[user triton] Ignore backend-specific args in the TTIR analysis (pytorch#141062)

Fixes pytorch#140800.

On AMD, backend-specific args like `matrix_instr_nonkdim`, `waves_per_eu` and `kpack` are passed either direclty to the kernel or via `triton.Config`, whereas they don't exist as kernel parameters. Native Triton code handles those excessive args [here](https://github.com/triton-lang/triton/blob/a6bb57d6285e723c58e87dd7cba263db6efff789/python/triton/runtime/jit.py#L594-L596). In this PR, we add similar handling to the TTIR analysis code to avoid bailing out.
Pull Request resolved: pytorch#141062
Approved by: https://github.com/oulgen

Author: aakhundov

torch/_higher_order_ops/triton_kernel_wrap.py

@@ -196,8 +196,23 @@ def generate_ttir(
 
     assert isinstance(kernel, JITFunction)
 
+    context = triton._C.libtriton.ir.context()
+    target = triton.runtime.driver.active.get_current_target()
+    backend = triton.compiler.compiler.make_backend(target)
+    options = backend.parse_options({})
+
+    # ignore backend-specific kwargs same way as in the native Triton code
+    # https://github.com/triton-lang/triton/blob/a6bb57d6285e723c58e87dd7cba263db6efff789/python/triton/runtime/jit.py#L594-L596
+    # why this is important for user-defined Triton kernels on AMD: https://github.com/pytorch/pytorch/issues/140800
+    for name in list(kwargs):
+        if name not in kernel.arg_names and name in options.__dict__:
+            kwargs.pop(name)
+
     if len(kwargs) != len(kernel.arg_names):
-        raise ValueError("Incorrect number of arguments passed to kernel")
+        raise ValueError(
+            "Incorrect number of arguments passed to kernel: "
+            f"passed {list(kwargs.keys())}, expected {kernel.arg_names}."
+        )
 
     # Replace all SymExprs with a regular value for TTIR generation
     # Replace all FakeTensor/TensorBox with real tensors
@@ -239,10 +254,6 @@ def _get_specialization(args):  # type: ignore[no-untyped-def]
         if i not in kernel.constexprs
     }
 
-    context = triton._C.libtriton.ir.context()
-    target = triton.runtime.driver.active.get_current_target()
-    backend = triton.compiler.compiler.make_backend(target)
-    options = backend.parse_options({})
     triton._C.libtriton.ir.load_dialects(context)
     backend.load_dialects(context)