[cherry-pick][inductor][triton] Update HAS_WARP_SPEC to check triton.Config params. Update Triton Hash to top of release/3.4.x stack (pytorch#158646)

* [inductor][triton] Update HAS_WARP_SPEC to check triton.Config params. Update Triton Hash to top of release/3.4.x stack (pytorch#158459)

Update triton commit hash to `11ec6354315768a85da41032535e3b7b99c5f706`, which is the new release/3.4.x branch in triton-lang/triton.

Also, update HAS_WARP_SPEC handling: In triton 3.4, warp spec will have a different interface: num_consumer_groups will be determined automatically by the compiler. This breaks the current Inductor integration, so for now, update HAS_WARP_SPEC to check whether triton.Config takes num_consumer_groups and num_buffers_warp_spec as parameters.

Pull Request resolved: pytorch#158459
Approved by: https://github.com/atalman

* dont_upde_hash

* Revert "dont_upde_hash"

This reverts commit 5fffb12.

* fix_docker_builds

---------

Co-authored-by: David Berard <[email protected]>

Author: atalman

.ci/docker/ci_commit_pins/triton.txt

@@ -1 +1 @@
-ae848267bebc65c6181e8cc5e64a6357d2679260
+11ec6354315768a85da41032535e3b7b99c5f706

.ci/docker/common/install_conda.sh

@@ -54,6 +54,8 @@ if [ -n "$ANACONDA_PYTHON_VERSION" ]; then
     export SYSROOT_DEP="sysroot_linux-64=2.17"
   fi
 
+  # Please see: https://github.com/pytorch/pytorch/pull/158370#issuecomment-3084705725
+  export CONDA_PLUGINS_AUTO_ACCEPT_TOS="yes"
   # Install correct Python version
   # Also ensure sysroot is using a modern GLIBC to match system compilers
   as_jenkins conda create -n py_$ANACONDA_PYTHON_VERSION -y\

test/inductor/test_static_cuda_launcher.py

@@ -14,7 +14,6 @@
 from torch._inductor.test_case import TestCase
 from torch.testing._internal.common_utils import skipIfRocm
 from torch.testing._internal.triton_utils import requires_cuda
-from torch.torch_version import TorchVersion
 
 
 @requires_cuda
@@ -140,36 +139,6 @@ def signed_integers(
         launcher.run(1, 1, 1, stream, new_arg0, 50, 50, 50, 50)
         self.assertEqual(new_arg0, arg0)
 
-    # TODO: floats don't work properly, triton seems to think they're all tl.float32
-    # despite type annotations.
-    # There's also not really a good way for me to make a float16 in python...
-    @skipIfRocm
-    def test_floats(self):
-        @triton.jit
-        def floats(arg0, arg1: tl.float16, arg2: tl.float32, arg3: tl.float64):
-            x = tl.load(arg0)
-            y = arg1 + arg2 + arg3
-            tl.store(arg0, x + y)
-
-        arg0 = torch.zeros(1, dtype=torch.float64, device="cuda")
-
-        args = (arg0, 1.0, 1.0, 1.0)
-
-        compiled_kernel = floats[1,](*args)
-        launcher = self._make_launcher(compiled_kernel)
-        if TorchVersion(triton.__version__) >= TorchVersion("3.4.0"):
-            self.assertEqual(launcher.arg_tys, "Offd")
-        else:
-            self.assertEqual(launcher.arg_tys, "Offf")
-        # TODO this line fails on Triton 3.4.0 (https://github.com/triton-lang/triton/issues/6176)
-        # Add the check back when this is fixed in Triton
-        # self.assertEqual(arg0, torch.tensor([3.0], dtype=torch.float64, device="cuda"))
-        new_arg0 = torch.zeros(1, dtype=torch.float64, device="cuda")
-        device_interface = get_interface_for_device("cuda")
-        stream = device_interface.get_raw_stream(device_interface.current_device())
-        launcher.run(1, 1, 1, stream, new_arg0, 1.0, 1.0, 1.0)
-        self.assertEqual(new_arg0, arg0)
-
     @skipIfRocm
     def test_basic_1arg(self):
         @triton.jit

torch/_inductor/runtime/triton_compat.py

@@ -69,7 +69,18 @@ def GPUTarget(
         def _log2(x: Any) -> Any:
             raise NotImplementedError
 
-    HAS_WARP_SPEC = hasattr(tl, "async_task")
+    def _triton_config_has(param_name: str) -> bool:
+        if not hasattr(triton, "Config"):
+            return False
+        if not hasattr(triton.Config, "__init__"):
+            return False
+        return param_name in inspect.signature(triton.Config.__init__).parameters
+
+    HAS_WARP_SPEC = (
+        hasattr(tl, "async_task")
+        and _triton_config_has("num_consumer_groups")
+        and _triton_config_has("num_buffers_warp_spec")
+    )
 
     try:
         from triton import knobs