openai.rs 74.5 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
    body::Body,
13
    extract::State,
14
    http::Request,
Ryan Olson's avatar
Ryan Olson committed
15
    http::{HeaderMap, StatusCode},
16
    middleware::{self, Next},
17
18
    response::{
        IntoResponse, Response,
19
        sse::{KeepAlive, Sse},
20
21
22
    },
    routing::{get, post},
};
23
use dynamo_runtime::config::environment_names::llm as env_llm;
Ryan Olson's avatar
Ryan Olson committed
24
25
26
27
use dynamo_runtime::{
    pipeline::{AsyncEngineContextProvider, Context},
    protocols::annotated::AnnotationsProvider,
};
28
use futures::{StreamExt, stream};
29
30
31
use serde::{Deserialize, Serialize};

use super::{
32
33
    RouteDoc,
    disconnect::{ConnectionHandle, create_connection_monitor, monitor_for_disconnects},
34
    error::HttpError,
35
36
37
38
    metrics::{
        Endpoint, EventConverter, process_response_and_observe_metrics,
        process_response_using_event_converter_and_observe_metrics,
    },
39
    service_v2,
40
};
41
use crate::engines::ValidateRequest;
42
use crate::protocols::openai::chat_completions::aggregator::ChatCompletionAggregator;
43
use crate::protocols::openai::{
44
45
46
47
    chat_completions::{NvCreateChatCompletionRequest, NvCreateChatCompletionResponse},
    completions::{NvCreateCompletionRequest, NvCreateCompletionResponse},
    embeddings::{NvCreateEmbeddingRequest, NvCreateEmbeddingResponse},
    responses::{NvCreateResponse, NvResponse},
48
};
49
use crate::request_template::RequestTemplate;
50
use crate::types::Annotated;
51
52
use dynamo_runtime::logging::get_distributed_tracing_context;
use tracing::Instrument;
53

Ryan Olson's avatar
Ryan Olson committed
54
55
56
57
58
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";

59
60
61
62
// 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 {
63
    std::env::var(env_llm::DYN_HTTP_BODY_LIMIT_MB)
64
65
66
67
68
69
        .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
70
71
pub type ErrorResponse = (StatusCode, Json<ErrorMessage>);

72
#[derive(Serialize, Deserialize, Debug)]
Ryan Olson's avatar
Ryan Olson committed
73
pub(crate) struct ErrorMessage {
74
75
76
77
78
79
80
81
82
83
84
    message: String,
    #[serde(rename = "type")]
    error_type: String,
    code: u16,
}

fn map_error_code_to_error_type(code: StatusCode) -> String {
    match code.canonical_reason() {
        Some(reason) => reason.to_string(),
        None => "UnknownError".to_string(),
    }
85
86
}

Ryan Olson's avatar
Ryan Olson committed
87
impl ErrorMessage {
88
    /// Not Found Error
Ryan Olson's avatar
Ryan Olson committed
89
    pub fn model_not_found() -> ErrorResponse {
90
91
        let code = StatusCode::NOT_FOUND;
        let error_type = map_error_code_to_error_type(code);
92
        (
93
            code,
Ryan Olson's avatar
Ryan Olson committed
94
            Json(ErrorMessage {
95
96
97
                message: "Model not found".to_string(),
                error_type,
                code: code.as_u16(),
98
99
100
101
102
103
            }),
        )
    }

    /// Service Unavailable
    /// This is returned when the service is live, but not ready.
Ryan Olson's avatar
Ryan Olson committed
104
    pub fn _service_unavailable() -> ErrorResponse {
105
106
        let code = StatusCode::SERVICE_UNAVAILABLE;
        let error_type = map_error_code_to_error_type(code);
107
        (
108
            code,
Ryan Olson's avatar
Ryan Olson committed
109
            Json(ErrorMessage {
110
111
112
                message: "Service is not ready".to_string(),
                error_type,
                code: code.as_u16(),
113
114
115
116
117
118
119
120
            }),
        )
    }

    /// 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
121
    pub fn internal_server_error(msg: &str) -> ErrorResponse {
122
        tracing::error!("Internal server error: {msg}");
123
124
        let code = StatusCode::INTERNAL_SERVER_ERROR;
        let error_type = map_error_code_to_error_type(code);
125
        (
126
            code,
Ryan Olson's avatar
Ryan Olson committed
127
            Json(ErrorMessage {
128
129
130
                message: msg.to_string(),
                error_type,
                code: code.as_u16(),
131
132
133
134
            }),
        )
    }

135
136
137
    /// 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
138
    pub fn not_implemented_error(msg: &str) -> ErrorResponse {
139
        tracing::error!("Not Implemented error: {msg}");
140
141
        let code = StatusCode::NOT_IMPLEMENTED;
        let error_type = map_error_code_to_error_type(code);
142
        (
143
            code,
Ryan Olson's avatar
Ryan Olson committed
144
            Json(ErrorMessage {
145
146
147
                message: msg.to_string(),
                error_type,
                code: code.as_u16(),
148
149
150
151
            }),
        )
    }

Neelay Shah's avatar
Neelay Shah committed
152
    /// The OAI endpoints call an [`dynamo.runtime::engine::AsyncEngine`] which are specialized to return
153
    /// an [`anyhow::Error`]. This method will convert the [`anyhow::Error`] into an [`HttpError`].
Ryan Olson's avatar
Ryan Olson committed
154
    /// If successful, it will return the [`HttpError`] as an [`ErrorMessage::internal_server_error`]
155
    /// with the details of the error.
Ryan Olson's avatar
Ryan Olson committed
156
    pub fn from_anyhow(err: anyhow::Error, alt_msg: &str) -> ErrorResponse {
157
158
159
        // First check for PipelineError::ServiceOverloaded
        if let Some(pipeline_err) =
            err.downcast_ref::<dynamo_runtime::pipeline::error::PipelineError>()
160
            && matches!(
161
162
                pipeline_err,
                dynamo_runtime::pipeline::error::PipelineError::ServiceOverloaded(_)
163
164
165
166
167
            )
        {
            return (
                StatusCode::SERVICE_UNAVAILABLE,
                Json(ErrorMessage {
168
169
170
                    message: pipeline_err.to_string(),
                    error_type: map_error_code_to_error_type(StatusCode::SERVICE_UNAVAILABLE),
                    code: StatusCode::SERVICE_UNAVAILABLE.as_u16(),
171
172
                }),
            );
173
174
175
        }

        // Then check for HttpError
176
        match err.downcast::<HttpError>() {
Ryan Olson's avatar
Ryan Olson committed
177
178
            Ok(http_error) => ErrorMessage::from_http_error(http_error),
            Err(err) => ErrorMessage::internal_server_error(&format!("{alt_msg}: {err}")),
179
180
181
182
        }
    }

    /// Implementers should only be able to throw 400-499 errors.
Ryan Olson's avatar
Ryan Olson committed
183
    pub fn from_http_error(err: HttpError) -> ErrorResponse {
184
        if err.code < 400 || err.code >= 500 {
Ryan Olson's avatar
Ryan Olson committed
185
            return ErrorMessage::internal_server_error(&err.message);
186
187
        }
        match StatusCode::from_u16(err.code) {
188
189
190
191
192
193
194
195
            Ok(code) => (
                code,
                Json(ErrorMessage {
                    message: err.message,
                    error_type: map_error_code_to_error_type(code),
                    code: code.as_u16(),
                }),
            ),
Ryan Olson's avatar
Ryan Olson committed
196
            Err(_) => ErrorMessage::internal_server_error(&err.message),
197
198
199
200
        }
    }
}

