preprocessor.rs 34.2 KB
Newer Older
Biswa Panda's avatar
Biswa Panda committed
1
2
3
4
5
6
// SPDX-FileCopyrightText: Copyright (c) 2024-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
// SPDX-License-Identifier: Apache-2.0

//! The Preprocessor consists of the following modules
//!
//! - `translation`: This module converts the allowed Ingress message types to the corresponding
7
//!   internal representation.
Biswa Panda's avatar
Biswa Panda committed
8
9
10
11
12
13
14
15
16
17
//! - `apply`: This module applies ModelConfig defaults to any empty optional fields specified
//! - `prompt`: This module applies any prompt template logic to the internal Request object.
//! - `tokenize`: This module tokenizes the formatted prompt string and returns the token ids.
//!
//! The Preprocessor will accept any IngressRequest and transform it to a BackendRequest.

pub mod prompt;
pub mod tools;

use anyhow::Result;
Ryan Olson's avatar
Ryan Olson committed
18
19
use dynamo_async_openai::types::{ChatCompletionToolChoiceOption, EncodingFormat};
use futures::Stream;
Biswa Panda's avatar
Biswa Panda committed
20
21
use futures::stream::{self, StreamExt};
use prompt::OAIPromptFormatter;
22
use rayon::iter::{IntoParallelRefIterator, ParallelIterator};
Ryan Olson's avatar
Ryan Olson committed
23
use std::{collections::HashMap, pin::Pin, sync::Arc};
Biswa Panda's avatar
Biswa Panda committed
24
25
use tracing;

26
use crate::model_card::{ModelDeploymentCard, ModelInfo};
Biswa Panda's avatar
Biswa Panda committed
27
use crate::preprocessor::prompt::OAIChatLikeRequest;
28
use crate::protocols::common::preprocessor::PreprocessedRequestBuilder;
29
use crate::tokenizers::Encoding;
Biswa Panda's avatar
Biswa Panda committed
30

Neelay Shah's avatar
Neelay Shah committed
31
32
use dynamo_runtime::engine::{AsyncEngine, AsyncEngineContextProvider, ResponseStream};
use dynamo_runtime::pipeline::{
33
    AsyncEngineContext, Error, ManyOut, Operator, SingleIn, async_trait,
Biswa Panda's avatar
Biswa Panda committed
34
};
Neelay Shah's avatar
Neelay Shah committed
35
use dynamo_runtime::protocols::annotated::{Annotated, AnnotationsProvider};
Biswa Panda's avatar
Biswa Panda committed
36
37

use crate::protocols::{
Greg Clark's avatar
Greg Clark committed
38
    common::{OutputOptionsProvider, SamplingOptionsProvider, StopConditionsProvider},
Biswa Panda's avatar
Biswa Panda committed
39
    openai::{
40
        DeltaGeneratorExt,
Ryan Olson's avatar
Ryan Olson committed
41
42
43
        chat_completions::{
            NvCreateChatCompletionRequest, NvCreateChatCompletionStreamResponse, jail::JailedStream,
        },
44
        completions::{NvCreateCompletionRequest, NvCreateCompletionResponse},
45
        embeddings::{NvCreateEmbeddingRequest, NvCreateEmbeddingResponse},
Biswa Panda's avatar
Biswa Panda committed
46
47
48
        nvext::NvExtProvider,
    },
};
49
use crate::tokenizers::{HuggingFaceTokenizer, traits::Tokenizer};
Biswa Panda's avatar
Biswa Panda committed
50

51
use crate::preprocessor::prompt::{PromptFormatter, PromptInput, TextInput, TokenInput};
Biswa Panda's avatar
Biswa Panda committed
52

