| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from dataclasses import dataclass |
| 4 | from typing import Any |
| 5 | |
| 6 | from vllm.config import VllmConfig |
| 7 | from vllm.distributed.kv_transfer.kv_connector.v1.metrics import ( |
| 8 | KVConnectorPromMetrics, |
| 9 | KVConnectorStats, |
| 10 | PromMetric, |
| 11 | PromMetricT, |
| 12 | ) |
| 13 | from vllm.v1.kv_offload.base import ( |
| 14 | OffloadingCounterMetadata, |
| 15 | OffloadingGaugeMetadata, |
| 16 | OffloadingHistogramMetadata, |
| 17 | OffloadingMetricMetadata, |
| 18 | ) |
| 19 | from vllm.v1.kv_offload.factory import OffloadingSpecFactory |
| 20 | |
| 21 | |
| 22 | class _TransferMetricName: |
| 23 | """Flat metric names for GPU↔offload-medium transfer operations.""" |
| 24 | |
| 25 | LOAD_BYTES = "vllm:kv_offload_load_bytes" |
| 26 | LOAD_TIME = "vllm:kv_offload_load_time" |
| 27 | LOAD_SIZE = "vllm:kv_offload_load_size" |
| 28 | STORE_BYTES = "vllm:kv_offload_store_bytes" |
| 29 | STORE_TIME = "vllm:kv_offload_store_time" |
| 30 | STORE_SIZE = "vllm:kv_offload_store_size" |
| 31 | |
| 32 | |
| 33 | class _ConnectorMetricName: |
| 34 | """Connector-side metrics emitted by scheduler-side offloading code.""" |
| 35 | |
| 36 | LOOKUP_SYNC_DELAY = "vllm:kv_offload_lookup_sync_delay_seconds" |
| 37 | LOOKUP_ASYNC_DELAY = "vllm:kv_offload_lookup_async_delay_seconds" |
| 38 | ALLOCATION_FAILURE = "vllm:kv_offload_allocation_failure" |
| 39 | |
| 40 | |
| 41 | class _TransferType: |
| 42 | """Transfer direction labels for deprecated CPU offload metrics.""" |
| 43 | |
| 44 | LOAD = "CPU_to_GPU" |
| 45 | STORE = "GPU_to_CPU" |
| 46 | ALL = (LOAD, STORE) |
| 47 | |
| 48 | |
| 49 | TRANSFER_SIZE_BUCKETS = ( |
| 50 | 1e6, |
| 51 | 5e6, |
| 52 | 10e6, |
| 53 | 20e6, |
| 54 | 40e6, |
| 55 | 60e6, |
| 56 | 80e6, |
| 57 | 100e6, |
| 58 | 150e6, |
| 59 | 200e6, |
| 60 | ) |
| 61 | |
| 62 | |
| 63 | def get_connector_metric_definitions() -> dict[str, OffloadingMetricMetadata]: |
| 64 | return { |
| 65 | _TransferMetricName.LOAD_BYTES: OffloadingCounterMetadata( |
| 66 | documentation="Total bytes loaded from offload storage to GPU.", |
| 67 | ), |
| 68 | _TransferMetricName.LOAD_TIME: OffloadingCounterMetadata( |
| 69 | documentation="Total load time from offload storage to GPU, in seconds.", |
| 70 | ), |
| 71 | _TransferMetricName.LOAD_SIZE: OffloadingHistogramMetadata( |
| 72 | documentation="Histogram of KV offload load operation size, in bytes.", |
| 73 | buckets=TRANSFER_SIZE_BUCKETS, |
| 74 | ), |
| 75 | _TransferMetricName.STORE_BYTES: OffloadingCounterMetadata( |
| 76 | documentation="Total bytes stored from GPU to offload storage.", |
| 77 | ), |
| 78 | _TransferMetricName.STORE_TIME: OffloadingCounterMetadata( |
| 79 | documentation="Total store time from GPU to offload storage, in seconds.", |
| 80 | ), |
| 81 | _TransferMetricName.STORE_SIZE: OffloadingHistogramMetadata( |
| 82 | documentation="Histogram of KV offload store operation size, in bytes.", |
| 83 | buckets=TRANSFER_SIZE_BUCKETS, |
| 84 | ), |
| 85 | _ConnectorMetricName.LOOKUP_SYNC_DELAY: OffloadingHistogramMetadata( |
| 86 | documentation=( |
| 87 | "Histogram of the time spent in a single offload lookup call, " |
| 88 | "in seconds." |
| 89 | ), |
| 90 | buckets=( |
| 91 | 0.00001, |
| 92 | 0.00005, |
| 93 | 0.0001, |
| 94 | 0.0005, |
| 95 | 0.001, |
| 96 | 0.005, |
| 97 | 0.01, |
| 98 | 0.05, |
