openai.rs 46.3 KB
Newer Older
1
2
3
// SPDX-FileCopyrightText: Copyright (c) 2024-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
// SPDX-License-Identifier: Apache-2.0

4
5
6
7
8
9
use std::{
    collections::HashSet,
    sync::Arc,
    time::{SystemTime, UNIX_EPOCH},
};

10
use axum::{
11
    Json, Router,
12
    extract::State,
Ryan Olson's avatar
Ryan Olson committed
13
    http::{HeaderMap, StatusCode},
14
15
    response::{
        IntoResponse, Response,
16
        sse::{Event, KeepAlive, Sse},
17
18
19
    },
    routing::{get, post},
};
Ryan Olson's avatar
Ryan Olson committed
20
21
22
23
use dynamo_runtime::{
    pipeline::{AsyncEngineContextProvider, Context},
    protocols::annotated::AnnotationsProvider,
};
24
use futures::{StreamExt, stream};
25
26
27
use serde::{Deserialize, Serialize};

use super::{
28
29
    RouteDoc,
    disconnect::{ConnectionHandle, create_connection_monitor, monitor_for_disconnects},
30
    error::HttpError,
Ryan Olson's avatar
Ryan Olson committed
31
    metrics::{Endpoint, ResponseMetricCollector},
32
    service_v2,
33
};
34
use crate::protocols::openai::chat_completions::aggregator::ChatCompletionAggregator;
35
use crate::protocols::openai::{
36
    ParsingOptions,
37
38
39
40
    chat_completions::{NvCreateChatCompletionRequest, NvCreateChatCompletionResponse},
    completions::{NvCreateCompletionRequest, NvCreateCompletionResponse},
    embeddings::{NvCreateEmbeddingRequest, NvCreateEmbeddingResponse},
    responses::{NvCreateResponse, NvResponse},
41
};
42
use crate::request_template::RequestTemplate;
43
use crate::types::Annotated;
GuanLuo's avatar
GuanLuo committed
44
use crate::{discovery::ModelManager, preprocessor::LLMMetricAnnotation};
45
46
use dynamo_runtime::logging::get_distributed_tracing_context;
use tracing::Instrument;
47

Ryan Olson's avatar
Ryan Olson committed
48
49
50
51
52
pub const DYNAMO_REQUEST_ID_HEADER: &str = "x-dynamo-request-id";

/// Dynamo Annotation for the request ID
pub const ANNOTATION_REQUEST_ID: &str = "request_id";

53
54
55
56
57
58
59
60
61
62
63
// Default axum max body limit without configuring is 2MB: https://docs.rs/axum/latest/axum/extract/struct.DefaultBodyLimit.html
/// Default body limit in bytes (45MB) to support 500k+ token payloads.
/// Can be configured at compile time using the DYN_FRONTEND_BODY_LIMIT_MB environment variable
fn get_body_limit() -> usize {
    std::env::var("DYN_HTTP_BODY_LIMIT_MB")
        .ok()
        .and_then(|s| s.parse::<usize>().ok())
        .map(|mb| mb * 1024 * 1024)
        .unwrap_or(45 * 1024 * 1024)
}

Ryan Olson's avatar
Ryan Olson committed
64
65
pub type ErrorResponse = (StatusCode, Json<ErrorMessage>);

66
#[derive(Serialize, Deserialize)]
Ryan Olson's avatar
Ryan Olson committed
67
pub(crate) struct ErrorMessage {
68
69
70
    error: String,
}

Ryan Olson's avatar
Ryan Olson committed
71
impl ErrorMessage {
72
    /// Not Found Error
Ryan Olson's avatar
Ryan Olson committed
73
    pub fn model_not_found() -> ErrorResponse {
74
75
        (
            StatusCode::NOT_FOUND,
Ryan Olson's avatar
Ryan Olson committed
76
            Json(ErrorMessage {
77
78
79
80
81
82
83
                error: "Model not found".to_string(),
            }),
        )
    }

    /// Service Unavailable
    /// This is returned when the service is live, but not ready.
Ryan Olson's avatar
Ryan Olson committed
84
    pub fn _service_unavailable() -> ErrorResponse {
85
86
        (
            StatusCode::SERVICE_UNAVAILABLE,
Ryan Olson's avatar
Ryan Olson committed
87
            Json(ErrorMessage {
88
89
90
91
92
93
94
95
96
                error: "Service is not ready".to_string(),
            }),
        )
    }

    /// Internal Service Error
    /// Return this error when the service encounters an internal error.
    /// We should return a generic message to the client instead of the real error.
    /// Internal Services errors are the result of misconfiguration or bugs in the service.
Ryan Olson's avatar
Ryan Olson committed
97
    pub fn internal_server_error(msg: &str) -> ErrorResponse {
98
99
100
        tracing::error!("Internal server error: {msg}");
        (
            StatusCode::INTERNAL_SERVER_ERROR,
Ryan Olson's avatar
Ryan Olson committed
101
            Json(ErrorMessage {
102
103
104
105
106
                error: msg.to_string(),
            }),
        )
    }

107
108
109
    /// Not Implemented Error
    /// Return this error when the client requests a feature that is not yet implemented.
    /// This should be used for features that are planned but not available.
Ryan Olson's avatar
Ryan Olson committed
110
    pub fn not_implemented_error(msg: &str) -> ErrorResponse {
111
112
113
        tracing::error!("Not Implemented error: {msg}");
        (
            StatusCode::NOT_IMPLEMENTED,
Ryan Olson's avatar
Ryan Olson committed
114
            Json(ErrorMessage {
115
116
117
118
119
                error: msg.to_string(),
            }),
        )
    }

Neelay Shah's avatar
Neelay Shah committed
120
    /// The OAI endpoints call an [`dynamo.runtime::engine::AsyncEngine`] which are specialized to return
121
    /// an [`anyhow::Error`]. This method will convert the [`anyhow::Error`] into an [`HttpError`].
Ryan Olson's avatar
Ryan Olson committed
122
    /// If successful, it will return the [`HttpError`] as an [`ErrorMessage::internal_server_error`]
123
    /// with the details of the error.
Ryan Olson's avatar
Ryan Olson committed
124
    pub fn from_anyhow(err: anyhow::Error, alt_msg: &str) -> ErrorResponse {
125
126
127
        // First check for PipelineError::ServiceOverloaded
        if let Some(pipeline_err) =
            err.downcast_ref::<dynamo_runtime::pipeline::error::PipelineError>()
128
            && matches!(
129
130
                pipeline_err,
                dynamo_runtime::pipeline::error::PipelineError::ServiceOverloaded(_)
131
132
133
134
135
136
137
138
            )
        {
            return (
                StatusCode::SERVICE_UNAVAILABLE,
                Json(ErrorMessage {
                    error: pipeline_err.to_string(),
                }),
            );
139
140
141
        }

