vllm-cpu-amxbf16 0.9.1__cp312-cp312-manylinux_2_17_x86_64.whl

vllm/v1/metrics/ray_wrappers.py

@@ -0,0 +1,131 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import time
+from typing import Optional, Union
+
+from vllm.config import VllmConfig
+from vllm.v1.metrics.loggers import PrometheusStatLogger
+from vllm.v1.spec_decode.metrics import SpecDecodingProm
+
+try:
+    from ray.util import metrics as ray_metrics
+    from ray.util.metrics import Metric
+except ImportError:
+    ray_metrics = None
+
+
+class RayPrometheusMetric:
+
+    def __init__(self):
+        if ray_metrics is None:
+            raise ImportError(
+                "RayPrometheusMetric requires Ray to be installed.")
+
+        self.metric: Metric = None
+
+    def labels(self, *labels, **labelskwargs):
+        if labelskwargs:
+            for k, v in labelskwargs.items():
+                if not isinstance(v, str):
+                    labelskwargs[k] = str(v)
+
+            self.metric.set_default_tags(labelskwargs)
+
+        if labels:
+            if len(labels) != len(self.metric._tag_keys):
+                raise ValueError(
+                    "Number of labels must match the number of tag keys. "
+                    f"Expected {len(self.metric._tag_keys)}, got {len(labels)}"
+                )
+
+            self.metric.set_default_tags(
+                dict(zip(self.metric._tag_keys, labels)))
+
+        return self
+
+
+class RayGaugeWrapper(RayPrometheusMetric):
+    """Wraps around ray.util.metrics.Gauge to provide same API as
+    prometheus_client.Gauge"""
+
+    def __init__(self,
+                 name: str,
+                 documentation: Optional[str] = "",
+                 labelnames: Optional[list[str]] = None):
+        labelnames_tuple = tuple(labelnames) if labelnames else None
+        self.metric = ray_metrics.Gauge(name=name,
+                                        description=documentation,
+                                        tag_keys=labelnames_tuple)
+
+    def set(self, value: Union[int, float]):
+        return self.metric.set(value)
+
+    def set_to_current_time(self):
+        # ray metrics doesn't have set_to_current time, https://docs.ray.io/en/latest/_modules/ray/util/metrics.html
+        return self.metric.set(time.time())
+
+
+class RayCounterWrapper(RayPrometheusMetric):
+    """Wraps around ray.util.metrics.Counter to provide same API as
+    prometheus_client.Counter"""
+
+    def __init__(self,
+                 name: str,
+                 documentation: Optional[str] = "",
+                 labelnames: Optional[list[str]] = None):
+        labelnames_tuple = tuple(labelnames) if labelnames else None
+        self.metric = ray_metrics.Counter(name=name,
+                                          description=documentation,
+                                          tag_keys=labelnames_tuple)
+
+    def inc(self, value: Union[int, float] = 1.0):
+        if value == 0:
+            return
+        return self.metric.inc(value)
+
+
+class RayHistogramWrapper(RayPrometheusMetric):
+    """Wraps around ray.util.metrics.Histogram to provide same API as
+    prometheus_client.Histogram"""
+
+    def __init__(self,
+                 name: str,
+                 documentation: Optional[str] = "",
+                 labelnames: Optional[list[str]] = None,
+                 buckets: Optional[list[float]] = None):
+        labelnames_tuple = tuple(labelnames) if labelnames else None
+        boundaries = buckets if buckets else []
+        self.metric = ray_metrics.Histogram(name=name,
+                                            description=documentation,
+                                            tag_keys=labelnames_tuple,
+                                            boundaries=boundaries)
+
+    def observe(self, value: Union[int, float]):
+        return self.metric.observe(value)
+
+
+class RaySpecDecodingProm(SpecDecodingProm):
+    """
+    RaySpecDecodingProm is used by RayMetrics to log to Ray metrics.
+    Provides the same metrics as SpecDecodingProm but uses Ray's
+    util.metrics library.
+    """
+
+    _counter_cls = RayCounterWrapper
+
+
+class RayPrometheusStatLogger(PrometheusStatLogger):
+    """RayPrometheusStatLogger uses Ray metrics instead."""
+
+    _gauge_cls = RayGaugeWrapper
+    _counter_cls = RayCounterWrapper
+    _histogram_cls = RayHistogramWrapper
+    _spec_decoding_cls = RaySpecDecodingProm
+
+    def __init__(self, vllm_config: VllmConfig, engine_index: int = 0):
+        super().__init__(vllm_config, engine_index)
+
+    @staticmethod
+    def _unregister_vllm_metrics():
+        # No-op on purpose
+        pass

vllm/v1/metrics/reader.py

@@ -0,0 +1,246 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from dataclasses import dataclass
+from typing import Optional
+
+from prometheus_client import REGISTRY
+from prometheus_client import Metric as PromMetric
+from prometheus_client.samples import Sample
+
+
+@dataclass
+class Metric:
+    """A base class for prometheus metrics.