Ryan Olson's avatar
Ryan Olson committed
201
impl From<HttpError> for ErrorMessage {
202
    fn from(err: HttpError) -> Self {
203
204
205
206
207
208
209
        ErrorMessage {
            message: err.message,
            error_type: map_error_code_to_error_type(
                StatusCode::from_u16(err.code).unwrap_or(StatusCode::INTERNAL_SERVER_ERROR),
            ),
            code: err.code,
        }
210
211
212
    }
}

213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
// Problem: Currently we are using JSON from axum as the request validator. Whenever there is an invalid JSON, it will return a 422.
// But all the downstream apps that relies on openai based APIs, expects to get 400 for all these cases otherwise they fail badly
// Solution: Intercept the response from handlers and convert ANY 422 status codes to 400 with the actual error message.
pub async fn smart_json_error_middleware(request: Request<Body>, next: Next) -> Response {
    let response = next.run(request).await;

    if response.status() == StatusCode::UNPROCESSABLE_ENTITY {
        let (_parts, body) = response.into_parts();
        let body_bytes = axum::body::to_bytes(body, usize::MAX)
            .await
            .unwrap_or_default();
        let error_message = String::from_utf8_lossy(&body_bytes).to_string();
        (
            StatusCode::BAD_REQUEST,
            Json(ErrorMessage {
228
229
230
                message: error_message,
                error_type: map_error_code_to_error_type(StatusCode::BAD_REQUEST),
                code: StatusCode::BAD_REQUEST.as_u16(),
231
232
233
234
235
236
237
238
239
            }),
        )
            .into_response()
    } else {
        // Pass through if it is not a 422
        response
    }
}

Ryan Olson's avatar
Ryan Olson committed
240
/// Get the request ID from a primary source, or next from the headers, or lastly create a new one if not present
241
// TODO: Similar function exists in lib/llm/src/grpc/service/openai.rs but with different signature and simpler logic
Ryan Olson's avatar
Ryan Olson committed
242
fn get_or_create_request_id(primary: Option<&str>, headers: &HeaderMap) -> String {
243
    // Try to get request id from trace context
244
245
246
247
    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;
248
249
    }

Ryan Olson's avatar
Ryan Olson committed
250
    // Try to get the request ID from the primary source
251
252
253
254
    if let Some(primary) = primary
        && let Ok(uuid) = uuid::Uuid::parse_str(primary)
    {
        return uuid.to_string();
Ryan Olson's avatar
Ryan Olson committed
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
    }

    // 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()
}

273
274
275
276
277
278
279
280
/// 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
281
async fn handler_completions(
282
    State(state): State<Arc<service_v2::State>>,
Ryan Olson's avatar
Ryan Olson committed
283
    headers: HeaderMap,
284
    Json(request): Json<NvCreateCompletionRequest>,
Ryan Olson's avatar
Ryan Olson committed
285
286
287
288
289
290
291
292
293
294
) -> 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
295
296
    let (mut connection_handle, stream_handle) =
        create_connection_monitor(context.clone(), Some(state.metrics_clone())).await;
Ryan Olson's avatar
Ryan Olson committed
297
298
299

    // possibly long running task
    // if this returns a streaming response, the stream handle will be armed and captured by the response stream
300
    let response = tokio::spawn(completions(state, request, stream_handle).in_current_span())
Ryan Olson's avatar
Ryan Olson committed
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
        .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> {
322
323
    use crate::protocols::openai::completions::get_prompt_batch_size;

324
325
326
    // return a 503 if the service is not ready
    check_ready(&state)?;

327
328
    validate_completion_fields_generic(&request)?;

329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
    // Detect batch prompts
    let batch_size = get_prompt_batch_size(&request.inner.prompt);
    let n = request.inner.n.unwrap_or(1);

    // If single prompt or single-element batch, use original flow
    if batch_size == 1 {
        return completions_single(state, request, stream_handle).await;
    }

    // Batch processing: handle multiple prompts
    completions_batch(state, request, stream_handle, batch_size, n).await
}

/// Handle single prompt completions (original logic)
#[tracing::instrument(skip_all)]
async fn completions_single(
    state: Arc<service_v2::State>,
    request: Context<NvCreateCompletionRequest>,
    stream_handle: ConnectionHandle,
) -> Result<Response, ErrorResponse> {
349
    let request_id = request.id().to_string();
350
351

    // todo - decide on default
352
    let streaming = request.inner.stream.unwrap_or(false);
353

354
355
356
357
358
359
    // todo - make the protocols be optional for model name
    // todo - when optional, if none, apply a default
    let model = request.inner.model.clone();
    // Create http_queue_guard early - tracks time waiting to be processed
    let http_queue_guard = state.metrics_clone().create_http_queue_guard(&model);

360
361
    // todo - error handling should be more robust
    let engine = state
362
        .manager()
363
        .get_completions_engine(&model)
Ryan Olson's avatar
Ryan Olson committed
364
        .map_err(|_| ErrorMessage::model_not_found())?;
365

366
    let parsing_options = state.manager().get_parsing_options(&model);
367

368
    let mut response_collector = state.metrics_clone().create_response_collector(&model);
369

Ryan Olson's avatar
Ryan Olson committed
370
371
    // prepare to process any annotations
    let annotations = request.annotations();
372

373
374
375
376
377
378
    // Create inflight_guard before calling engine to ensure errors are counted
    let mut inflight_guard =
        state
            .metrics_clone()
            .create_inflight_guard(&model, Endpoint::Completions, streaming);

379
380
381
382
    // issue the generate call on the engine
    let stream = engine
        .generate(request)
        .await
Ryan Olson's avatar
Ryan Olson committed
383
        .map_err(|e| ErrorMessage::from_anyhow(e, "Failed to generate completions"))?;
384
385
386
387

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

Ryan Olson's avatar
Ryan Olson committed
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
    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);
407
408

    if streaming {
409
410
        // For streaming, we'll drop the http_queue_guard on the first token
        let mut http_queue_guard = Some(http_queue_guard);
411
        let stream = stream.map(move |response| {
412
413
414
415
416
417
            // Calls observe_response() on each token
            process_response_using_event_converter_and_observe_metrics(
                EventConverter::from(response),
                &mut response_collector,
                &mut http_queue_guard,
            )
418
        });
Ryan Olson's avatar
Ryan Olson committed
419
        let stream = monitor_for_disconnects(stream, ctx, inflight_guard, stream_handle);
420

421
422
        let mut sse_stream = Sse::new(stream);

423
        if let Some(keep_alive) = state.sse_keep_alive() {
424
425
426
427
            sse_stream = sse_stream.keep_alive(KeepAlive::default().interval(keep_alive));
        }

        Ok(sse_stream.into_response())
428
    } else {
429
        // Tap the stream to collect metrics for non-streaming requests without altering items
430
        let mut http_queue_guard = Some(http_queue_guard);
431
        let stream = stream.inspect(move |response| {
432
433
434
435
436
437
            // Calls observe_response() on each token - drops http_queue_guard on first token
            process_response_and_observe_metrics(
                response,
                &mut response_collector,
                &mut http_queue_guard,
            );
438
439
        });

440
        let response = NvCreateCompletionResponse::from_annotated_stream(stream, parsing_options)
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
            .await
            .map_err(|e| {
                tracing::error!(
                    "Failed to fold completions stream for {}: {:?}",
                    request_id,
                    e
                );
                ErrorMessage::internal_server_error(&format!(
                    "Failed to fold completions stream for {}: {:?}",
                    request_id, e
                ))
            })?;

        inflight_guard.mark_ok();
        Ok(Json(response).into_response())
    }
}

/// Handle batch prompt completions (multiple prompts with n choices each)
#[tracing::instrument(skip_all)]
async fn completions_batch(
    state: Arc<service_v2::State>,
    request: Context<NvCreateCompletionRequest>,
    stream_handle: ConnectionHandle,
    batch_size: usize,
    n: u8,
) -> Result<Response, ErrorResponse> {
    use crate::protocols::openai::completions::extract_single_prompt;
    use futures::stream::{self, StreamExt};

    let request_id = request.id().to_string();
    let streaming = request.inner.stream.unwrap_or(false);
    let model = request.inner.model.clone();

    // Create http_queue_guard early - tracks time waiting to be processed
    let http_queue_guard = state.metrics_clone().create_http_queue_guard(&model);

    let engine = state
        .manager()
        .get_completions_engine(&model)
        .map_err(|_| ErrorMessage::model_not_found())?;

    let parsing_options = state.manager().get_parsing_options(&model);

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

    // prepare to process any annotations
    let annotations = request.annotations();

    // Create inflight_guard before calling engine to ensure errors are counted
    let mut inflight_guard =
        state
            .metrics_clone()
            .create_inflight_guard(&model, Endpoint::Completions, streaming);

    // Generate streams for each prompt in the batch
    let mut all_streams = Vec::new();
    let mut first_ctx = None;

    for prompt_idx in 0..batch_size {
        // Extract single prompt at this index
        let single_prompt = extract_single_prompt(&request.inner.prompt, prompt_idx);

        // Create a new request with this single prompt
        let mut single_request = request.content().clone();
        single_request.inner.prompt = single_prompt;

        // Generate unique request_id for each prompt: original_id-{prompt_idx}
        let unique_request_id = format!("{}-{}", request.id(), prompt_idx);
        let single_request_context = Context::with_id(single_request, unique_request_id);

        // Generate stream for this prompt
        let stream = engine
            .generate(single_request_context)
            .await
            .map_err(|e| ErrorMessage::from_anyhow(e, "Failed to generate completions"))?;

        // Capture context from first stream
        if first_ctx.is_none() {
            first_ctx = Some(stream.context());
        }

        // Remap choice indices: choice.index += prompt_idx * n
        let prompt_idx_u32 = prompt_idx as u32;
        let n_u32 = n as u32;
        let remapped_stream = stream.map(move |mut response| {
            if let Some(ref mut data) = response.data {
                for choice in &mut data.inner.choices {
                    choice.index += prompt_idx_u32 * n_u32;
                }
            }
            response
        });

        all_streams.push(remapped_stream);
    }

    // Merge all streams
    let merged_stream = stream::select_all(all_streams);

    // capture the context to cancel the stream if the client disconnects
    let ctx = first_ctx.expect("At least one stream should be generated");

    let annotations_vec = 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 merged_stream = stream::iter(annotations_vec).chain(merged_stream);

    if streaming {
        // For streaming, we'll drop the http_queue_guard on the first token
        let mut http_queue_guard = Some(http_queue_guard);
        let stream = merged_stream.map(move |response| {
            // Calls observe_response() on each token
            process_response_using_event_converter_and_observe_metrics(
                EventConverter::from(response),
                &mut response_collector,
                &mut http_queue_guard,
            )
        });
        let stream = monitor_for_disconnects(stream, ctx, inflight_guard, stream_handle);

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

        if let Some(keep_alive) = state.sse_keep_alive() {
            sse_stream = sse_stream.keep_alive(KeepAlive::default().interval(keep_alive));
        }

        Ok(sse_stream.into_response())
    } else {
        // Tap the stream to collect metrics for non-streaming requests without altering items
        let mut http_queue_guard = Some(http_queue_guard);
        let stream = merged_stream.inspect(move |response| {
            // Calls observe_response() on each token - drops http_queue_guard on first token
            process_response_and_observe_metrics(
                response,
                &mut response_collector,
                &mut http_queue_guard,
            );
        });

        let response = NvCreateCompletionResponse::from_annotated_stream(stream, parsing_options)
597
598
599
600
601
602
603
            .await
            .map_err(|e| {
                tracing::error!(
                    "Failed to fold completions stream for {}: {:?}",
                    request_id,
                    e
                );
604
605
606
607
                ErrorMessage::internal_server_error(&format!(
                    "Failed to fold completions stream for {}: {:?}",
                    request_id, e
                ))
608
609
            })?;

610
        inflight_guard.mark_ok();
611
612
613
614
        Ok(Json(response).into_response())
    }
}

615
616
#[tracing::instrument(skip_all)]
async fn embeddings(
617
    State(state): State<Arc<service_v2::State>>,
618
    headers: HeaderMap,
619
    Json(request): Json<NvCreateEmbeddingRequest>,
Ryan Olson's avatar
Ryan Olson committed
620
) -> Result<Response, ErrorResponse> {
621
622
623
    // return a 503 if the service is not ready
    check_ready(&state)?;

624
625
626
    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();
627
628
629
630
631
632
633
634

    // 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;

635
636
637
    // Create http_queue_guard early - tracks time waiting to be processed
    let http_queue_guard = state.metrics_clone().create_http_queue_guard(model);

638
639
640
641
    // todo - error handling should be more robust
    let engine = state
        .manager()
        .get_embeddings_engine(model)
Ryan Olson's avatar
Ryan Olson committed
642
        .map_err(|_| ErrorMessage::model_not_found())?;
643
644
645
646
647
648
649

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

650
651
    let mut response_collector = state.metrics_clone().create_response_collector(model);

652
653
654
655
    // issue the generate call on the engine
    let stream = engine
        .generate(request)
        .await
Ryan Olson's avatar
Ryan Olson committed
656
        .map_err(|e| ErrorMessage::from_anyhow(e, "Failed to generate embeddings"))?;
657

658
659
660
661
662
663
664
665
666
667
668
    // Process stream to collect metrics and drop http_queue_guard on first token
    let mut http_queue_guard = Some(http_queue_guard);
    let stream = stream.inspect(move |response| {
        // Calls observe_response() on each token - drops http_queue_guard on first token
        process_response_and_observe_metrics(
            response,
            &mut response_collector,
            &mut http_queue_guard,
        );
    });

669
670
    // 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
671
    let response = NvCreateEmbeddingResponse::from_annotated_stream(stream)
672
673
674
675
676
677
678
        .await
        .map_err(|e| {
            tracing::error!(
                "Failed to fold embeddings stream for {}: {:?}",
                request_id,
                e
            );
Ryan Olson's avatar
Ryan Olson committed
679
            ErrorMessage::internal_server_error("Failed to fold embeddings stream")
680
681
682
683
        })?;

    inflight.mark_ok();
    Ok(Json(response).into_response())
684
685
}

Ryan Olson's avatar
Ryan Olson committed
686
687
688
689
690
691
692
693
694
695
696
697
698
699
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
700
701
    let (mut connection_handle, stream_handle) =
        create_connection_monitor(context.clone(), Some(state.metrics_clone())).await;
Ryan Olson's avatar
Ryan Olson committed
702

703
704
705
706
707
708
709
710
711
    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
712
713
714
715
716
717
718
719

    // 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
}

720
721
722
723
724
725
726
727
728
/// 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
729
730
731
732
733
    state: Arc<service_v2::State>,
    template: Option<RequestTemplate>,
    mut request: Context<NvCreateChatCompletionRequest>,
    mut stream_handle: ConnectionHandle,
) -> Result<Response, ErrorResponse> {
734
735
736
    // return a 503 if the service is not ready
    check_ready(&state)?;

