A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 15, 11:00 UTC
[P/D] Add some more debug logs to `NixlConnector` (#18102)
Signed-off-by: Nick Hill <nhill@redhat.com>
Nick Hill committed
May 13, 2025 at 19:33 UTC
d5af47a149b8ffbcb8249638681740f1d0787cb0
1 file changed
+16
vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
+16
@@ -253,6 +253,11 @@ class NixlConnectorScheduler:
253
asynchronously (between scheduler steps).
254
"""
255
256
+ logger.debug(
257
+ "NIXLConnector get_num_new_matched_tokens: "
258
+ "num_computed_tokens=%s, kv_transfer_params=%s",
259
+ num_computed_tokens, request.kv_transfer_params)
260
+
261
# No KVTransfer for this request.
262
if request.kv_transfer_params is None:
263
return 0, False
@@ -271,6 +276,12 @@ class NixlConnectorScheduler:
276
def update_state_after_alloc(self, request: "Request",
277
blocks: "KVCacheBlocks",
278
num_external_tokens: int):
279
+
280
+ logger.debug(
281
+ "NIXLConnector update_state_after_alloc: "
282
+ "num_external_tokens=%s, kv_transfer_params=%s",
283
+ num_external_tokens, request.kv_transfer_params)
284
+
285
if request.kv_transfer_params is None:
286
return
287
@@ -319,6 +330,11 @@ class NixlConnectorScheduler:
330
should be freed now or will be sent asynchronously and freed later.
331
"""
332
333
+ logger.debug(
334
+ "NIXLConnector request_finished, "
335
+ "request_status=%s, kv_transfer_params=%s", request.status,
336
+ request.kv_transfer_params)
337
+
338
if request.kv_transfer_params is None:
339
return False, None
340
assert isinstance(request.kv_transfer_params, NixlKVTransferParams)