Skip to content

Commit 5100ee1

Browse files
committed
Prefer layer reshape cache event
1 parent bdfc4db commit 5100ee1

3 files changed

Lines changed: 6 additions & 9 deletions

File tree

docs/source/user-guide/metrics/metrics.md

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -224,7 +224,7 @@ The default metrics configuration contains the following UCM metrics.
224224
| `ucm:load_bytes_total` | Total bytes loaded through the UCM connector. |
225225
| `ucm:save_bytes_total` | Total bytes saved through the UCM connector. |
226226
| `ucm:dump_event_reshape_cache_direct_used_total` | Dump submissions synchronized with `attn_metadata.reshape_cache_event`. |
227-
| `ucm:dump_event_reshape_cache_mla_layer_used_total` | Dump submissions synchronized with `attn_metadata[layer_name].reshape_cache_event`. |
227+
| `ucm:dump_event_reshape_cache_layer_used_total` | Dump submissions synchronized with `attn_metadata[layer_name].reshape_cache_event`. |
228228
| `ucm:dump_event_current_stream_used_total` | Dump submissions synchronized with a UCM-recorded current stream event fallback. |
229229
| `ucm:dump_event_sync_fallback_used_total` | Dump submissions that fell back to device synchronization because no event handle was available. |
230230

examples/metrics/metrics_configs.yaml

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -65,7 +65,7 @@ counter:
6565
documentation: "Total bytes saved through the UCM connector (summed across all wait_for_save calls)"
6666
- name: "dump_event_reshape_cache_direct_used_total"
6767
documentation: "Number of dump submissions synchronized with attn_metadata.reshape_cache_event"
68-
- name: "dump_event_reshape_cache_mla_layer_used_total"
68+
- name: "dump_event_reshape_cache_layer_used_total"
6969
documentation: "Number of dump submissions synchronized with attn_metadata[layer_name].reshape_cache_event"
7070
- name: "dump_event_current_stream_used_total"
7171
documentation: "Number of dump submissions synchronized with a UCM-recorded current stream event fallback"

ucm/integration/vllm/ucm_connector.py

Lines changed: 4 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -836,17 +836,14 @@ def _get_reshape_cache_event(
836836
if attn_metadata is None:
837837
return None, "none"
838838

839-
if self.use_mla:
840-
if not layer_name:
841-
return None, "mla_missing_layer"
839+
if layer_name:
842840
try:
843841
layer_metadata = attn_metadata[layer_name]
844842
event = getattr(layer_metadata, "reshape_cache_event", None)
845843
if event is not None:
846-
return event, "mla_layer"
844+
return event, "layer"
847845
except (KeyError, TypeError, AttributeError):
848846
pass
849-
return None, "mla_missing"
850847

851848
event = getattr(attn_metadata, "reshape_cache_event", None)
852849
if event is not None:
@@ -876,9 +873,9 @@ def _get_dump_event_handle(
876873
ucmmetrics.update_stats(
877874
"dump_event_reshape_cache_direct_used_total", 1.0
878875
)
879-
elif event_source == "mla_layer":
876+
elif event_source == "layer":
880877
ucmmetrics.update_stats(
881-
"dump_event_reshape_cache_mla_layer_used_total", 1.0
878+
"dump_event_reshape_cache_layer_used_total", 1.0
882879
)
883880
return event_handle
884881

0 commit comments

Comments
 (0)