Ryan Olson's avatar
Ryan Olson committed
737
738
    let request_id = request.id().to_string();

739
740
741
742
    // 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
743
    validate_chat_completion_unsupported_fields(&request)?;
744

745
746
747
    // Handle required fields like messages shouldn't be empty.
    validate_chat_completion_required_fields(&request)?;

748
749
750
    // Handle Rest of Validation Errors
    validate_chat_completion_fields_generic(&request)?;

751
752
753
754
755
756
757
758
759
760
761
762
    // 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
763
    tracing::trace!("Received chat completions request: {:?}", request.content());
764
765

    // todo - decide on default
Paul Hendricks's avatar
Paul Hendricks committed
766
    let streaming = request.inner.stream.unwrap_or(false);
767
768
769
770

    // todo - make the protocols be optional for model name
    // todo - when optional, if none, apply a default
    // todo - determine the proper error code for when a request model is not present
771
772
773
774
775
    let model = request.inner.model.clone();

    // Create HTTP queue guard after template resolution so labels are correct
    let http_queue_guard = state.metrics_clone().create_http_queue_guard(&model);

776
777
778
    tracing::trace!("Getting chat completions engine for model: {}", model);

    let engine = state
779
        .manager()
780
        .get_chat_completions_engine(&model)
Ryan Olson's avatar
Ryan Olson committed
781
        .map_err(|_| ErrorMessage::model_not_found())?;
782

783
    let parsing_options = state.manager().get_parsing_options(&model);
784

785
786
787
788
789
    let mut response_collector = state.metrics_clone().create_response_collector(&model);

    let annotations = request.annotations();

    // Create inflight_guard before calling engine to ensure errors are counted
790
    let mut inflight_guard =
791
792
        state
            .metrics_clone()
793
            .create_inflight_guard(&model, Endpoint::ChatCompletions, streaming);
794
795
796
797
798

    // issue the generate call on the engine
    let stream = engine
        .generate(request)
        .await
Ryan Olson's avatar
Ryan Olson committed
799
        .map_err(|e| ErrorMessage::from_anyhow(e, "Failed to generate completions"))?;
800
801
802
803

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

Ryan Olson's avatar
Ryan Olson committed
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
    // 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);

821
822
823
824
    // 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 {
825
        stream_handle.arm(); // allows the system to detect client disconnects and cancel the LLM generation
Ryan Olson's avatar
Ryan Olson committed
826

827
        let mut http_queue_guard = Some(http_queue_guard);
828
        let stream = stream.map(move |response| {
829
830
831
832
833
834
            // Calls observe_response() on each token
            process_response_using_event_converter_and_observe_metrics(
                EventConverter::from(response),
                &mut response_collector,
                &mut http_queue_guard,
            )
835
        });
Ryan Olson's avatar
Ryan Olson committed
836
        let stream = monitor_for_disconnects(stream, ctx, inflight_guard, stream_handle);
837

838
839
        let mut sse_stream = Sse::new(stream);

840
        if let Some(keep_alive) = state.sse_keep_alive() {
841
842
843
844
            sse_stream = sse_stream.keep_alive(KeepAlive::default().interval(keep_alive));
        }

        Ok(sse_stream.into_response())
845
    } else {
846
        let mut http_queue_guard = Some(http_queue_guard);
847
        let stream = stream.inspect(move |response| {
848
849
850
851
852
853
            // Calls observe_response() on each token - drops http_queue_guard on first token
            process_response_and_observe_metrics(
                response,
                &mut response_collector,
                &mut http_queue_guard,
            );
854
855
        });

856
857
858
859
860
861
862
863
864
865
866
867
868
869
        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
                    ))
                })?;
870

871
        inflight_guard.mark_ok();
872
873
874
875
        Ok(Json(response).into_response())
    }
}

876
877
878
879
880
/// 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
881
) -> Result<(), ErrorResponse> {
882
883
884
    let inner = &request.inner;

    if inner.function_call.is_some() {
Ryan Olson's avatar
Ryan Olson committed
885
        return Err(ErrorMessage::not_implemented_error(
886
887
888
889
890
            "`function_call` is deprecated. Please migrate to use `tool_choice` instead.",
        ));
    }

    if inner.functions.is_some() {
Ryan Olson's avatar
Ryan Olson committed
891
        return Err(ErrorMessage::not_implemented_error(
892
893
894
895
            "`functions` is deprecated. Please migrate to use `tools` instead.",
        ));
    }

Ryan Olson's avatar
Ryan Olson committed
896
    Ok(())
897
898
}

899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
/// 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(())
}

916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
/// Validates a chat completion request and returns an error response if validation fails.
///
/// This function calls the `validate` method implemented for `NvCreateChatCompletionRequest`.
/// If validation fails, it maps the error into an OpenAI-compatible error response.
pub fn validate_chat_completion_fields_generic(
    request: &NvCreateChatCompletionRequest,
) -> Result<(), ErrorResponse> {
    request.validate().map_err(|e| {
        ErrorMessage::from_http_error(HttpError {
            code: 400,
            message: e.to_string(),
        })
    })
}

/// Validates a completion request and returns an error response if validation fails.
///
/// This function calls the `validate` method implemented for `NvCreateCompletionRequest`.
/// If validation fails, it maps the error into an OpenAI-compatible error response.
pub fn validate_completion_fields_generic(
    request: &NvCreateCompletionRequest,
) -> Result<(), ErrorResponse> {
    request.validate().map_err(|e| {
        ErrorMessage::from_http_error(HttpError {
            code: 400,
            message: e.to_string(),
        })
    })
}

946
947
948
/// OpenAI Responses Request Handler
///
/// This method will handle the incoming request for the /v1/responses endpoint.
Ryan Olson's avatar
Ryan Olson committed
949
async fn handler_responses(
950
    State((state, template)): State<(Arc<service_v2::State>, Option<RequestTemplate>)>,
Ryan Olson's avatar
Ryan Olson committed
951
952
953
954
955
956
957
958
959
960
961
962
    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
963
964
    let (mut connection_handle, _stream_handle) =
        create_connection_monitor(context.clone(), Some(state.metrics_clone())).await;
Ryan Olson's avatar
Ryan Olson committed
965

966
    let response = tokio::spawn(responses(state, template, request).in_current_span())
Ryan Olson's avatar
Ryan Olson committed
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
        .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> {
988
989
990
    // return a 503 if the service is not ready
    check_ready(&state)?;

991
992
993
994
    // Create http_queue_guard early - tracks time waiting to be processed
    let model = request.inner.model.clone();
    let http_queue_guard = state.metrics_clone().create_http_queue_guard(&model);

995
996
997
    // 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.
998
    if let Some(resp) = validate_response_unsupported_fields(&request) {
999
1000
1001
1002
1003
1004
1005
        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.
1006
    if let Some(resp) = validate_response_input_is_text_only(&request) {
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
        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
1024
1025
    let request_id = request.id().to_string();
    let (request, context) = request.into_parts();
1026

Ryan Olson's avatar
Ryan Olson committed
1027
    let mut request: NvCreateChatCompletionRequest = request.try_into().map_err(|e| {
1028
1029
1030
1031
1032
        tracing::error!(
            request_id,
            "Failed to convert NvCreateResponse to NvCreateChatCompletionRequest: {:?}",
            e
        );
Ryan Olson's avatar
Ryan Olson committed
1033
        ErrorMessage::not_implemented_error(&format!(
1034
1035
1036
1037
1038
            "Only Input::Text(_) is currently supported: {}",
            e
        ))
    })?;

