vLLM Ascend vllm_ascend/observability/ - an MS Service Metric provider, not an OTel or Prometheus surface
repo
Fields
id
sources/w5e-vllm-ascend-observability-providerPublisher
vllm-project
Kind
repo
Accessed
2026-10-05
notes
Read at main 2026-10-05. get_metric_provider() imports ms_service_metric.provider_api.MetricProvider and registers YAML config from the sibling config/ directory (base_metrics.yaml, eplb_metrics.yaml) under framework_package vllm_ascend and handler_module_prefixes vllm_ascend.observability. handlers.py registers five EPLB expert-hotness gauges (eplb:expert_hotness:current_mean/current_max/update_mean/update_max/imbalance) through get_metric_recorder(). The provider is imported lazily so vLLM Ascend has no hard Core dependency. This is a vendor-internal metrics bridge, not a Prometheus endpoint and not a tracing surface.
Notes
Read at main 2026-10-05. get_metric_provider() imports ms_service_metric.provider_api.MetricProvider and registers YAML config from the sibling config/ directory (base_metrics.yaml, eplb_metrics.yaml) under framework_package vllm_ascend and handler_module_prefixes vllm_ascend.observability. handlers.py registers five EPLB expert-hotness gauges (eplb:expert_hotness:current_mean/current_max/update_mean/update_max/imbalance) through get_metric_recorder(). The provider is imported lazily so vLLM Ascend has no hard Core dependency. This is a vendor-internal metrics bridge, not a Prometheus endpoint and not a tracing surface.
Cross-references
This record references no other records.