fix(ninetoothed): bound RmsNorm autotune shapes

src/ninetoothed/ops/rms_norm/build.py

@@ -1,3 +1,4 @@
+import functools
 import itertools
 
 import ninetoothed
@@ -9,6 +10,9 @@
 
 _DEFAULT_NDIMS = (2, 3, 4)
 
+_BATCH_DIM_AUTO_TUNE_SIZE = 1
+_NORMALIZED_DIM_AUTO_TUNE_SIZE = 256
+
 _CONFIGS = tuple(
     (
         (),
@@ -30,11 +34,54 @@
 )
 
 
+def _shape_options(ndim, num_normalized_dims):
+    batch_dims = ndim - num_normalized_dims
+
+    return (
+        *({"upper_bound": _BATCH_DIM_AUTO_TUNE_SIZE} for _ in range(batch_dims)),
+        *(
+            {"upper_bound": _NORMALIZED_DIM_AUTO_TUNE_SIZE}
+            for _ in range(num_normalized_dims)
+        ),
+    )
+
+
+def _premake(
+    ndim,
+    num_normalized_dims,
+    input_dtype=None,
+    weight_dtype=None,
+    output_dtype=None,
+    block_size=None,
+):
+    dims = tuple(-(dim + 1) for dim in range(num_normalized_dims))
+    arrangement = functools.partial(
+        ntops.kernels.reduction.arrangement,
+        dim=dims,
+        block_size=block_size,
+    )
+    shape_options = _shape_options(ndim, num_normalized_dims)
+    tensors = (
+        ninetoothed.Tensor(
+            ndim,
+            other=0,
+            dtype=input_dtype,
+            shape_options=shape_options,
+        ),
+        ninetoothed.Tensor(ndim, dtype=weight_dtype, shape_options=shape_options),
+        ninetoothed.Tensor(0, dtype=ninetoothed.float32),
+        ninetoothed.Tensor(ndim, dtype=output_dtype, shape_options=shape_options),
+        ninetoothed.Tensor(0, dtype=ninetoothed.uint64),
+    )
+
+    return arrangement, ntops.kernels.rms_norm.application, tensors
+
+
 def build(output_dir):
     variant_dir = output_dir / "rms_norm"
     variant_dir.mkdir(parents=True, exist_ok=True)
     ninetoothed.build(
-        ntops.kernels.rms_norm.premake,
+        _premake,
         _CONFIGS,
         meta_parameters=("block_size",),
         caller="cuda",