Ryan Olson's avatar
Ryan Olson committed
1039
1040
1041
1042
1043
    let request = context.map(|mut _req| {
        request.inner.stream = Some(false);
        request
    });

1044
1045
1046
1047
    tracing::trace!("Getting chat completions engine for model: {}", model);

    let engine = state
        .manager()
1048
        .get_chat_completions_engine(&model)
Ryan Olson's avatar
Ryan Olson committed
1049
        .map_err(|_| ErrorMessage::model_not_found())?;
1050

1051
    let parsing_options = state.manager().get_parsing_options(&model);
1052

1053
    let mut response_collector = state.metrics_clone().create_response_collector(&model);
1054
1055
1056
1057
1058
1059
1060

    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
1061
        .map_err(|e| ErrorMessage::from_anyhow(e, "Failed to generate completions"))?;
1062

1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
    // Create inflight_guard now that actual processing has begun
    let mut inflight_guard =
        state
            .metrics_clone()
            .create_inflight_guard(&model, Endpoint::Responses, false);

    // Process stream to collect metrics and drop http_queue_guard on first token
    let mut http_queue_guard = Some(http_queue_guard);
    let stream = stream.inspect(move |response| {
        // Calls observe_response() on each token - drops http_queue_guard on first token
        process_response_and_observe_metrics(
            response,
            &mut response_collector,
            &mut http_queue_guard,
        );
    });

1080
    // TODO: handle streaming, currently just unary
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
    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
                ))
            })?;
1095
1096
1097
1098
1099
1100
1101
1102

    // 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
1103
        ErrorMessage::internal_server_error("Failed to convert internal response")
1104
1105
1106
1107
1108
1109
1110
    })?;

    inflight_guard.mark_ok();

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

1111
1112
1113
pub fn validate_response_input_is_text_only(
    request: &NvCreateResponse,
) -> Option<impl IntoResponse> {
1114
    match &request.inner.input {
1115
        dynamo_async_openai::types::responses::Input::Text(_) => None,
1116
1117
1118
        _ => Some(ErrorMessage::not_implemented_error(
            "Only `Input::Text` is supported. Structured, multimedia, or custom input types are not yet implemented.",
        )),
1119
1120
1121
1122
1123
    }
}

/// Checks for unsupported fields in the request.
/// Returns Some(response) if unsupported fields are present.
1124
1125
1126
pub fn validate_response_unsupported_fields(
    request: &NvCreateResponse,
) -> Option<impl IntoResponse> {
1127
1128
1129
    let inner = &request.inner;

    if inner.background == Some(true) {
Ryan Olson's avatar
Ryan Olson committed
1130
        return Some(ErrorMessage::not_implemented_error(
1131
1132
1133
1134
            "`background: true` is not supported.",
        ));
    }
    if inner.include.is_some() {
Ryan Olson's avatar
Ryan Olson committed
1135
        return Some(ErrorMessage::not_implemented_error(
1136
1137
1138
1139
            "`include` is not supported.",
        ));
    }
    if inner.instructions.is_some() {
Ryan Olson's avatar
Ryan Olson committed
1140
        return Some(ErrorMessage::not_implemented_error(
1141
1142
1143
1144
            "`instructions` is not supported.",
        ));
    }
    if inner.max_tool_calls.is_some() {
Ryan Olson's avatar
Ryan Olson committed
1145
        return Some(ErrorMessage::not_implemented_error(
1146
1147
1148
1149
            "`max_tool_calls` is not supported.",
        ));
    }
    if inner.previous_response_id.is_some() {
Ryan Olson's avatar
Ryan Olson committed
1150
        return Some(ErrorMessage::not_implemented_error(
1151
1152
1153
1154
            "`previous_response_id` is not supported.",
        ));
    }
    if inner.prompt.is_some() {
Ryan Olson's avatar
Ryan Olson committed
1155
        return Some(ErrorMessage::not_implemented_error(
1156
1157
1158
1159
            "`prompt` is not supported.",
        ));
    }
    if inner.reasoning.is_some() {
Ryan Olson's avatar
Ryan Olson committed
1160
        return Some(ErrorMessage::not_implemented_error(
1161
1162
1163
1164
            "`reasoning` is not supported.",
        ));
    }
    if inner.service_tier.is_some() {
Ryan Olson's avatar
Ryan Olson committed
1165
        return Some(ErrorMessage::not_implemented_error(
1166
1167
1168
1169
            "`service_tier` is not supported.",
        ));
    }
    if inner.store == Some(true) {
Ryan Olson's avatar
Ryan Olson committed
1170
        return Some(ErrorMessage::not_implemented_error(
1171
1172
1173
1174
            "`store: true` is not supported.",
        ));
    }
    if inner.stream == Some(true) {
Ryan Olson's avatar
Ryan Olson committed
1175
        return Some(ErrorMessage::not_implemented_error(
1176
1177
1178
1179
            "`stream: true` is not supported.",
        ));
    }
    if inner.text.is_some() {
Ryan Olson's avatar
Ryan Olson committed
1180
        return Some(ErrorMessage::not_implemented_error(
1181
1182
1183
1184
            "`text` is not supported.",
        ));
    }
    if inner.tool_choice.is_some() {
Ryan Olson's avatar
Ryan Olson committed
1185
        return Some(ErrorMessage::not_implemented_error(
1186
1187
1188
1189
            "`tool_choice` is not supported.",
        ));
    }
    if inner.tools.is_some() {
Ryan Olson's avatar
Ryan Olson committed
1190
        return Some(ErrorMessage::not_implemented_error(
1191
1192
1193
1194
            "`tools` is not supported.",
        ));
    }
    if inner.truncation.is_some() {
Ryan Olson's avatar
Ryan Olson committed
1195
        return Some(ErrorMessage::not_implemented_error(
1196
1197
1198
1199
            "`truncation` is not supported.",
        ));
    }
    if inner.user.is_some() {
Ryan Olson's avatar
Ryan Olson committed
1200
        return Some(ErrorMessage::not_implemented_error(
1201
1202
1203
1204
1205
1206
1207
            "`user` is not supported.",
        ));
    }

    None
}

1208
1209
// 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
1210
fn check_ready(_state: &Arc<service_v2::State>) -> Result<(), ErrorResponse> {
1211
    // if state.service_observer.stage() != ServiceStage::Ready {
Ryan Olson's avatar
Ryan Olson committed
1212
    //     return Err(ErrorMessage::service_unavailable());
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
    // }
    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(
1231
    State(state): State<Arc<service_v2::State>>,
Ryan Olson's avatar
Ryan Olson committed
1232
) -> Result<Response, ErrorResponse> {
1233
1234
1235
1236
1237
1238
1239
1240
    check_ready(&state)?;

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

1241
1242
    let models: HashSet<String> = state.manager().model_display_names();
    for model_name in models {
1243
        data.push(ModelListing {
1244
            id: model_name.clone(),
1245
            object: "object",
1246
1247
            created,                        // Where would this come from?
            owned_by: "nvidia".to_string(), // Get organization from config
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
        });
    }

