[Frontend] Expose per-request spec decode metrics in generate API (#43310)

Signed-off-by: aoshen02 <aoshen@inferact.ai>
Co-authored-by: aoshen02 <aoshen@inferact.ai>
Co-authored-by: Codex <codex@openai.com>
This commit is contained in:
Hongxin Xu
2026-09-20 12:21:33 +00:00
committed by GitHub
co-authored by aoshen02 Codex
parent bf01fc4a31
commit 10e6a7f210
12 changed files with 516 additions and 9 deletions
@@ -130,10 +130,36 @@ pub struct EngineCoreOutput {
#[serde(default)]
pub new_sampling_mask: Option<MaybeWireSamplingMask>,
/// Per-request speculative-decoding acceptance metrics, set on the final
/// output when `--per-request-spec-decode-metrics` is enabled. Opaque here;
/// the Rust frontend does not yet surface it in responses.
/// output when `--per-request-spec-decode-metrics` is enabled.
#[serde(default)]
pub spec_decode_metrics: Option<OpaqueValue>,
pub spec_decode_metrics: Option<RequestSpecDecodeMetrics>,
}
/// Raw per-sequence speculative-decoding accumulator.
///
/// Python models this as a plain `@dataclass`, so it is serialized by msgspec
/// as a map (named fields). Derived values such as acceptance rates are
/// computed by the frontend, not carried on the wire.
///
/// Original Python definition: `RequestSpecDecodeMetrics` in
/// `vllm/v1/metrics/stats.py`.
#[derive(Debug, Clone, Default, PartialEq, Eq, Serialize, Deserialize)]
pub struct RequestSpecDecodeMetrics {
/// Configured `num_speculative_tokens`.
#[serde(default)]
pub num_spec_tokens: u64,
/// Verify-step counts indexed by accepted draft-token count.
#[serde(default)]
pub histogram: Vec<u64>,
/// Total proposed draft tokens.
#[serde(default)]
pub num_draft_tokens: u64,
/// Accepted draft count per verify step; empty unless `detailed`.
#[serde(default)]
pub per_step_accepted: Vec<u64>,
/// Proposed draft count per verify step; empty unless `detailed`.
#[serde(default)]
pub per_step_drafted: Vec<u64>,
}
impl EngineCoreOutput {
+11 -1
View File
@@ -11,7 +11,9 @@ use futures::stream::FusedStream;
use futures::{Stream, StreamExt as _, pin_mut};
use serde::{Deserialize, Serialize};
use vllm_engine_core_client::protocol::logprobs::Logprobs;
use vllm_engine_core_client::protocol::output::{EngineCoreFinishReason, StopReason};
use vllm_engine_core_client::protocol::output::{
EngineCoreFinishReason, RequestSpecDecodeMetrics, StopReason,
};
use vllm_engine_core_client::protocol::sampling_mask::SamplingMask;
use vllm_engine_core_client::{AbortCause, EngineCoreOutputStream};
@@ -47,6 +49,8 @@ pub struct CollectedGenerateOutput {
pub ec_transfer_params: Option<serde_json::Value>,
/// Sampling support sets aligned one-to-one with generated token positions.
pub sampling_mask: Option<SamplingMask>,
/// Per-request speculative-decoding metrics from the terminal output.
pub spec_decode_metrics: Option<RequestSpecDecodeMetrics>,
}
/// Prompt-scoped metadata emitted only once on the first [`GenerateOutput`] for
@@ -160,6 +164,8 @@ pub struct GenerateOutput {
pub ec_transfer_params: Option<serde_json::Value>,
/// Sampling support sets aligned one-to-one with `token_ids`.
pub sampling_mask: Option<SamplingMask>,
/// Per-request speculative-decoding metrics, present on terminal outputs.
