[copy of #45208] CuMem slept-L1 fragmentation accounting (#49208)

Signed-off-by: haosdent <haosdent@gmail.com>
Signed-off-by: Justin Wood <justin.m.wood@me.com>
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Co-authored-by: haosdent <haosdent@gmail.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-authored-by: Justin Wood <jwood@me.com>
This commit is contained in:
Matthew Bonanni
2026-07-20 23:08:11 +00:00
committed by GitHub
co-authored by haosdent Claude Opus 4.7 Justin Wood
parent b7c20d0cfa
commit 58b2012aa2
4 changed files with 73 additions and 46 deletions
+13 -2
View File
@@ -9,6 +9,7 @@ import torch.nn.functional as F
from transformers import AutoModel
from vllm.platforms import current_platform
from vllm.utils.mem_constants import MiB_bytes
from ....conftest import HfRunner
from ....utils import VLLM_PATH
@@ -106,7 +107,12 @@ def test_prm_models(
if current_platform.is_cpu():
pytest.skip("CPU only supports V1")
with vllm_runner(model, max_model_len=1024, dtype=dtype) as vllm_model:
with vllm_runner(
model,
max_model_len=1024,
dtype=dtype,
kv_cache_memory_bytes=64 * MiB_bytes,
) as vllm_model:
vllm_outputs = vllm_model.token_classify(math_step_prompts)
with hf_runner(model, dtype=dtype, auto_cls=AutoModel) as hf_model:
@@ -145,7 +151,12 @@ def test_prm_models_with_golden_outputs(
if not FIXTURE_REWARD_RESULT.get(model):
pytest.skip(f"No available golden outputs for {model}.")
with vllm_runner(model, max_model_len=1024, dtype=dtype) as vllm_model:
with vllm_runner(
model,
max_model_len=1024,
dtype=dtype,
kv_cache_memory_bytes=64 * MiB_bytes,
) as vllm_model:
vllm_outputs = vllm_model.token_classify(math_step_prompts)
golden_outputs = load_reward_outputs(FIXTURE_REWARD_RESULT[model])
+17
View File
@@ -67,6 +67,23 @@ def test_memory_profiling():
non_torch_ratio = result.non_torch_increase / (256 * 1024 * 1024) # noqa
assert abs(non_torch_ratio - 1) <= 0.05
assert result.torch_peak_increase == 1024 * 1024 * 1024
expected_total_consumed = (256 + 512) * 1024 * 1024
total_consumed_ratio = result.total_consumed / expected_total_consumed
assert abs(total_consumed_ratio - 1) <= 0.05, (
f"total_consumed={result.total_consumed}, "
f"expected={expected_total_consumed}, "
f"ratio={total_consumed_ratio}"
)
expected_non_kv = expected_total_consumed + 1024 * 1024 * 1024
non_kv_ratio = result.non_kv_cache_memory / expected_non_kv
assert abs(non_kv_ratio - 1) <= 0.05, (
f"non_kv_cache_memory={result.non_kv_cache_memory}, "
f"expected={expected_non_kv}, "
f"ratio={non_kv_ratio}"
)
del weights
lib.cudaFree(handle1)
lib.cudaFree(handle2)