    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,
}

/// 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(
1275
    state: Arc<service_v2::State>,
1276
1277
1278
1279
1280
    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
1281
        .route(&path, post(handler_completions))
1282
        .layer(middleware::from_fn(smart_json_error_middleware))
1283
        .layer(axum::extract::DefaultBodyLimit::max(get_body_limit()))
1284
1285
1286
1287
1288
1289
1290
        .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(
1291
    state: Arc<service_v2::State>,
1292
    template: Option<RequestTemplate>,
1293
1294
1295
1296
1297
    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
1298
        .route(&path, post(handler_chat_completions))
1299
        .layer(middleware::from_fn(smart_json_error_middleware))
1300
        .layer(axum::extract::DefaultBodyLimit::max(get_body_limit()))
1301
        .with_state((state, template));
1302
1303
1304
    (vec![doc], router)
}

1305
1306
1307
/// 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(
1308
    state: Arc<service_v2::State>,
1309
1310
1311
1312
1313
1314
    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))
1315
        .layer(middleware::from_fn(smart_json_error_middleware))
1316
        .layer(axum::extract::DefaultBodyLimit::max(get_body_limit()))
1317
1318
1319
1320
        .with_state(state);
    (vec![doc], router)
}

1321
1322
/// List Models
pub fn list_models_router(
1323
    state: Arc<service_v2::State>,
1324
1325
1326
    path: Option<String>,
) -> (Vec<RouteDoc>, Router) {
    // Standard OpenAI compatible list models endpoint
1327
    let openai_path = path.unwrap_or("/v1/models".to_string());
1328
1329
1330
1331
1332
1333
    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);

1334
    (vec![doc_for_openai], router)
1335
1336
}

1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
/// 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
1347
        .route(&path, post(handler_responses))
1348
        .layer(middleware::from_fn(smart_json_error_middleware))
1349
1350
1351
1352
        .with_state((state, template));
    (vec![doc], router)
}

1353
1354
#[cfg(test)]
mod tests {
1355

1356
1357
1358
1359
1360
1361
    use super::*;
    use crate::discovery::ModelManagerError;
    use crate::protocols::openai::chat_completions::NvCreateChatCompletionRequest;
    use crate::protocols::openai::common_ext::CommonExt;
    use crate::protocols::openai::completions::NvCreateCompletionRequest;
    use crate::protocols::openai::responses::NvCreateResponse;
1362
    use dynamo_async_openai::types::responses::{
1363
1364
1365
1366
        CreateResponse, Input, InputContent, InputItem, InputMessage, PromptConfig,
        Role as ResponseRole, ServiceTier, TextConfig, TextResponseFormat, ToolChoice,
        ToolChoiceMode, Truncation,
    };
1367
    use dynamo_async_openai::types::{
1368
1369
        ChatCompletionRequestMessage, ChatCompletionRequestUserMessage,
        ChatCompletionRequestUserMessageContent, CreateChatCompletionRequest,
1370
        CreateCompletionRequest,
1371
    };
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382

    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> {
1383
        Err(ModelManagerError::ModelNotFound("foo".to_string()))?
1384
1385
    }

1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
    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,
        }
    }

1417
1418
1419
    #[test]
    fn test_http_error_response_from_anyhow() {
        let err = http_error_from_engine(400).unwrap_err();
1420
1421
1422
        let response = ErrorMessage::from_anyhow(err, BACKUP_ERROR_MESSAGE);
        assert_eq!(response.0, StatusCode::BAD_REQUEST);
        assert_eq!(response.1.message, "custom error message");
1423
1424
1425
1426
1427
    }

    #[test]
    fn test_error_response_from_anyhow_out_of_range() {
        let err = http_error_from_engine(399).unwrap_err();
1428
1429
1430
        let response = ErrorMessage::from_anyhow(err, BACKUP_ERROR_MESSAGE);
        assert_eq!(response.0, StatusCode::INTERNAL_SERVER_ERROR);
        assert_eq!(response.1.message, "custom error message");
1431
1432

        let err = http_error_from_engine(500).unwrap_err();
1433
1434
1435
        let response = ErrorMessage::from_anyhow(err, BACKUP_ERROR_MESSAGE);
        assert_eq!(response.0, StatusCode::INTERNAL_SERVER_ERROR);
        assert_eq!(response.1.message, "custom error message");
1436
1437

        let err = http_error_from_engine(501).unwrap_err();
1438
1439
1440
        let response = ErrorMessage::from_anyhow(err, BACKUP_ERROR_MESSAGE);
        assert_eq!(response.0, StatusCode::INTERNAL_SERVER_ERROR);
        assert_eq!(response.1.message, "custom error message");
1441
1442
1443
1444
1445
    }

    #[test]
    fn test_other_error_response_from_anyhow() {
        let err = other_error_from_engine().unwrap_err();
1446
1447
        let response = ErrorMessage::from_anyhow(err, BACKUP_ERROR_MESSAGE);
        assert_eq!(response.0, StatusCode::INTERNAL_SERVER_ERROR);
1448
        assert_eq!(
1449
            response.1.message,
1450
1451
1452
1453
1454
1455
1456
            format!(
                "{}: {}",
                BACKUP_ERROR_MESSAGE,
                other_error_from_engine().unwrap_err()
            )
        );
    }
1457

1458
1459
1460
1461
1462
1463
1464
1465
    #[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();
1466
1467
        let response = ErrorMessage::from_anyhow(err, BACKUP_ERROR_MESSAGE);
        assert_eq!(response.0, StatusCode::SERVICE_UNAVAILABLE);
1468
        assert_eq!(
1469
            response.1.message,
1470
1471
1472
1473
            "Service temporarily unavailable: All workers are busy, please retry later"
        );
    }

1474
1475
1476
    #[test]
    fn test_validate_input_is_text_only_accepts_text() {
        let request = make_base_request();
1477
        let result = validate_response_input_is_text_only(&request);
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
        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()),
        })]);
1489
        let result = validate_response_input_is_text_only(&request);
1490
1491
1492
1493
1494
1495
        assert!(result.is_some());
    }

    #[test]
    fn test_validate_unsupported_fields_accepts_clean_request() {
        let request = make_base_request();
1496
        let result = validate_response_unsupported_fields(&request);
1497
1498
1499
        assert!(result.is_none());
    }

1500
1501
1502
1503
1504
1505
1506
1507
    #[test]
    fn test_validate_unsupported_fields_accepts_parallel_tool_calls() {
        let mut request = make_base_request();
        request.inner.parallel_tool_calls = Some(true);
        let result = validate_response_unsupported_fields(&request);
        assert!(result.is_none(), "parallel_tool_calls should be supported");
    }

1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
    #[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))),
            (
                "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);
1569
            let result = validate_response_unsupported_fields(&req);
1570
1571
1572
            assert!(result.is_some(), "Expected rejection for `{field}`");
        }
    }
1573
1574
1575
1576
1577
1578
1579
1580
1581

    #[test]
    fn test_validate_chat_completion_required_fields_empty_messages() {
        let request = NvCreateChatCompletionRequest {
            inner: CreateChatCompletionRequest {
                model: "test-model".to_string(),
                messages: vec![],
                ..Default::default()
            },
1582
            common: Default::default(),
1583
            nvext: None,
1584
            chat_template_args: None,
1585
            unsupported_fields: Default::default(),
1586
1587
1588
        };
        let result = validate_chat_completion_required_fields(&request);
        assert!(result.is_err());
1589
1590
        if let Err(error_response) = result {
            assert_eq!(error_response.0, StatusCode::BAD_REQUEST);
1591
            assert_eq!(
1592
                error_response.1.message,
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
                "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()
            },
1611
            common: Default::default(),
1612
            nvext: None,
1613
            chat_template_args: None,
1614
            unsupported_fields: Default::default(),
1615
1616
1617
1618
        };
        let result = validate_chat_completion_required_fields(&request);
        assert!(result.is_ok());
    }
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634