        // Then check for HttpError
142
        match err.downcast::<HttpError>() {
Ryan Olson's avatar
Ryan Olson committed
143
144
            Ok(http_error) => ErrorMessage::from_http_error(http_error),
            Err(err) => ErrorMessage::internal_server_error(&format!("{alt_msg}: {err}")),
145
146
147
148
        }
    }

    /// Implementers should only be able to throw 400-499 errors.
Ryan Olson's avatar
Ryan Olson committed
149
    pub fn from_http_error(err: HttpError) -> ErrorResponse {
150
        if err.code < 400 || err.code >= 500 {
Ryan Olson's avatar
Ryan Olson committed
151
            return ErrorMessage::internal_server_error(&err.message);
152
153
        }
        match StatusCode::from_u16(err.code) {
Ryan Olson's avatar
Ryan Olson committed
154
155
            Ok(code) => (code, Json(ErrorMessage { error: err.message })),
            Err(_) => ErrorMessage::internal_server_error(&err.message),
156
157
158
159
        }
    }
}

Ryan Olson's avatar
Ryan Olson committed
160
impl From<HttpError> for ErrorMessage {
161
    fn from(err: HttpError) -> Self {
Ryan Olson's avatar
Ryan Olson committed
162
        ErrorMessage { error: err.message }
163
164
165
    }
}

Ryan Olson's avatar
Ryan Olson committed
166
167
/// Get the request ID from a primary source, or next from the headers, or lastly create a new one if not present
fn get_or_create_request_id(primary: Option<&str>, headers: &HeaderMap) -> String {
168
    // Try to get request id from trace context
169
170
171
172
    if let Some(trace_context) = get_distributed_tracing_context()
        && let Some(x_dynamo_request_id) = trace_context.x_dynamo_request_id
    {
        return x_dynamo_request_id;
173
174
    }

Ryan Olson's avatar
Ryan Olson committed
175
    // Try to get the request ID from the primary source
176
177
178
179
    if let Some(primary) = primary
        && let Ok(uuid) = uuid::Uuid::parse_str(primary)
    {
        return uuid.to_string();
Ryan Olson's avatar
Ryan Olson committed
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
    }

    // Try to get the request ID header as a string slice
    let request_id_opt = headers
        .get(DYNAMO_REQUEST_ID_HEADER)
        .and_then(|h| h.to_str().ok());

    // Try to parse the request ID as a UUID, or generate a new one if missing/invalid
    let uuid = match request_id_opt {
        Some(request_id) => {
            uuid::Uuid::parse_str(request_id).unwrap_or_else(|_| uuid::Uuid::new_v4())
        }
        None => uuid::Uuid::new_v4(),
    };

    uuid.to_string()
}

GuanLuo's avatar
GuanLuo committed
198
199
fn get_parsing_options(manager: &ModelManager, model: &str) -> ParsingOptions {
    let tool_call_parser = manager.get_model_tool_call_parser(model);
200
201
202
203
204
    let reasoning_parser = None; // TODO: Implement reasoning parser

    ParsingOptions::new(tool_call_parser, reasoning_parser)
}

205
206
207
208
209
210
211
212
/// OpenAI Completions Request Handler
///
/// This method will handle the incoming request for the `/v1/completions endpoint`. The endpoint is a "source"
/// for an [`super::OpenAICompletionsStreamingEngine`] and will return a stream of
/// responses which will be forward to the client.
///
/// Note: For all requests, streaming or non-streaming, we always call the engine with streaming enabled. For
/// non-streaming requests, we will fold the stream into a single response as part of this handler.
Ryan Olson's avatar
Ryan Olson committed
213
async fn handler_completions(
214
    State(state): State<Arc<service_v2::State>>,
Ryan Olson's avatar
Ryan Olson committed
215
    headers: HeaderMap,
216
    Json(request): Json<NvCreateCompletionRequest>,
Ryan Olson's avatar
Ryan Olson committed
217
218
219
220
221
222
223
224
225
226
) -> Result<Response, ErrorResponse> {
    // return a 503 if the service is not ready
    check_ready(&state)?;

    // create the context for the request
    let request_id = get_or_create_request_id(request.inner.user.as_deref(), &headers);
    let request = Context::with_id(request, request_id);
    let context = request.context();

    // create the connection handles
227
228
    let (mut connection_handle, stream_handle) =
        create_connection_monitor(context.clone(), Some(state.metrics_clone())).await;
Ryan Olson's avatar
Ryan Olson committed
229
230
231

    // possibly long running task
    // if this returns a streaming response, the stream handle will be armed and captured by the response stream
232
    let response = tokio::spawn(completions(state, request, stream_handle).in_current_span())
Ryan Olson's avatar
Ryan Olson committed
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
        .await
        .map_err(|e| {
            ErrorMessage::internal_server_error(&format!(
                "Failed to await chat completions task: {:?}",
                e,
            ))
        })?;

    // if we got here, then we will return a response and the potentially long running task has completed successfully
    // without need to be cancelled.
    connection_handle.disarm();

    response
}

#[tracing::instrument(skip_all)]
async fn completions(
    state: Arc<service_v2::State>,
    request: Context<NvCreateCompletionRequest>,
    stream_handle: ConnectionHandle,
) -> Result<Response, ErrorResponse> {
254
255
256
    // return a 503 if the service is not ready
    check_ready(&state)?;

257
    let request_id = request.id().to_string();
258
259

    // todo - decide on default
260
    let streaming = request.inner.stream.unwrap_or(false);
261
262

    // update the request to always stream
Ryan Olson's avatar
Ryan Olson committed
263
264
265
266
    let request = request.map(|mut req| {
        req.inner.stream = Some(true);
        req
    });
267

268
269
    // todo - make the protocols be optional for model name
    // todo - when optional, if none, apply a default
270
    let model = &request.inner.model;
271
272
273

    // todo - error handling should be more robust
    let engine = state
274
        .manager()
275
        .get_completions_engine(model)
Ryan Olson's avatar
Ryan Olson committed
276
        .map_err(|_| ErrorMessage::model_not_found())?;
277

GuanLuo's avatar
GuanLuo committed
278
    let parsing_options = get_parsing_options(state.manager(), model);
279

280
    let mut inflight_guard =
281
282
283
        state
            .metrics_clone()
            .create_inflight_guard(model, Endpoint::Completions, streaming);
284

285
286
    let mut response_collector = state.metrics_clone().create_response_collector(model);

Ryan Olson's avatar
Ryan Olson committed
287
288
    // prepare to process any annotations
    let annotations = request.annotations();
289
290
291
292
293

    // issue the generate call on the engine
    let stream = engine
        .generate(request)
        .await
Ryan Olson's avatar
Ryan Olson committed
294
        .map_err(|e| ErrorMessage::from_anyhow(e, "Failed to generate completions"))?;
295
296
297
298

    // capture the context to cancel the stream if the client disconnects
    let ctx = stream.context();

