huggingface/text-generation-inference
5
0
Fork 0
mirror of https://github.com/huggingface/text-generation-inference.git synced 2025-09-09 19:34:53 +00:00
Code Issues Packages Projects Releases Wiki Activity

force as_secs

Browse Source
This commit is contained in:
OlivierDehaene 2023-04-09 18:30:08 +02:00
parent 23b55861bc
commit a3bdaca014
2 changed files with 104 additions and 91 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
router/src/infer.rs
View File

@ -335,11 +335,12 @@ async fn prefill(
) -> Option<Batch> { ) -> Option<Batch> {
let start_time = Instant::now(); let start_time = Instant::now();
let batch_id = batch.id; let batch_id = batch.id;
metrics::increment_counter!("tgi_batch_inference_count", "method" => "prefill");
match client.prefill(batch).await { match client.prefill(batch).await {
Ok((generations, next_batch)) => { Ok((generations, next_batch)) => {
send_generations(generations, entries); send_generations(generations, entries);
metrics::histogram!("tgi_batch_inference_duration", start_time.elapsed(), "method" => "prefill"); metrics::histogram!("tgi_batch_inference_duration", start_time.elapsed().as_secs_f64(), "method" => "prefill");
metrics::increment_counter!("tgi_batch_inference_success", "method" => "prefill"); metrics::increment_counter!("tgi_batch_inference_success", "method" => "prefill");
next_batch next_batch
} }
@ -360,11 +361,12 @@ async fn decode(
entries: &mut IntMap<u64, Entry>, entries: &mut IntMap<u64, Entry>,
) -> Option<Batch> { ) -> Option<Batch> {
let start_time = Instant::now(); let start_time = Instant::now();
metrics::increment_counter!("tgi_batch_inference_count", "method" => "decode");
match client.decode(batches).await { match client.decode(batches).await {
Ok((generations, next_batch)) => { Ok((generations, next_batch)) => {
send_generations(generations, entries); send_generations(generations, entries);
metrics::histogram!("tgi_batch_inference_duration", start_time.elapsed(), "method" => "decode"); metrics::histogram!("tgi_batch_inference_duration", start_time.elapsed().as_secs_f64(), "method" => "decode");
metrics::increment_counter!("tgi_batch_inference_success", "method" => "decode"); metrics::increment_counter!("tgi_batch_inference_success", "method" => "decode");
next_batch next_batch
} }

