openai.rs 43.2 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
11
use axum::{
    extract::State,
Ryan Olson's avatar
Ryan Olson committed
12
    http::{HeaderMap, StatusCode},
13
14
15
16
17
18
19
    response::{
        sse::{Event, KeepAlive, Sse},
        IntoResponse, Response,
    },
    routing::{get, post},
    Json, Router,
};
Ryan Olson's avatar
Ryan Olson committed
20
21
22
23
24
use dynamo_runtime::{
    pipeline::{AsyncEngineContextProvider, Context},
    protocols::annotated::AnnotationsProvider,
};
use futures::{stream, StreamExt};
25
26
27
use serde::{Deserialize, Serialize};

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

Ryan Olson's avatar
Ryan Olson committed
45
46
47
48
49
50
51
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";

pub type ErrorResponse = (StatusCode, Json<ErrorMessage>);

52
#[derive(Serialize, Deserialize)]
Ryan Olson's avatar
Ryan Olson committed
53
pub(crate) struct ErrorMessage {
54
55
56
    error: String,
}

Ryan Olson's avatar
Ryan Olson committed
57
impl ErrorMessage {
58
    /// Not Found Error
Ryan Olson's avatar
Ryan Olson committed
59
    pub fn model_not_found() -> ErrorResponse {
60
61
        (
            StatusCode::NOT_FOUND,
Ryan Olson's avatar
Ryan Olson committed
62
            Json(ErrorMessage {
63
64
65
66
67
68
69
                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
70
    pub fn _service_unavailable() -> ErrorResponse {
71
72
        (
            StatusCode::SERVICE_UNAVAILABLE,
Ryan Olson's avatar
Ryan Olson committed
73
            Json(ErrorMessage {
74
75
76
77
78
79
80
81
82
                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
83
    pub fn internal_server_error(msg: &str) -> ErrorResponse {
84
85
86
        tracing::error!("Internal server error: {msg}");
        (
            StatusCode::INTERNAL_SERVER_ERROR,
Ryan Olson's avatar
Ryan Olson committed
87
            Json(ErrorMessage {
88
89
90
91
92
                error: msg.to_string(),
            }),
        )
    }

93
94
95
    /// 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
96
    pub fn not_implemented_error(msg: &str) -> ErrorResponse {
97
98
99
        tracing::error!("Not Implemented error: {msg}");
        (
            StatusCode::NOT_IMPLEMENTED,
Ryan Olson's avatar
Ryan Olson committed
100
            Json(ErrorMessage {
101
102
103
104
105
                error: msg.to_string(),
            }),
        )
    }

Neelay Shah's avatar
Neelay Shah committed
106
    /// The OAI endpoints call an [`dynamo.runtime::engine::AsyncEngine`] which are specialized to return
107
    /// an [`anyhow::Error`]. This method will convert the [`anyhow::Error`] into an [`HttpError`].
Ryan Olson's avatar
Ryan Olson committed
108
    /// If successful, it will return the [`HttpError`] as an [`ErrorMessage::internal_server_error`]
109
    /// with the details of the error.
Ryan Olson's avatar
Ryan Olson committed
110
    pub fn from_anyhow(err: anyhow::Error, alt_msg: &str) -> ErrorResponse {
111
        match err.downcast::<HttpError>() {
Ryan Olson's avatar
Ryan Olson committed
112
113
            Ok(http_error) => ErrorMessage::from_http_error(http_error),
            Err(err) => ErrorMessage::internal_server_error(&format!("{alt_msg}: {err}")),
114
115
116
117
        }
    }

    /// Implementers should only be able to throw 400-499 errors.
Ryan Olson's avatar
Ryan Olson committed
118
    pub fn from_http_error(err: HttpError) -> ErrorResponse {
119
        if err.code < 400 || err.code >= 500 {
Ryan Olson's avatar
Ryan Olson committed
120
            return ErrorMessage::internal_server_error(&err.message);
121
122
        }
        match StatusCode::from_u16(err.code) {
Ryan Olson's avatar
Ryan Olson committed
123
124
            Ok(code) => (code, Json(ErrorMessage { error: err.message })),
            Err(_) => ErrorMessage::internal_server_error(&err.message),
125
126
127
128
        }
    }
}