53
pub use crate::protocols::common::llm_backend::{BackendOutput, PreprocessedRequest};
54
55
56
pub use crate::protocols::common::preprocessor::PreprocessedEmbeddingRequest;

use crate::protocols::common::llm_backend::EmbeddingsEngineOutput;
Biswa Panda's avatar
Biswa Panda committed
57
58
59

pub const ANNOTATION_FORMATTED_PROMPT: &str = "formatted_prompt";
pub const ANNOTATION_TOKEN_IDS: &str = "token_ids";
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
pub const ANNOTATION_LLM_METRICS: &str = "llm_metrics";
#[derive(Debug, Clone, serde::Serialize, serde::Deserialize)]
pub struct LLMMetricAnnotation {
    pub input_tokens: usize,
    pub output_tokens: usize,
    pub chunk_tokens: usize,
}

impl LLMMetricAnnotation {
    /// Convert this metrics struct to an Annotated event
    pub fn to_annotation<T>(&self) -> Result<Annotated<T>, serde_json::Error> {
        Annotated::from_annotation(ANNOTATION_LLM_METRICS, self)
    }

    /// Extract LLM metrics from an Annotated event, if present
    pub fn from_annotation<T>(
        annotation: &Annotated<T>,
    ) -> Result<Option<LLMMetricAnnotation>, Box<dyn std::error::Error>> {
        if annotation.event.is_none() {
            return Ok(None);
        }
        if annotation.event.as_ref().unwrap() != ANNOTATION_LLM_METRICS {
            return Ok(None);
        }
        let comments = annotation
            .comment
            .as_ref()
            .ok_or("missing comments block")?;
        if comments.len() != 1 {
            return Err("malformed comments block - expected exactly 1 comment".into());
        }
        let metrics: LLMMetricAnnotation = serde_json::from_str(&comments[0])?;
        Ok(Some(metrics))
    }
}
Biswa Panda's avatar
Biswa Panda committed
95
96
97
98
99
100

pub struct OpenAIPreprocessor {
    mdcsum: String,
    formatter: Arc<dyn OAIPromptFormatter>,
    tokenizer: Arc<dyn Tokenizer>,
    model_info: Arc<dyn ModelInfo>,
101
102
    /// Per-model runtime configuration propagated to response generator (e.g., reasoning/tool parser)
    runtime_config: crate::local_model::runtime_config::ModelRuntimeConfig,
103
    tool_call_parser: Option<String>,
Biswa Panda's avatar
Biswa Panda committed
104
105
106
}

impl OpenAIPreprocessor {
107
108
109
    pub fn new(mdc: ModelDeploymentCard) -> Result<Arc<Self>> {
        let formatter = PromptFormatter::from_mdc(&mdc)?;
        let tokenizer = mdc.tokenizer_hf()?;
110
        match formatter {
111
            PromptFormatter::OAI(formatter) => Self::new_with_parts(mdc, formatter, tokenizer),
112
113
114
        }
    }

115
    pub fn new_with_parts(
116
117
        mdc: ModelDeploymentCard,
        formatter: Arc<dyn OAIPromptFormatter>,
118
        hf_tokenizer: tokenizers::Tokenizer,
119
120
    ) -> Result<Arc<Self>> {
        let mdcsum = mdc.mdcsum();
121
        let tokenizer = Arc::new(HuggingFaceTokenizer::from_tokenizer(hf_tokenizer));
122
123
124
125
126
        let Some(model_info) = mdc.model_info else {
            anyhow::bail!(
                "Blank ModelDeploymentCard cannot be used for pre-processing, no model_info"
            );
        };
127
        let model_info = model_info.get_model_info()?;
128
        let tool_call_parser = mdc.runtime_config.tool_call_parser.clone();
Biswa Panda's avatar
Biswa Panda committed
129

130
131
132
        // // Initialize runtime config from the ModelDeploymentCard
        let runtime_config = mdc.runtime_config.clone();

Biswa Panda's avatar
Biswa Panda committed
133
134
135
136
137
        Ok(Arc::new(Self {
            formatter,
            tokenizer,
            model_info,
            mdcsum,
138
            runtime_config,
139
            tool_call_parser,
Biswa Panda's avatar
Biswa Panda committed
140
141
        }))
    }
142
143
144
145
146
    /// Encode a string to it's tokens
    pub fn tokenize(&self, s: &str) -> anyhow::Result<Encoding> {
        self.tokenizer.encode(s)
    }

147
    /// Translate a [`NvCreateChatCompletionRequest`] request to a common completion request.
Biswa Panda's avatar
Biswa Panda committed
148
149
150
151
152
153
154
155
156
157
    /// Returns both the common completion request and a hashmap of annotations.
    ///
    /// Annotations evaluated by this method include:
    /// - `formatted_prompt`
    /// - `token_ids`
    pub fn preprocess_request<
        R: OAIChatLikeRequest
            + AnnotationsProvider
            + SamplingOptionsProvider
            + StopConditionsProvider
Greg Clark's avatar
Greg Clark committed
158
            + OutputOptionsProvider
Biswa Panda's avatar
Biswa Panda committed
159
160
161
162
            + NvExtProvider,
    >(
        &self,
        request: &R,
163
    ) -> Result<(PreprocessedRequest, HashMap<String, String>)> {
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
        let mut builder = self.builder(request)?;
        let formatted_prompt = self.apply_template(request)?;
        let annotations = self.gather_tokens(request, &mut builder, formatted_prompt)?;

        Ok((builder.build()?, annotations))
    }

    pub fn builder<
        R: OAIChatLikeRequest
            + AnnotationsProvider
            + SamplingOptionsProvider
            + StopConditionsProvider
            + OutputOptionsProvider
            + NvExtProvider,
    >(
        &self,
        request: &R,
    ) -> Result<PreprocessedRequestBuilder> {
182
        let mut builder = PreprocessedRequest::builder();
183
        builder.model(request.model());
Biswa Panda's avatar
Biswa Panda committed
184

185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
        let mut stop_conditions = request.extract_stop_conditions()?;
        if let Some(stop_tokens) = &mut stop_conditions.stop_token_ids_hidden {
            for eos_token in self.model_info.eos_token_ids() {
                if !stop_tokens.contains(&eos_token) {
                    stop_tokens.push(eos_token);
                }
            }
        } else {
            stop_conditions.stop_token_ids_hidden = Some(self.model_info.eos_token_ids());
        }

        // apply ignore eos if not already set
        stop_conditions.apply_ignore_eos();

        if !stop_conditions.ignore_eos.unwrap_or(false) {
            builder.eos_token_ids(self.model_info.eos_token_ids());
        }

        builder.stop_conditions(stop_conditions);
        builder.sampling_options(request.extract_sampling_options()?);
        builder.output_options(request.extract_output_options()?);
        builder.annotations(request.annotations().unwrap_or_default());
        builder.mdc_sum(Some(self.mdcsum.clone()));
        builder.estimated_prefix_hit_num_blocks(None);
        // Extract backend_instance_id from nvext if present
        if let Some(nvext) = request.nvext() {
            builder.backend_instance_id(nvext.backend_instance_id);
        }

        Ok(builder)
    }

    pub fn apply_template<
        R: OAIChatLikeRequest
            + AnnotationsProvider
            + SamplingOptionsProvider
            + StopConditionsProvider
            + OutputOptionsProvider
            + NvExtProvider,
    >(
        &self,
        request: &R,
    ) -> Result<Option<String>> {
        if let PromptInput::Text(_) = request.prompt_input_type()
            && let Some(TextInput::Single(_)) = request.extract_text()
        {
            let use_raw_prompt = request
                .nvext()
                .is_some_and(|ext| ext.use_raw_prompt.unwrap_or(false));

            let formatted_prompt = if use_raw_prompt {
                match request.raw_prompt() {
                    Some(prompt) => prompt,
                    None => {
                        tracing::warn!("Raw prompt requested but not available");
                        self.formatter.render(request)?
                    }
                }
            } else {
                self.formatter.render(request)?
            };
            Ok(Some(formatted_prompt))
        } else {
            Ok(None)
        }
    }