| 99 | 0.1, |
| 100 | 0.5, |
| 101 | 1, |
| 102 | ), |
| 103 | ), |
| 104 | _ConnectorMetricName.LOOKUP_ASYNC_DELAY: OffloadingHistogramMetadata( |
| 105 | documentation=( |
| 106 | "Histogram of time between a request's offload lookup first " |
| 107 | "deferring and the following lookup resolving, or request " |
| 108 | "finish, in seconds." |
| 109 | ), |
| 110 | buckets=( |
| 111 | 0.0001, |
| 112 | 0.0005, |
| 113 | 0.001, |
| 114 | 0.005, |
| 115 | 0.01, |
| 116 | 0.05, |
| 117 | 0.1, |
| 118 | 0.5, |
| 119 | 1, |
| 120 | 5, |
| 121 | 10, |
| 122 | ), |
| 123 | ), |
| 124 | _ConnectorMetricName.ALLOCATION_FAILURE: OffloadingCounterMetadata( |
| 125 | documentation=( |
| 126 | "Number of KV offload store allocation attempts that failed." |
| 127 | ), |
| 128 | ), |
| 129 | } |
| 130 | |
| 131 | |
| 132 | _DEPRECATED_TOTAL_BYTES = "vllm:kv_offload_total_bytes" |
| 133 | _DEPRECATED_TOTAL_TIME = "vllm:kv_offload_total_time" |
| 134 | _DEPRECATED_SIZE = "vllm:kv_offload_size" |
| 135 | |
| 136 | # Deprecated legacy transfer metrics, kept during the migration to the flat |
| 137 | # metric names above. These stay in a separate definition block because they |
| 138 | # use a transfer_type label, but are emitted from the same flat stats payload |
| 139 | # for compatibility. |
| 140 | _DEPRECATED_CONNECTOR_METRIC_DEFINITIONS: dict[str, OffloadingMetricMetadata] = { |
| 141 | _DEPRECATED_TOTAL_BYTES: OffloadingCounterMetadata( |
| 142 | documentation="Number of bytes offloaded by KV connector", |
| 143 | ), |
| 144 | _DEPRECATED_TOTAL_TIME: OffloadingCounterMetadata( |
| 145 | documentation="Total time measured by all KV offloading operations", |
| 146 | ), |
| 147 | _DEPRECATED_SIZE: OffloadingHistogramMetadata( |
| 148 | documentation="Histogram of KV offload transfer size, in bytes.", |
| 149 | buckets=TRANSFER_SIZE_BUCKETS, |
| 150 | ), |
| 151 | } |
| 152 | |
| 153 | |
| 154 | class _MetricType: |
| 155 | """Type tags embedded in the serialized stats payload.""" |
| 156 | |
| 157 | COUNTER = "counter" |
| 158 | GAUGE = "gauge" |
| 159 | HISTOGRAM = "histogram" |
| 160 | |
| 161 | |
| 162 | class _StatsKey: |
| 163 | """Top-level keys in the serialized stats dict.""" |
| 164 | |
| 165 | # Maps metric name -> _MetricType value |
| 166 | TYPES = "types" |
| 167 | # Maps metric name -> {label values tuple -> observed value (number or list)} |
| 168 | DATA = "data" |
| 169 | |
| 170 | |
| 171 | @dataclass |
| 172 | class OffloadingConnectorStats(KVConnectorStats): |
| 173 | """Offloading connector stats use flat metric names as keys. |
| 174 | |
| 175 | The ``data`` dict is structured using ``_StatsKey`` / ``_MetricType``:: |
| 176 | |
| 177 | { |
| 178 | _StatsKey.TYPES: {name: _MetricType.*, ...}, |
| 179 | _StatsKey.DATA: {name: {labelvalues: value, ...}, ...}, |
| 180 | } |
| 181 | |
| 182 | This structure is self-describing: it survives IPC serialization |
| 183 | without needing the full ``OffloadingMetricMetadata`` objects on the |
| 184 | receiving side. |
| 185 | |
| 186 | Counter values are aggregated by summing per-label-tuple, gauge values |
| 187 | use the latest snapshot per-label-tuple, and histogram values are lists of |
| 188 | observed samples per-label-tuple. Unlabeled metrics use ``()`` as their |
| 189 | labelvalues tuple. |
| 190 | """ |
| 191 | |
| 192 | def __post_init__(self): |
| 193 | if _StatsKey.DATA not in self.data: |
