[COMPAT] Require Triton 3.8

pyproject.toml

@@ -22,7 +22,7 @@ classifiers = [
 ]
 dependencies = [
   "setuptools",
-  "triton>=3.6.0",
+  "triton>=3.8.0",
   "pyarrow",
   "pre-commit",
   "z3-solver==4.15.3.0",

tests/end_to_end/test_gluon.py

@@ -6,31 +6,23 @@
 from triton import knobs
 from triton.experimental import gluon
 from triton.experimental.gluon import language as gl
+from triton.experimental.gluon.language.amd.cdna4 import async_copy as amd_cdna4_cp
 from triton.experimental.gluon.language.nvidia import blackwell, hopper
 from triton.experimental.gluon.language.nvidia.blackwell import tma as blackwell_tma
 from triton.experimental.gluon.language.nvidia.hopper import (
     mbarrier,
     tma,
 )
-from triton.experimental.gluon.nvidia.hopper import TensorDescriptor
+from triton.experimental.gluon.nvidia.hopper import (
+    TensorDescriptor,
+    TensorDescriptorIm2Col,
+)
 
 import tilelens
 from tilelens.clients.sanitizer.sanitizer import SymbolicSanitizer
 from tilelens.core.data import Load
 from tilelens.core.simulation.gluon import GluonInterpretedFunction, gluon_builder
 
-try:
-    from triton.experimental.gluon.language.amd.cdna4 import (
-        async_copy as amd_cdna4_cp,
-    )
-except ImportError:
-    amd_cdna4_cp = None
-
-try:
-    from triton.experimental.gluon.nvidia.hopper import TensorDescriptorIm2Col
-except ImportError:
-    TensorDescriptorIm2Col = None
-
 try:
     from triton.experimental.gluon.language.nvidia.ampere import async_copy as cp
 except ImportError:
@@ -39,11 +31,6 @@
 _HAS_AMPERE_ASYNC_COPY = (
     cp is not None and getattr(cp, "async_copy_global_to_local", None) is not None
 )
-_HAS_AMD_CDNA4_ASYNC_COPY = (
-    amd_cdna4_cp is not None
-    and getattr(amd_cdna4_cp, "global_load_to_shared", None) is not None
-)
-_HAS_TMA_IM2COL = TensorDescriptorIm2Col is not None
 
 
 def _run_gluon_on_cpu(fn, grid, *args, **kwargs):
@@ -942,10 +929,6 @@ def test_gluon_async_copy_runs_masked_1d_copy_on_cpu():
     torch.testing.assert_close(out, inp, atol=0, rtol=0)
 
 
-@pytest.mark.skipif(
-    not _HAS_AMD_CDNA4_ASYNC_COPY,
-    reason="Gluon AMD CDNA4 async copy builtins are unavailable",
-)
 def test_gluon_amd_async_copy_preserves_masked_other_on_cpu():
     inp = torch.arange(40, dtype=torch.float32)
     out = torch.full((64,), -1, dtype=torch.float32)
@@ -1058,10 +1041,6 @@ def test_gluon_tma_runs_float_atomics_on_cpu():
     torch.testing.assert_close(max_dst, expected_max, atol=0, rtol=0)
 
 
-@pytest.mark.skipif(
-    not _HAS_TMA_IM2COL,
-    reason="Gluon TensorDescriptorIm2Col is unavailable in this Triton build",
-)
 def test_gluon_tma_im2col_runs_simple_tile_on_cpu():
     inp = torch.arange(1, 17, dtype=torch.float32).unsqueeze(1).repeat(1, 32)
     inp = inp.reshape(1, 4, 4, 32)
@@ -1071,10 +1050,6 @@ def test_gluon_tma_im2col_runs_simple_tile_on_cpu():
     torch.testing.assert_close(out, inp.reshape(16, 32), atol=0, rtol=0)
 
 
-@pytest.mark.skipif(
-    not _HAS_TMA_IM2COL,
-    reason="Gluon TensorDescriptorIm2Col is unavailable in this Triton build",
-)
 def test_gluon_tma_im2col_zero_fills_padded_pixels_on_cpu():
     inp = torch.arange(1, 17, dtype=torch.float32).unsqueeze(1).repeat(1, 32)
     inp = inp.reshape(1, 4, 4, 32)
@@ -1088,10 +1063,6 @@ def test_gluon_tma_im2col_zero_fills_padded_pixels_on_cpu():
     torch.testing.assert_close(out[:, 0], expected_first_channel, atol=0, rtol=0)
 
 
-@pytest.mark.skipif(
-    not _HAS_TMA_IM2COL,
-    reason="Gluon TensorDescriptorIm2Col is unavailable in this Triton build",
-)
 def test_gluon_tma_im2col_honors_runtime_offsets_on_cpu():
     inp = torch.arange(1, 17, dtype=torch.float32).unsqueeze(1).repeat(1, 32)
     inp = inp.reshape(1, 4, 4, 32)
@@ -1125,8 +1096,6 @@ def test_gluon_builder_preserves_tensor_memory_fp4_padding():
         False,
         True,
     )
-    if not hasattr(layout, "fp4_padded"):
-        pytest.skip("Gluon TensorMemoryLayout has no fp4_padded field")
     assert layout.fp4_padded is True
 