pub spec_decode_metrics: Option<RequestSpecDecodeMetrics>,
}
impl GenerateOutput {
@@ -208,6 +214,7 @@ impl GenerateOutput {
kv_transfer_params: None,
ec_transfer_params: None,
sampling_mask: None,
spec_decode_metrics: None,
}
}
}
@@ -308,6 +315,7 @@ impl Stream for GenerateOutputStream {
kv_transfer_params: raw.kv_transfer_params,
ec_transfer_params: raw.ec_transfer_params,
sampling_mask,
spec_decode_metrics: raw.spec_decode_metrics,
};
Poll::Ready(Some(Ok(output)))
@@ -396,6 +404,7 @@ impl<T: Stream<Item = Result<GenerateOutput>> + Send> T {
kv_transfer_params: None,
ec_transfer_params: None,
sampling_mask,
spec_decode_metrics: None,
});
}
@@ -409,6 +418,7 @@ impl<T: Stream<Item = Result<GenerateOutput>> + Send> T {
};
collected.kv_transfer_params = output.kv_transfer_params;
collected.ec_transfer_params = output.ec_transfer_params;
collected.spec_decode_metrics = output.spec_decode_metrics;
if let Some(mask) = collected.sampling_mask.as_ref()
&& mask.rows.len() != collected.token_ids.len()
{
+1
View File
@@ -413,6 +413,7 @@ async fn collect_output_rejects_partial_sampling_mask() {
sampling_mask: Some(SamplingMask {
rows: vec![vec![1, 33, 99]],
}),
spec_decode_metrics: None,
};
let error = futures::stream::iter([Ok(output)]).collect_output().await.unwrap_err();
@@ -29,7 +29,8 @@ use vllm_llm::{
use self::convert::{ResponseOptions, prepare_generate_request};
use self::types::{
GenerateLogprob, GenerateResponse, GenerateResponseChoice, GenerateResponseStreamChoice,
GenerateStreamResponse, MultiModalPlaceholders, PlaceholderRangeInfo,
GenerateStreamResponse, MultiModalPlaceholders, PerRequestMetrics, PlaceholderRangeInfo,
SpeculativeDecodingMetrics, StreamingSpeculativeDecodingMetrics,
};
pub(crate) use self::types::{GenerateRequest, GenerateSamplingParams};
pub(crate) use self::validate::validate_request_compat;
@@ -160,10 +161,14 @@ async fn generate_chunk_stream(
let mut prompt_tokens = None;
let mut prompt_token_ids = None;
let mut usage = TokenUsage::default();
let mut spec_decode_metrics = None;
while let Some(next) = stream.next().await {
match next {
Ok(output) => {
if let Some(metrics) = output.spec_decode_metrics {
spec_decode_metrics = Some(SpeculativeDecodingMetrics::from(metrics));
}
if prompt_tokens.is_none()
&& let Some(info) = output.prompt_info.as_ref()
{
@@ -223,6 +228,7 @@ async fn generate_chunk_stream(
.then(|| Usage::from_token_usage(usage, enable_prompt_tokens_details)),
mm_placeholders: prompt_token_ids.as_ref().and_then(|_| mm_placeholders.take()),
prompt_token_ids,
metrics: None,
})
.await;
}
@@ -243,6 +249,11 @@ async fn generate_chunk_stream(
usage: Some(Usage::from_token_usage(usage, enable_prompt_tokens_details)),
prompt_token_ids: None,
mm_placeholders: None,
metrics: spec_decode_metrics.map(|speculative_decoding| PerRequestMetrics {
speculative_decoding: StreamingSpeculativeDecodingMetrics::from(
speculative_decoding,
),
}),
})
.await;
}
@@ -318,6 +329,9 @@ fn collect_generate(
mm_placeholders: return_token_ids.then_some(mm_placeholders).flatten(),
kv_transfer_params: collected.kv_transfer_params,
ec_transfer_params: collected.ec_transfer_params,
metrics: collected.spec_decode_metrics.map(|metrics| PerRequestMetrics {
speculative_decoding: SpeculativeDecodingMetrics::from(metrics),
}),
})
}
@@ -465,6 +479,7 @@ mod tests {
use futures::{TryStreamExt as _, stream};
use vllm_engine_core_client::protocol::multimodal::{MmModality, PlaceholderRange};
use vllm_engine_core_client::protocol::output::RequestSpecDecodeMetrics;
use vllm_llm::GeneratePromptInfo;
use super::*;