| 194 | self.reset() |
| 195 | |
| 196 | def reset(self): |
| 197 | self.data: dict[str, Any] = { |
| 198 | _StatsKey.TYPES: {}, |
| 199 | _StatsKey.DATA: {}, |
| 200 | } |
| 201 | |
| 202 | @property |
| 203 | def _types(self) -> dict[str, str]: |
| 204 | return self.data[_StatsKey.TYPES] |
| 205 | |
| 206 | @property |
| 207 | def _values(self) -> dict[str, Any]: |
| 208 | return self.data[_StatsKey.DATA] |
| 209 | |
| 210 | def aggregate(self, other: "KVConnectorStats") -> "KVConnectorStats": |
| 211 | if other.is_empty(): |
| 212 | return self |
| 213 | assert isinstance(other, OffloadingConnectorStats) |
| 214 | other_types = other._types |
| 215 | other_values = other._values |
| 216 | for key, other_label_values in other_values.items(): |
| 217 | type_str = other_types.get(key) |
| 218 | if type_str is None: |
| 219 | raise AssertionError(f"Unknown offloading stats key: {key}") |
| 220 | self._types.setdefault(key, type_str) |
| 221 | current_label_values = self._values.setdefault(key, {}) |
| 222 | for labelvalues, value in other_label_values.items(): |
| 223 | if type_str == _MetricType.HISTOGRAM: |
| 224 | assert isinstance(value, list) |
| 225 | if labelvalues not in current_label_values: |
| 226 | current_label_values[labelvalues] = list(value) |
| 227 | else: |
| 228 | assert isinstance(current_label_values[labelvalues], list) |
| 229 | current_label_values[labelvalues].extend(value) |
| 230 | elif type_str == _MetricType.COUNTER: |
| 231 | assert isinstance(value, int | float) |
| 232 | current_label_values[labelvalues] = ( |
| 233 | current_label_values.get(labelvalues, 0) + value |
| 234 | ) |
| 235 | elif type_str == _MetricType.GAUGE: |
| 236 | assert isinstance(value, int | float) |
| 237 | current_label_values[labelvalues] = value |
| 238 | else: |
| 239 | raise AssertionError( |
| 240 | f"Unknown metric type '{type_str}' for key: {key}" |
| 241 | ) |
| 242 | return self |
| 243 | |
| 244 | def reduce(self) -> dict[str, int | float]: |
| 245 | """Reduce the observations collected during a time interval to one or |
| 246 | more representative values (eg avg/median/sum of the series). |
| 247 | This is meant to be called by the logger to produce a summary of the |
| 248 | stats for the last time interval. |
| 249 | """ |
| 250 | return_dict: dict[str, int | float] = {} |
| 251 | for key, label_value_map in self._values.items(): |
| 252 | type_str = self._types.get(key) |
| 253 | if type_str is None: |
| 254 | raise AssertionError(f"Unknown offloading stats key: {key}") |
| 255 | for labelvalues, value in label_value_map.items(): |
| 256 | key_with_labels = f"{key}:{labelvalues}" if labelvalues else key |
| 257 | if type_str == _MetricType.HISTOGRAM: |
| 258 | assert isinstance(value, list) |
| 259 | return_dict[f"{key_with_labels}_count"] = len(value) |
| 260 | return_dict[f"{key_with_labels}_sum"] = sum(value) |
| 261 | elif type_str in (_MetricType.COUNTER, _MetricType.GAUGE): |
| 262 | assert isinstance(value, int | float) |
| 263 | return_dict[key_with_labels] = value |
| 264 | else: |
| 265 | raise AssertionError( |
| 266 | f"Unknown metric type '{type_str}' for key: {key}" |
| 267 | ) |
| 268 | return return_dict |
| 269 | |
| 270 | def is_empty(self) -> bool: |
| 271 | return not self.data.get(_StatsKey.DATA) |
| 272 | |
| 273 | def increase_counter( |
| 274 | self, |
| 275 | counter_name: str, |
