Ignore autotune runs failed with PTXAS error (#5017)

Author: htyu

python/triton/runtime/autotuner.py

@@ -7,7 +7,7 @@
 from typing import Dict
 
 from .jit import KernelInterface
-from .errors import OutOfResources
+from .errors import OutOfResources, PTXASError
 from .driver import driver
 
 
@@ -157,7 +157,7 @@ def kernel_call():
 
         try:
             return self.do_bench(kernel_call, quantiles=(0.5, 0.2, 0.8))
-        except (OutOfResources, CompileTimeAssertionFailure):
+        except (OutOfResources, CompileTimeAssertionFailure, PTXASError):
             return [float("inf"), float("inf"), float("inf")]
 
     def run(self, *args, **kwargs):

python/triton/runtime/errors.py

@@ -24,3 +24,13 @@ def __str__(self) -> str:
     def __reduce__(self):
         # this is necessary to make CompilationError picklable
         return (type(self), (self.required, self.limit, self.name))
+
+
+class PTXASError(TritonError):
+
+    def __init__(self, error_message: Optional[str] = None):
+        self.error_message = error_message
+
+    def __str__(self) -> str:
+        error_message = self.error_message or ""
+        return f"PTXAS error: {error_message}"

third_party/nvidia/backend/compiler.py

@@ -1,5 +1,6 @@
 from triton.backends.compiler import BaseBackend, GPUTarget
 from triton._C.libtriton import ir, passes, llvm, nvidia
+from triton.runtime.errors import PTXASError
 
 from dataclasses import dataclass
 import functools
@@ -361,9 +362,9 @@ def make_cubin(src, metadata, opt, capability):
                 else:
                     error = f'`ptxas` failed with error code {e.returncode}'
 
-                raise RuntimeError(f'{error}\n'
-                                   f'`ptxas` stderr:\n{log}\n'
-                                   f'Repro command: {" ".join(ptxas_cmd)}\n')
+                raise PTXASError(f"{error}\n"
+                                 f"`ptxas` stderr:\n{log}\n"
+                                 f'Repro command: {" ".join(ptxas_cmd)}\n')
 
             with open(fbin, 'rb') as f:
                 cubin = f.read()