129
router/src/server.rs
View File

@ -87,21 +87,21 @@ async fn health(infer: Extension<Infer>) -> Result<(), (StatusCode, Json<ErrorRe
/// Generate tokens /// Generate tokens
#[utoipa::path( #[utoipa::path(
post, post,
tag = "Text Generation Inference", tag = "Text Generation Inference",
path = "/generate", path = "/generate",
request_body = GenerateRequest, request_body = GenerateRequest,
responses( responses(
(status = 200, description = "Generated Text", body = GenerateResponse), (status = 200, description = "Generated Text", body = GenerateResponse),
(status = 424, description = "Generation Error", body = ErrorResponse, (status = 424, description = "Generation Error", body = ErrorResponse,
example = json ! ({"error": "Request failed during generation"})), example = json ! ({"error": "Request failed during generation"})),
(status = 429, description = "Model is overloaded", body = ErrorResponse, (status = 429, description = "Model is overloaded", body = ErrorResponse,
example = json ! ({"error": "Model is overloaded"})), example = json ! ({"error": "Model is overloaded"})),
(status = 422, description = "Input validation error", body = ErrorResponse, (status = 422, description = "Input validation error", body = ErrorResponse,
example = json ! ({"error": "Input validation error"})), example = json ! ({"error": "Input validation error"})),
(status = 500, description = "Incomplete generation", body = ErrorResponse, (status = 500, description = "Incomplete generation", body = ErrorResponse,
example = json ! ({"error": "Incomplete generation"})), example = json ! ({"error": "Incomplete generation"})),
) )
)] )]
#[instrument( #[instrument(
skip(infer), skip(infer),
@ -186,6 +186,14 @@ async fn generate(
let inference_time = Instant::now() - response.start; let inference_time = Instant::now() - response.start;
let time_per_token = inference_time / response.generated_text.generated_tokens; let time_per_token = inference_time / response.generated_text.generated_tokens;
// Tracing metadata
span.record("total_time", format!("{total_time:?}"));
span.record("validation_time", format!("{validation_time:?}"));
span.record("queue_time", format!("{queue_time:?}"));
span.record("inference_time", format!("{inference_time:?}"));
span.record("time_per_token", format!("{time_per_token:?}"));
span.record("seed", format!("{:?}", response.generated_text.seed));
// Headers // Headers
let mut headers = HeaderMap::new(); let mut headers = HeaderMap::new();
headers.insert("x-compute-type", "gpu+optimized".parse().unwrap()); headers.insert("x-compute-type", "gpu+optimized".parse().unwrap());
@ -218,22 +226,22 @@ async fn generate(
time_per_token.as_millis().to_string().parse().unwrap(), time_per_token.as_millis().to_string().parse().unwrap(),
); );
// Tracing metadata
span.record("total_time", format!("{total_time:?}"));
span.record("validation_time", format!("{validation_time:?}"));
span.record("queue_time", format!("{queue_time:?}"));
span.record("inference_time", format!("{inference_time:?}"));
span.record("time_per_token", format!("{time_per_token:?}"));
span.record("seed", format!("{:?}", response.generated_text.seed));
tracing::info!("Output: {}", response.generated_text.text);
// Metrics // Metrics
metrics::increment_counter!("tgi_request_success"); metrics::increment_counter!("tgi_request_success");
metrics::histogram!("tgi_request_duration", total_time); metrics::histogram!("tgi_request_duration", total_time.as_secs_f64());
metrics::histogram!("tgi_request_validation_duration", validation_time); metrics::histogram!(
metrics::histogram!("tgi_request_queue_duration", queue_time); "tgi_request_validation_duration",
metrics::histogram!("tgi_request_inference_duration", inference_time); validation_time.as_secs_f64()
metrics::histogram!("tgi_request_mean_time_per_token_duration", time_per_token); );
metrics::histogram!("tgi_request_queue_duration", queue_time.as_secs_f64());
metrics::histogram!(
"tgi_request_inference_duration",
inference_time.as_secs_f64()
);
metrics::histogram!(
"tgi_request_mean_time_per_token_duration",
time_per_token.as_secs_f64()
);
metrics::histogram!( metrics::histogram!(
"tgi_request_generated_tokens", "tgi_request_generated_tokens",
response.generated_text.generated_tokens as f64 response.generated_text.generated_tokens as f64
@ -245,6 +253,8 @@ async fn generate(
output_text = prompt + &output_text; output_text = prompt + &output_text;
} }
tracing::info!("Output: {}", output_text);
let response = GenerateResponse { let response = GenerateResponse {
generated_text: output_text, generated_text: output_text,
details, details,
@ -254,26 +264,26 @@ async fn generate(
/// Generate a stream of token using Server-Sent Events /// Generate a stream of token using Server-Sent Events
#[utoipa::path( #[utoipa::path(
post, post,
tag = "Text Generation Inference", tag = "Text Generation Inference",
path = "/generate_stream", path = "/generate_stream",
request_body = GenerateRequest, request_body = GenerateRequest,
responses( responses(
(status = 200, description = "Generated Text", body = StreamResponse, (status = 200, description = "Generated Text", body = StreamResponse,
content_type = "text/event-stream"), content_type = "text/event-stream"),
(status = 424, description = "Generation Error", body = ErrorResponse, (status = 424, description = "Generation Error", body = ErrorResponse,
example = json ! ({"error": "Request failed during generation"}), example = json ! ({"error": "Request failed during generation"}),
content_type = "text/event-stream"), content_type = "text/event-stream"),
(status = 429, description = "Model is overloaded", body = ErrorResponse, (status = 429, description = "Model is overloaded", body = ErrorResponse,
example = json ! ({"error": "Model is overloaded"}), example = json ! ({"error": "Model is overloaded"}),
content_type = "text/event-stream"), content_type = "text/event-stream"),
(status = 422, description = "Input validation error", body = ErrorResponse, (status = 422, description = "Input validation error", body = ErrorResponse,
example = json ! ({"error": "Input validation error"}), example = json ! ({"error": "Input validation error"}),
content_type = "text/event-stream"), content_type = "text/event-stream"),
(status = 500, description = "Incomplete generation", body = ErrorResponse, (status = 500, description = "Incomplete generation", body = ErrorResponse,
example = json ! ({"error": "Incomplete generation"}), example = json ! ({"error": "Incomplete generation"}),
content_type = "text/event-stream"), content_type = "text/event-stream"),
) )
)] )]
#[instrument( #[instrument(
skip(infer), skip(infer),
@ -370,15 +380,14 @@ async fn generate_stream(
span.record("inference_time", format!("{inference_time:?}")); span.record("inference_time", format!("{inference_time:?}"));
span.record("time_per_token", format!("{time_per_token:?}")); span.record("time_per_token", format!("{time_per_token:?}"));
span.record("seed", format!("{:?}", generated_text.seed)); span.record("seed", format!("{:?}", generated_text.seed));
tracing::info!(parent: &span, "Output: {}", generated_text.text);
// Metrics // Metrics
metrics::increment_counter!("tgi_request_success"); metrics::increment_counter!("tgi_request_success");
metrics::histogram!("tgi_request_duration", total_time); metrics::histogram!("tgi_request_duration", total_time.as_secs_f64());
metrics::histogram!("tgi_request_validation_duration", validation_time); metrics::histogram!("tgi_request_validation_duration", validation_time.as_secs_f64());
metrics::histogram!("tgi_request_queue_duration", queue_time); metrics::histogram!("tgi_request_queue_duration", queue_time.as_secs_f64());
metrics::histogram!("tgi_request_inference_duration", inference_time); metrics::histogram!("tgi_request_inference_duration", inference_time.as_secs_f64());
metrics::histogram!("tgi_request_mean_time_per_token_duration", time_per_token); metrics::histogram!("tgi_request_mean_time_per_token_duration", time_per_token.as_secs_f64());
metrics::histogram!("tgi_request_generated_tokens", generated_text.generated_tokens as f64); metrics::histogram!("tgi_request_generated_tokens", generated_text.generated_tokens as f64);
// StreamResponse // StreamResponse
@ -389,6 +398,8 @@ async fn generate_stream(
output_text = prompt + &output_text; output_text = prompt + &output_text;
} }
tracing::info!(parent: &span, "Output: {}", output_text);
let stream_token = StreamResponse { let stream_token = StreamResponse {
token, token,
generated_text: Some(output_text), generated_text: Some(output_text),
@ -436,10 +447,10 @@ async fn generate_stream(
/// Prometheus metrics scrape endpoint /// Prometheus metrics scrape endpoint
#[utoipa::path( #[utoipa::path(
get, get,
tag = "Text Generation Inference", tag = "Text Generation Inference",
path = "/metrics", path = "/metrics",
responses((status = 200, description = "Prometheus Metrics", body = String)) responses((status = 200, description = "Prometheus Metrics", body = String))
)] )]
async fn metrics(prom_handle: Extension<PrometheusHandle>) -> String { async fn metrics(prom_handle: Extension<PrometheusHandle>) -> String {
prom_handle.render() prom_handle.render()