@@ -482,6 +497,7 @@ mod tests {
kv_transfer_params: None,
ec_transfer_params: None,
sampling_mask: None,
spec_decode_metrics: None,
}),
Ok(GenerateOutput {
request_id: String::new(),
@@ -496,6 +512,7 @@ mod tests {
kv_transfer_params: None,
ec_transfer_params: None,
sampling_mask: None,
spec_decode_metrics: None,
}),
]);
@@ -578,6 +595,7 @@ mod tests {
kv_transfer_params: None,
ec_transfer_params: None,
sampling_mask: None,
spec_decode_metrics: None,
}
}
@@ -645,6 +663,80 @@ mod tests {
assert!(chunks[1].mm_placeholders.is_none());
}
#[tokio::test]
async fn generate_chunk_stream_returns_terminal_spec_decode_metrics() {
let metrics = RequestSpecDecodeMetrics {
num_spec_tokens: 3,
histogram: vec![0, 1, 1, 0],
num_draft_tokens: 6,
..Default::default()
};
let mut output = stream_output(None, Vec::new(), Some(FinishReason::stop_eos()));
output.spec_decode_metrics = Some(metrics.clone());
for include_usage in [false, true] {
let chunks: Vec<_> = generate_chunk_stream(
stream::iter(vec![Ok(output.clone())]),
"raw-stream".to_string(),
ApiServerOptions::default(),
ResponseOptions {
include_usage,
..Default::default()
},
None,
)
.try_collect()
.await
.expect("collect chunks");
assert_eq!(chunks.len(), 1 + usize::from(include_usage));
assert!(chunks[0].metrics.is_none());
if include_usage {
assert!(chunks[1].choices.is_empty());
assert_eq!(
chunks[1].metrics,
Some(PerRequestMetrics {
speculative_decoding: StreamingSpeculativeDecodingMetrics::from(
SpeculativeDecodingMetrics::from(metrics.clone())
)
})
);
let json = serde_json::to_value(&chunks[1]).expect("serialize chunk");
let stats = &json["metrics"]["speculative_decoding"];
assert_eq!(stats.as_object().expect("metrics object").len(), 7);
assert_eq!(stats["num_draft_tokens"], 6);
assert_eq!(stats["num_accepted_draft_tokens"], 3);
assert_eq!(stats["mean_acceptance_length"], 2.5);
assert_eq!(
stats["acceptance_histogram"],
serde_json::json!([0, 1, 1, 0])
);
assert!(stats.get("per_step_accepted").is_none());
}
}
}
#[test]
fn spec_decode_metrics_zero_steps_and_detailed() {
let empty = SpeculativeDecodingMetrics::from(RequestSpecDecodeMetrics {
num_spec_tokens: 2,
histogram: vec![0, 0, 0],
..Default::default()
});
assert_eq!(empty.mean_acceptance_length, 1.0);
assert_eq!(empty.draft_acceptance_rate, 0.0);
assert_eq!(empty.per_step_accepted, None);
let detailed = SpeculativeDecodingMetrics::from(RequestSpecDecodeMetrics {
num_spec_tokens: 2,
histogram: vec![0, 1, 1],
num_draft_tokens: 4,
per_step_accepted: vec![1, 2],
per_step_drafted: vec![2, 2],
});
assert_eq!(detailed.per_step_accepted, Some(vec![1, 2]));
assert_eq!(detailed.per_step_drafted, Some(vec![2, 2]));
}
#[test]
fn collect_generate_omits_prompt_metadata_by_default() {
let output = CollectedGenerateOutput {
@@ -662,6 +754,7 @@ mod tests {
ec_transfer_params: None,
prompt_token_ids: vec![10, 20],
sampling_mask: None,
spec_decode_metrics: None,
};
let response = collect_generate(
@@ -683,6 +776,50 @@ mod tests {
assert!(response.mm_placeholders.is_none());
}
#[test]
fn collect_generate_returns_spec_decode_metrics() {
let metrics = RequestSpecDecodeMetrics {
num_spec_tokens: 3,
histogram: vec![0, 1, 1, 0],
num_draft_tokens: 6,
..Default::default()
};
let output = CollectedGenerateOutput {
request_id: "raw-metrics".to_string(),
prompt_logprobs: None,
token_ids: vec![30],
logprobs: None,
finish_reason: FinishReason::stop_eos(),
usage: TokenUsage::default(),
kv_transfer_params: None,
ec_transfer_params: None,
prompt_token_ids: vec![10],
sampling_mask: None,
spec_decode_metrics: Some(metrics.clone()),
};
let response = collect_generate(
output,
"raw-metrics".to_string(),
ApiServerOptions::default(),
ResponseOptions::default(),
None,
)
.expect("response");
assert_eq!(
response.metrics,
Some(PerRequestMetrics {
speculative_decoding: SpeculativeDecodingMetrics::from(metrics)
})
);
let json = serde_json::to_value(&response).expect("serialize response");
let stats = &json["metrics"]["speculative_decoding"];
assert_eq!(stats.as_object().expect("metrics object").len(), 9);
assert!(stats["per_step_accepted"].is_null());
assert!(stats["per_step_drafted"].is_null());
}
#[test]
fn extract_mm_placeholders_returns_none_without_features() {
assert!(extract_mm_placeholders(None).is_none());
@@ -749,6 +886,7 @@ mod tests {
ec_transfer_params: None,
prompt_token_ids: vec![10, 20],
sampling_mask: None,
spec_decode_metrics: None,
};
let response = collect_generate(
@@ -782,6 +920,7 @@ mod tests {
kv_transfer_params: None,
ec_transfer_params: None,
sampling_mask: None,
spec_decode_metrics: None,
prompt_token_ids,
};
@@ -7,6 +7,7 @@ use llm_multimodal::MediaContentPart;
use serde::{Deserialize, Serialize};
use serde_json::{Map, Value};
use validator::Validate;
use vllm_engine_core_client::protocol::output::RequestSpecDecodeMetrics;
use vllm_text::SamplingParams;
use crate::routes::openai::utils::types::{ChatLogProbs, Normalizable, StreamOptions, Usage};
@@ -84,6 +85,7 @@ pub(super) struct GenerateStreamResponse {
pub usage: Option<Usage>,
pub prompt_token_ids: Option<Vec<u32>>,
pub mm_placeholders: Option<MultiModalPlaceholders>,
pub metrics: Option<PerRequestMetrics<StreamingSpeculativeDecodingMetrics>>,
}
/// Mirrors the Python vLLM `GenerateResponse` class.
@@ -96,6 +98,91 @@ pub(super) struct GenerateResponse {
pub mm_placeholders: Option<MultiModalPlaceholders>,
pub kv_transfer_params: Option<Value>,
pub ec_transfer_params: Option<Value>,
pub metrics: Option<PerRequestMetrics<SpeculativeDecodingMetrics>>,
}
#[derive(Debug, Clone, PartialEq, Serialize)]
pub(super) struct PerRequestMetrics<T> {
pub speculative_decoding: T,
}
/// Mirrors the Python vLLM `SpeculativeDecodingMetrics` class.
///
/// Derived from the raw engine accumulator the same way as Python
/// `RequestSpecDecodeMetrics.to_dict`.
#[derive(Debug, Clone, PartialEq, Serialize)]
pub(super) struct SpeculativeDecodingMetrics {
pub mean_acceptance_length: f64,
pub draft_acceptance_rate: f64,
pub acceptance_histogram: Vec<u64>,
pub num_spec_steps: u64,
pub num_accepted_draft_tokens: u64,
pub num_draft_tokens: u64,
pub num_spec_tokens: u64,
pub per_step_accepted: Option<Vec<u64>>,
pub per_step_drafted: Option<Vec<u64>>,
}
impl From<RequestSpecDecodeMetrics> for SpeculativeDecodingMetrics {
fn from(raw: RequestSpecDecodeMetrics) -> Self {
let num_spec_steps: u64 = raw.histogram.iter().sum();
let num_accepted_draft_tokens: u64 =
(0u64..).zip(&raw.histogram).map(|(accepted, count)| accepted * count).sum();
let ratio = |num: u64, den: u64| {
if den == 0 {
0.0
} else {
num as f64 / den as f64
}
};
let detailed = !raw.per_step_accepted.is_empty();
Self {
mean_acceptance_length: if num_spec_steps == 0 {
1.0
} else {
1.0 + ratio(num_accepted_draft_tokens, num_spec_steps)
},
draft_acceptance_rate: ratio(num_accepted_draft_tokens, raw.num_draft_tokens),
acceptance_histogram: raw.histogram,
num_spec_steps,
num_accepted_draft_tokens,
num_draft_tokens: raw.num_draft_tokens,
num_spec_tokens: raw.num_spec_tokens,
per_step_accepted: detailed.then_some(raw.per_step_accepted),
per_step_drafted: detailed.then_some(raw.per_step_drafted),
}
}
}
/// Streaming form omits detailed fields when summary metrics are requested.
#[serde_with::skip_serializing_none]
#[derive(Debug, Clone, PartialEq, Serialize)]
pub(super) struct StreamingSpeculativeDecodingMetrics {
pub mean_acceptance_length: f64,
pub draft_acceptance_rate: f64,
pub acceptance_histogram: Vec<u64>,
pub num_spec_steps: u64,
pub num_accepted_draft_tokens: u64,
pub num_draft_tokens: u64,
pub num_spec_tokens: u64,
pub per_step_accepted: Option<Vec<u64>>,
pub per_step_drafted: Option<Vec<u64>>,
}
impl From<SpeculativeDecodingMetrics> for StreamingSpeculativeDecodingMetrics {
fn from(metrics: SpeculativeDecodingMetrics) -> Self {
Self {
mean_acceptance_length: metrics.mean_acceptance_length,
draft_acceptance_rate: metrics.draft_acceptance_rate,
acceptance_histogram: metrics.acceptance_histogram,
num_spec_steps: metrics.num_spec_steps,
num_accepted_draft_tokens: metrics.num_accepted_draft_tokens,
num_draft_tokens: metrics.num_draft_tokens,
num_spec_tokens: metrics.num_spec_tokens,
per_step_accepted: metrics.per_step_accepted,
per_step_drafted: metrics.per_step_drafted,
}
}
}
pub(super) type MultiModalPlaceholders = HashMap<String, Vec<PlaceholderRangeInfo>>;
@@ -52,6 +52,7 @@ def _make_generate_response(
logprobs: dict | None = None,
prompt_logprobs: list | None = None,
kv_transfer_params: dict | None = None,
metrics: dict | None = None,
) -> dict:
choice: dict = {
"index": 0,
@@ -64,6 +65,7 @@ def _make_generate_response(
"choices": [choice],
"prompt_logprobs": prompt_logprobs,
"kv_transfer_params": kv_transfer_params,
"metrics": metrics,
}
@@ -266,6 +268,49 @@ async def test_derender_chat_kv_transfer_params_passthrough(client):
assert response.json()["kv_transfer_params"] == kv
@pytest.mark.asyncio
async def test_derender_chat_metrics_passthrough(client):
gen_req = await _render_chat(client)
metrics = {
"speculative_decoding": {
"mean_acceptance_length": 2.0,
"draft_acceptance_rate": 0.5,
"acceptance_histogram": [0, 1],
"num_spec_steps": 1,
"num_accepted_draft_tokens": 1,
"num_draft_tokens": 2,
"num_spec_tokens": 1,
}
}
response = await client.post(
"/v1/chat/completions/derender",
json={
"model": MODEL_NAME,
"generate_response": _make_generate_response(
gen_req["token_ids"][:3], metrics=metrics
),
},
)
assert response.status_code == 200
actual = response.json()["metrics"]
assert set(actual) == {
"time_to_first_token_ms",
"generation_time_ms",
"queue_time_ms",
"mean_itl_ms",
"tokens_per_second",
"speculative_decoding",
}
assert all(
actual[name] is None for name in actual if name != "speculative_decoding"
)
assert actual["speculative_decoding"] == {
**metrics["speculative_decoding"],
"per_step_accepted": None,
"per_step_drafted": None,
}
@pytest.mark.asyncio
async def test_derender_chat_empty_token_ids(client):
"""Empty token_ids list returns 400."""
@@ -27,6 +27,8 @@ from vllm.entrypoints.generate.base.protocol import (
DeltaFunctionCall,
DeltaMessage,
DeltaToolCall,
PerRequestMetrics,
SpeculativeDecodingMetrics,
)
from vllm.entrypoints.scale_out.token_in_token_out.protocol import (
DerenderStreamState,
@@ -164,6 +166,7 @@ def _make_stream_chunk(
finish_reason: str | None = None,
request_id: str = "test-req",
usage: dict | None = None,
metrics: PerRequestMetrics | None = None,
) -> GenerateStreamResponse:
"""Build a GenerateStreamResponse SSE chunk."""
return GenerateStreamResponse(
@@ -176,6 +179,7 @@ def _make_stream_chunk(
)
],
usage=UsageInfo(**usage) if usage else None,
metrics=metrics,
)
@@ -512,6 +516,26 @@ class TestDerenderChatStream:
assert chunk1.choices[0].delta.role == "assistant"
assert chunk2.choices[0].delta.role is None
@pytest.mark.asyncio
async def test_metrics_passthrough(self, derenderer):
metrics = PerRequestMetrics(
speculative_decoding=SpeculativeDecodingMetrics(
mean_acceptance_length=2.0,
draft_acceptance_rate=0.5,
acceptance_histogram=[0, 1],
num_spec_steps=1,
num_accepted_draft_tokens=1,
num_draft_tokens=2,
num_spec_tokens=1,
)
)
chunk, _ = await derenderer.derender_chat_stream(
model=MODEL_NAME,
generate_chunk=_make_stream_chunk([], metrics=metrics),
state=DerenderStreamState(),
)
assert chunk.metrics == metrics
@pytest.mark.asyncio
async def test_chunked_equals_oneshot(self, derenderer, tokenizer):
"""Sum of streaming content deltas == one shot decode."""
@@ -28,6 +28,7 @@ from vllm.renderers import renderer_from_config
from vllm.renderers.online_renderer import OnlineRenderer
from vllm.sampling_params import SamplingParams
from vllm.v1.engine.async_llm import AsyncLLM
from vllm.v1.metrics.stats import RequestSpecDecodeMetrics
MODEL_NAME = "openai-community/gpt2"
BASE_MODEL_PATHS = [
@@ -130,6 +131,7 @@ def _make_request_output(
logprobs: list[dict[int, Any] | None] | None = None,
num_cached_tokens: int | None = None,
index: int = 0,
spec_decode_metrics: RequestSpecDecodeMetrics | None = None,
) -> RequestOutput:
return RequestOutput(
request_id=request_id,
@@ -144,6 +146,7 @@ def _make_request_output(
cumulative_logprob=None,
logprobs=logprobs,
finish_reason=finish_reason,
spec_decode_metrics=spec_decode_metrics,
)
],
finished=finished,
@@ -321,6 +324,145 @@ async def test_stream_basic():
assert data_chunks[2]["choices"][0]["finish_reason"] == "stop"
@pytest.mark.asyncio
@pytest.mark.parametrize("stream", [False, True])
async def test_serve_tokens_returns_spec_decode_metrics(stream: bool):
engine = _mock_engine()
metrics = RequestSpecDecodeMetrics.new(num_spec_tokens=3)
metrics.observe(num_draft_tokens=3, num_accepted=2)
async def mock_generate(*args, **kwargs):
yield _make_request_output(
"req-1",
token_ids=[10],
finish_reason="stop",
finished=True,
spec_decode_metrics=metrics,
)
engine.generate = MagicMock(side_effect=mock_generate)
serving = _build_serving_tokens(engine)
request = GenerateRequest(
token_ids=[1, 2, 3],
sampling_params=SamplingParams(max_tokens=1),
model=MODEL_NAME,
stream=stream,
stream_options=StreamOptions(include_usage=True) if stream else None,
)
response = await serving.serve_tokens(request)
if stream:
chunks = [chunk async for chunk in response]
payload = next(
chunk
for chunk in _parse_sse_chunks(chunks)
if isinstance(chunk, dict) and chunk.get("metrics")
)["metrics"]["speculative_decoding"]
else:
assert isinstance(response, GenerateResponse)
assert response.metrics is not None
assert response.metrics.speculative_decoding is not None
payload = response.metrics.speculative_decoding.model_dump()
assert payload["num_draft_tokens"] == 3
assert payload["num_accepted_draft_tokens"] == 2
assert payload["num_spec_steps"] == 1
@pytest.mark.asyncio
@pytest.mark.parametrize("include_usage", [False, True])
async def test_stream_returns_spec_decode_metrics_on_empty_terminal_output(
include_usage,
):
engine = _mock_engine()
metrics = RequestSpecDecodeMetrics.new(num_spec_tokens=3)
metrics.observe(num_draft_tokens=3, num_accepted=2)
async def mock_generate(*args, **kwargs):
yield _make_request_output("req-1", token_ids=[10])
yield _make_request_output(
"req-1",
token_ids=[],
finish_reason="stop",
finished=True,
spec_decode_metrics=metrics,
)
engine.generate = MagicMock(side_effect=mock_generate)
serving = _build_serving_tokens(engine)
request = GenerateRequest(
token_ids=[1, 2, 3],
sampling_params=SamplingParams(max_tokens=1),
model=MODEL_NAME,
stream=True,
stream_options=StreamOptions(include_usage=include_usage),
)
chunks = [chunk async for chunk in await serving.serve_tokens(request)]
data_chunks = [chunk for chunk in _parse_sse_chunks(chunks) if chunk != "[DONE]"]
assert "metrics" not in data_chunks[0]
assert len(data_chunks) == 1 + int(include_usage)
if include_usage:
assert data_chunks[1]["choices"] == []
assert (
data_chunks[1]["metrics"]["speculative_decoding"]["num_draft_tokens"] == 3
)
@pytest.mark.asyncio
@pytest.mark.parametrize("stream", [False, True])
async def test_serve_tokens_omits_spec_decode_metrics_for_parallel_sampling(
stream: bool,
):
engine = _mock_engine()
metrics = RequestSpecDecodeMetrics.new(num_spec_tokens=3)
metrics.observe(num_draft_tokens=3, num_accepted=2)
result = _make_request_output(
"req-1",
token_ids=[10],
finish_reason="stop",
finished=True,
spec_decode_metrics=metrics,
)
result.outputs.append(
CompletionOutput(
index=1,
text="",
token_ids=[20],
cumulative_logprob=None,
logprobs=None,
finish_reason="stop",
spec_decode_metrics=metrics,
)
)
async def mock_generate(*args, **kwargs):
yield result
engine.generate = MagicMock(side_effect=mock_generate)
serving = _build_serving_tokens(engine)
request = GenerateRequest(
token_ids=[1, 2, 3],
sampling_params=SamplingParams(max_tokens=1, n=2),
model=MODEL_NAME,
stream=stream,
stream_options=StreamOptions(include_usage=True) if stream else None,
)
response = await serving.serve_tokens(request)
if stream:
chunks = [chunk async for chunk in response]
assert all(
"metrics" not in chunk
for chunk in _parse_sse_chunks(chunks)
if isinstance(chunk, dict)
)
else:
assert isinstance(response, GenerateResponse)
assert response.metrics is None
@pytest.mark.asyncio
async def test_non_stream_returns_final_prompt_token_ids_when_requested():
engine = _mock_engine()
@@ -179,6 +179,7 @@ class ServingDerender(BaseServing):
usage=usage,
prompt_logprobs=gen.prompt_logprobs,
kv_transfer_params=gen.kv_transfer_params,
metrics=gen.metrics,
)
async def derender_completion_response(
@@ -245,6 +246,9 @@ class ServingDerender(BaseServing):
choices=choices,
usage=usage,
kv_transfer_params=kv_params,
# Metrics describe one prompt. Parallel samples are already
# suppressed by /generate; multi-prompt responses cannot be merged.
metrics=first.metrics if len(request.generate_responses) == 1 else None,
)
async def derender_chat_stream_response(
@@ -11,7 +11,11 @@ from pydantic import (
)
from vllm.config import ModelConfig
from vllm.entrypoints.generate.base.protocol import StreamOptions, validate_cache_salt
from vllm.entrypoints.generate.base.protocol import (
PerRequestMetrics,
StreamOptions,
validate_cache_salt,
)
from vllm.entrypoints.openai.chat_completion.protocol import (
ChatCompletionLogProbs,
ChatCompletionRequest,
@@ -360,6 +364,7 @@ class GenerateStreamResponse(BaseModel):
usage: UsageInfo | None = Field(default=None)
prompt_token_ids: list[int] | None = None
mm_placeholders: dict[str, list[PlaceholderRangeInfo]] | None = None
metrics: PerRequestMetrics | None = None
class GenerateResponse(BaseModel):
@@ -378,6 +383,7 @@ class GenerateResponse(BaseModel):
prompt_logprobs: list[dict[int, Logprob] | None] | None = None
prompt_token_ids: list[int] | None = None
mm_placeholders: dict[str, list[PlaceholderRangeInfo]] | None = None
metrics: PerRequestMetrics | None = None
kv_transfer_params: dict[str, Any] | None = Field(
default=None,
@@ -12,9 +12,13 @@ from fastapi import Request
from vllm.engine.protocol import EngineClient
from vllm.entrypoints.chat_utils import AsyncMultiModalItemTracker
from vllm.entrypoints.generate.base.protocol import RequestResponseMetadata
from vllm.entrypoints.generate.base.protocol import (
PerRequestMetrics,
RequestResponseMetadata,
)
from vllm.entrypoints.generate.base.serving import (
GenerateBaseServing,
build_spec_decoding_metrics,
clamp_prompt_logprobs,
)
from vllm.entrypoints.openai.chat_completion.protocol import (
@@ -373,6 +377,11 @@ class ServingTokens(GenerateBaseServing):
request_metadata.final_usage_info = usage
per_request_metrics = None
if request.sampling_params.n == 1:
spec_stats = build_spec_decoding_metrics(final_res)
if spec_stats is not None:
per_request_metrics = PerRequestMetrics(speculative_decoding=spec_stats)
response = GenerateResponse(
request_id=request_id,
created=created_time,
@@ -384,6 +393,7 @@ class ServingTokens(GenerateBaseServing):
final_res.prompt_token_ids if request.return_token_ids else None
),
mm_placeholders=request._response_mm_placeholders,
metrics=per_request_metrics,
kv_transfer_params=final_res.kv_transfer_params,
ec_transfer_params=final_res.ec_transfer_params,
)
@@ -423,6 +433,7 @@ class ServingTokens(GenerateBaseServing):
prompt_token_ids: list[int] | None = None
num_cached_tokens = None
sampling_params: SamplingParams = request.sampling_params
last_res: RequestOutput | None = None
include_usage, include_continuous_usage = should_include_usage(
request.stream_options, False
@@ -430,6 +441,7 @@ class ServingTokens(GenerateBaseServing):
try:
async for res in result_generator:
last_res = res
if first_iteration:
if res.prompt_token_ids is not None:
num_prompt_tokens = len(res.prompt_token_ids)
@@ -496,10 +508,10 @@ class ServingTokens(GenerateBaseServing):
total_tokens=(num_prompt_tokens + num_generated_tokens[i]),
)
# Omit absent prompt metadata (Rust skips None fields).
# Omit fields that are absent from token-bearing chunks.
exclude = {
name
for name in ("prompt_token_ids", "mm_placeholders")
for name in ("prompt_token_ids", "mm_placeholders", "metrics")
if getattr(chunk, name) is None
}
yield f"data: {chunk.model_dump_json(exclude=exclude)}\n\n"
@@ -517,10 +529,18 @@ class ServingTokens(GenerateBaseServing):
)
if include_usage:
per_request_metrics = None
if sampling_params.n == 1:
spec_stats = build_spec_decoding_metrics(last_res)
if spec_stats is not None:
per_request_metrics = PerRequestMetrics(
speculative_decoding=spec_stats
)
final_chunk = GenerateStreamResponse(
request_id=request_id,
choices=[],
usage=final_usage_info,
metrics=per_request_metrics,
)
yield f"data: {final_chunk.model_dump_json(exclude_none=True)}\n\n"
+3
View File
@@ -398,6 +398,7 @@ class OnlineDerenderer:
model=model,
choices=stream_choices,
usage=usage,
metrics=generate_chunk.metrics,
)
return chunk, updated_state
@@ -595,6 +596,7 @@ class OnlineDerenderer:
model=model,
choices=stream_choices,
usage=usage,
metrics=generate_chunk.metrics,
)
return chunk, updated_state
@@ -741,6 +743,7 @@ class OnlineDerenderer:
model=model,
choices=stream_choices,
usage=usage,
metrics=generate_chunk.metrics,
)
return chunk, updated_state