+
+    Each metric may be associated with key=value labels, and
+    in some cases a single vLLM instance may have multiple
+    metrics with the same name but different sets of labels.
+    """
+    name: str
+    labels: dict[str, str]
+
+
+@dataclass
+class Counter(Metric):
+    """A monotonically increasing integer counter."""
+    value: int
+
+
+@dataclass
+class Vector(Metric):
+    """An ordered array of integer counters.
+
+    This type - which doesn't exist in Prometheus - models one very
+    specific metric, vllm:spec_decode_num_accepted_tokens_per_pos.
+    """
+    values: list[int]
+
+
+@dataclass
+class Gauge(Metric):
+    """A numerical value that can go up or down."""
+    value: float
+
+
+@dataclass
+class Histogram(Metric):
+    """Observations recorded in configurable buckets.
+
+    Buckets are represented by a dictionary. The key is
+    the upper limit of the bucket, and the value is the
+    observed count in that bucket. A '+Inf' key always
+    exists.
+
+    The count property is the total count across all
+    buckets, identical to the count of the '+Inf' bucket.
+
+    The sum property is the total sum of all observed
+    values.
+    """
+    count: int
+    sum: float
+    buckets: dict[str, int]
+
+
+def get_metrics_snapshot() -> list[Metric]:
+    """An API for accessing in-memory Prometheus metrics.
+
+    Example:
+        >>> for metric in llm.get_metrics():
+        ...     if isinstance(metric, Counter):
+        ...         print(f"{metric} = {metric.value}")
+        ...     elif isinstance(metric, Gauge):
+        ...         print(f"{metric} = {metric.value}")
+        ...     elif isinstance(metric, Histogram):
+        ...         print(f"{metric}")
+        ...         print(f"    sum = {metric.sum}")
+        ...         print(f"    count = {metric.count}")
+        ...         for bucket_le, value in metrics.buckets.items():
+        ...             print(f"    {bucket_le} = {value}")
+    """
+    collected: list[Metric] = []
+    for metric in REGISTRY.collect():
+        if not metric.name.startswith("vllm:"):
+            continue
+        if metric.type == "gauge":
+            samples = _get_samples(metric)
+            for s in samples:
+                collected.append(
+                    Gauge(name=metric.name, labels=s.labels, value=s.value))
+        elif metric.type == "counter":
+            samples = _get_samples(metric, "_total")
+            if metric.name == "vllm:spec_decode_num_accepted_tokens_per_pos":
+                #
+                # Ugly vllm:num_accepted_tokens_per_pos special case.
+                #
+                # This metric is a vector of counters - for each spec
+                # decoding token position, we observe the number of
+                # accepted tokens using a Counter labeled with 'position'.
+                # We convert these into a vector of integer values.
+                #
+                for labels, values in _digest_num_accepted_by_pos_samples(
+                        samples):
+                    collected.append(
+                        Vector(name=metric.name, labels=labels, values=values))
+            else:
+                for s in samples:
+                    collected.append(
+                        Counter(name=metric.name,
+                                labels=s.labels,
+                                value=int(s.value)))
+
+        elif metric.type == "histogram":
+            #
+            # A histogram has a number of '_bucket' samples where
+            # the 'le' label represents the upper limit of the bucket.
+            # We convert these bucketized values into a dict of values
+            # indexed by the value of the 'le' label. The 'le=+Inf'
+            # label is a special case, catching all values observed.