Ryan Olson's avatar
Ryan Olson committed
129
impl From<HttpError> for ErrorMessage {
130
    fn from(err: HttpError) -> Self {
Ryan Olson's avatar
Ryan Olson committed
131
        ErrorMessage { error: err.message }
132
133
134
    }
}

Ryan Olson's avatar
Ryan Olson committed
135
136
/// 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 {
137
138
139
140
141
142
143
    // Try to get request id from trace context
    if let Some(trace_context) = get_distributed_tracing_context() {
        if let Some(x_dynamo_request_id) = trace_context.x_dynamo_request_id {
            return x_dynamo_request_id;
        }
    }

Ryan Olson's avatar
Ryan Olson committed
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
    // Try to get the request ID from the primary source
    if let Some(primary) = primary {
        if let Ok(uuid) = uuid::Uuid::parse_str(primary) {
            return uuid.to_string();
        }
    }

    // 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()
}

167
168
169
170
171
172
173
174
/// 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
175
async fn handler_completions(
176
    State(state): State<Arc<service_v2::State>>,
Ryan Olson's avatar
Ryan Olson committed
177
    headers: HeaderMap,
178
    Json(request): Json<NvCreateCompletionRequest>,
Ryan Olson's avatar
Ryan Olson committed
179
180
181
182
183
184
185
186
187
188
189
190
191
192
) -> 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
    let (mut connection_handle, stream_handle) = create_connection_monitor(context.clone()).await;

    // possibly long running task
    // if this returns a streaming response, the stream handle will be armed and captured by the response stream
193
    let response = tokio::spawn(completions(state, request, stream_handle).in_current_span())
Ryan Olson's avatar
Ryan Olson committed
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
        .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> {
215
216
217
218
219
220
221
    // return a 503 if the service is not ready
    check_ready(&state)?;

    // todo - extract distributed tracing id and context id from headers
    let request_id = uuid::Uuid::new_v4().to_string();

    // todo - decide on default
222
    let streaming = request.inner.stream.unwrap_or(false);
223
224

    // update the request to always stream
Ryan Olson's avatar
Ryan Olson committed
225
226
227
228
    let request = request.map(|mut req| {
        req.inner.stream = Some(true);
        req
    });
229

230
231
    // todo - make the protocols be optional for model name
    // todo - when optional, if none, apply a default
232
    let model = &request.inner.model;
233
234
235

    // todo - error handling should be more robust
    let engine = state
236
        .manager()
237
        .get_completions_engine(model)
Ryan Olson's avatar
Ryan Olson committed
238
        .map_err(|_| ErrorMessage::model_not_found())?;
239

240
    let mut inflight_guard =
241
242
243
        state
            .metrics_clone()
            .create_inflight_guard(model, Endpoint::Completions, streaming);
244

245
246
    let mut response_collector = state.metrics_clone().create_response_collector(model);

Ryan Olson's avatar
Ryan Olson committed
247
248
    // prepare to process any annotations
    let annotations = request.annotations();
249
250
251
252
253

    // issue the generate call on the engine
    let stream = engine
        .generate(request)
        .await
Ryan Olson's avatar
Ryan Olson committed
254
        .map_err(|e| ErrorMessage::from_anyhow(e, "Failed to generate completions"))?;
255
256
257
258

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

Ryan Olson's avatar
Ryan Olson committed
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
    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);
278
279

    if streaming {
280
281
282
        let stream = stream.map(move |response| {
            process_event_converter(EventConverter::from(response), &mut response_collector)
        });
Ryan Olson's avatar
Ryan Olson committed
283
        let stream = monitor_for_disconnects(stream, ctx, inflight_guard, stream_handle);
284

285
286
        let mut sse_stream = Sse::new(stream);

287
        if let Some(keep_alive) = state.sse_keep_alive() {
288
289
290
291
            sse_stream = sse_stream.keep_alive(KeepAlive::default().interval(keep_alive));
        }

        Ok(sse_stream.into_response())
292
    } else {
293
        // TODO: report ISL/OSL for non-streaming requests
Ryan Olson's avatar
Ryan Olson committed
294
        let response = NvCreateCompletionResponse::from_annotated_stream(stream)
295
296
297
298
299
300
301
            .await
            .map_err(|e| {
                tracing::error!(
                    "Failed to fold completions stream for {}: {:?}",
                    request_id,
                    e
                );
Ryan Olson's avatar
Ryan Olson committed
302
                ErrorMessage::internal_server_error("Failed to fold completions stream")
303
304
            })?;

