Apply mask to gather in gelu

[zxjml]

TileGym/src/tilegym/ops/cutile/activation/gelu.py

Line 125 in 98175fd

mask = ct.less(offsets, n_elements) # new var

  # Create offset tile
  offsets = ct.add(ct.arange(BLOCK_SIZE, dtype=ct.int32), block_start)  # new var
  mask = ct.less(offsets, n_elements)  # new var

  # Load input data
  x_tile = ct.gather(x, offsets)  # new var

[hannahli-nv]

Thanks for flagging this! We took a look and it turns out ct.gather and ct.scatter both default to check_bounds=True (see cuda.tile API), so the kernel was already safe from out-of-bounds access.

The mask variable was indeed a dead code (computed but never used), and the bounds-checking behavior was not obvious from reading the code. We've cleaned this up in #99 — removed the unused mask and made the bounds checking explicit with padding_value=0 on gather and check_bounds=True on scatter.