Merge branch 'main' into dcu

examples/blocksparse_attention/heuristic.py

 import math
 
 
-def num_splits_heuristic(total_mblocks, num_SMs, num_n_blocks, num_m_blocks, size_one_kv_head,
-                         is_causal_or_local, max_splits):
+def num_splits_heuristic(total_mblocks, num_SMs, num_n_blocks, num_m_blocks, size_one_kv_head, is_causal_or_local, max_splits):
     """
     Determines the optimal number of splits for maximizing GPU occupancy while balancing memory efficiency.
 

examples/cast/example_triton_cast_to_fp8.py

@@ -128,9 +128,7 @@ def per_token_group_quant_fp8(
         Tuple[torch.Tensor, torch.Tensor]: The quantized tensor and the
         scaling factor for quantization.
     """
-    assert (x.shape[-1] %
-            group_size == 0), (f"the last dimension of `x` {x.shape[-1]} must be divisible "
-                               f"by `group_size` {group_size}")
+    assert x.shape[-1] % group_size == 0, f"the last dimension of `x` {x.shape[-1]} must be divisible by `group_size` {group_size}"
     assert x.stride(-1) == 1, "`x` groups must be contiguous"
 
     finfo = torch.finfo(dtype)