chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

Add SGLANG_CP_TRANSFER_LOG: decouple CP shared-KV transfer-partition dumps from debug flag

Browse Source 
SGLANG_DEBUG_CP_SHARED_KV also disables tai IPC materialize -> NSA index fail-fast at
warmup (unusable in this config). Add a logging-only flag that emits the sender/worker
transfer-partition dumps (main-KV pages/positions vs NSA-state pages/positions) without
that side effect, and lift the 64-log cap, so we can diff the state-vs-main-KV partition
for a cache-hit vs cache-miss.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
孙远航
2026-06-17 16:29:31 +00:00
parent 268bcf78da
commit b34e7cb932
2 changed files with 9 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
python/sglang/srt/disaggregation/mooncake/conn.py
View File

@@ -43,8 +43,8 @@ logger = logging.getLogger(__name__)
_CP_SHARED_DEBUG_COUNTS: dict[str, int] = {}
def _cp_shared_debug_log(key: str, message: str, *args, limit: int = 64) -> None:
if not envs.SGLANG_DEBUG_CP_SHARED_KV.get():
def _cp_shared_debug_log(key: str, message: str, *args, limit: int = 1_000_000) -> None:
if not (envs.SGLANG_DEBUG_CP_SHARED_KV.get() or envs.SGLANG_CP_TRANSFER_LOG.get()):
return
count = _CP_SHARED_DEBUG_COUNTS.get(key, 0)
if count >= limit:
@@ -1241,7 +1241,7 @@ class MooncakeKVManager(CommonKVManager):
_np_summary(chunked_dst_kv_indice),
kv_chunk.is_last_chunk,
)
if envs.SGLANG_DEBUG_CP_SHARED_KV.get():
if envs.SGLANG_DEBUG_CP_SHARED_KV.get() or envs.SGLANG_CP_TRANSFER_LOG.get():
_cp_shared_debug_log(
"transfer_worker_kv",
"transfer worker cp_rank=%s room=%s prefill_pages=%s "
@@ -1350,7 +1350,7 @@ class MooncakeKVManager(CommonKVManager):
req.dst_state_indices,
kv_chunk.state_logical_page_positions,
)
if envs.SGLANG_DEBUG_CP_SHARED_KV.get():
if envs.SGLANG_DEBUG_CP_SHARED_KV.get() or envs.SGLANG_CP_TRANSFER_LOG.get():
_cp_shared_debug_log(
"transfer_worker_state",
"transfer worker state cp_rank=%s room=%s prefill_state_pages=%s "
@@ -1731,7 +1731,7 @@ class MooncakeKVSender(CommonKVSender):
)
)
index_slice = None
if envs.SGLANG_DEBUG_CP_SHARED_KV.get():
if envs.SGLANG_DEBUG_CP_SHARED_KV.get() or envs.SGLANG_CP_TRANSFER_LOG.get():
_cp_shared_debug_log(
"sender_filter",
"sender filter cp_rank=%s room=%s page_start=%s orig_kv_pages=%s "

4
python/sglang/srt/environ.py
View File

@@ -274,6 +274,10 @@ class Envs:
# CP HiCache round-trip KV-corruption tracing. 0=off, 1=structural lifecycle
# (rid_map/backup/split/evict/reload), 2=+compose/free/ack timing.
SGLANG_CP_HICACHE_KV_TRACE = EnvInt(0)
# Logging-only: emit the CP shared-KV sender/worker transfer-partition dumps
# (main-KV pages/positions vs NSA-state pages/positions) WITHOUT the side effects
# of SGLANG_DEBUG_CP_SHARED_KV (which disables tai materialize -> fail-fast).
SGLANG_CP_TRANSFER_LOG = EnvBool(False)
# Dir to torch.save per-stage attention tensors for L1-hit vs L2-reload diff.
# Empty=off. Only requests whose rid starts with "dump-" are dumped.
SGLANG_NSA_DUMP_DIR = EnvStr("")