[XPU] sleep mode: fix KV cache release test (#57485)

Signed-off-by: Yan Ma <yan.ma@intel.com>
Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
This commit is contained in:
Yan Ma
2026-09-18 15:49:17 +00:00
committed by GitHub
co-authored by Kunshang Ji
parent 017dced6a6
commit 4991f97669
+17 -8
View File
@@ -128,31 +128,40 @@ def test_release_kv_cache_memory_preserves_generation(full_sleep, monkeypatch):
sampling_params = SamplingParams(temperature=0, max_tokens=10)
expected = llm.generate(prompt, sampling_params)[0].outputs[0].token_ids
def get_mapped_bytes(worker):
return mapped_usage(get_mem_allocator_instance())
def get_mapped_bytes_by_tag(worker):
allocator = get_mem_allocator_instance()
mapped_bytes = mapped_usage(allocator)
kv_cache_bytes = sum(
data.handle[1]
for data in allocator.pointer_to_data.values()
if data.tag == "kv_cache" and not data.is_asleep
)
return mapped_bytes, kv_cache_bytes
mapped_before = llm.collective_rpc(get_mapped_bytes)[0]
mapped_before, kv_cache_before = llm.collective_rpc(get_mapped_bytes_by_tag)[0]
# Utility RPC transports the engine error as a plain Exception.
with pytest.raises(Exception, match="requires a completed pause"):
llm.release_kv_cache_memory()
assert llm.collective_rpc(get_mapped_bytes)[0] == mapped_before
assert llm.collective_rpc(get_mapped_bytes_by_tag)[0][0] == mapped_before
assert not llm.llm_engine.is_sleeping()
assert llm.generate(prompt, sampling_params)[0].outputs[0].token_ids == expected
llm.sleep(level=0)
llm.release_kv_cache_memory()
mapped_after = llm.collective_rpc(get_mapped_bytes)[0]
assert mapped_before - mapped_after >= kv_cache_memory_bytes * 0.99
mapped_after, kv_cache_after = llm.collective_rpc(get_mapped_bytes_by_tag)[0]
assert kv_cache_before > 0
assert kv_cache_after == 0
assert mapped_before - mapped_after >= kv_cache_before
assert mapped_after > 0
assert llm.llm_engine.is_sleeping()
if full_sleep:
llm.sleep(level=full_sleep)
assert llm.collective_rpc(get_mapped_bytes)[0] == 0
assert llm.collective_rpc(get_mapped_bytes_by_tag)[0][0] == 0
llm.wake_up(tags=None if full_sleep else ["kv_cache"])
if full_sleep == 2:
llm.collective_rpc("reload_weights")
assert llm.collective_rpc(get_mapped_bytes)[0] == mapped_before
assert llm.collective_rpc(get_mapped_bytes_by_tag)[0][0] == mapped_before
assert not llm.llm_engine.is_sleeping()
actual = llm.generate(prompt, sampling_params)[0].outputs[0].token_ids
assert actual == expected