305
        inflight_guard.mark_ok();
306
307
308
309
        Ok(Json(response).into_response())
    }
}

310
311
#[tracing::instrument(skip_all)]
async fn embeddings(
312
313
    State(state): State<Arc<service_v2::State>>,
    Json(request): Json<NvCreateEmbeddingRequest>,
Ryan Olson's avatar
Ryan Olson committed
314
) -> Result<Response, ErrorResponse> {
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
    // return a 503 if the service is not ready
    check_ready(&state)?;

    // todo - extract distributed tracing id and context id from headers
    let request_id = uuid::Uuid::new_v4().to_string();

    // 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
332
        .map_err(|_| ErrorMessage::model_not_found())?;
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347

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

    // setup context
    // todo - inherit request_id from distributed trace details
    let request = Context::with_id(request, request_id.clone());

    // issue the generate call on the engine
    let stream = engine
        .generate(request)
        .await
Ryan Olson's avatar
Ryan Olson committed
348
        .map_err(|e| ErrorMessage::from_anyhow(e, "Failed to generate embeddings"))?;
349
350
351

    // 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
352
    let response = NvCreateEmbeddingResponse::from_annotated_stream(stream)
353
354
355
356
357
358
359
        .await
        .map_err(|e| {
            tracing::error!(
                "Failed to fold embeddings stream for {}: {:?}",
                request_id,
                e
            );
Ryan Olson's avatar
Ryan Olson committed
360
            ErrorMessage::internal_server_error("Failed to fold embeddings stream")
361
362
363
364
        })?;

    inflight.mark_ok();
    Ok(Json(response).into_response())
365
366
}

Ryan Olson's avatar
Ryan Olson committed
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
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
    let (mut connection_handle, stream_handle) = create_connection_monitor(context.clone()).await;

383
384
385
386
387
388
389
390
391
    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
392
393
394
395
396
397
398
399

    // 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
}

400
401
402
403
404
405
406
407
408
/// 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
409
410
411
412
413
    state: Arc<service_v2::State>,
    template: Option<RequestTemplate>,
    mut request: Context<NvCreateChatCompletionRequest>,
    mut stream_handle: ConnectionHandle,
) -> Result<Response, ErrorResponse> {
414
415
416
    // return a 503 if the service is not ready
    check_ready(&state)?;

Ryan Olson's avatar
Ryan Olson committed
417
418
    let request_id = request.id().to_string();

419
420
421
422
    // 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
423
    validate_chat_completion_unsupported_fields(&request)?;
424

425
426
427
    // Handle required fields like messages shouldn't be empty.
    validate_chat_completion_required_fields(&request)?;

428
429
430
431
432
433
434
435
436
437
438
439
    // 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
440
    tracing::trace!("Received chat completions request: {:?}", request.content());
441
442

    // todo - decide on default
Paul Hendricks's avatar
Paul Hendricks committed
443
    let streaming = request.inner.stream.unwrap_or(false);
444
445

    // update the request to always stream
Ryan Olson's avatar
Ryan Olson committed
446
447
448
449
    let request = request.map(|mut req| {
        req.inner.stream = Some(true);
        req
    });
450
451
452

    // todo - make the protocols be optional for model name
    // todo - when optional, if none, apply a default
Paul Hendricks's avatar
Paul Hendricks committed
453
    let model = &request.inner.model;
454
455
456
457
458

    // 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
459
        .manager()
460
        .get_chat_completions_engine(model)
Ryan Olson's avatar
Ryan Olson committed
461
        .map_err(|_| ErrorMessage::model_not_found())?;
462

463
    let mut inflight_guard =
464
465
466
        state
            .metrics_clone()
            .create_inflight_guard(model, Endpoint::ChatCompletions, streaming);
467

468
469
    let mut response_collector = state.metrics_clone().create_response_collector(model);

470
    tracing::trace!("Issuing generate call for chat completions");
Ryan Olson's avatar
Ryan Olson committed
471
    let annotations = request.annotations();
472
473
474
475
476