| 276 | counter_increase_value: int | float = 1, |
| 277 | labelvalues: tuple[str, ...] = (), |
| 278 | ) -> None: |
| 279 | """Increase a counter on the stats payload.""" |
| 280 | self._types.setdefault(counter_name, _MetricType.COUNTER) |
| 281 | counter_values = self._values.setdefault(counter_name, {}) |
| 282 | counter_values[labelvalues] = ( |
| 283 | counter_values.get(labelvalues, 0) + counter_increase_value |
| 284 | ) |
| 285 | |
| 286 | def set_gauge( |
| 287 | self, |
| 288 | gauge_name: str, |
| 289 | gauge_value: int | float, |
| 290 | labelvalues: tuple[str, ...] = (), |
| 291 | ) -> None: |
| 292 | """Set a gauge snapshot on the stats payload.""" |
| 293 | self._types.setdefault(gauge_name, _MetricType.GAUGE) |
| 294 | gauge_values = self._values.setdefault(gauge_name, {}) |
| 295 | gauge_values[labelvalues] = gauge_value |
| 296 | |
| 297 | def observe_histogram( |
| 298 | self, |
| 299 | histogram_name: str, |
| 300 | histogram_value: int | float, |
| 301 | labelvalues: tuple[str, ...] = (), |
| 302 | ) -> None: |
| 303 | """Record a histogram observation on the stats payload.""" |
| 304 | self._types.setdefault(histogram_name, _MetricType.HISTOGRAM) |
| 305 | histogram_values = self._values.setdefault(histogram_name, {}) |
| 306 | histogram_values.setdefault(labelvalues, []).append(histogram_value) |
| 307 | |
| 308 | |
| 309 | class OffloadPromMetrics(KVConnectorPromMetrics): |
| 310 | def __init__( |
| 311 | self, |
| 312 | vllm_config: VllmConfig, |
| 313 | metric_types: dict[type[PromMetric], type[PromMetricT]], |
| 314 | labelnames: list[str], |
| 315 | per_engine_labelvalues: dict[int, list[object]], |
| 316 | ): |
| 317 | super().__init__(vllm_config, metric_types, labelnames, per_engine_labelvalues) |
| 318 | # (engine_idx, transfer_type) -> (metric with bounded labels) |
| 319 | self.histogram_transfer_size: dict[tuple[int, str], PromMetricT] = {} |
| 320 | self.counter_kv_bytes: dict[tuple[int, str], PromMetricT] = {} |
| 321 | self.counter_kv_transfer_time: dict[tuple[int, str], PromMetricT] = {} |
| 322 | kv_transfer_config = vllm_config.kv_transfer_config |
| 323 | assert kv_transfer_config is not None |
| 324 | extra_config = kv_transfer_config.kv_connector_extra_config |
| 325 | spec_cls = OffloadingSpecFactory.get_spec_cls(extra_config) |
| 326 | self._offloading_metric_metadata: dict[str, OffloadingMetricMetadata] = { |
| 327 | **spec_cls.build_metric_definitions(extra_config), |
| 328 | **get_connector_metric_definitions(), |
| 329 | } |
| 330 | from vllm.v1.kv_offload.cpu.spec import CPUOffloadingSpec |
| 331 | |
| 332 | self._observe_deprecated_metrics = issubclass(spec_cls, CPUOffloadingSpec) |
| 333 | self._offloading_metric_defs: dict[str, PromMetricT] = {} |
| 334 | # (engine_idx, metric_name, labelvalues) -> metric with bound labels |
| 335 | self.offloading_metrics: dict[ |
| 336 | tuple[int, str, tuple[str, ...]], PromMetricT |
| 337 | ] = {} |
| 338 | |
| 339 | self._counter_kv_bytes = self._counter_cls( |
| 340 | name=_DEPRECATED_TOTAL_BYTES, |
| 341 | documentation=_DEPRECATED_CONNECTOR_METRIC_DEFINITIONS[ |
| 342 | _DEPRECATED_TOTAL_BYTES |
| 343 | ].documentation, |
| 344 | labelnames=labelnames + ["transfer_type"], |
| 345 | ) |
| 346 | |
| 347 | self._counter_kv_transfer_time = self._counter_cls( |
| 348 | name=_DEPRECATED_TOTAL_TIME, |
| 349 | documentation=_DEPRECATED_CONNECTOR_METRIC_DEFINITIONS[ |
