undefined

```bash

No layout flags — the read driver walks every `.npy` header to recover
per-file row counts, the trailing shape, and the dtype, then derives
`tile_rows` from the available processor count.

`--min-gpu-chunk 1` is only needed when the per-tile element count is
below Legate's default minimum chunk size for GPU launches (e.g. the
worked example's defaults — total rows split across 4 GPUs at
`~1M` per tile — fall below the threshold and would otherwise be
folded onto a single GPU). For production-sized datasets (tens of
millions of elements per tile or larger) you can drop the flag and
let Legate use its default. Bumping it to a moderate value (e.g.
`--min-gpu-chunk 1024`) is fine when each tile is large enough that
per-task overhead matters more than getting *every* GPU a tile.

paths = sorted(SHARD_DIR.glob("shard_*.npy"))

per_file_rows = []                       # rows along axis 0 per file
trailing_shape = None                    # shape[1:], must match across files
dtype = None
for p in paths:
    hdr = np.load(p, mmap_mode="r")
    if trailing_shape is None:
        trailing_shape = tuple(hdr.shape[1:])
        dtype = hdr.dtype
    elif tuple(hdr.shape[1:]) != trailing_shape or hdr.dtype != dtype:
        raise RuntimeError(
            f"{p.name}: trailing shape / dtype mismatch "
            f"({hdr.shape[1:]}/{hdr.dtype} vs {trailing_shape}/{dtype})"
        )
    per_file_rows.append(int(hdr.shape[0]))

cum_rows = np.cumsum([0] + per_file_rows, dtype=np.int64)  # length N+1
total_rows = int(cum_rows[-1])

import cupynumeric as cn

total_shape = (total_rows,) + trailing_shape
out = cn.empty(total_shape, dtype=dtype)

from legate.core import TaskTarget, get_legate_runtime
from legate.core.data_interface import as_logical_array

runtime = get_legate_runtime()
machine = runtime.get_machine()
num_processors = max(
    machine.count(TaskTarget.GPU),
    machine.count(TaskTarget.OMP),
    machine.count(TaskTarget.CPU),
    1,
)

tile_rows = max(1, (total_rows + num_processors - 1) // num_processors)
tile_shape = (tile_rows,) + trailing_shape
partition = as_logical_array(out).data.partition_by_tiling(tile_shape)

num_tasks = (total_rows + tile_rows - 1) // tile_rows  # match partition tile count

import bisect
from legate.core import TaskContext, VariantCode
from legate.core.task import OutputStore, task

@task(variants=(VariantCode.CPU, VariantCode.OMP, VariantCode.GPU))
def load_tile(ctx: TaskContext, dst: OutputStore) -> None:
    t = ctx.task_index[0]                              # tile index 0..num_tasks-1

    variant = ctx.get_variant_kind()
    if variant == VariantCode.GPU:
        import cupy as cp                              # lazy: only on GPU
        view = cp.from_dlpack(dst)
    else:
        view = np.asarray(dst)                         # zero-copy numpy view

    tile_rows_actual = view.shape[0]                   # short on the last tile
    row_start = t * TILE_ROWS                          # global axis-0 start
    row_end = row_start + tile_rows_actual

    # Find the half-open range of file indices that overlap [row_start, row_end).
    first_file = bisect.bisect_right(CUM_ROWS, row_start) - 1
    last_file = bisect.bisect_right(CUM_ROWS, row_end - 1) - 1

    for f in range(first_file, last_file + 1):
        # Intersection of tile [row_start, row_end) with file [cum[f], cum[f+1]).
        lo = max(row_start, int(CUM_ROWS[f]))
        hi = min(row_end, int(CUM_ROWS[f + 1]))
        file_lo = lo - int(CUM_ROWS[f])
        file_hi = hi - int(CUM_ROWS[f])
        dst_lo = lo - row_start
        dst_hi = hi - row_start
        chunk = np.ascontiguousarray(
            np.load(PATHS[f], mmap_mode="r")[file_lo:file_hi]
        )
        if variant == VariantCode.GPU:
            view[dst_lo:dst_hi].set(chunk)             # cudaMemcpyAsync H2D
        else:
            view[dst_lo:dst_hi] = chunk                # zero-copy numpy write

manual_task = runtime.create_manual_task(
    load_tile.library,
    load_tile.task_id,
    (num_tasks,),                                      # launch domain == tile count
)
manual_task.add_output(partition)
manual_task.execute()

get_legate_runtime().issue_execution_fence(block=True)

if all(r == per_file_rows[0] for r in per_file_rows) and num_shards == num_processors:
    tile_rows = per_file_rows[0]
else:
    tile_rows = max(1, (total_rows + num_processors - 1) // num_processors)

tile_rows = max(1, (total_rows + K * num_processors - 1) // (K * num_processors))

LEGION API USAGE EXCEPTION: Invalid task context passed to runtime call
create_index_space

tile_rows = max(1, (total_rows + num_processors - 1) // num_processors)
num_tasks = (total_rows + tile_rows - 1) // tile_rows
partition = ...partition_by_tiling((tile_rows,) + trailing_shape)
runtime.create_manual_task(load_tile.library, load_tile.task_id, (num_tasks,))