    // issue the generate call on the engine
    let stream = engine
        .generate(request)
        .await
Ryan Olson's avatar
Ryan Olson committed
477
        .map_err(|e| ErrorMessage::from_anyhow(e, "Failed to generate completions"))?;
478
479
480
481

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

Ryan Olson's avatar
Ryan Olson committed
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
    // 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);

499
500
501
502
    // 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
503
504
        stream_handle.arm();

505
506
507
        let stream = stream.map(move |response| {
            process_event_converter(EventConverter::from(response), &mut response_collector)
        });
Ryan Olson's avatar
Ryan Olson committed
508
        let stream = monitor_for_disconnects(stream, ctx, inflight_guard, stream_handle);
509

510
511
        let mut sse_stream = Sse::new(stream);

512
        if let Some(keep_alive) = state.sse_keep_alive() {
513
514
515
516
            sse_stream = sse_stream.keep_alive(KeepAlive::default().interval(keep_alive));
        }

        Ok(sse_stream.into_response())
517
    } else {
518
        // TODO: report ISL/OSL for non-streaming requests
Ryan Olson's avatar
Ryan Olson committed
519
        let response = NvCreateChatCompletionResponse::from_annotated_stream(stream)
520
521
522
523
524
525
526
            .await
            .map_err(|e| {
                tracing::error!(
                    request_id,
                    "Failed to fold chat completions stream for: {:?}",
                    e
                );
Ryan Olson's avatar
Ryan Olson committed
527
                ErrorMessage::internal_server_error(&format!(
528
529
530
531
532
                    "Failed to fold chat completions stream: {}",
                    e
                ))
            })?;

533
        inflight_guard.mark_ok();
534
535
536
537
        Ok(Json(response).into_response())
    }
}

538
539
540
541
542
/// 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
543
) -> Result<(), ErrorResponse> {
544
545
546
    let inner = &request.inner;

    if inner.parallel_tool_calls == Some(true) {
Ryan Olson's avatar
Ryan Olson committed
547
        return Err(ErrorMessage::not_implemented_error(
548
549
550
551
552
            "`parallel_tool_calls: true` is not supported.",
        ));
    }

    if inner.stream == Some(true) && inner.tools.is_some() {
Ryan Olson's avatar
Ryan Olson committed
553
        return Err(ErrorMessage::not_implemented_error(
554
555
556
557
558
            "`stream: true` is not supported when `tools` are provided.",
        ));
    }

    if inner.function_call.is_some() {
Ryan Olson's avatar
Ryan Olson committed
559
        return Err(ErrorMessage::not_implemented_error(
560
561
562
563
564
            "`function_call` is deprecated. Please migrate to use `tool_choice` instead.",
        ));
    }

    if inner.functions.is_some() {
Ryan Olson's avatar
Ryan Olson committed
565
        return Err(ErrorMessage::not_implemented_error(
566
567
568
569
            "`functions` is deprecated. Please migrate to use `tools` instead.",
        ));
    }

Ryan Olson's avatar
Ryan Olson committed
570
    Ok(())
571
572
}

573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
/// 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(())
}

590
591
592
/// OpenAI Responses Request Handler
///
/// This method will handle the incoming request for the /v1/responses endpoint.
Ryan Olson's avatar
Ryan Olson committed
593
async fn handler_responses(
594
    State((state, template)): State<(Arc<service_v2::State>, Option<RequestTemplate>)>,
Ryan Olson's avatar
Ryan Olson committed
595
596
597
598
599
600
601
602
603
604
605
606
607
608
    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
    let (mut connection_handle, _stream_handle) = create_connection_monitor(context.clone()).await;

609
    let response = tokio::spawn(responses(state, template, request).in_current_span())
Ryan Olson's avatar
Ryan Olson committed
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
        .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> {
631
632
633
634
635
636
    // 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.
637
    if let Some(resp) = validate_response_unsupported_fields(&request) {
638
639
640
641
642
643
644
        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.
645
    if let Some(resp) = validate_response_input_is_text_only(&request) {
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
        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
663
664
    let request_id = request.id().to_string();
    let (request, context) = request.into_parts();
665

