Use IEEE precision for conv_transpose2d fp32 dot

Author: goldenfox2025

src/flag_gems/ops/conv_transpose2d.py

@@ -429,7 +429,7 @@ def conv_transpose2d_forward_kernel(
 
         input_block = tl.load(curr_input_pointer, mask=input_mask, other=0.0)
         weight_block = tl.load(curr_weight_pointer, mask=weight_mask, other=0.0)
-        accum += tl.dot(input_block, weight_block, allow_tf32=False)
+        accum += tl.dot(input_block, weight_block, input_precision="ieee")
 
     bias_pointer += (pid_group[None] * out_per_group_c)[None, :] + output_c_offset[
         None, :
@@ -571,7 +571,9 @@ def _conv_transpose2d_forward_residue_body(
                         weight_block = tl.load(
                             curr_weight_pointer, mask=weight_mask, other=0.0
                         )
-                        accum += tl.dot(input_block, weight_block, allow_tf32=False)
+                        accum += tl.dot(
+                            input_block, weight_block, input_precision="ieee"
+                        )
 
     bias_pointer += (pid_group[None] * out_per_group_c)[None, :] + output_c_offset[
         None, :
@@ -1088,7 +1090,7 @@ def conv_transpose2d_forward_scatter_kernel(
 
         input_block = tl.load(curr_input_pointer, mask=input_mask, other=0.0)
         weight_block = tl.load(curr_weight_pointer, mask=weight_mask, other=0.0)
-        accum += tl.dot(input_block, weight_block, allow_tf32=False)
+        accum += tl.dot(input_block, weight_block, input_precision="ieee")
 
     bias_pointer += (pid_group[None] * out_per_group_c)[None, :] + output_c_offset[
         None, :