    pub fn gather_tokens<
        R: OAIChatLikeRequest
            + AnnotationsProvider
            + SamplingOptionsProvider
            + StopConditionsProvider
            + OutputOptionsProvider
            + NvExtProvider,
    >(
        &self,
        request: &R,
        builder: &mut PreprocessedRequestBuilder,
        formatted_prompt: Option<String>,
    ) -> Result<HashMap<String, String>> {
        let mut annotations = HashMap::new();
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
        // match request type before any conversion/processing
        match request.prompt_input_type() {
            PromptInput::Tokens(_) => {
                if let Some(token_input) = request.extract_tokens() {
                    match token_input {
                        TokenInput::Single(tokens) => {
                            builder.token_ids(tokens);
                        }
                        TokenInput::Batch(token_batches) => {
                            if token_batches.len() == 1 {
                                builder.token_ids(token_batches[0].clone());
                            } else {
                                builder.batch_token_ids(Some(token_batches));
                                builder.token_ids(vec![]);
                            }
                        }
                    }
Biswa Panda's avatar
Biswa Panda committed
283
284
                }
            }
285
286
287
            PromptInput::Text(_) => {
                if let Some(text_input) = request.extract_text() {
                    match text_input {
288
289
290
291
292
293
294
295
296
297
                        TextInput::Single(raw_prompt) => {
                            if let Some(f) = formatted_prompt.as_ref()
                                && request.has_annotation(ANNOTATION_FORMATTED_PROMPT)
                            {
                                annotations
                                    .insert(ANNOTATION_FORMATTED_PROMPT.to_string(), f.to_string());
                            }

                            // Completions will use raw_prompt, no template
                            let prompt = formatted_prompt.unwrap_or(raw_prompt);
Biswa Panda's avatar
Biswa Panda committed
298

299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
                            // Check if backend_instance_id is present and token_data is provided
                            let has_backend_instance_id = request
                                .nvext()
                                .and_then(|ext| ext.backend_instance_id)
                                .is_some();

                            let token_data =
                                request.nvext().and_then(|ext| ext.token_data.as_ref());

                            let (tokens_vec, skip_token_annotation) = if has_backend_instance_id {
                                if let Some(tokens) = token_data {
                                    tracing::trace!(
                                        "Using provided tokens from EPP: {} ids",
                                        tokens.len()
                                    );
                                    // need ownership for the builder, so clone.
                                    (tokens.clone(), true)
                                } else {
                                    tracing::warn!(
                                        "backend_instance_id provided but no token_data; tokenizing prompt"
                                    );
320
                                    let encoding = self.tokenizer.encode(&prompt)?;
321
322
323
324
                                    (encoding.token_ids().to_vec(), false)
                                }
                            } else {
                                // No backend_instance_id provided, continue the normal flow.
325
                                let encoding = self.tokenizer.encode(&prompt)?;
326
327
                                (encoding.token_ids().to_vec(), false)
                            };
Biswa Panda's avatar
Biswa Panda committed
328

329
330
331
                            if request.has_annotation(ANNOTATION_TOKEN_IDS)
                                && !skip_token_annotation
                            {
332
333
                                annotations.insert(
                                    ANNOTATION_TOKEN_IDS.to_string(),
334
                                    serde_json::to_string(&tokens_vec)?,
335
336
337
                                );
                            }

338
                            builder.token_ids(tokens_vec);
339
340
                        }
                        TextInput::Batch(texts) => {
341
                            let token_batches: Vec<Vec<u32>> = texts
342
343
                                .par_iter()
                                .map(|text| {
344
345
346
                                    self.tokenizer
                                        .encode(text)
                                        .map(|encoded| encoded.token_ids().to_vec())
347
                                })
348
                                .collect::<Result<Vec<_>>>()?;
349
350
351
352
353
354
                            builder.batch_token_ids(Some(token_batches));
                            builder.token_ids(vec![]);
                        }
                    }
                }
            }
Biswa Panda's avatar
Biswa Panda committed
355
        }
356
        Ok(annotations)
Biswa Panda's avatar
Biswa Panda committed
357
358
    }

359
360
361
362
363
364
365
366
367
368
369
370
371
372
    /// Preprocess an embedding request, handling both text and token ID inputs.
    ///
    /// For text inputs, tokenizes the text using the configured tokenizer.
    /// For token ID inputs, uses the provided token IDs directly and skips tokenization.
    ///
    /// Returns both the preprocessed request and a hashmap of annotations.
    pub async fn preprocess_embedding_request(
        &self,
        request: &NvCreateEmbeddingRequest,
    ) -> Result<(PreprocessedEmbeddingRequest, HashMap<String, String>)> {
        let mut annotations = HashMap::new();
        let mut builder = PreprocessedEmbeddingRequest::builder();

        let all_token_ids = match &request.inner.input {
373
            dynamo_async_openai::types::EmbeddingInput::String(s) => {
374
375
                let encoding = self.tokenizer.encode(s)?;
                vec![encoding.token_ids().to_vec()]
376
            }
377
            dynamo_async_openai::types::EmbeddingInput::StringArray(arr) => {
378
379
380
381
382
383
384
385
386
387
388
                let input_strs: Vec<String> = arr.to_vec();
                let encodings = tokio::task::spawn_blocking({
                    let tokenizer = self.tokenizer.clone();
                    let strs = input_strs.clone();
                    move || {
                        tokenizer.encode_batch(&strs.iter().map(|s| s.as_str()).collect::<Vec<_>>())
                    }
                })
                .await??;
                let token_arrays: Vec<Vec<u32>> = encodings
                    .into_iter()
389
                    .map(|encoding| encoding.token_ids().to_vec())
390
391
392
                    .collect();
                token_arrays
            }
393
394
395
396
            dynamo_async_openai::types::EmbeddingInput::IntegerArray(token_ids) => {
                vec![token_ids.clone()]
            }
            dynamo_async_openai::types::EmbeddingInput::ArrayOfIntegerArray(token_arrays) => {
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
                token_arrays.clone()
            }
        };

        // Handle annotations
        if request.has_annotation(ANNOTATION_TOKEN_IDS) {
            annotations.insert(
                ANNOTATION_TOKEN_IDS.to_string(),
                serde_json::to_string(&all_token_ids)?,
            );
        }

        builder.token_ids(all_token_ids);
        builder.model(request.inner.model.clone());
        builder.encoding_format(request.inner.encoding_format.as_ref().map(|f| match f {
            EncodingFormat::Float => "float".to_string(),
            EncodingFormat::Base64 => "base64".to_string(),
        }));
        builder.dimensions(request.inner.dimensions);

        builder.annotations(request.annotations().unwrap_or_default());
        builder.mdc_sum(Some(self.mdcsum.clone()));

        Ok((builder.build()?, annotations))
    }

Ryan Olson's avatar
Ryan Olson committed
423
424
    pub fn transform_postprocessor_stream<S, Resp>(
        stream: S,
Biswa Panda's avatar
Biswa Panda committed
425
        generator: Box<dyn DeltaGeneratorExt<Resp>>,
Ryan Olson's avatar
Ryan Olson committed
426
427
428
429
430
431
432
433
434
435
436
        context: Arc<dyn AsyncEngineContext>,
    ) -> impl Stream<Item = Annotated<Resp>> + Send
    where
        S: Stream<Item = Annotated<BackendOutput>> + Send + 'static,
        Resp: Send + Sync + 'static + std::fmt::Debug,
    {
        struct State<Resp>
        where
            Resp: Send + Sync + 'static + std::fmt::Debug,
        {
            response_stream: Pin<Box<dyn Stream<Item = Annotated<BackendOutput>> + Send>>,
Biswa Panda's avatar
Biswa Panda committed
437
438
439
            response_generator: Box<dyn DeltaGeneratorExt<Resp>>,
            context: Arc<dyn AsyncEngineContext>,
            cancelled: bool,
440
            cumulative_output_tokens: usize,
441
442
            finish_reason_sent: bool,
            usage_chunk_sent: bool,
Ryan Olson's avatar
Ryan Olson committed
443
            finished: bool,
Biswa Panda's avatar
Biswa Panda committed
444
445
446
        }

        let state = State {
Ryan Olson's avatar
Ryan Olson committed
447
            response_stream: Box::pin(stream),
Biswa Panda's avatar
Biswa Panda committed
448
449
450
            response_generator: generator,
            context: context.clone(),
            cancelled: false,
451
            cumulative_output_tokens: 0,
452
453
            finish_reason_sent: false,
            usage_chunk_sent: false,
Ryan Olson's avatar
Ryan Olson committed
454
            finished: false,
Biswa Panda's avatar
Biswa Panda committed
455
456
457
        };