| 350 | _DEPRECATED_TOTAL_TIME |
| 351 | ].documentation, |
| 352 | labelnames=labelnames + ["transfer_type"], |
| 353 | ) |
| 354 | |
| 355 | deprecated_size_metadata = _DEPRECATED_CONNECTOR_METRIC_DEFINITIONS[ |
| 356 | _DEPRECATED_SIZE |
| 357 | ] |
| 358 | assert isinstance(deprecated_size_metadata, OffloadingHistogramMetadata) |
| 359 | self._histogram_transfer_size = self._histogram_cls( |
| 360 | name=_DEPRECATED_SIZE, |
| 361 | documentation=deprecated_size_metadata.documentation, |
| 362 | buckets=deprecated_size_metadata.buckets, |
| 363 | labelnames=labelnames + ["transfer_type"], |
| 364 | ) |
| 365 | |
| 366 | for engine_idx, labelvalues in per_engine_labelvalues.items(): |
| 367 | for transfer_type in _TransferType.ALL: |
| 368 | bounded_labelvalues = labelvalues + [transfer_type] |
| 369 | self.histogram_transfer_size[(engine_idx, transfer_type)] = ( |
| 370 | self._histogram_transfer_size.labels(*bounded_labelvalues) |
| 371 | ) |
| 372 | self.counter_kv_bytes[(engine_idx, transfer_type)] = ( |
| 373 | self._counter_kv_bytes.labels(*bounded_labelvalues) |
| 374 | ) |
| 375 | self.counter_kv_transfer_time[(engine_idx, transfer_type)] = ( |
| 376 | self._counter_kv_transfer_time.labels(*bounded_labelvalues) |
| 377 | ) |
| 378 | |
| 379 | for metric_name, metadata in self._offloading_metric_metadata.items(): |
| 380 | self._offloading_metric_defs[metric_name] = self._create_metric( |
| 381 | metric_name, metadata |
| 382 | ) |
| 383 | |
| 384 | def _create_metric( |
| 385 | self, metric_name: str, metadata: OffloadingMetricMetadata |
| 386 | ) -> Any: |
| 387 | kwargs: dict[str, Any] = { |
| 388 | "name": metric_name, |
| 389 | "documentation": metadata.documentation, |
| 390 | "labelnames": self._labelnames + list(metadata.labelnames), |
| 391 | } |
| 392 | if isinstance(metadata, OffloadingCounterMetadata): |
| 393 | metric_cls = self._counter_cls |
| 394 | elif isinstance(metadata, OffloadingGaugeMetadata): |
| 395 | metric_cls = self._gauge_cls |
| 396 | elif isinstance(metadata, OffloadingHistogramMetadata): |
| 397 | metric_cls = self._histogram_cls |
| 398 | if metadata.buckets is not None: |
| 399 | kwargs["buckets"] = metadata.buckets |
| 400 | else: |
| 401 | raise AssertionError(f"Unknown offloading metric metadata: {metadata}") |
| 402 | return metric_cls(**kwargs) |
| 403 | |
| 404 | def _get_prometheus_metric( |
| 405 | self, |
| 406 | metric_name: str, |
| 407 | labelvalues: tuple[str, ...], |
| 408 | engine_idx: int, |
| 409 | ) -> PromMetric: |
| 410 | metadata = self._offloading_metric_metadata[metric_name] |
| 411 | if len(labelvalues) != len(metadata.labelnames): |
| 412 | raise AssertionError( |
| 413 | f"Metric {metric_name} expects {len(metadata.labelnames)} labels, " |
| 414 | f"got {len(labelvalues)}" |
| 415 | ) |
| 416 | key = (engine_idx, metric_name, labelvalues) |
| 417 | prom_metric = self.offloading_metrics.get(key) |
| 418 | if prom_metric is None: |
| 419 | engine_labelvalues = self.per_engine_labelvalues[engine_idx] |
| 420 | prom_metric = self._offloading_metric_defs[metric_name].labels( |
| 421 | *(engine_labelvalues + list(labelvalues)) |
| 422 | ) |
| 423 | self.offloading_metrics[key] = prom_metric |
| 424 | return prom_metric |
| 425 | |
| 426 | def _increase_counter( |
| 427 | self, |
| 428 | metric_name: str, |
| 429 | value: int | float, |
| 430 | labelvalues: tuple[str, ...], |