Ryan Olson's avatar
Ryan Olson committed
666
    let mut request: NvCreateChatCompletionRequest = request.try_into().map_err(|e| {
667
668
669
670
671
        tracing::error!(
            request_id,
            "Failed to convert NvCreateResponse to NvCreateChatCompletionRequest: {:?}",
            e
        );
Ryan Olson's avatar
Ryan Olson committed
672
        ErrorMessage::not_implemented_error(&format!(
673
674
675
676
677
            "Only Input::Text(_) is currently supported: {}",
            e
        ))
    })?;

Ryan Olson's avatar
Ryan Olson committed
678
679
680
681
682
    let request = context.map(|mut _req| {
        request.inner.stream = Some(false);
        request
    });

683
684
685
686
687
688
689
    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
690
        .map_err(|_| ErrorMessage::model_not_found())?;
691
692
693
694
695
696
697
698
699
700
701
702
703
704

    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
705
        .map_err(|e| ErrorMessage::from_anyhow(e, "Failed to generate completions"))?;
706
707

    // TODO: handle streaming, currently just unary
Ryan Olson's avatar
Ryan Olson committed
708
    let response = NvCreateChatCompletionResponse::from_annotated_stream(stream)
709
710
711
712
713
714
715
        .await
        .map_err(|e| {
            tracing::error!(
                request_id,
                "Failed to fold chat completions stream for: {:?}",
                e
            );
Ryan Olson's avatar
Ryan Olson committed
716
            ErrorMessage::internal_server_error(&format!(
717
718
719
720
721
722
723
724
725
726
727
728
                "Failed to fold chat completions stream: {}",
                e
            ))
        })?;

    // 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
729
        ErrorMessage::internal_server_error("Failed to convert internal response")
730
731
732
733
734
735
736
    })?;

    inflight_guard.mark_ok();

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

737
738
739
pub fn validate_response_input_is_text_only(
    request: &NvCreateResponse,
) -> Option<impl IntoResponse> {
740
741
    match &request.inner.input {
        async_openai::types::responses::Input::Text(_) => None,
Ryan Olson's avatar
Ryan Olson committed
742
        _ => Some(ErrorMessage::not_implemented_error("Only `Input::Text` is supported. Structured, multimedia, or custom input types are not yet implemented.")),
743
744
745
746
747
    }
}

/// Checks for unsupported fields in the request.
/// Returns Some(response) if unsupported fields are present.
748
749
750
pub fn validate_response_unsupported_fields(
    request: &NvCreateResponse,
) -> Option<impl IntoResponse> {
751
752
753
    let inner = &request.inner;

