[dLLM] Keep the request row across FDFO blocks (#42846)

This commit is contained in:
Liangsheng Yin
2026-10-07 15:39:33 -07:00
committed by GitHub
parent 685d4d453d
commit b1ffe5c16a
2 changed files with 37 additions and 3 deletions
+2 -2
View File
@@ -215,9 +215,9 @@ class SchedulerDllmMixin:
elif self.dllm_config.requires_separate_context_encoding:
self._stash_dllm_context(req)
else:
# The row stays with the request between blocks: a later abort
# releases its KV and tree lock through it.
self.stash_chunked_request(req)
if fdfo_mode:
self.req_to_token_pool.free(req)
def _fetch_waiting_reqs(self: Scheduler):
# Calculate how many requests can be added to DLLM manager
@@ -3,10 +3,11 @@
import unittest
from array import array
from types import SimpleNamespace
from unittest.mock import Mock
import torch
from sglang.srt.dllm.mixin.scheduler import DllmManager
from sglang.srt.dllm.mixin.scheduler import DllmManager, SchedulerDllmMixin
from sglang.srt.managers.schedule_batch import ReqKvInfo
from sglang.srt.mem_cache.allocation import alloc_for_extend
from sglang.srt.mem_cache.memory_pool import ReqToTokenPool
@@ -224,5 +225,38 @@ class TestDllmFdfoKvReuse(unittest.TestCase):
self.assertEqual(manager.staging_queue, [])
class TestDllmFdfoResolvedBlockKeepsRow(unittest.TestCase):
def test_resolved_block_keeps_row_until_next_block(self):
"""A resolved FDFO block used to hand its row back to the pool while the
request kept running. An abort before the next block then skipped
release_kv_cache (it only runs for row holders), leaking the request's
tree lock and any KV the tree does not own."""
pool = ReqToTokenPool(
size=4, max_context_len=16, device="cpu", enable_memory_saver=False
)
req = SimpleNamespace(
dllm_incomplete_ids=array("q"),
is_dllm_prefill=lambda: False,
kv=ReqKvInfo(kv_allocated_len=8, kv_committed_len=8),
)
pool.alloc([req])
row = req.kv.req_pool_idx
scheduler = SimpleNamespace(
dllm_config=SimpleNamespace(
first_done_first_out_mode=True,
requires_separate_context_encoding=False,
),
req_to_token_pool=pool,
stash_chunked_request=Mock(),
)
SchedulerDllmMixin.finish_dllm_forward(scheduler, req)
scheduler.stash_chunked_request.assert_called_once_with(req)
self.assertTrue(req.kv.holds_kv)
self.assertEqual(req.kv.req_pool_idx, row)
self.assertNotIn(row, pool.free_slots)
if __name__ == "__main__":
unittest.main()