+            #
+            bucket_samples = _get_samples(metric, "_bucket")
+            count_samples = _get_samples(metric, "_count")
+            sum_samples = _get_samples(metric, "_sum")
+            for labels, buckets, count_value, sum_value in _digest_histogram(
+                    bucket_samples, count_samples, sum_samples):
+                collected.append(
+                    Histogram(name=metric.name,
+                              labels=labels,
+                              buckets=buckets,
+                              count=count_value,
+                              sum=sum_value))
+        else:
+            raise AssertionError(f"Unknown metric type {metric.type}")
+
+    return collected
+
+
+def _get_samples(metric: PromMetric,
+                 suffix: Optional[str] = None) -> list[Sample]:
+    name = (metric.name + suffix) if suffix is not None else metric.name
+    return [s for s in metric.samples if s.name == name]
+
+
+def _strip_label(labels: dict[str, str], key_to_remove: str) -> dict[str, str]:
+    labels_copy = labels.copy()
+    labels_copy.pop(key_to_remove)
+    return labels_copy
+
+
+def _digest_histogram(
+    bucket_samples: list[Sample], count_samples: list[Sample],
+    sum_samples: list[Sample]
+) -> list[tuple[dict[str, str], dict[str, int], int, float]]:
+    #
+    # In the case of DP, we have an indigestable
+    # per-bucket-per-engine count as a list of labelled
+    # samples, along with total and sum samples
+    #
+    # bucket_samples (in):
+    #   labels = {bucket: 100, idx: 0}, value = 2
+    #   labels = {bucket: 200, idx: 0}, value = 4
+    #   labels = {bucket: Inf, idx: 0}, value = 10
+    #   labels = {bucket: 100, idx: 1}, value = 1
+    #   labels = {bucket: 200, idx: 2}, value = 5
+    #   labels = {bucket: Inf, idx: 3}, value = 7
+    # count_samples (in):
+    #   labels = {idx: 0}, value = 10
+    #   labels = {idx: 1}, value = 7
+    # sum_samples (in):
+    #   labels = {idx: 0}, value = 2000
+    #   labels = {idx: 1}, value = 1200
+    #
+    # output: [
+    #   {idx: 0}, {"100": 2, "200": 4, "Inf": 10}, 10, 2000
+    #   {idx: 1}, {"100": 1, "200": 5, "Inf": 7},   7, 1200
+    # ]
+    buckets_by_labels: dict[frozenset[tuple[str, str]], dict[str, int]] = {}
+    for s in bucket_samples:
+        bucket = s.labels["le"]
+        labels_key = frozenset(_strip_label(s.labels, "le").items())
+        if labels_key not in buckets_by_labels:
+            buckets_by_labels[labels_key] = {}
+        buckets_by_labels[labels_key][bucket] = int(s.value)
+
+    counts_by_labels: dict[frozenset[tuple[str, str]], int] = {}
+    for s in count_samples:
+        labels_key = frozenset(s.labels.items())
+        counts_by_labels[labels_key] = int(s.value)
+
+    sums_by_labels: dict[frozenset[tuple[str, str]], float] = {}
+    for s in sum_samples:
+        labels_key = frozenset(s.labels.items())
+        sums_by_labels[labels_key] = s.value
+
+    assert set(buckets_by_labels.keys()) == set(
+        counts_by_labels.keys()) == set(sums_by_labels.keys())
+
+    output = []
+    label_keys = list(buckets_by_labels.keys())
+    for k in label_keys:
+        labels = dict(k)
+        output.append((labels, buckets_by_labels[k], counts_by_labels[k],
+                       sums_by_labels[k]))
+    return output
+
+
+def _digest_num_accepted_by_pos_samples(
+        samples: list[Sample]) -> list[tuple[dict[str, str], list[int]]]:
+    #
+    # In the case of DP, we have an indigestable
+    # per-position-per-engine count as a list of
+    # labelled samples
+    #
+    # samples (in):
+    #   labels = {pos: 0, idx: 0}, value = 10
+    #   labels = {pos: 1, idx: 0}, value = 7
+    #   labels = {pos: 2, idx: 0}, value = 2
+    #   labels = {pos: 0, idx: 1}, value = 5
+    #   labels = {pos: 1, idx: 1}, value = 3
+    #   labels = {pos: 2, idx: 1}, value = 1
+    #
+    # output: [
+    #   {idx: 0}, [10, 7, 2]
+    #   {idx: 1}, [5, 3, 1]
+    # ]
+    #
+    max_pos = 0
+    values_by_labels: dict[frozenset[tuple[str, str]], dict[int, int]] = {}
+
+    for s in samples:
+        position = int(s.labels["position"])
+        max_pos = max(max_pos, position)
+
+        labels_key = frozenset(_strip_label(s.labels, "position").items())
+        if labels_key not in values_by_labels:
+            values_by_labels[labels_key] = {}
+        values_by_labels[labels_key][position] = int(s.value)
+
+    output = []
+    for labels_key, values_by_position in values_by_labels.items():
+        labels = dict(labels_key)
+        values = [0] * (max_pos + 1)
+        for pos, val in values_by_position.items():
+            values[pos] = val
+        output.append((labels, values))
+    return output

vllm/v1/metrics/stats.py

@@ -0,0 +1,239 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import time
+from dataclasses import dataclass, field
+from typing import TYPE_CHECKING, Optional
+
+from vllm.v1.spec_decode.metrics import SpecDecodingStats
+
+if TYPE_CHECKING:
+    from vllm.v1.engine import EngineCoreEvent, EngineCoreOutput, FinishReason
+    from vllm.v1.engine.output_processor import RequestState
+
+
+@dataclass
+class PrefixCacheStats:
+    """Stores prefix cache hit statistics."""
