Merge commit '1b27b93fa5a4efdfef285581df4524189da81bba'

Author: anmyachev

.github/workflows/wheels.yml

@@ -88,8 +88,8 @@ jobs:
             export CIBW_MANYLINUX_AARCH64_IMAGE="quay.io/pypa/manylinux_2_28_${{ matrix.config.arch }}:latest"
           fi
 
-          export CIBW_BUILD="cp3{9,10,11,12,13,13t,14,14t}-manylinux_${{ matrix.config.arch }}"
-          export CIBW_SKIP="cp{35,36,37,38}-*"
+          export CIBW_BUILD="cp3{10,11,12,13,13t,14,14t}-manylinux_${{ matrix.config.arch }}"
+          export CIBW_SKIP="cp{35,36,37,38,39}-*"
           export CIBW_ENABLE=cpython-freethreading
           python3 -m cibuildwheel . --output-dir wheelhouse
 

README.md

@@ -38,7 +38,7 @@ You can install the latest stable release of Triton from pip:
 pip install triton
 ```
 
-Binary wheels are available for CPython 3.10-3.13.
+Binary wheels are available for CPython 3.10-3.14.
 
 # Install from source
 

docs/getting-started/installation.rst

@@ -14,7 +14,7 @@ You can install the latest stable release of Triton from pip:
 
       pip install triton
 
-Binary wheels are available for CPython 3.10-3.13.
+Binary wheels are available for CPython 3.10-3.14.
 
 -----------
 From Source

python/test/gluon/test_frontend.py

@@ -2634,3 +2634,39 @@ def test_get_num_warps():
     print_num_warps()
     ttgl.warp_specialize((), print_num_warps, (), [print_num_warps, print_num_warps, print_num_warps], [1, 2, 8],
                          [24, 24, 24])
+
+
+def test_non_scalar_loop_bounds():
+
+    @gluon.jit
+    def kernel():
+        x = ttgl.full([32], 0, ttgl.int32, layout=ttgl.BlockedLayout([1], [32], [1], [0]))
+        for _ in range(x, 10, 1):
+            pass
+
+    with pytest.raises(CompilationError) as e:
+        run_parser(kernel)
+
+    assert "For lower bound must be a scalar, got" in str(e.value)
+
+    @gluon.jit
+    def kernel():
+        x = ttgl.full([32], 0, ttgl.int32, layout=ttgl.BlockedLayout([1], [32], [1], [0]))
+        for _ in range(1, x, 1):
+            pass
+
+    with pytest.raises(CompilationError) as e:
+        run_parser(kernel)
+
+    assert "For upper bound must be a scalar, got" in str(e.value)
+
+    @gluon.jit
+    def kernel():
+        x = ttgl.full([32], 0, ttgl.int32, layout=ttgl.BlockedLayout([1], [32], [1], [0]))
+        for _ in range(1, 10, x):
+            pass
+
+    with pytest.raises(CompilationError) as e:
+        run_parser(kernel)
+
+    assert "For step must be a scalar, got" in str(e.value)

python/triton/compiler/code_generator.py

@@ -1178,6 +1178,12 @@ def visit_For(self, node):
         # induction variable type
         if not lb.dtype.is_int() or not ub.dtype.is_int() or not step.dtype.is_int():
             raise TypeError(f"For loop bounds and step must all be ints, are ({lb.dtype}, {ub.dtype}, {step.dtype})")
+        if _is_non_scalar_tensor(lb):
+            raise TypeError(f"For lower bound must be a scalar, got {lb.type}")
+        if _is_non_scalar_tensor(ub):
+            raise TypeError(f"For upper bound must be a scalar, got {ub.type}")
+        if _is_non_scalar_tensor(step):
+            raise TypeError(f"For step must be a scalar, got {step.type}")
         iv_type = self.semantic.integer_promote_impl(lb.dtype, ub.dtype)
         iv_type = self.semantic.integer_promote_impl(iv_type, step.dtype)
         iv_ir_type = iv_type.to_ir(self.builder)

python/triton_kernels/triton_kernels/roofline.py

@@ -266,7 +266,7 @@ def interp(yxs, yys, x):
     xmin, xmax = xs[0], xs[-1]
     dx = 0.05 * (xmax - xmin) if xmax > xmin else 1.0
     ax.set_xlim(xmin - dx, xmax + dx)
-    ax.set_ylim(min(y_roof) * 0.8 if y_roof else 0.0, max_tflops * 1.05)
+    ax.set_ylim(min(min(perf) for perf in series_perf) * 0.8 if series_perf else 0.0, max_tflops * 1.05)
 
     # Points of interest
     if points_of_interest:

third_party/proton/csrc/lib/Context/Python.cpp

@@ -25,32 +25,15 @@ PyObject *_Py_XNewRef(PyObject *obj) {
 #define Py_XNewRef(obj) _Py_XNewRef((PyObject *)(obj))
 #endif
 
-// bpo-40421 added PyFrame_GetCode() to Python 3.9.0b1
-#if PY_VERSION_HEX < 0x030900B1
-PyCodeObject *getFrameCodeObject(PyFrameObject *frame) {
-  assert(frame != nullptr);
-  assert(frame->f_code != nullptr);
-  return (PyCodeObject *)(Py_NewRef(frame->f_code));
-}
-#else
 PyCodeObject *getFrameCodeObject(PyFrameObject *frame) {
   assert(frame != nullptr);
   return PyFrame_GetCode(frame);
 }
-#endif
 
-// bpo-40421 added PyFrame_GetBack() to Python 3.9.0b1
-#if PY_VERSION_HEX < 0x030900B1
-PyFrameObject *getFrameBack(PyFrameObject *frame) {
-  assert(frame != nullptr);
-  return (PyFrameObject *)(Py_XNewRef(frame->f_back));
-}
-#else
 PyFrameObject *getFrameBack(PyFrameObject *frame) {
   assert(frame != nullptr);
   return PyFrame_GetBack(frame);
 }
-#endif
 
 std::string unpackPyobject(PyObject *pyObject) {
   if (PyBytes_Check(pyObject)) {