Ryan Olson's avatar
Ryan Olson committed
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
    let annotations = annotations.map_or(Vec::new(), |annotations| {
        annotations
            .iter()
            .filter_map(|annotation| {
                if annotation == ANNOTATION_REQUEST_ID {
                    Annotated::<NvCreateCompletionResponse>::from_annotation(
                        ANNOTATION_REQUEST_ID,
                        &request_id,
                    )
                    .ok()
                } else {
                    None
                }
            })
            .collect::<Vec<_>>()
    });

    // apply any annotations to the front of the stream
    let stream = stream::iter(annotations).chain(stream);
318
319

    if streaming {
320
321
322
        let stream = stream.map(move |response| {
            process_event_converter(EventConverter::from(response), &mut response_collector)
        });
Ryan Olson's avatar
Ryan Olson committed
323
        let stream = monitor_for_disconnects(stream, ctx, inflight_guard, stream_handle);
324

325
326
        let mut sse_stream = Sse::new(stream);

327
        if let Some(keep_alive) = state.sse_keep_alive() {
328
329
330
331
            sse_stream = sse_stream.keep_alive(KeepAlive::default().interval(keep_alive));
        }

        Ok(sse_stream.into_response())
332
    } else {
333
334
335
336
337
        // Tap the stream to collect metrics for non-streaming requests without altering items
        let stream = stream.inspect(move |response| {
            process_metrics_only(response, &mut response_collector);
        });

338
        let response = NvCreateCompletionResponse::from_annotated_stream(stream, parsing_options)
339
340
341
342
343
344
345
            .await
            .map_err(|e| {
                tracing::error!(
                    "Failed to fold completions stream for {}: {:?}",
                    request_id,
                    e
                );
Ryan Olson's avatar
Ryan Olson committed
346
                ErrorMessage::internal_server_error("Failed to fold completions stream")
347
348
            })?;

349
        inflight_guard.mark_ok();
350
351
352
353
        Ok(Json(response).into_response())
    }
}

354
355
#[tracing::instrument(skip_all)]
async fn embeddings(
356
    State(state): State<Arc<service_v2::State>>,
357
    headers: HeaderMap,
358
    Json(request): Json<NvCreateEmbeddingRequest>,
Ryan Olson's avatar
Ryan Olson committed
359
) -> Result<Response, ErrorResponse> {
360
361
362
    // return a 503 if the service is not ready
    check_ready(&state)?;

363
364
365
    let request_id = get_or_create_request_id(request.inner.user.as_deref(), &headers);
    let request = Context::with_id(request, request_id);
    let request_id = request.id().to_string();
366
367
368
369
370
371
372
373
374
375
376
377

    // Embeddings are typically not streamed, so we default to non-streaming
    let streaming = false;

    // todo - make the protocols be optional for model name
    // todo - when optional, if none, apply a default
    let model = &request.inner.model;

    // todo - error handling should be more robust
    let engine = state
        .manager()
        .get_embeddings_engine(model)
Ryan Olson's avatar
Ryan Olson committed
378
        .map_err(|_| ErrorMessage::model_not_found())?;
379
380
381
382
383
384
385
386
387
388
389

    // this will increment the inflight gauge for the model
    let mut inflight =
        state
            .metrics_clone()
            .create_inflight_guard(model, Endpoint::Embeddings, streaming);

    // issue the generate call on the engine
    let stream = engine
        .generate(request)
        .await
Ryan Olson's avatar
Ryan Olson committed
390
        .map_err(|e| ErrorMessage::from_anyhow(e, "Failed to generate embeddings"))?;
391
392
393

    // Embeddings are typically returned as a single response (non-streaming)
    // so we fold the stream into a single response
Ryan Olson's avatar
Ryan Olson committed
394
    let response = NvCreateEmbeddingResponse::from_annotated_stream(stream)
395
396
397
398
399
400
401
        .await
        .map_err(|e| {
            tracing::error!(
                "Failed to fold embeddings stream for {}: {:?}",
                request_id,
                e
            );
Ryan Olson's avatar
Ryan Olson committed
402
            ErrorMessage::internal_server_error("Failed to fold embeddings stream")
403
404
405
406
        })?;

    inflight.mark_ok();
    Ok(Json(response).into_response())
407
408
}

Ryan Olson's avatar
Ryan Olson committed
409
410
411
412
413
414
415
416
417
418
419
420
421
422
async fn handler_chat_completions(
    State((state, template)): State<(Arc<service_v2::State>, Option<RequestTemplate>)>,
    headers: HeaderMap,
    Json(request): Json<NvCreateChatCompletionRequest>,
) -> Result<Response, ErrorResponse> {
    // return a 503 if the service is not ready
    check_ready(&state)?;

    // create the context for the request
    let request_id = get_or_create_request_id(request.inner.user.as_deref(), &headers);
    let request = Context::with_id(request, request_id);
    let context = request.context();

    // create the connection handles
423
424
    let (mut connection_handle, stream_handle) =
        create_connection_monitor(context.clone(), Some(state.metrics_clone())).await;
Ryan Olson's avatar
Ryan Olson committed
425

426
427
428
429
430
431
432
433
434
    let response =
        tokio::spawn(chat_completions(state, template, request, stream_handle).in_current_span())
            .await
            .map_err(|e| {
                ErrorMessage::internal_server_error(&format!(
                    "Failed to await chat completions task: {:?}",
                    e,
                ))
            })?;
Ryan Olson's avatar
Ryan Olson committed
435
436
437
438
439
440
441
442

    // if we got here, then we will return a response and the potentially long running task has completed successfully
    // without need to be cancelled.
    connection_handle.disarm();

    response
}

443
444
445
446
447
448
449
450
451
/// OpenAI Chat Completions Request Handler
///
/// This method will handle the incoming request for the /v1/chat/completions endpoint. The endpoint is a "source"
/// for an [`super::OpenAIChatCompletionsStreamingEngine`] and will return a stream of responses which will be
/// forward to the client.
///
/// Note: For all requests, streaming or non-streaming, we always call the engine with streaming enabled. For
/// non-streaming requests, we will fold the stream into a single response as part of this handler.
async fn chat_completions(
Ryan Olson's avatar
Ryan Olson committed
452
453
454
455
456
    state: Arc<service_v2::State>,
    template: Option<RequestTemplate>,
    mut request: Context<NvCreateChatCompletionRequest>,
    mut stream_handle: ConnectionHandle,
) -> Result<Response, ErrorResponse> {
457
458
459
    // return a 503 if the service is not ready
    check_ready(&state)?;

Ryan Olson's avatar
Ryan Olson committed
460
461
    let request_id = request.id().to_string();

462
463
464
465
    // Handle unsupported fields - if Some(resp) is returned by
    // validate_chat_completion_unsupported_fields,
    // then a field was used that is unsupported. We will log an error message
    // and early return a 501 NOT_IMPLEMENTED status code. Otherwise, proceeed.
Ryan Olson's avatar
Ryan Olson committed
466
    validate_chat_completion_unsupported_fields(&request)?;
467