+    # Whether reset_prefix_cache was invoked.
+    reset: bool = False
+    # The number of requests in this update.
+    requests: int = 0
+    # The number of queries in these requests. Note that "queries" here
+    # means the number of tokens that were queried from the cache.
+    queries: int = 0
+    # The number of hits in these requests.
+    hits: int = 0
+
+
+@dataclass
+class SchedulerStats:
+    """Stats associated with the scheduler."""
+
+    num_running_reqs: int = 0
+    num_waiting_reqs: int = 0
+
+    gpu_cache_usage: float = 0.0
+
+    prefix_cache_stats: PrefixCacheStats = field(
+        default_factory=PrefixCacheStats)
+
+    spec_decoding_stats: Optional[SpecDecodingStats] = None
+
+
+@dataclass
+class LoRAStats:
+    waiting_requests: set[str] = field(default_factory=set)
+    running_requests: set[str] = field(default_factory=set)
+
+
+@dataclass
+class RequestStateStats:
+    """Stats that need to be tracked across delta updates."""
+
+    num_generation_tokens: int = 0
+
+    # This is a engine frontend timestamp (wall-clock)
+    arrival_time: float = 0.0
+
+    # These are engine core timestamps (monotonic)
+    queued_ts: float = 0.0
+    scheduled_ts: float = 0.0
+    first_token_ts: float = 0.0
+    last_token_ts: float = 0.0
+
+
+@dataclass
+class FinishedRequestStats:
+    """Stats associated with a finished request."""
+
+    finish_reason: "FinishReason"
+    e2e_latency: float = 0.0
+    num_prompt_tokens: int = 0
+    num_generation_tokens: int = 0
+    max_tokens_param: Optional[int] = None
+    queued_time: float = 0.0
+    prefill_time: float = 0.0
+    inference_time: float = 0.0
+    decode_time: float = 0.0
+
+
+class IterationStats:
+    """Stats associated with a single set of EngineCoreOutputs."""
+
+    def __init__(self):
+        self.iteration_timestamp = time.time()
+        self.num_generation_tokens = 0
+        self.num_prompt_tokens = 0
+        self.num_preempted_reqs = 0
+        self.finished_requests: list[FinishedRequestStats] = []
+        self.max_num_generation_tokens_iter: list[int] = []
+        self.n_params_iter: list[int] = []
+        self.time_to_first_tokens_iter: list[float] = []
+        self.time_per_output_tokens_iter: list[float] = []
+        self.waiting_lora_adapters: dict[str, int] = {}
+        self.running_lora_adapters: dict[str, int] = {}
+
+    def _time_since(self, start: float) -> float:
+        """Calculate an interval relative to this iteration's timestamp."""
+        return self.iteration_timestamp - start
+
+    def update_from_output(self, output: "EngineCoreOutput",
+                           engine_core_timestamp: float, is_prefilling: bool,
+                           prompt_len: int, req_stats: RequestStateStats,
+                           lora_stats: Optional[LoRAStats]):
+        num_new_generation_tokens = len(output.new_token_ids)
+
+        self.num_generation_tokens += num_new_generation_tokens
+        if is_prefilling:
+            assert num_new_generation_tokens > 0
+            self.num_prompt_tokens += prompt_len
+
+            first_token_latency = self._time_since(req_stats.arrival_time)
+            self.time_to_first_tokens_iter.append(first_token_latency)
+
+        req_stats.num_generation_tokens += num_new_generation_tokens
+
+        # Process request-level engine core events
+        if output.events is not None:
+            self.update_from_events(output.request_id, output.events,
+                                    is_prefilling, req_stats, lora_stats)
+
+        # Process the batch-level "new tokens" engine core event
+        if is_prefilling:
+            req_stats.first_token_ts = engine_core_timestamp
+        else:
+            tpot = engine_core_timestamp - req_stats.last_token_ts
+            self.time_per_output_tokens_iter.append(tpot)
+
+        req_stats.last_token_ts = engine_core_timestamp
+
+    def update_from_events(self, req_id: str, events: list["EngineCoreEvent"],
+                           is_prefilling: bool, req_stats: RequestStateStats,
+                           lora_stats: Optional[LoRAStats]):
+        # Avoid circular dependency
+        from vllm.v1.engine import EngineCoreEventType
+        for event in events:
+            if event.type == EngineCoreEventType.QUEUED:
+                req_stats.queued_ts = event.timestamp
+                if lora_stats is not None:
+                    lora_stats.waiting_requests.add(req_id)
+            elif event.type == EngineCoreEventType.SCHEDULED:
+                if req_stats.scheduled_ts == 0.0:  # ignore preemptions
