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 15, 21:00 UTC
[V0][Metrics] Deprecate some questionable request time metrics (#14135)
Signed-off-by: Mark McLoughlin <markmc@redhat.com>
Mark McLoughlin committed
Mar 4, 2025 at 15:11 UTC
c8525f06fcc99dbc3564eaed9985edf16d287b47
1 file changed
+17
-6
vllm/engine/metrics.py
+17
-6
@@ -197,24 +197,35 @@ class Metrics:
197
"Histogram of time spent in DECODE phase for request.",
198
labelnames=labelnames,
199
buckets=request_latency_buckets)
200
+ # Deprecated in 0.8 - duplicates vllm:request_queue_time_seconds:
201
+ # TODO: in 0.9, only enable if show_hidden_metrics=True
202
self.histogram_time_in_queue_request = self._histogram_cls(
203
name="vllm:time_in_queue_requests",
202
- documentation=
203
- "Histogram of time the request spent in the queue in seconds.",
204
+ documentation=(
205
+ "Histogram of time the request spent in the queue in seconds. "
206
+ "DEPRECATED: use vllm:request_queue_time_seconds instead."),
207
labelnames=labelnames,
208
buckets=request_latency_buckets)
209
+
210
+ # Deprecated in 0.8 - use prefill/decode/inference time metrics
211
+ # TODO: in 0.9, only enable if show_hidden_metrics=True
212
self.histogram_model_forward_time_request = self._histogram_cls(
213
name="vllm:model_forward_time_milliseconds",
208
- documentation=
209
- "Histogram of time spent in the model forward pass in ms.",
214
+ documentation=(
215
+ "Histogram of time spent in the model forward pass in ms. "
216
+ "DEPRECATED: use prefill/decode/inference time metrics instead."
217
+ ),
218
labelnames=labelnames,
219
buckets=build_1_2_3_5_8_buckets(3000))
220
self.histogram_model_execute_time_request = self._histogram_cls(
221
name="vllm:model_execute_time_milliseconds",
214
- documentation=
215
- "Histogram of time spent in the model execute function in ms.",
222
+ documentation=(
223
+ "Histogram of time spent in the model execute function in ms."
224
+ "DEPRECATED: use prefill/decode/inference time metrics instead."
225
+ ),
226
labelnames=labelnames,
227
buckets=build_1_2_3_5_8_buckets(3000))
228
+
229
# Metadata
230
self.histogram_num_prompt_tokens_request = self._histogram_cls(
231
name="vllm:request_prompt_tokens",