468
469
470
    // Handle required fields like messages shouldn't be empty.
    validate_chat_completion_required_fields(&request)?;

471
472
473
474
475
476
477
478
479
480
481
482
    // Apply template values if present
    if let Some(template) = template {
        if request.inner.model.is_empty() {
            request.inner.model = template.model.clone();
        }
        if request.inner.temperature.unwrap_or(0.0) == 0.0 {
            request.inner.temperature = Some(template.temperature);
        }
        if request.inner.max_completion_tokens.unwrap_or(0) == 0 {
            request.inner.max_completion_tokens = Some(template.max_completion_tokens);
        }
    }
Ryan Olson's avatar
Ryan Olson committed
483
    tracing::trace!("Received chat completions request: {:?}", request.content());
484
485

    // todo - decide on default
Paul Hendricks's avatar
Paul Hendricks committed
486
    let streaming = request.inner.stream.unwrap_or(false);
487
488

    // update the request to always stream
Ryan Olson's avatar
Ryan Olson committed
489
490
491
492
    let request = request.map(|mut req| {
        req.inner.stream = Some(true);
        req
    });
493
494
495

    // todo - make the protocols be optional for model name
    // todo - when optional, if none, apply a default
Paul Hendricks's avatar
Paul Hendricks committed
496
    let model = &request.inner.model;
497
498
499
500
501

    // todo - determine the proper error code for when a request model is not present
    tracing::trace!("Getting chat completions engine for model: {}", model);

    let engine = state
502
        .manager()
503
        .get_chat_completions_engine(model)
Ryan Olson's avatar
Ryan Olson committed
504
        .map_err(|_| ErrorMessage::model_not_found())?;
505

GuanLuo's avatar
GuanLuo committed
506
    let parsing_options = get_parsing_options(state.manager(), model);
507

508
    let mut inflight_guard =
509
510
511
        state
            .metrics_clone()
            .create_inflight_guard(model, Endpoint::ChatCompletions, streaming);
512

513
514
    let mut response_collector = state.metrics_clone().create_response_collector(model);

515
    tracing::trace!("Issuing generate call for chat completions");
Ryan Olson's avatar
Ryan Olson committed
516
    let annotations = request.annotations();
517
518
519
520
521

    // issue the generate call on the engine
    let stream = engine
        .generate(request)
        .await
Ryan Olson's avatar
Ryan Olson committed
522
        .map_err(|e| ErrorMessage::from_anyhow(e, "Failed to generate completions"))?;
523
524
525
526

    // capture the context to cancel the stream if the client disconnects
    let ctx = stream.context();

Ryan Olson's avatar
Ryan Olson committed
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
    // prepare any requested annotations
    let annotations = annotations.map_or(Vec::new(), |annotations| {
        annotations
            .iter()
            .filter_map(|annotation| {
                if annotation == ANNOTATION_REQUEST_ID {
                    Annotated::from_annotation(ANNOTATION_REQUEST_ID, &request_id).ok()
                } else {
                    None
                }
            })
            .collect::<Vec<_>>()
    });

    // apply any annotations to the front of the stream
    let stream = stream::iter(annotations).chain(stream);

544
545
546
547
    // todo - tap the stream and propagate request level metrics
    // note - we might do this as part of the post processing set to make it more generic

    if streaming {
Ryan Olson's avatar
Ryan Olson committed
548
549
        stream_handle.arm();

550
551
552
        let stream = stream.map(move |response| {
            process_event_converter(EventConverter::from(response), &mut response_collector)
        });
Ryan Olson's avatar
Ryan Olson committed
553
        let stream = monitor_for_disconnects(stream, ctx, inflight_guard, stream_handle);
554

555
556
        let mut sse_stream = Sse::new(stream);

557
        if let Some(keep_alive) = state.sse_keep_alive() {
558
559
560
561
            sse_stream = sse_stream.keep_alive(KeepAlive::default().interval(keep_alive));
        }

        Ok(sse_stream.into_response())
562
    } else {
563
564
565
566
        let stream = stream.inspect(move |response| {
            process_metrics_only(response, &mut response_collector);
        });

567
568
569
570
571
572
573
574
575
576
577
578
579
580
        let response =
            NvCreateChatCompletionResponse::from_annotated_stream(stream, parsing_options.clone())
                .await
                .map_err(|e| {
                    tracing::error!(
                        request_id,
                        "Failed to fold chat completions stream for: {:?}",
                        e
                    );
                    ErrorMessage::internal_server_error(&format!(
                        "Failed to fold chat completions stream: {}",
                        e
                    ))
                })?;
581

582
        inflight_guard.mark_ok();
583
584
585
586
        Ok(Json(response).into_response())
    }
}

587
588
589
590
591
/// Checks for unsupported fields in the request.
/// Returns Some(response) if unsupported fields are present.
#[allow(deprecated)]
pub fn validate_chat_completion_unsupported_fields(
    request: &NvCreateChatCompletionRequest,
Ryan Olson's avatar
Ryan Olson committed
592
) -> Result<(), ErrorResponse> {
593
594
595
    let inner = &request.inner;

    if inner.parallel_tool_calls == Some(true) {
Ryan Olson's avatar
Ryan Olson committed
596
        return Err(ErrorMessage::not_implemented_error(
597
598
599
600
601
            "`parallel_tool_calls: true` is not supported.",
        ));
    }

    if inner.function_call.is_some() {
Ryan Olson's avatar
Ryan Olson committed
602
        return Err(ErrorMessage::not_implemented_error(
603
604
605
606
607
            "`function_call` is deprecated. Please migrate to use `tool_choice` instead.",
        ));
    }

    if inner.functions.is_some() {
Ryan Olson's avatar
Ryan Olson committed
608
        return Err(ErrorMessage::not_implemented_error(
609
610
611
612
            "`functions` is deprecated. Please migrate to use `tools` instead.",
        ));
    }

Ryan Olson's avatar
Ryan Olson committed
613
    Ok(())
614
615
}

616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
/// Validates that required fields are present and valid in the chat completion request
pub fn validate_chat_completion_required_fields(
    request: &NvCreateChatCompletionRequest,
) -> Result<(), ErrorResponse> {
    let inner = &request.inner;

    if inner.messages.is_empty() {
        return Err(ErrorMessage::from_http_error(HttpError {
            code: 400,
            message: "The 'messages' field cannot be empty. At least one message is required."
                .to_string(),
        }));
    }

    Ok(())
}

633
634
635
/// OpenAI Responses Request Handler
///
/// This method will handle the incoming request for the /v1/responses endpoint.
Ryan Olson's avatar
Ryan Olson committed
636
async fn handler_responses(
637
    State((state, template)): State<(Arc<service_v2::State>, Option<RequestTemplate>)>,
Ryan Olson's avatar
Ryan Olson committed
638
639
640
641
642
643
644
645
646
647
648
649
    headers: HeaderMap,
    Json(request): Json<NvCreateResponse>,
) -> Result<Response, ErrorResponse> {
    // return a 503 if the service is not ready
    check_ready(&state)?;

    // create the context for the request
    let request_id = get_or_create_request_id(request.inner.user.as_deref(), &headers);
    let request = Context::with_id(request, request_id);
    let context = request.context();

    // create the connection handles
650
651
    let (mut connection_handle, _stream_handle) =
        create_connection_monitor(context.clone(), Some(state.metrics_clone())).await;
Ryan Olson's avatar
Ryan Olson committed
652

653
    let response = tokio::spawn(responses(state, template, request).in_current_span())
Ryan Olson's avatar
Ryan Olson committed
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
        .await
        .map_err(|e| {
            ErrorMessage::internal_server_error(&format!(
                "Failed to await chat completions task: {:?}",
                e,
            ))
        })?;

    // if we got here, then we will return a response and the potentially long running task has completed successfully
    // without need to be cancelled.
    connection_handle.disarm();

    response
}

#[tracing::instrument(level = "debug", skip_all, fields(request_id = %request.id()))]
async fn responses(
    state: Arc<service_v2::State>,
    template: Option<RequestTemplate>,
    mut request: Context<NvCreateResponse>,
) -> Result<Response, ErrorResponse> {
675
676
677
678
679
680
    // return a 503 if the service is not ready
    check_ready(&state)?;

    // Handle unsupported fields - if Some(resp) is returned by validate_unsupported_fields,
    // then a field was used that is unsupported. We will log an error message
    // and early return a 501 NOT_IMPLEMENTED status code. Otherwise, proceeed.
681
    if let Some(resp) = validate_response_unsupported_fields(&request) {
682
683
684
685
686
687
688
        return Ok(resp.into_response());
    }

    // Handle non-text (image, audio, file) inputs - if Some(resp) is returned by
    // validate_input_is_text_only, then we are handling something other than Input::Text(_).
    // We will log an error message and early return a 501 NOT_IMPLEMENTED status code.
    // Otherwise, proceeed.
689
    if let Some(resp) = validate_response_input_is_text_only(&request) {
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
        return Ok(resp.into_response());
    }

    // Apply template values if present
    if let Some(template) = template {
        if request.inner.model.is_empty() {
            request.inner.model = template.model.clone();
        }
        if request.inner.temperature.unwrap_or(0.0) == 0.0 {
            request.inner.temperature = Some(template.temperature);
        }
        if request.inner.max_output_tokens.unwrap_or(0) == 0 {
            request.inner.max_output_tokens = Some(template.max_completion_tokens);
        }
    }
    tracing::trace!("Received chat completions request: {:?}", request.inner);

Ryan Olson's avatar
Ryan Olson committed
707
708
    let request_id = request.id().to_string();
    let (request, context) = request.into_parts();
709

Ryan Olson's avatar
Ryan Olson committed
710
    let mut request: NvCreateChatCompletionRequest = request.try_into().map_err(|e| {
711
712
713
714
715
        tracing::error!(
            request_id,
            "Failed to convert NvCreateResponse to NvCreateChatCompletionRequest: {:?}",
            e
        );
Ryan Olson's avatar
Ryan Olson committed
716
        ErrorMessage::not_implemented_error(&format!(
717
718
719
720
721
            "Only Input::Text(_) is currently supported: {}",
            e
        ))
    })?;

Ryan Olson's avatar
Ryan Olson committed
722
723
724
725
726
    let request = context.map(|mut _req| {
        request.inner.stream = Some(false);
        request
    });

727
728
729
730
731
732
733
    let model = &request.inner.model;

    tracing::trace!("Getting chat completions engine for model: {}", model);

    let engine = state
        .manager()
        .get_chat_completions_engine(model)
Ryan Olson's avatar
Ryan Olson committed
734
        .map_err(|_| ErrorMessage::model_not_found())?;
735

GuanLuo's avatar
GuanLuo committed
736
    let parsing_options = get_parsing_options(state.manager(), model);
737

738
739
740
741
742
743
744
745
746
747
748
749
750
    let mut inflight_guard =
        state
            .metrics_clone()
            .create_inflight_guard(model, Endpoint::Responses, false);

    let _response_collector = state.metrics_clone().create_response_collector(model);

    tracing::trace!("Issuing generate call for chat completions");

    // issue the generate call on the engine
    let stream = engine
        .generate(request)
        .await
Ryan Olson's avatar
Ryan Olson committed
751
        .map_err(|e| ErrorMessage::from_anyhow(e, "Failed to generate completions"))?;
752
753

    // TODO: handle streaming, currently just unary
754
755
756
757
758
759
760
761
762
763
764
765
766
767
    let response =
        NvCreateChatCompletionResponse::from_annotated_stream(stream, parsing_options.clone())
            .await
            .map_err(|e| {
                tracing::error!(
                    request_id,
                    "Failed to fold chat completions stream for: {:?}",
                    e
                );
                ErrorMessage::internal_server_error(&format!(
                    "Failed to fold chat completions stream: {}",
                    e
                ))
            })?;
768
769
770
771
772
773
774
775

    // Convert NvCreateChatCompletionResponse --> NvResponse
    let response: NvResponse = response.try_into().map_err(|e| {
        tracing::error!(
            request_id,
            "Failed to convert NvCreateChatCompletionResponse to NvResponse: {:?}",
            e
        );
Ryan Olson's avatar
Ryan Olson committed
776
        ErrorMessage::internal_server_error("Failed to convert internal response")
777
778
779
780
781
782
783
    })?;

    inflight_guard.mark_ok();

    Ok(Json(response).into_response())
}

784
785
786
pub fn validate_response_input_is_text_only(
    request: &NvCreateResponse,
) -> Option<impl IntoResponse> {
787
    match &request.inner.input {
788
        dynamo_async_openai::types::responses::Input::Text(_) => None,
789
790
791
        _ => Some(ErrorMessage::not_implemented_error(
            "Only `Input::Text` is supported. Structured, multimedia, or custom input types are not yet implemented.",
        )),
792
793
794
795
796
    }
}