    #[test]
    // Test for all Bad Requests Example for Chat Completion
    // 1. Echo:  Should be a boolean : Not Done
    // 2. Frequency Penalty: Should be a float between -2.0 and 2.0 : Done
    // 3. logprobs: Done
    // 4. Model Format: Should be a string : Not Done
    // 5. Prompt or Messages Validation
    // 6. Max Tokens: Should be a positive integer
    // 7. Presence Penalty: Should be a float between -2.0 and 2.0 : Done
    // 8. Stop : Should be a string or an array of strings : Not Done
    // 9. Invalid or Out of range temperature: Done
    // 10.Invalid or out of range top_p: Done
    // 11. Repetition Penalty: Should be a float between 0.0 and 2.0 : Done
    // 12. Logprobs: Should be a positive integer between 0 and 5 : Done
    // invalid or non existing user : Only empty string is not allowed validation is there. How can we check non-extisting user ?
1635
    // Unknown fields : Done (rejected via extra_fields catch-all)
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
    // guided_whitespace_pattern null or invalid : Not Done
    // "response_format": { "type": "invalid_format" } : Not Done
    // "logit_bias": { "invalid_token": "not_a_number" }, : Partial Validation is already there
    fn test_bad_base_request_for_completion() {
        // Frequency Penalty: Should be a float between -2.0 and 2.0
        let request = NvCreateCompletionRequest {
            inner: CreateCompletionRequest {
                model: "test-model".to_string(),
                prompt: "Hello".into(),
                frequency_penalty: Some(-3.0),
                ..Default::default()
            },
            common: Default::default(),
            nvext: None,
1650
            metadata: None,
1651
            unsupported_fields: Default::default(),
1652
1653
1654
1655
        };

        let result = validate_completion_fields_generic(&request);
        assert!(result.is_err());
1656
1657
        if let Err(error_response) = result {
            assert_eq!(error_response.0, StatusCode::BAD_REQUEST);
1658
            assert_eq!(
1659
                error_response.1.message,
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
                "Frequency penalty must be between -2 and 2, got -3"
            );
        }

        // Presence Penalty: Should be a float between -2.0 and 2.0
        let request = NvCreateCompletionRequest {
            inner: CreateCompletionRequest {
                model: "test-model".to_string(),
                prompt: "Hello".into(),
                presence_penalty: Some(-3.0),
                ..Default::default()
            },
            common: Default::default(),
            nvext: None,
1674
            metadata: None,
1675
            unsupported_fields: Default::default(),
1676
1677
1678
        };
        let result = validate_completion_fields_generic(&request);
        assert!(result.is_err());
1679
1680
        if let Err(error_response) = result {
            assert_eq!(error_response.0, StatusCode::BAD_REQUEST);
1681
            assert_eq!(
1682
                error_response.1.message,
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
                "Presence penalty must be between -2 and 2, got -3"
            );
        }

        // Temperature: Should be a float between 0.0 and 2.0
        let request = NvCreateCompletionRequest {
            inner: CreateCompletionRequest {
                model: "test-model".to_string(),
                prompt: "Hello".into(),
                temperature: Some(-3.0),
                ..Default::default()
            },
            common: Default::default(),
            nvext: None,
1697
            metadata: None,
1698
            unsupported_fields: Default::default(),
1699
1700
1701
        };
        let result = validate_completion_fields_generic(&request);
        assert!(result.is_err());
1702
1703
        if let Err(error_response) = result {
            assert_eq!(error_response.0, StatusCode::BAD_REQUEST);
1704
            assert_eq!(
1705
                error_response.1.message,
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
                "Temperature must be between 0 and 2, got -3"
            );
        }

        // Top P: Should be a float between 0.0 and 1.0
        let request = NvCreateCompletionRequest {
            inner: CreateCompletionRequest {
                model: "test-model".to_string(),
                prompt: "Hello".into(),
                top_p: Some(-3.0),
                ..Default::default()
            },
            common: Default::default(),
            nvext: None,
1720
            metadata: None,
1721
            unsupported_fields: Default::default(),
1722
1723
1724
        };
        let result = validate_completion_fields_generic(&request);
        assert!(result.is_err());
1725
1726
        if let Err(error_response) = result {
            assert_eq!(error_response.0, StatusCode::BAD_REQUEST);
1727
            assert_eq!(
1728
                error_response.1.message,
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
                "Top_p must be between 0 and 1, got -3"
            );
        }

        // Repetition Penalty: Should be a float between 0.0 and 2.0
        let request = NvCreateCompletionRequest {
            inner: CreateCompletionRequest {
                model: "test-model".to_string(),
                prompt: "Hello".into(),
                ..Default::default()
            },
            common: CommonExt::builder()
                .repetition_penalty(-3.0)
                .build()
                .unwrap(),
            nvext: None,
1745
            metadata: None,
1746
            unsupported_fields: Default::default(),
1747
1748
1749
        };
        let result = validate_completion_fields_generic(&request);
        assert!(result.is_err());
1750
1751
        if let Err(error_response) = result {
            assert_eq!(error_response.0, StatusCode::BAD_REQUEST);
1752
            assert_eq!(
1753
                error_response.1.message,
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
                "Repetition penalty must be between 0 and 2, got -3"
            );
        }

        // Logprobs: Should be a positive integer between 0 and 5
        let request = NvCreateCompletionRequest {
            inner: CreateCompletionRequest {
                model: "test-model".to_string(),
                prompt: "Hello".into(),
                logprobs: Some(6),
                ..Default::default()
            },
            common: Default::default(),
            nvext: None,
1768
            metadata: None,
1769
            unsupported_fields: Default::default(),
1770
1771
1772
        };
        let result = validate_completion_fields_generic(&request);
        assert!(result.is_err());
1773
1774
        if let Err(error_response) = result {
            assert_eq!(error_response.0, StatusCode::BAD_REQUEST);
1775
            assert_eq!(
1776
                error_response.1.message,
1777
1778
1779
1780
1781
                "Logprobs must be between 0 and 5, got 6"
            );
        }
    }

1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
    #[test]
    fn test_metadata_field_nested() {
        use serde_json::json;

        // Test metadata field with nested object
        let request = NvCreateCompletionRequest {
            inner: CreateCompletionRequest {
                model: "test-model".to_string(),
                prompt: "Hello".into(),
                ..Default::default()
            },
            common: Default::default(),
            nvext: None,
            metadata: json!({
                "user": {"id": 1, "name": "user-1"},
                "session": {"id": "session-1", "timestamp": 1640995200}
            })
            .into(),
1800
            unsupported_fields: Default::default(),
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
        };

        let result = validate_completion_fields_generic(&request);
        assert!(result.is_ok());

        // Verify metadata is accessible
        assert!(request.metadata.is_some());
        assert_eq!(request.metadata.as_ref().unwrap()["user"]["id"], 1);
    }

1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
    #[test]
    fn test_bad_base_request_for_chatcompletion() {
        // Frequency Penalty: Should be a float between -2.0 and 2.0
        let request = NvCreateChatCompletionRequest {
            inner: CreateChatCompletionRequest {
                model: "test-model".to_string(),
                messages: vec![ChatCompletionRequestMessage::User(
                    ChatCompletionRequestUserMessage {
                        content: ChatCompletionRequestUserMessageContent::Text("Hello".to_string()),
                        name: None,
                    },
                )],
                frequency_penalty: Some(-3.0),
                ..Default::default()
            },
            common: Default::default(),
            nvext: None,
1828
            chat_template_args: None,
1829
            unsupported_fields: Default::default(),
1830
1831
1832
1833
        };

        let result = validate_chat_completion_fields_generic(&request);
        assert!(result.is_err());
1834
1835
        if let Err(error_response) = result {
            assert_eq!(error_response.0, StatusCode::BAD_REQUEST);
1836
            assert_eq!(
1837
                error_response.1.message,
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
                "Frequency penalty must be between -2 and 2, got -3"
            );
        }

        // Presence Penalty: Should be a float between -2.0 and 2.0
        let request = NvCreateChatCompletionRequest {
            inner: CreateChatCompletionRequest {
                model: "test-model".to_string(),
                messages: vec![ChatCompletionRequestMessage::User(
                    ChatCompletionRequestUserMessage {
                        content: ChatCompletionRequestUserMessageContent::Text("Hello".to_string()),
                        name: None,
                    },
                )],
                presence_penalty: Some(-3.0),
                ..Default::default()
            },
            common: Default::default(),
            nvext: None,
1857
            chat_template_args: None,
1858
            unsupported_fields: Default::default(),
1859
1860
1861
        };
        let result = validate_chat_completion_fields_generic(&request);
        assert!(result.is_err());
1862
1863
        if let Err(error_response) = result {
            assert_eq!(error_response.0, StatusCode::BAD_REQUEST);
1864
            assert_eq!(
1865
                error_response.1.message,
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
                "Presence penalty must be between -2 and 2, got -3"
            );
        }

        // Temperature: Should be a float between 0.0 and 2.0
        let request = NvCreateChatCompletionRequest {
            inner: CreateChatCompletionRequest {
                model: "test-model".to_string(),
                messages: vec![ChatCompletionRequestMessage::User(
                    ChatCompletionRequestUserMessage {
                        content: ChatCompletionRequestUserMessageContent::Text("Hello".to_string()),
                        name: None,
                    },
                )],
                temperature: Some(-3.0),
                ..Default::default()
            },
            common: Default::default(),
            nvext: None,
1885
            chat_template_args: None,
1886
            unsupported_fields: Default::default(),
1887
1888
1889
        };
        let result = validate_chat_completion_fields_generic(&request);
        assert!(result.is_err());
1890
1891
        if let Err(error_response) = result {
            assert_eq!(error_response.0, StatusCode::BAD_REQUEST);
1892
            assert_eq!(
1893
                error_response.1.message,
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
                "Temperature must be between 0 and 2, got -3"
            );
        }

        // Top P: Should be a float between 0.0 and 1.0
        let request = NvCreateChatCompletionRequest {
            inner: CreateChatCompletionRequest {
                model: "test-model".to_string(),
                messages: vec![ChatCompletionRequestMessage::User(
                    ChatCompletionRequestUserMessage {
                        content: ChatCompletionRequestUserMessageContent::Text("Hello".to_string()),
                        name: None,
                    },
                )],
                top_p: Some(-3.0),
                ..Default::default()
            },
            common: Default::default(),
            nvext: None,
1913
            chat_template_args: None,
1914
            unsupported_fields: Default::default(),
1915
1916
1917
        };
        let result = validate_chat_completion_fields_generic(&request);
        assert!(result.is_err());
1918
1919
        if let Err(error_response) = result {
            assert_eq!(error_response.0, StatusCode::BAD_REQUEST);
1920
            assert_eq!(
1921
                error_response.1.message,
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
                "Top_p must be between 0 and 1, got -3"
            );
        }

        // Repetition Penalty: Should be a float between 0.0 and 2.0
        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()
            },
            common: CommonExt::builder()
                .repetition_penalty(-3.0)
                .build()
                .unwrap(),
            nvext: None,
1943
            chat_template_args: None,
1944
            unsupported_fields: Default::default(),
1945
1946
1947
        };
        let result = validate_chat_completion_fields_generic(&request);
        assert!(result.is_err());
1948
1949
        if let Err(error_response) = result {
            assert_eq!(error_response.0, StatusCode::BAD_REQUEST);
1950
            assert_eq!(
1951
                error_response.1.message,
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
                "Repetition penalty must be between 0 and 2, got -3"
            );
        }

        // Top Logprobs: Should be a positive integer between 0 and 20
        let request = NvCreateChatCompletionRequest {
            inner: CreateChatCompletionRequest {
                model: "test-model".to_string(),
                messages: vec![ChatCompletionRequestMessage::User(
                    ChatCompletionRequestUserMessage {
                        content: ChatCompletionRequestUserMessageContent::Text("Hello".to_string()),
                        name: None,
                    },
                )],
                top_logprobs: Some(25),
                ..Default::default()
            },
            common: Default::default(),
            nvext: None,
1971
            chat_template_args: None,
1972
            unsupported_fields: Default::default(),
1973
1974
1975
        };
        let result = validate_chat_completion_fields_generic(&request);
        assert!(result.is_err());
1976
1977
        if let Err(error_response) = result {
            assert_eq!(error_response.0, StatusCode::BAD_REQUEST);
1978
            assert_eq!(
1979
                error_response.1.message,
1980
1981
1982
1983
                "Top_logprobs must be between 0 and 20, got 25"
            );
        }
    }
1984
1985

    #[test]
1986
1987
    fn test_chat_completions_unknown_fields_rejected() {
        // Test that known unsupported fields are rejected and all shown in error message
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
        let json = r#"{
            "messages": [{"role": "user", "content": "Hello"}],
            "model": "test-model",
            "add_special_tokens": true,
            "documents": ["doc1"],
            "chat_template": "custom",
            "chat_template_kwargs": {"key": "val"}
        }"#;

        let request: NvCreateChatCompletionRequest = serde_json::from_str(json).unwrap();

        // Verify all unsupported fields were captured
        assert!(
            request
                .unsupported_fields
                .contains_key("add_special_tokens")
        );
        assert!(request.unsupported_fields.contains_key("documents"));
        assert!(request.unsupported_fields.contains_key("chat_template"));
        assert!(
            request
                .unsupported_fields
                .contains_key("chat_template_kwargs")
        );

        let result = validate_chat_completion_fields_generic(&request);
        assert!(result.is_err());
        if let Err(error_response) = result {
            assert_eq!(error_response.0, StatusCode::BAD_REQUEST);
            let msg = &error_response.1.message;
            assert!(msg.contains("Unsupported parameter"));
            // Verify all fields appear in the error message
            assert!(msg.contains("add_special_tokens"));
            assert!(msg.contains("documents"));
            assert!(msg.contains("chat_template"));
            assert!(msg.contains("chat_template_kwargs"));
        }
    }
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057

    #[test]
    fn test_completions_unsupported_fields_rejected() {
        // Test that known unsupported fields are rejected and all shown in error message
        let json = r#"{
            "model": "test-model",
            "prompt": "Hello",
            "add_special_tokens": true,
            "response_format": {"type": "json_object"}
        }"#;

        let request: NvCreateCompletionRequest = serde_json::from_str(json).unwrap();

        // Verify both unsupported fields were captured
        assert!(
            request
                .unsupported_fields
                .contains_key("add_special_tokens")
        );
        assert!(request.unsupported_fields.contains_key("response_format"));

        let result = validate_completion_fields_generic(&request);
        assert!(result.is_err());
        if let Err(error_response) = result {
            assert_eq!(error_response.0, StatusCode::BAD_REQUEST);
            let msg = &error_response.1.message;
            assert!(msg.contains("Unsupported parameter"));
            // Verify both fields appear in error message
            assert!(msg.contains("add_special_tokens"));
            assert!(msg.contains("response_format"));
        }
    }
2058
}