Files
sglang/test/registered/chunked_prefill/test_scripted_swa_1gpu.py
T

108 lines
3.2 KiB
Python

import unittest
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.scripted_runtime.context import ScriptedContext
from sglang.test.scripted_runtime.test_case import ScriptedTestCase
from sglang.test.scripted_runtime_chunked_helpers import base_engine_kwargs
register_cuda_ci(est_time=130, stage="extra-a", runner_config="1-gpu-large")
_SWA_MODEL = "openai/gpt-oss-20b"
_MAX_TOTAL_TOKENS = 4096
_SWA_FULL_TOKENS_RATIO = 0.1
_CHUNK_SIZE = 64
_N_DECODERS = 6
_DECODER_PROMPT = 64
_DECODER_MAX_NEW = 512
_DECODER_WARMUP_RUNNING = 3
_CHUNKED_PROMPT = 384
_CHUNKED_MAX_NEW = 2
_N_CANDIDATES = 24
_STEPS_PER_CANDIDATE = 120
_DECODER_WARMUP_STEPS = 60
_DRAIN_STEPS = 400
class TestScriptedSwaChunkedReqEarlyReturn(ScriptedTestCase):
ENGINE_KWARGS = base_engine_kwargs(
model_path=_SWA_MODEL,
chunked_prefill_size=_CHUNK_SIZE,
max_total_tokens=_MAX_TOTAL_TOKENS,
swa_full_tokens_ratio=_SWA_FULL_TOKENS_RATIO,
page_size=1,
mem_fraction_static=0.70,
)
def test_swa_chunked_req_early_return_no_double_free(self):
self.server.execute_script(
self._script_swa_chunked_req_early_return_no_double_free
)
@staticmethod
def _script_swa_chunked_req_early_return_no_double_free(t: ScriptedContext):
s = t.scheduler
for i in range(_N_DECODERS):
t.start_req(
prompt_len=_DECODER_PROMPT,
max_new_tokens=_DECODER_MAX_NEW,
ignore_eos=True,
prompt_token=10 + i,
)
for _ in range(_DECODER_WARMUP_STEPS):
if len(s.running_batch.reqs) >= _DECODER_WARMUP_RUNNING:
break
yield
candidates = []
parked = False
for _ in range(_N_CANDIDATES):
candidates.append(
t.start_req(
prompt_len=_CHUNKED_PROMPT,
max_new_tokens=_CHUNKED_MAX_NEW,
prompt_token=2,
)
)
for _ in range(_STEPS_PER_CANDIDATE):
if any(t.chunked_parks(c.rid) > 0 for c in candidates):
parked = True
break
if candidates[-1].finished:
break
yield
if parked:
break
parked = parked or any(t.chunked_parks(c.rid) > 0 for c in candidates)
assert parked, (
"no chunked candidate was ever parked by add_chunked_req's hybrid-SWA "
"early-return; the test never exercised the stash gate"
)
t.abort_all()
for _ in range(_DRAIN_STEPS):
if (
s.chunked_req is None
and len(s.waiting_queue) == 0
and s.running_batch.is_empty()
):
break
yield
for _ in range(20):
yield
locked = {nid: lr for nid, lr in t.get_all_node_lock_refs().items() if lr != 0}
assert not locked, (
f"radix nodes left locked after drain {locked} -- stash gate let an "
"un-scheduled chunked req commit partial KV"
)
if __name__ == "__main__":
unittest.main()