Fix a few operators on AMD

Summary:
Need to make a few changes to enable more operators on AMD:
1. use `HAS_LIGER_KERNEL` to enable liger layer_norm on demand
2. fix a bug when running with `--isolate`. `--isolate` will run each operator in subprocess, to avoid interference between operators.
3. we need to explicitly add some dependencies on AMD

Reviewed By: adamomainz

Differential Revision: D65450899

fbshipit-source-id: 9cd1406123b0d2a5fbddf15f4a711e88bbd977d6

test/test_gpu/main.py

@@ -39,7 +39,7 @@ def check_ci_output(op):
 
     output = op.output
     output_impls = output.result[0][1].keys()
-    skiped_impls = op.tb_args.skip
+    skiped_impls = op.tb_args.skip if op.tb_args.skip else []
     ci_enabled_impls = [
         x for x in REGISTERED_BENCHMARKS[output.op_name].keys() if x not in skiped_impls
     ]

tritonbench/operators/layer_norm/operator.py

@@ -39,7 +39,7 @@ def inner(*args):
 
         return lambda: inner(*args)
 
-    @register_benchmark()
+    @register_benchmark(enabled=HAS_LIGER_KERNEL)
     def liger_layer_norm(self, *args):
         (x, w_shape, weight, bias, eps) = args
         return lambda: LigerLayerNormFunction.apply(x, weight, bias, eps)

tritonbench/utils/parser.py

@@ -190,6 +190,8 @@ def _find_param_loc(params, key: str) -> int:
 def _remove_params(params, loc):
     if loc == -1:
         return params
+    if params[loc + 1].startswith("--"):
+        return params[:loc] + params[loc + 1 :]
     return params[:loc] + params[loc + 2 :]