/// Checks for unsupported fields in the request.
/// Returns Some(response) if unsupported fields are present.
797
798
799
pub fn validate_response_unsupported_fields(
    request: &NvCreateResponse,
) -> Option<impl IntoResponse> {
800
801
802
    let inner = &request.inner;

    if inner.background == Some(true) {
Ryan Olson's avatar
Ryan Olson committed
803
        return Some(ErrorMessage::not_implemented_error(
804
805
806
807
            "`background: true` is not supported.",
        ));
    }
    if inner.include.is_some() {
Ryan Olson's avatar
Ryan Olson committed
808
        return Some(ErrorMessage::not_implemented_error(
809
810
811
812
            "`include` is not supported.",
        ));
    }
    if inner.instructions.is_some() {
Ryan Olson's avatar
Ryan Olson committed
813
        return Some(ErrorMessage::not_implemented_error(
814
815
816
817
            "`instructions` is not supported.",
        ));
    }
    if inner.max_tool_calls.is_some() {
Ryan Olson's avatar
Ryan Olson committed
818
        return Some(ErrorMessage::not_implemented_error(
819
820
821
822
            "`max_tool_calls` is not supported.",
        ));
    }
    if inner.metadata.is_some() {
Ryan Olson's avatar
Ryan Olson committed
823
        return Some(ErrorMessage::not_implemented_error(
824
825
826
827
            "`metadata` is not supported.",
        ));
    }
    if inner.parallel_tool_calls == Some(true) {
Ryan Olson's avatar
Ryan Olson committed
828
        return Some(ErrorMessage::not_implemented_error(
829
830
831
832
            "`parallel_tool_calls: true` is not supported.",
        ));
    }
    if inner.previous_response_id.is_some() {
Ryan Olson's avatar
Ryan Olson committed
833
        return Some(ErrorMessage::not_implemented_error(
834
835
836
837
            "`previous_response_id` is not supported.",
        ));
    }
    if inner.prompt.is_some() {
Ryan Olson's avatar
Ryan Olson committed
838
        return Some(ErrorMessage::not_implemented_error(
839
840
841
842
            "`prompt` is not supported.",
        ));
    }
    if inner.reasoning.is_some() {
Ryan Olson's avatar
Ryan Olson committed
843
        return Some(ErrorMessage::not_implemented_error(
844
845
846
847
            "`reasoning` is not supported.",
        ));
    }
    if inner.service_tier.is_some() {
Ryan Olson's avatar
Ryan Olson committed
848
        return Some(ErrorMessage::not_implemented_error(
849
850
851
852
            "`service_tier` is not supported.",
        ));
    }
    if inner.store == Some(true) {
Ryan Olson's avatar
Ryan Olson committed
853
        return Some(ErrorMessage::not_implemented_error(
854
855
856
857
            "`store: true` is not supported.",
        ));
    }
    if inner.stream == Some(true) {
Ryan Olson's avatar
Ryan Olson committed
858
        return Some(ErrorMessage::not_implemented_error(
859
860
861
862
            "`stream: true` is not supported.",
        ));
    }
    if inner.text.is_some() {
Ryan Olson's avatar
Ryan Olson committed
863
        return Some(ErrorMessage::not_implemented_error(
864
865
866
867
            "`text` is not supported.",
        ));
    }
    if inner.tool_choice.is_some() {
Ryan Olson's avatar
Ryan Olson committed
868
        return Some(ErrorMessage::not_implemented_error(
869
870
871
872
            "`tool_choice` is not supported.",
        ));
    }
    if inner.tools.is_some() {
Ryan Olson's avatar
Ryan Olson committed
873
        return Some(ErrorMessage::not_implemented_error(
874
875
876
877
            "`tools` is not supported.",
        ));
    }
    if inner.truncation.is_some() {
Ryan Olson's avatar
Ryan Olson committed
878
        return Some(ErrorMessage::not_implemented_error(
879
880
881
882
            "`truncation` is not supported.",
        ));
    }
    if inner.user.is_some() {
Ryan Olson's avatar
Ryan Olson committed
883
        return Some(ErrorMessage::not_implemented_error(
884
885
886
887
888
889
890
            "`user` is not supported.",
        ));
    }

    None
}

891
892
// todo - abstract this to the top level lib.rs to be reused
// todo - move the service_observer to its own state/arc
Ryan Olson's avatar
Ryan Olson committed
893
fn check_ready(_state: &Arc<service_v2::State>) -> Result<(), ErrorResponse> {
894
    // if state.service_observer.stage() != ServiceStage::Ready {
Ryan Olson's avatar
Ryan Olson committed
895
    //     return Err(ErrorMessage::service_unavailable());
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
    // }
    Ok(())
}

/// openai compatible format
/// Example:
/// {
///  "object": "list",
///  "data": [
///    {
///      "id": "model-id-0",
///      "object": "model",
///      "created": 1686935002,
///      "owned_by": "organization-owner"
///    },
///    ]
/// }
async fn list_models_openai(
914
    State(state): State<Arc<service_v2::State>>,
Ryan Olson's avatar
Ryan Olson committed
915
) -> Result<Response, ErrorResponse> {
916
917
918
919
920
921
922
923
    check_ready(&state)?;

    let created = SystemTime::now()
        .duration_since(UNIX_EPOCH)
        .unwrap()
        .as_secs();
    let mut data = Vec::new();

924
925
    let models: HashSet<String> = state.manager().model_display_names();
    for model_name in models {
926
        data.push(ModelListing {
927
            id: model_name.clone(),
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
            object: "object",
            created,                        // Where would this come from? The GGUF?
            owned_by: "nvidia".to_string(), // Get organization from GGUF
        });
    }

    let out = ListModelOpenAI {
        object: "list",
        data,
    };
    Ok(Json(out).into_response())
}

#[derive(Serialize)]
struct ListModelOpenAI {
    object: &'static str, // always "list"
    data: Vec<ModelListing>,
}

#[derive(Serialize)]
struct ModelListing {
    id: String,
    object: &'static str, // always "object"
    created: u64,         //  Seconds since epoch
    owned_by: String,
}

struct EventConverter<T>(Annotated<T>);

impl<T> From<Annotated<T>> for EventConverter<T> {
    fn from(annotated: Annotated<T>) -> Self {
        EventConverter(annotated)
    }
}

963
964
965
966
967
968
969
970
971
972
973
fn process_metrics_only<T>(
    annotated: &Annotated<T>,
    response_collector: &mut ResponseMetricCollector,
) {
    // update metrics
    if let Ok(Some(metrics)) = LLMMetricAnnotation::from_annotation(annotated) {
        response_collector.observe_current_osl(metrics.output_tokens);
        response_collector.observe_response(metrics.input_tokens, metrics.chunk_tokens);
    }
}

974
975
976
977
fn process_event_converter<T: Serialize>(
    annotated: EventConverter<T>,
    response_collector: &mut ResponseMetricCollector,
) -> Result<Event, axum::Error> {
978
    let mut annotated = annotated.0;
979

980
981
982
983
    // update metrics
    if let Ok(Some(metrics)) = LLMMetricAnnotation::from_annotation(&annotated) {
        response_collector.observe_current_osl(metrics.output_tokens);
        response_collector.observe_response(metrics.input_tokens, metrics.chunk_tokens);
984
985
986
987
988
989
990

        // Chomp the LLMMetricAnnotation so it's not returned in the response stream
        // TODO: add a flag to control what is returned in the SSE stream
        if annotated.event.as_deref() == Some(crate::preprocessor::ANNOTATION_LLM_METRICS) {
            annotated.event = None;
            annotated.comment = None;
        }
991
992
    }

