mirror of
https://github.com/katanemo/plano.git
synced 2026-06-17 15:25:17 +02:00
add in tpot stat
setup check for first token as well as time per token after that
This commit is contained in:
parent
5543aa543f
commit
bf39fecd6d
2 changed files with 52 additions and 2 deletions
|
|
@ -18,6 +18,7 @@ pub struct WasmMetrics {
|
|||
pub active_http_calls: Gauge,
|
||||
pub ratelimited_rq: Counter,
|
||||
pub time_to_first_token: Histogram,
|
||||
pub time_per_output_token: Histogram,
|
||||
}
|
||||
|
||||
impl WasmMetrics {
|
||||
|
|
@ -26,6 +27,7 @@ impl WasmMetrics {
|
|||
active_http_calls: Gauge::new(String::from("active_http_calls")),
|
||||
ratelimited_rq: Counter::new(String::from("ratelimited_rq")),
|
||||
time_to_first_token: Histogram::new(String::from("time_to_first_token")),
|
||||
time_per_output_token: Histogram::new(String::from("time_per_output_token")),
|
||||
}
|
||||
}
|
||||
}
|
||||
|
|
|
|||
|
|
@ -38,6 +38,8 @@ pub struct StreamContext {
|
|||
start_time: Option<SystemTime>,
|
||||
ttft_recorded: bool,
|
||||
ttft_duration: Option<Duration>, // Store the duration directly
|
||||
first_token_processed: bool,
|
||||
last_token_time: Option<SystemTime>,
|
||||
}
|
||||
|
||||
impl StreamContext {
|
||||
|
|
@ -55,6 +57,8 @@ impl StreamContext {
|
|||
start_time: None,
|
||||
ttft_recorded: false,
|
||||
ttft_duration: None,
|
||||
first_token_processed: false,
|
||||
last_token_time: None,
|
||||
}
|
||||
}
|
||||
fn llm_provider(&self) -> &LlmProvider {
|
||||
|
|
@ -354,14 +358,14 @@ impl HttpContext for StreamContext {
|
|||
};
|
||||
self.response_tokens += token_count;
|
||||
|
||||
// Compute TFT if not already recorded
|
||||
// Compute TTFT if not already recorded
|
||||
if !self.ttft_recorded {
|
||||
if let Some(start_time) = self.start_time {
|
||||
match get_current_time() {
|
||||
Ok(current_time) => match current_time.duration_since(start_time) {
|
||||
Ok(duration) => {
|
||||
let duration_ms = duration.as_millis();
|
||||
debug!("Time to First Token (TFT): {} milliseconds", duration_ms);
|
||||
debug!("Time to First Token (TTFT): {} milliseconds", duration_ms);
|
||||
self.ttft_duration = Some(duration);
|
||||
self.metrics.time_to_first_token.record(duration_ms as u64);
|
||||
}
|
||||
|
|
@ -378,6 +382,50 @@ impl HttpContext for StreamContext {
|
|||
warn!("Start time was not recorded");
|
||||
}
|
||||
}
|
||||
// Check if first token was not processed yet, and if there are tokens in the response.
|
||||
// If so, set the last_token_time to now and set first_token_processed to true
|
||||
if !self.first_token_processed && token_count > 0 {
|
||||
self.first_token_processed = true;
|
||||
// Set last_token_time to now
|
||||
match get_current_time() {
|
||||
Ok(current_time) => {
|
||||
self.last_token_time = Some(current_time);
|
||||
}
|
||||
Err(e) => {
|
||||
warn!("Failed to get current time: {:?}", e);
|
||||
}
|
||||
}
|
||||
} else if self.first_token_processed && token_count == 0 {
|
||||
if let Some(last_token_time) = self.last_token_time {
|
||||
match get_current_time() {
|
||||
Ok(current_time) => {
|
||||
// record the time for the current output token and calculate the time per output token
|
||||
match current_time.duration_since(last_token_time) {
|
||||
Ok(duration) => {
|
||||
// Convert the duration to milliseconds
|
||||
let duration_ms = duration.as_millis();
|
||||
debug!(
|
||||
"Time for Current Output Token: {} milliseconds",
|
||||
duration_ms
|
||||
);
|
||||
// Record TPOT metric for historgram
|
||||
self.metrics
|
||||
.time_per_output_token
|
||||
.record(duration_ms as u64);
|
||||
}
|
||||
Err(e) => {
|
||||
warn!("SystemTime error: {:?}", e);
|
||||
}
|
||||
}
|
||||
// Set last_token_time to now
|
||||
self.last_token_time = Some(current_time);
|
||||
}
|
||||
Err(e) => {
|
||||
warn!("Failed to get current time: {:?}", e);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
} else {
|
||||
debug!("non streaming response");
|
||||
let chat_completions_response: ChatCompletionsResponse =
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue