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 8, 20:15 UTC
[Misc] Add debug logs (#33931)
Signed-off-by: NickLucche <nlucches@redhat.com>
Nicolò Lucchesi committed
Feb 5, 2026 at 18:42 UTC
7d8c6804e2654873cb25d0b23fef178fa5f37237
2 files changed
+5
vllm/distributed/kv_transfer/kv_connector/utils.py
+2
@@ -324,6 +324,7 @@ class TpKVTopology:
324
kv_cache_shape = self.attn_backend.get_kv_cache_shape(
325
num_blocks=1, block_size=_MOCK_BLOCK_SIZE, num_kv_heads=1, head_size=1
326
)
327
+ logger.debug("Test kv_cache_shape: %s", kv_cache_shape)
328
# Non-MLA backends caches have 5 dims [2, num_blocks, H,N,D],
329
# we just mock num_blocks to 1 for the dimension check below.
330
self._is_kv_layout_blocks_first = (
@@ -337,6 +338,7 @@ class TpKVTopology:
338
)
339
340
if self._cross_layers_blocks:
341
+ logger.debug("Using cross-layer KV cache")
342
# prepend layers dimension
343
_MOCK_NUM_LAYERS = 80
344
kv_cache_shape = (_MOCK_NUM_LAYERS,) + kv_cache_shape
vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
+3
@@ -1354,6 +1354,9 @@ class NixlConnectorWorker:
1354
if base_addr in seen_base_addresses:
1355
continue
1356
1357
+ logger.debug(
1358
+ "Registering layer %s with cache shape: %s", layer_name, cache.shape
1359
+ )
1360
kernel_block_size = cache.shape[self.kv_topo.block_size_position]
1361
if self.block_size != kernel_block_size:
1362
logger.info_once(