Add token-offset based selective offload in OffloadConnector (#39983)

Signed-off-by: Angelo Ruocco <ang@zurich.ibm.com>
Co-authored-by: Or Ozeri <or@ozery.com>
This commit is contained in:
Angelo Ruocco
2026-05-28 14:11:02 +00:00
committed by GitHub
co-authored by Or Ozeri
parent f3b2a819f7
commit 864990e8d9
2 changed files with 113 additions and 0 deletions
@@ -936,6 +936,99 @@ def test_complete_store_called_per_job(request_runner, async_scheduling: bool):
assert runner.manager.complete_store.call_count == 0
@pytest.mark.parametrize("async_scheduling", [True, False])
def test_max_offload_tokens_validation(request_runner, async_scheduling: bool):
"""Validates max_offload_tokens: type coercion, boundary values, and capping.
Setup: 3 offloaded blocks × 3 GPU blocks each = 9 GPU block offsets (0–8).
"""
gpu_block_size = 4
block_size_factor = 3
offloaded_block_size = gpu_block_size * block_size_factor # 12
num_gpu_blocks = 100
all_offsets = (0, 1, 2, 3, 4, 5, 6, 7, 8)
def make_runner():
return request_runner(
block_size=gpu_block_size,
num_gpu_blocks=num_gpu_blocks,
async_scheduling=async_scheduling,
block_size_factor=block_size_factor,
)
def setup(r, max_offload_tokens):
r.new_request(token_ids=[0] * offloaded_block_size * 3)
req = r.scheduler.requests[str(r.req_id)]
req.kv_transfer_params = {"max_offload_tokens": max_offload_tokens}
r.manager.prepare_store.side_effect = (
lambda keys, req_context: generate_store_output(keys)
)
# With sync scheduling, the connector flushes completed stores when the
# request finishes; async scheduling defers the flush to the next step.
flushed_all = all_offsets if not async_scheduling else ()
flushed_two = (0, 1, 2, 3, 4, 5) if not async_scheduling else ()
# None -> no cap, all 9 offsets stored
r = make_runner()
setup(r, None)
r.run(
decoded_tokens=[EOS_TOKEN_ID],
expected_stored=all_offsets,
expected_flushed=flushed_all,
)
# str -> warn and fall back to no cap
r = make_runner()
setup(r, "24")
r.run(
decoded_tokens=[EOS_TOKEN_ID],
expected_stored=all_offsets,
expected_flushed=flushed_all,
)
# float -> warn and fall back to no cap
r = make_runner()
setup(r, 24.5)
r.run(
decoded_tokens=[EOS_TOKEN_ID],
expected_stored=all_offsets,
expected_flushed=flushed_all,
)
# negative -> warn and fall back to no cap
r = make_runner()
setup(r, -1)
r.run(
decoded_tokens=[EOS_TOKEN_ID],
expected_stored=all_offsets,
expected_flushed=flushed_all,
)
# bool -> rejected (type(True) is bool, not int), falls back to no cap
r = make_runner()
setup(r, True)
r.run(
decoded_tokens=[EOS_TOKEN_ID],
expected_stored=all_offsets,
expected_flushed=flushed_all,
)
# zero -> valid, no blocks offloaded
r = make_runner()
setup(r, 0)
r.run(decoded_tokens=[EOS_TOKEN_ID], expected_stored=())
# positive int cap -> limits offload to first 2 offloaded blocks (offsets 0–5)
r = make_runner()
setup(r, 24) # 24 tokens = 2 offloaded blocks × 12 tokens each
r.run(
decoded_tokens=[EOS_TOKEN_ID],
expected_stored=(0, 1, 2, 3, 4, 5),
expected_flushed=flushed_two,
)
def test_flush_all_jobs_when_no_requests_remain(request_runner):
"""When all tracked requests are finished, build_connector_meta flushes
all pending jobs since there will be no future step to complete them."""