993
    let mut event = Event::default();
994

995
996
997
    if let Some(data) = annotated.data {
        event = event.json_data(data)?;
    }
998

999
1000
1001
1002
1003
1004
    if let Some(msg) = annotated.event {
        if msg == "error" {
            let msgs = annotated
                .comment
                .unwrap_or_else(|| vec!["unspecified error".to_string()]);
            return Err(axum::Error::new(msgs.join(" -- ")));
1005
        }
1006
1007
        event = event.event(msg);
    }
1008

1009
1010
1011
1012
    if let Some(comments) = annotated.comment {
        for comment in comments {
            event = event.comment(comment);
        }
1013
    }
1014
1015

    Ok(event)
1016
1017
1018
1019
1020
}

/// Create an Axum [`Router`] for the OpenAI API Completions endpoint
/// If not path is provided, the default path is `/v1/completions`
pub fn completions_router(
1021
    state: Arc<service_v2::State>,
1022
1023
1024
1025
1026
    path: Option<String>,
) -> (Vec<RouteDoc>, Router) {
    let path = path.unwrap_or("/v1/completions".to_string());
    let doc = RouteDoc::new(axum::http::Method::POST, &path);
    let router = Router::new()
Ryan Olson's avatar
Ryan Olson committed
1027
        .route(&path, post(handler_completions))
1028
        .layer(axum::extract::DefaultBodyLimit::max(get_body_limit()))
1029
1030
1031
1032
1033
1034
1035
        .with_state(state);
    (vec![doc], router)
}

/// Create an Axum [`Router`] for the OpenAI API Chat Completions endpoint
/// If not path is provided, the default path is `/v1/chat/completions`
pub fn chat_completions_router(
1036
    state: Arc<service_v2::State>,
1037
    template: Option<RequestTemplate>,
1038
1039
1040
1041
1042
    path: Option<String>,
) -> (Vec<RouteDoc>, Router) {
    let path = path.unwrap_or("/v1/chat/completions".to_string());
    let doc = RouteDoc::new(axum::http::Method::POST, &path);
    let router = Router::new()
Ryan Olson's avatar
Ryan Olson committed
1043
        .route(&path, post(handler_chat_completions))
1044
        .layer(axum::extract::DefaultBodyLimit::max(get_body_limit()))
1045
        .with_state((state, template));
1046
1047
1048
    (vec![doc], router)
}

1049
1050
1051
/// Create an Axum [`Router`] for the OpenAI API Embeddings endpoint
/// If not path is provided, the default path is `/v1/embeddings`
pub fn embeddings_router(
1052
    state: Arc<service_v2::State>,
1053
1054
1055
1056
1057
1058
    path: Option<String>,
) -> (Vec<RouteDoc>, Router) {
    let path = path.unwrap_or("/v1/embeddings".to_string());
    let doc = RouteDoc::new(axum::http::Method::POST, &path);
    let router = Router::new()
        .route(&path, post(embeddings))
1059
        .layer(axum::extract::DefaultBodyLimit::max(get_body_limit()))
1060
1061
1062
1063
        .with_state(state);
    (vec![doc], router)
}

1064
1065
/// List Models
pub fn list_models_router(
1066
    state: Arc<service_v2::State>,
1067
1068
1069
    path: Option<String>,
) -> (Vec<RouteDoc>, Router) {
    // Standard OpenAI compatible list models endpoint
1070
    let openai_path = path.unwrap_or("/v1/models".to_string());
1071
1072
1073
1074
1075
1076
    let doc_for_openai = RouteDoc::new(axum::http::Method::GET, &openai_path);

    let router = Router::new()
        .route(&openai_path, get(list_models_openai))
        .with_state(state);

1077
    (vec![doc_for_openai], router)
1078
1079
}

1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
/// Create an Axum [`Router`] for the OpenAI API Responses endpoint
/// If not path is provided, the default path is `/v1/responses`
pub fn responses_router(
    state: Arc<service_v2::State>,
    template: Option<RequestTemplate>,
    path: Option<String>,
) -> (Vec<RouteDoc>, Router) {
    let path = path.unwrap_or("/v1/responses".to_string());
    let doc = RouteDoc::new(axum::http::Method::POST, &path);
    let router = Router::new()
Ryan Olson's avatar
Ryan Olson committed
1090
        .route(&path, post(handler_responses))
1091
1092
1093
1094
        .with_state((state, template));
    (vec![doc], router)
}

1095
1096
#[cfg(test)]
mod tests {
1097
1098
    use std::collections::HashMap;

1099
    use dynamo_async_openai::types::responses::{
1100
1101
1102
1103
        CreateResponse, Input, InputContent, InputItem, InputMessage, PromptConfig,
        Role as ResponseRole, ServiceTier, TextConfig, TextResponseFormat, ToolChoice,
        ToolChoiceMode, Truncation,
    };
1104
    use dynamo_async_openai::types::{
1105
1106
1107
        ChatCompletionRequestMessage, ChatCompletionRequestUserMessage,
        ChatCompletionRequestUserMessageContent, CreateChatCompletionRequest,
    };
1108
1109

    use super::*;
1110
1111
    use crate::discovery::ModelManagerError;
    use crate::protocols::openai::responses::NvCreateResponse;
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122

    const BACKUP_ERROR_MESSAGE: &str = "Failed to generate completions";

    fn http_error_from_engine(code: u16) -> Result<(), anyhow::Error> {
        Err(HttpError {
            code,
            message: "custom error message".to_string(),
        })?
    }

    fn other_error_from_engine() -> Result<(), anyhow::Error> {
1123
        Err(ModelManagerError::ModelNotFound("foo".to_string()))?
1124
1125
    }

1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
    fn make_base_request() -> NvCreateResponse {
        NvCreateResponse {
            inner: CreateResponse {
                input: Input::Text("hello".into()),
                model: "test-model".into(),
                background: None,
                include: None,
                instructions: None,
                max_output_tokens: None,
                max_tool_calls: None,
                metadata: None,
                parallel_tool_calls: None,
                previous_response_id: None,
                prompt: None,
                reasoning: None,
                service_tier: None,
                store: None,
                stream: None,
                text: None,
                tool_choice: None,
                tools: None,
                truncation: None,
                user: None,
                temperature: None,
                top_logprobs: None,
                top_p: None,
            },
            nvext: None,
        }
    }

1157
1158
1159
    #[test]
    fn test_http_error_response_from_anyhow() {
        let err = http_error_from_engine(400).unwrap_err();
Ryan Olson's avatar
Ryan Olson committed
1160
        let (status, response) = ErrorMessage::from_anyhow(err, BACKUP_ERROR_MESSAGE);
1161
1162
1163
1164
1165
1166
1167
        assert_eq!(status, StatusCode::BAD_REQUEST);
        assert_eq!(response.error, "custom error message");
    }