+                    req_stats.scheduled_ts = event.timestamp
+                LoRARequestStates.scheduled_request(lora_stats, req_id)
+            elif event.type == EngineCoreEventType.PREEMPTED:
+                self.num_preempted_reqs += 1
+                LoRARequestStates.preempted_request(lora_stats, req_id)
+
+    def update_from_finished_request(self, finish_reason: "FinishReason",
+                                     num_prompt_tokens: int,
+                                     max_tokens_param: Optional[int],
+                                     req_stats: RequestStateStats):
+        e2e_latency = self._time_since(req_stats.arrival_time)
+
+        # Queued interval is from first QUEUED event to first SCHEDULED
+        queued_time = req_stats.scheduled_ts - req_stats.queued_ts
+
+        # Prefill interval is from first SCHEDULED to first NEW_TOKEN
+        # Any preemptions during prefill is included in the interval
+        prefill_time = req_stats.first_token_ts - req_stats.scheduled_ts
+
+        # Decode interval is from first NEW_TOKEN to last NEW_TOKEN
+        # Any preemptions during decode are included
+        decode_time = req_stats.last_token_ts - req_stats.first_token_ts
+
+        # Inference interval is from first SCHEDULED to last NEW_TOKEN
+        # Any preemptions during prefill or decode are included
+        inference_time = req_stats.last_token_ts - req_stats.scheduled_ts
+
+        finished_req = \
+            FinishedRequestStats(finish_reason=finish_reason,
+                                 e2e_latency=e2e_latency,
+                                 num_prompt_tokens=num_prompt_tokens,
+                                 num_generation_tokens=req_stats.num_generation_tokens,
+                                 max_tokens_param=max_tokens_param,
+                                 queued_time=queued_time,
+                                 prefill_time=prefill_time,
+                                 inference_time=inference_time,
+                                 decode_time=decode_time)
+        self.finished_requests.append(finished_req)
+
+
+class LoRARequestStates:
+    """Per-LoRA request state stats."""
+
+    def __init__(self):
+        self.lora_name_to_stats: dict[str, LoRAStats] = {}
+
+    def get_stats(self, req_state: 'RequestState') -> Optional[LoRAStats]:
+        if req_state.lora_name is None:
+            return None
+        if req_state.lora_name not in self.lora_name_to_stats:
+            self.lora_name_to_stats[req_state.lora_name] = LoRAStats()
+        return self.lora_name_to_stats[req_state.lora_name]
+
+    def add_request(self, req_state: 'RequestState'):
+        if (lora_stats := self.get_stats(req_state)) is not None:
+            lora_stats.waiting_requests.add(req_state.request_id)
+
+    def finish_request(self, req_state: 'RequestState'):
+        if req_state.lora_name is None:
+            return
+        lora_stats = self.lora_name_to_stats[req_state.lora_name]
+        lora_stats.running_requests.remove(req_state.request_id)
+
+    def abort_request(self, req_state: 'RequestState'):
+        if req_state.lora_name is None:
+            return
+        lora_stats = self.lora_name_to_stats[req_state.lora_name]
+        lora_stats.waiting_requests.discard(req_state.request_id)
+        lora_stats.running_requests.discard(req_state.request_id)
+
+    # Break the pattern for this lifecycle methods so we can
+    # call this from IterationStats.update_from_events()
+    @staticmethod
+    def scheduled_request(lora_stats: Optional[LoRAStats], request_id: str):
+        if lora_stats is None:
+            return
+        lora_stats.waiting_requests.remove(request_id)
+        lora_stats.running_requests.add(request_id)
+
+    @staticmethod
+    def preempted_request(lora_stats: Optional[LoRAStats], request_id: str):
+        if lora_stats is None:
+            return
+        lora_stats.running_requests.remove(request_id)
+        lora_stats.waiting_requests.add(request_id)
+
+    def update_iteration_stats(self,
+                               iteration_stats: Optional[IterationStats]):
+        if iteration_stats is None:
+            return
+        for lora_name, stats in self.lora_name_to_stats.items():
+            if stats.waiting_requests:
+                iteration_stats.waiting_lora_adapters[lora_name] = \
+                    len(stats.waiting_requests)
+            if stats.running_requests:
+                iteration_stats.running_lora_adapters[lora_name] = \
+                    len(stats.running_requests)