Disable cuDNN benchmark on ROCm and add cudnn_benchmark_limit option

modules/devices.py

@@ -395,10 +395,10 @@ def set_cudnn_params():
             torch.use_deterministic_algorithms(opts.cudnn_deterministic)
             if opts.cudnn_deterministic:
                 os.environ.setdefault('CUBLAS_WORKSPACE_CONFIG', ':4096:8')
-            torch.backends.cudnn.benchmark = True
+            torch.backends.cudnn.benchmark = opts.cudnn_benchmark
             if opts.cudnn_benchmark:
                 log.debug('Torch cuDNN: enable benchmark')
-                torch.backends.cudnn.benchmark_limit = 0
+            torch.backends.cudnn.benchmark_limit = opts.cudnn_benchmark_limit
             torch.backends.cudnn.allow_tf32 = True
         except Exception as e:
             log.warning(f'Torch cudnn: {e}')

modules/shared.py

@@ -482,9 +482,10 @@ options_templates.update(options_section(('backends', "Backend Settings"), {
     "other_sep": OptionInfo("<h2>Torch Options</h2>", "", gr.HTML),
     "opt_channelslast": OptionInfo(False, "Channels last "),
     "cudnn_deterministic": OptionInfo(False, "Deterministic mode"),
-    "cudnn_benchmark": OptionInfo(False, "Full-depth cuDNN benchmark"),
     "diffusers_fuse_projections": OptionInfo(False, "Fused projections"),
     "torch_expandable_segments": OptionInfo(False, "Expandable segments"),
+    "cudnn_benchmark": OptionInfo(devices.backend != "rocm", "Full-depth cuDNN benchmark"),
+    "cudnn_benchmark_limit": OptionInfo(10, "cuDNN benchmark limit", gr.Slider, {"minimum": 0, "maximum": 100, "step": 1}),
     "torch_tunable_ops": OptionInfo("default", "Tunable ops", gr.Radio, {"choices": ["default", "true", "false"]}),
     "torch_tunable_limit": OptionInfo(30, "Tunable ops limit", gr.Slider, {"minimum": 1, "maximum": 100, "step": 1}),
     "cuda_mem_fraction": OptionInfo(0.0, "Memory limit", gr.Slider, {"minimum": 0, "maximum": 2.0, "step": 0.05}),