 

tests/unit/test_patch_scope.py

@@ -1,4 +1,3 @@
-import pytest
 import warnings
 from types import SimpleNamespace
 
@@ -85,13 +84,8 @@ def test_scope_removes_tensor_magic_methods_added_by_interpreter():
 
 def test_scope_restores_tensor_descriptor_base_builtins(monkeypatch):
     """Scope must restore descriptor builtins on tensor_descriptor_base."""
-    if not hasattr(tl.core, "tensor_descriptor_base"):
-        pytest.skip("tensor_descriptor_base is not available")
-
     descriptor = tl.core.tensor_descriptor_base
     attr = "load"
-    if not hasattr(descriptor, attr):
-        pytest.skip("descriptor load is not available")
 
     original = getattr(descriptor, attr)
     sentinel = object()

tilelens/core/frontend/gluon.py

@@ -9,11 +9,18 @@
 from triton.experimental.gluon.language import _semantic as gluon_semantic  # type: ignore
 from triton.experimental.gluon.language.amd import cdna3 as gluon_amd_cdna3  # type: ignore
 from triton.experimental.gluon.language.amd import cdna4 as gluon_amd_cdna4  # type: ignore
+from triton.experimental.gluon.language.amd import gfx1250 as gluon_amd_gfx1250  # type: ignore
 from triton.experimental.gluon.language.amd import rdna3 as gluon_amd_rdna3  # type: ignore
 from triton.experimental.gluon.language.amd import rdna4 as gluon_amd_rdna4  # type: ignore
 from triton.experimental.gluon.language.amd.cdna4 import (  # type: ignore
     async_copy as gluon_amd_cdna4_async_copy,
 )
+from triton.experimental.gluon.language.amd.gfx1250 import (  # type: ignore
+    async_copy as gluon_amd_async_copy,
+)
+from triton.experimental.gluon.language.amd.gfx1250 import (  # type: ignore
+    tdm as gluon_amd_tdm,
+)
 from triton.experimental.gluon.language.nvidia.blackwell import (  # type: ignore
     tma as gluon_blackwell_tma,
 )
@@ -52,23 +59,6 @@
 from .base import AdapterResult, Frontend, _LangPatchScope, register_frontend
 from .triton import TritonFrontend
 
-try:
-    from triton.experimental.gluon.language.amd import gfx1250 as gluon_amd_gfx1250  # type: ignore
-    from triton.experimental.gluon.language.amd.gfx1250 import (  # type: ignore
-        async_copy as gluon_amd_async_copy,
-    )
-    from triton.experimental.gluon.language.amd.gfx1250 import (  # type: ignore
-        tdm as gluon_amd_tdm,
-    )
-except ImportError as exc:
-    if "is_hip_gfx1250" not in str(exc) or "triton.language.target_info" not in str(
-        exc
-    ):
-        raise
-    gluon_amd_gfx1250 = None
-    gluon_amd_async_copy = None
-    gluon_amd_tdm = None
-
 _WARP_SPECIALIZE_SCHEDULER: Any = None
 _MISSING = object()
 
@@ -309,8 +299,6 @@ def _gluon_make_range_adapter(start: Any, end: Any, *_args: Any, **_kwargs: Any)
 
 
 def _existing_ops(namespace: Any, attrs: dict[str, type[Op]]) -> dict[str, type[Op]]:
-    if namespace is None:
-        return {}
     namespace_attrs = vars(namespace)
     return {attr: op_type for attr, op_type in attrs.items() if attr in namespace_attrs}
 

tilelens/core/frontend/triton.py

@@ -422,16 +422,13 @@ def _triton_language_attr_targets() -> list[tuple[Any, tuple[str, ...]]]:
             targets.append((lang.dtype, ("to_ir",)))
             if lang == tl:
                 targets.append((lang.math, ()))
-        if hasattr(tl.core, "tensor_descriptor_base"):
-            targets.append((tl.core.tensor_descriptor_base, ()))
+        targets.append((tl.core.tensor_descriptor_base, ()))
         return targets
 
     def _triton_snapshot_scope(self, fn: Callable[..., Any]) -> _LangPatchScope:
         """
         Stores Triton attributes into a LangPatchScope for later unpatching.
         This is to be run before patching with the interpreter.
-        This is equivalent to what triton>=3.6.0 does natively
-        but also works for triton<3.6.0.
         """
 
         scope = _LangPatchScope()
@@ -905,8 +902,8 @@ def unpatch_for_loop(self) -> None:
 
     def patch_lang(self, fn, client_manager=None) -> _LangPatchScope:
         # Snapshot before calling Triton's patcher because Triton mutates many
-        # attributes in-place and older Triton versions do not retain enough
-        # restore metadata for nested/generated kernels.
+        # attributes in-place but only records the language modules visible
+        # from `fn`; nested/generated kernels without a `tl` global need both.
         scope = self._triton_snapshot_scope(fn)
         triton_patch_lang(fn)
         for module in self._triton_extra_modules: