torchbench.py

@@ -138,6 +138,10 @@ class TorchBenchmarkRunner(BenchmarkRunner):
     def skip_models_for_cpu(self):
         return self._skip["device"]["cpu"]
 
+    @property
+    def skip_models_for_cpu_aarch64(self):
+        return self._skip["device"]["cpu_aarch64"]
+
     @property
     def skip_models_for_cuda(self):
         return self._skip["device"]["cuda"]

benchmarks/dynamo/torchbench.yaml

@@ -213,6 +213,12 @@ skip:
       - llava
       - moco
 
+    # Skip these additional models when running on aarch64
+    cpu_aarch64:
+      # timeout on aarch64
+      - timm_regnet
+      - timm_nfnet
+
     cuda: []
 
   test: