[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:
co-authored by
aoshen02
Codex
parent
bf01fc4a31
commit
10e6a7f210
@@ -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 {
|
||||
|
||||
@@ -11,7 +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()
|
||||
{
|
||||
|
||||
@@ -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>>;
|
||||
|
||||
Reference in New Issue
Block a user