[KV Offload] Split the CPU offload region across shm files (#59546)

Signed-off-by: Itay Etelis <92247226+Etelis@users.noreply.github.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
Co-authored-by: Kevin H. Luu <khluu000@gmail.com>
This commit is contained in:
Itay Etelis
2026-10-15 20:17:40 +00:00
committed by GitHub
co-authored by Robert Shaw Codex Kevin H. Luu
parent 16532876e7
commit 21f4b6dc84
3 changed files with 116 additions and 9 deletions
@@ -653,6 +653,41 @@ def test_madvise_unexpected_oserror_propagates(iid, monkeypatch):
assert exc_info.value.errno == errno.EIO
def test_segmented_region_maps_files_back_to_back(iid, monkeypatch):
"""Rows split over three shm files mapped back to back: each worker sees
the other's writes in every file, and the creator removes every file."""
monkeypatch.setattr(region_module, "_MIN_SEGMENT_BYTES", PAGE_SIZE)
regions = [
SharedOffloadRegion(
engine_id=iid,
num_chunks=5,
rank=rank,
kv_bytes_per_chunk=2 * PAGE_SIZE,
cpu_page_size=PAGE_SIZE,
num_segments=3,
)
for rank in range(2)
]
segments = regions[0]._segments
try:
assert [length // PAGE_SIZE for _, _, length in segments] == [4, 4, 2]
for rank, region in enumerate(regions):
ranges = list(region.iter_worker_slot_ranges())
first_row = rank * 5 // 2
assert ranges[0] == ((first_row * 2 + rank) * PAGE_SIZE, PAGE_SIZE)
assert sorted(ranges) == [
((row * 2 + rank) * PAGE_SIZE, PAGE_SIZE) for row in range(5)
]
region.create_next_worker_view(PAGE_SIZE).fill_(rank + 1)
rows = regions[0].base_tensor.view(5, 2, PAGE_SIZE)
assert (rows[:, 0] == 1).all() and (rows[:, 1] == 2).all()
del rows
finally:
for region in regions:
region.cleanup()
assert not any(os.path.exists(path) for path, _, _ in segments)
# ---------------------------------------------------------------------------
# Multi-worker race — concurrent construction
# ---------------------------------------------------------------------------
@@ -899,6 +934,7 @@ def _unmapped_registration_region(num_chunks, slot_size, num_workers):
region.cpu_page_size = slot_size
region._row_stride = num_workers * slot_size
region.total_size_bytes = region.num_chunks * region._row_stride
region._segments = [("unused", 0, region.total_size_bytes)]
region._base = SimpleNamespace(data_ptr=lambda: 0x100000)
region.pinned_addresses = []
region.is_pinned = False