[Bugfix][Blackwell] Avoid fmha_bwd workspace int32 overflow

[XFDG]

Summary

BlackwellFusedMultiHeadAttentionBackward._get_workspace_size combines user-shape parameters for workspace sizing with integer math that can be narrowed to int32 in the Cute/CUDA path for large sequence and batch dimensions. That can truncate the size and trigger allocation-related failures.

Fix

In
examples/python/CuTeDSL/cute/blackwell/kernel/attention/fmha/fmha_bwd.py function
BlackwellFusedMultiHeadAttentionBackward._get_workspace_size, explicitly use Python int intermediates for all multipliers and dimensions:

• round and cast q / d with int((q + 7) // 8 * 8)
• compute workspace bytes from b_i32, h_i32, q_i32, d_i32, and acc_bytes
• raise OverflowError if computed workspace bytes is negative

This keeps workspace arithmetic in Python big-int space and avoids unintentional int32 narrowing before CUDA allocation.

Testing

• python3 - <<'PY' seq_len = 46341 ws = int(seq_len) * int(seq_len) * 4 assert ws > 0 print('workspace size:', ws) PY

Fixes #2886