    if inner.background == Some(true) {
Ryan Olson's avatar
Ryan Olson committed
754
        return Some(ErrorMessage::not_implemented_error(
755
756
757
758
            "`background: true` is not supported.",
        ));
    }
    if inner.include.is_some() {
Ryan Olson's avatar
Ryan Olson committed
759
        return Some(ErrorMessage::not_implemented_error(
760
761
762
763
            "`include` is not supported.",
        ));
    }
    if inner.instructions.is_some() {
Ryan Olson's avatar
Ryan Olson committed
764
        return Some(ErrorMessage::not_implemented_error(
765
766
767
768
            "`instructions` is not supported.",
        ));
    }
    if inner.max_tool_calls.is_some() {
Ryan Olson's avatar
Ryan Olson committed
769
        return Some(ErrorMessage::not_implemented_error(
770
771
772
773
            "`max_tool_calls` is not supported.",
        ));
    }
    if inner.metadata.is_some() {
Ryan Olson's avatar
Ryan Olson committed
774
        return Some(ErrorMessage::not_implemented_error(
775
776
777
778
            "`metadata` is not supported.",
        ));
    }
    if inner.parallel_tool_calls == Some(true) {
Ryan Olson's avatar
Ryan Olson committed
779
        return Some(ErrorMessage::not_implemented_error(
780
781
782
783
            "`parallel_tool_calls: true` is not supported.",
        ));
    }
    if inner.previous_response_id.is_some() {
Ryan Olson's avatar
Ryan Olson committed
784
        return Some(ErrorMessage::not_implemented_error(
785
786
787
788
            "`previous_response_id` is not supported.",
        ));
    }
    if inner.prompt.is_some() {
Ryan Olson's avatar
Ryan Olson committed
789
        return Some(ErrorMessage::not_implemented_error(
790
791
792
793
            "`prompt` is not supported.",
        ));
    }
    if inner.reasoning.is_some() {
Ryan Olson's avatar
Ryan Olson committed
794
        return Some(ErrorMessage::not_implemented_error(
795
796
797
798
            "`reasoning` is not supported.",
        ));
    }
    if inner.service_tier.is_some() {
Ryan Olson's avatar
Ryan Olson committed
799
        return Some(ErrorMessage::not_implemented_error(
800
801
802
803
            "`service_tier` is not supported.",
        ));
    }
    if inner.store == Some(true) {
Ryan Olson's avatar
Ryan Olson committed
804
        return Some(ErrorMessage::not_implemented_error(
805
806
807
808
            "`store: true` is not supported.",
        ));
    }
    if inner.stream == Some(true) {
Ryan Olson's avatar
Ryan Olson committed
809
        return Some(ErrorMessage::not_implemented_error(
810
811
812
813
            "`stream: true` is not supported.",
        ));
    }
    if inner.text.is_some() {
Ryan Olson's avatar
Ryan Olson committed
814
        return Some(ErrorMessage::not_implemented_error(
815
816
817
818
            "`text` is not supported.",
        ));
    }
    if inner.tool_choice.is_some() {
Ryan Olson's avatar
Ryan Olson committed
819
        return Some(ErrorMessage::not_implemented_error(
820
821
822
823
            "`tool_choice` is not supported.",
        ));
    }
    if inner.tools.is_some() {
Ryan Olson's avatar
Ryan Olson committed
824
        return Some(ErrorMessage::not_implemented_error(
825
826
827
828
            "`tools` is not supported.",
        ));
    }
    if inner.truncation.is_some() {
Ryan Olson's avatar
Ryan Olson committed
829
        return Some(ErrorMessage::not_implemented_error(
830
831
832
833
            "`truncation` is not supported.",
        ));
    }
    if inner.user.is_some() {
Ryan Olson's avatar
Ryan Olson committed
834
        return Some(ErrorMessage::not_implemented_error(
835
836
837
838
839
840
841
            "`user` is not supported.",
        ));
    }

    None
}

842
843
// 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
844
fn check_ready(_state: &Arc<service_v2::State>) -> Result<(), ErrorResponse> {
845
    // if state.service_observer.stage() != ServiceStage::Ready {
Ryan Olson's avatar
Ryan Olson committed
846
    //     return Err(ErrorMessage::service_unavailable());
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
    // }
    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(
865
    State(state): State<Arc<service_v2::State>>,
Ryan Olson's avatar
Ryan Olson committed
866
) -> Result<Response, ErrorResponse> {
867
868
869
870
871
872
873
874
    check_ready(&state)?;

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

875
876
    let models: HashSet<String> = state.manager().model_display_names();
    for model_name in models {
877
        data.push(ModelListing {
878
            id: model_name.clone(),
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
            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)
    }
}

914
915
916
917
fn process_event_converter<T: Serialize>(
    annotated: EventConverter<T>,
    response_collector: &mut ResponseMetricCollector,
) -> Result<Event, axum::Error> {
918
    let mut annotated = annotated.0;
919

920
921
922
923
    // 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);
924
925
926
927
928
929
930

        // 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;
        }
931
932
    }

933
    let mut event = Event::default();
934

935
936
937
    if let Some(data) = annotated.data {
        event = event.json_data(data)?;
    }
938

939
940
941
942
943
944
    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(" -- ")));
945
        }
946
947
        event = event.event(msg);
    }
948

949
950
951
952
    if let Some(comments) = annotated.comment {
        for comment in comments {
            event = event.comment(comment);
        }
953
    }
954
955

    Ok(event)
956
957
958
959
960
}

/// 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(
961
    state: Arc<service_v2::State>,
962
963
964
965
966
    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
967
        .route(&path, post(handler_completions))
968
969
970
971
972
973
974
        .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(
975
    state: Arc<service_v2::State>,
976
    template: Option<RequestTemplate>,
977
978
979
980
981
    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
982
        .route(&path, post(handler_chat_completions))
983
        .with_state((state, template));
984
985
986
    (vec![doc], router)
}

