modify triton_scale_swizzle kernel to accept column major inputs

stack-info: PR: #2553, branch: danielvegamyhre/stack/9

Author: danielvegamyhre

torchao/prototype/mx_formats/kernels.py

@@ -1405,6 +1405,7 @@ def triton_scale_swizzle(
         scale_cols,
         output_ptr,
         input_row_stride,
+        input_col_stride,
         output_block_stride,
         BLOCK_ROWS: tl.constexpr,
         BLOCK_COLS: tl.constexpr,
@@ -1424,7 +1425,7 @@ def triton_scale_swizzle(
         mask = (global_rows < scale_rows) & (global_cols < scale_cols)
 
         input_scales = tl.load(
-            scale_ptr + global_rows * input_row_stride + global_cols,
+            scale_ptr + global_rows * input_row_stride + global_cols * input_col_stride,
             mask=mask,
             other=0.0,
         )
@@ -1464,7 +1465,6 @@ def triton_mx_block_rearrange(scale_tensor: torch.Tensor) -> torch.Tensor:
         assert scale_tensor.element_size() == 1, (
             "Expected element size to be 1 byte (8 bits)"
         )
-        assert scale_tensor.is_contiguous(), "Input tensor must be contiguous"
 
         rows, cols = scale_tensor.shape
 
@@ -1477,7 +1477,8 @@ def triton_mx_block_rearrange(scale_tensor: torch.Tensor) -> torch.Tensor:
         out = scale_tensor.new_empty((padded_rows, padded_cols))
 
         # Input stride (for row-major format)
-        input_row_stride = cols
+        input_row_stride = scale_tensor.stride()[0]
+        input_col_stride = scale_tensor.stride()[1]
 
         # We probably want handle multiple blocks per tile but for now keep it simple
         BLOCK_ROWS, BLOCK_COLS = 128, 4
@@ -1496,6 +1497,7 @@ def triton_mx_block_rearrange(scale_tensor: torch.Tensor) -> torch.Tensor:
             cols,
             out.view(torch.uint8),
             input_row_stride,
+            input_col_stride,
             output_block_stride,
             BLOCK_ROWS=BLOCK_ROWS,
             BLOCK_COLS=BLOCK_COLS,