    #[test]
    fn test_error_response_from_anyhow_out_of_range() {
        let err = http_error_from_engine(399).unwrap_err();
Ryan Olson's avatar
Ryan Olson committed
1168
        let (status, response) = ErrorMessage::from_anyhow(err, BACKUP_ERROR_MESSAGE);
1169
1170
1171
1172
        assert_eq!(status, StatusCode::INTERNAL_SERVER_ERROR);
        assert_eq!(response.error, "custom error message");

        let err = http_error_from_engine(500).unwrap_err();
Ryan Olson's avatar
Ryan Olson committed
1173
        let (status, response) = ErrorMessage::from_anyhow(err, BACKUP_ERROR_MESSAGE);
1174
1175
1176
1177
        assert_eq!(status, StatusCode::INTERNAL_SERVER_ERROR);
        assert_eq!(response.error, "custom error message");

        let err = http_error_from_engine(501).unwrap_err();
Ryan Olson's avatar
Ryan Olson committed
1178
        let (status, response) = ErrorMessage::from_anyhow(err, BACKUP_ERROR_MESSAGE);
1179
1180
1181
1182
1183
1184
1185
        assert_eq!(status, StatusCode::INTERNAL_SERVER_ERROR);
        assert_eq!(response.error, "custom error message");
    }

    #[test]
    fn test_other_error_response_from_anyhow() {
        let err = other_error_from_engine().unwrap_err();
Ryan Olson's avatar
Ryan Olson committed
1186
        let (status, response) = ErrorMessage::from_anyhow(err, BACKUP_ERROR_MESSAGE);
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
        assert_eq!(status, StatusCode::INTERNAL_SERVER_ERROR);
        assert_eq!(
            response.error,
            format!(
                "{}: {}",
                BACKUP_ERROR_MESSAGE,
                other_error_from_engine().unwrap_err()
            )
        );
    }
1197

1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
    #[test]
    fn test_service_overloaded_error_response_from_anyhow() {
        use dynamo_runtime::pipeline::error::PipelineError;

        let err: anyhow::Error = PipelineError::ServiceOverloaded(
            "All workers are busy, please retry later".to_string(),
        )
        .into();
        let (status, response) = ErrorMessage::from_anyhow(err, BACKUP_ERROR_MESSAGE);
        assert_eq!(status, StatusCode::SERVICE_UNAVAILABLE);
        assert_eq!(
            response.error,
            "Service temporarily unavailable: All workers are busy, please retry later"
        );
    }

1214
1215
1216
    #[test]
    fn test_validate_input_is_text_only_accepts_text() {
        let request = make_base_request();
1217
        let result = validate_response_input_is_text_only(&request);
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
        assert!(result.is_none());
    }

    #[test]
    fn test_validate_input_is_text_only_rejects_items() {
        let mut request = make_base_request();
        request.inner.input = Input::Items(vec![InputItem::Message(InputMessage {
            kind: Default::default(),
            role: ResponseRole::User,
            content: InputContent::TextInput("structured".into()),
        })]);
1229
        let result = validate_response_input_is_text_only(&request);
1230
1231
1232
1233
1234
1235
        assert!(result.is_some());
    }

    #[test]
    fn test_validate_unsupported_fields_accepts_clean_request() {
        let request = make_base_request();
1236
        let result = validate_response_unsupported_fields(&request);
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
        assert!(result.is_none());
    }

    #[test]
    fn test_validate_unsupported_fields_detects_flags() {
        #[allow(clippy::type_complexity)]
        let unsupported_cases: Vec<(&str, Box<dyn FnOnce(&mut CreateResponse)>)> = vec![
            ("background", Box::new(|r| r.background = Some(true))),
            (
                "include",
                Box::new(|r| r.include = Some(vec!["file_search_call.results".into()])),
            ),
            (
                "instructions",
                Box::new(|r| r.instructions = Some("System prompt".into())),
            ),
            ("max_tool_calls", Box::new(|r| r.max_tool_calls = Some(3))),
            ("metadata", Box::new(|r| r.metadata = Some(HashMap::new()))),
            (
                "parallel_tool_calls",
                Box::new(|r| r.parallel_tool_calls = Some(true)),
            ),
            (
                "previous_response_id",
                Box::new(|r| r.previous_response_id = Some("prev-id".into())),
            ),
            (
                "prompt",
                Box::new(|r| {
                    r.prompt = Some(PromptConfig {
                        id: "template-id".into(),
                        version: None,
                        variables: None,
                    })
                }),
            ),
            (
                "reasoning",
                Box::new(|r| r.reasoning = Some(Default::default())),
            ),
            (
                "service_tier",
                Box::new(|r| r.service_tier = Some(ServiceTier::Auto)),
            ),
            ("store", Box::new(|r| r.store = Some(true))),
            ("stream", Box::new(|r| r.stream = Some(true))),
            (
                "text",
                Box::new(|r| {
                    r.text = Some(TextConfig {
                        format: TextResponseFormat::Text,
                    })
                }),
            ),
            (
                "tool_choice",
                Box::new(|r| r.tool_choice = Some(ToolChoice::Mode(ToolChoiceMode::Required))),
            ),
            ("tools", Box::new(|r| r.tools = Some(vec![]))),
            (
                "truncation",
                Box::new(|r| r.truncation = Some(Truncation::Auto)),
            ),
            ("user", Box::new(|r| r.user = Some("user-id".into()))),
        ];

        for (field, set_field) in unsupported_cases {
            let mut req = make_base_request();
            (set_field)(&mut req.inner);
1306
            let result = validate_response_unsupported_fields(&req);
1307
1308
1309
            assert!(result.is_some(), "Expected rejection for `{field}`");
        }
    }
1310
1311
1312
1313
1314
1315
1316
1317
1318

    #[test]
    fn test_validate_chat_completion_required_fields_empty_messages() {
        let request = NvCreateChatCompletionRequest {
            inner: CreateChatCompletionRequest {
                model: "test-model".to_string(),
                messages: vec![],
                ..Default::default()
            },
1319
            common: Default::default(),
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
            nvext: None,
        };
        let result = validate_chat_completion_required_fields(&request);
        assert!(result.is_err());
        if let Err((status, error_response)) = result {
            assert_eq!(status, StatusCode::BAD_REQUEST);
            assert_eq!(
                error_response.error,
                "The 'messages' field cannot be empty. At least one message is required."
            );
        }
    }

    #[test]
    fn test_validate_chat_completion_required_fields_with_messages() {
        let request = NvCreateChatCompletionRequest {
            inner: CreateChatCompletionRequest {
                model: "test-model".to_string(),
                messages: vec![ChatCompletionRequestMessage::User(
                    ChatCompletionRequestUserMessage {
                        content: ChatCompletionRequestUserMessageContent::Text("Hello".to_string()),
                        name: None,
                    },
                )],
                ..Default::default()
            },
1346
            common: Default::default(),
1347
1348
1349
1350
1351
            nvext: None,
        };
        let result = validate_chat_completion_required_fields(&request);
        assert!(result.is_ok());
    }
1352
}