987
988
989
/// 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(
990
    state: Arc<service_v2::State>,
991
992
993
994
995
996
997
998
999
1000
    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))
        .with_state(state);
    (vec![doc], router)
}

1001
1002
/// List Models
pub fn list_models_router(
1003
    state: Arc<service_v2::State>,
1004
1005
1006
    path: Option<String>,
) -> (Vec<RouteDoc>, Router) {
    // Standard OpenAI compatible list models endpoint
1007
    let openai_path = path.unwrap_or("/v1/models".to_string());
1008
1009
1010
1011
1012
1013
    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);

1014
    (vec![doc_for_openai], router)
1015
1016
}

1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
/// 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
1027
        .route(&path, post(handler_responses))
1028
1029
1030
1031
        .with_state((state, template));
    (vec![doc], router)
}

1032
1033
#[cfg(test)]
mod tests {
1034
1035
1036
1037
1038
1039
1040
    use std::collections::HashMap;

    use async_openai::types::responses::{
        CreateResponse, Input, InputContent, InputItem, InputMessage, PromptConfig,
        Role as ResponseRole, ServiceTier, TextConfig, TextResponseFormat, ToolChoice,
        ToolChoiceMode, Truncation,
    };
1041
1042
1043
1044
    use async_openai::types::{
        ChatCompletionRequestMessage, ChatCompletionRequestUserMessage,
        ChatCompletionRequestUserMessageContent, CreateChatCompletionRequest,
    };
1045
1046

    use super::*;
1047
1048
    use crate::discovery::ModelManagerError;
    use crate::protocols::openai::responses::NvCreateResponse;
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059

    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> {
1060
        Err(ModelManagerError::ModelNotFound("foo".to_string()))?
1061
1062
    }

1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
    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,
        }
    }

1094
1095
1096
    #[test]
    fn test_http_error_response_from_anyhow() {
        let err = http_error_from_engine(400).unwrap_err();
Ryan Olson's avatar
Ryan Olson committed
1097
        let (status, response) = ErrorMessage::from_anyhow(err, BACKUP_ERROR_MESSAGE);
1098
1099
1100
1101
1102
1103
1104
        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
1105
        let (status, response) = ErrorMessage::from_anyhow(err, BACKUP_ERROR_MESSAGE);
1106
1107
1108
1109
        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
1110
        let (status, response) = ErrorMessage::from_anyhow(err, BACKUP_ERROR_MESSAGE);
1111
1112
1113
1114
        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
1115
        let (status, response) = ErrorMessage::from_anyhow(err, BACKUP_ERROR_MESSAGE);
1116
1117
1118
1119
1120
1121
1122
        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
1123
        let (status, response) = ErrorMessage::from_anyhow(err, BACKUP_ERROR_MESSAGE);
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
        assert_eq!(status, StatusCode::INTERNAL_SERVER_ERROR);
        assert_eq!(
            response.error,
            format!(
                "{}: {}",
                BACKUP_ERROR_MESSAGE,
                other_error_from_engine().unwrap_err()
            )
        );
    }
1134
1135
1136
1137

    #[test]
    fn test_validate_input_is_text_only_accepts_text() {
        let request = make_base_request();
1138
        let result = validate_response_input_is_text_only(&request);
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
        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()),
        })]);
1150
        let result = validate_response_input_is_text_only(&request);
1151
1152
1153
1154
1155
1156
        assert!(result.is_some());
    }

    #[test]
    fn test_validate_unsupported_fields_accepts_clean_request() {
        let request = make_base_request();
1157
        let result = validate_response_unsupported_fields(&request);
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
        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);
1227
            let result = validate_response_unsupported_fields(&req);
1228
1229
1230
            assert!(result.is_some(), "Expected rejection for `{field}`");
        }
    }
1231
1232
1233
1234
1235
1236
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

    #[test]
    fn test_validate_chat_completion_required_fields_empty_messages() {
        let request = NvCreateChatCompletionRequest {
            inner: CreateChatCompletionRequest {
                model: "test-model".to_string(),
                messages: vec![],
                ..Default::default()
            },
            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()
            },
            nvext: None,
        };
        let result = validate_chat_completion_required_fields(&request);
        assert!(result.is_ok());
    }
1271
}