| 431 | engine_idx: int, |
| 432 | ) -> None: |
| 433 | self._get_prometheus_metric(metric_name, labelvalues, engine_idx).inc(value) |
| 434 | if labelvalues or not self._observe_deprecated_metrics: |
| 435 | return |
| 436 | # Keep deprecated CPU offload transfer metrics updated during the |
| 437 | # transition to flat metric names. |
| 438 | if metric_name == _TransferMetricName.LOAD_BYTES: |
| 439 | self.counter_kv_bytes[(engine_idx, _TransferType.LOAD)].inc(value) |
| 440 | elif metric_name == _TransferMetricName.LOAD_TIME: |
| 441 | self.counter_kv_transfer_time[(engine_idx, _TransferType.LOAD)].inc(value) |
| 442 | elif metric_name == _TransferMetricName.STORE_BYTES: |
| 443 | self.counter_kv_bytes[(engine_idx, _TransferType.STORE)].inc(value) |
| 444 | elif metric_name == _TransferMetricName.STORE_TIME: |
| 445 | self.counter_kv_transfer_time[(engine_idx, _TransferType.STORE)].inc(value) |
| 446 | |
| 447 | def _set_gauge( |
| 448 | self, |
| 449 | metric_name: str, |
| 450 | value: int | float, |
| 451 | labelvalues: tuple[str, ...], |
| 452 | engine_idx: int, |
| 453 | ) -> None: |
| 454 | self._get_prometheus_metric(metric_name, labelvalues, engine_idx).set(value) |
| 455 | |
| 456 | def _observe_histogram( |
| 457 | self, |
| 458 | metric_name: str, |
| 459 | value: list[int | float], |
| 460 | labelvalues: tuple[str, ...], |
| 461 | engine_idx: int, |
| 462 | ) -> None: |
| 463 | prom_metric = self._get_prometheus_metric(metric_name, labelvalues, engine_idx) |
| 464 | for observation in value: |
| 465 | prom_metric.observe(observation) |
| 466 | if labelvalues or not self._observe_deprecated_metrics: |
| 467 | continue |
| 468 | # Keep deprecated CPU offload transfer metrics updated during the |
| 469 | # transition to flat metric names. |
| 470 | if metric_name == _TransferMetricName.LOAD_SIZE: |
| 471 | self.histogram_transfer_size[(engine_idx, _TransferType.LOAD)].observe( |
| 472 | observation |
| 473 | ) |
| 474 | elif metric_name == _TransferMetricName.STORE_SIZE: |
| 475 | self.histogram_transfer_size[(engine_idx, _TransferType.STORE)].observe( |
| 476 | observation |
| 477 | ) |
| 478 | |
| 479 | def observe(self, transfer_stats_data: dict[str, Any], engine_idx: int = 0): |
| 480 | """Observe transfer statistics.""" |
| 481 | metric_types = transfer_stats_data.get(_StatsKey.TYPES, {}) |
| 482 | metric_data = transfer_stats_data.get(_StatsKey.DATA, {}) |
| 483 | for key, label_value_map in metric_data.items(): |
| 484 | type_str = metric_types.get(key) |
| 485 | if type_str is None: |
| 486 | raise AssertionError(f"Unknown offloading stats key: {key}") |
| 487 | assert key in self._offloading_metric_defs |
| 488 | for labelvalues, value in label_value_map.items(): |
| 489 | if type_str == _MetricType.COUNTER: |
| 490 | assert isinstance(value, int | float) |
| 491 | self._increase_counter(key, value, labelvalues, engine_idx) |
| 492 | elif type_str == _MetricType.GAUGE: |
| 493 | assert isinstance(value, int | float) |
| 494 | self._set_gauge(key, value, labelvalues, engine_idx) |
| 495 | elif type_str == _MetricType.HISTOGRAM: |
| 496 | assert isinstance(value, list) |
| 497 | assert all(isinstance(v, int | float) for v in value) |
| 498 | self._observe_histogram(key, value, labelvalues, engine_idx) |
| 499 | else: |
| 500 | raise AssertionError( |
| 501 | f"Unknown metric type '{type_str}' for key: {key}" |
| 502 | ) |