        // transform the common response stream into a chat response stream
Ryan Olson's avatar
Ryan Olson committed
458
459

        stream::unfold(state, |mut inner| {
Biswa Panda's avatar
Biswa Panda committed
460
            async move {
461
462
463
464
465
                // If already finished, return None immediately
                if inner.finished {
                    return None;
                }

Biswa Panda's avatar
Biswa Panda committed
466
467
468
469
470
471
                if let Some(response) = inner.response_stream.next().await {
                    if inner.cancelled {
                        tracing::debug!(
                            request_id = inner.context.id(),
                            "Cancellation issued last message; closing stream"
                        );
472
                        inner.finished = true; // Mark as finished
Biswa Panda's avatar
Biswa Panda committed
473
474
475
476
477
478
479
480
481
                        return None;
                    }

                    tracing::trace!(
                        request_id = inner.context.id(),
                        "Processing common response: {:?}",
                        response
                    );

482
483
484
485
486
487
488
                    // Check if this response has a finish_reason
                    let has_finish_reason = response
                        .data
                        .as_ref()
                        .map(|d| d.finish_reason.is_some())
                        .unwrap_or(false);

489
490
491
492
493
494
495
496
497
498
499
500
501
502
                    let (chunk_tokens, isl) = if let Some(ref backend_output) = response.data {
                        let chunk_tokens = backend_output.token_ids.len();
                        inner.cumulative_output_tokens += chunk_tokens;

                        let isl = inner.response_generator.get_isl().unwrap_or(0) as usize;

                        (chunk_tokens, isl)
                    } else {
                        (0, 0)
                    };

                    let current_osl = inner.cumulative_output_tokens;

                    let mut response = response.map_data(|data| {
Biswa Panda's avatar
Biswa Panda committed
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
                        inner
                            .response_generator
                            .choice_from_postprocessor(data)
                            .inspect_err(|e| {
                                tracing::error!(
                                    request_id = inner.context.id(),
                                    "Error processing common response: {:?}",
                                    e
                                );
                                inner.cancelled = true;
                                inner.context.stop_generating();
                            })
                            .map_err(|e| e.to_string())
                    });

518
519
520
521
522
523
524
525
526
527
528
                    // Create LLM metrics annotation
                    let llm_metrics = LLMMetricAnnotation {
                        input_tokens: isl,
                        output_tokens: current_osl,
                        chunk_tokens,
                    };

                    if let Ok(metrics_annotated) = llm_metrics.to_annotation::<()>() {
                        // Only set event if not already set to avoid overriding existing events (like errors)
                        if response.event.is_none() {
                            response.event = metrics_annotated.event;
529
                            response.comment = metrics_annotated.comment;
530
531
                        }
                    }
532

533
534
535
536
537
                    // Mark if we've seen a finish_reason
                    if has_finish_reason {
                        inner.finish_reason_sent = true;
                    }

Biswa Panda's avatar
Biswa Panda committed
538
539
                    tracing::trace!(
                        request_id = inner.context.id(),
540
                        "OpenAI NvCreateChatCompletionStreamResponse: {:?}",
Biswa Panda's avatar
Biswa Panda committed
541
542
543
544
545
                        response
                    );

                    Some((response, inner))
                } else {
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
                    // Stream has ended - check if we need to send a usage chunk
                    if inner.response_generator.is_usage_enabled()
                        && inner.finish_reason_sent
                        && !inner.usage_chunk_sent
                        && !inner.finished
                    {
                        inner.usage_chunk_sent = true;

                        // Create the final usage chunk
                        let usage_chunk = inner.response_generator.create_usage_chunk();
                        let annotated_usage = Annotated::<Resp> {
                            id: None,
                            data: Some(usage_chunk),
                            event: Some(ANNOTATION_LLM_METRICS.to_string()),
                            comment: None,
                        };

                        tracing::trace!(
                            request_id = inner.context.id(),
                            "Sending final usage chunk for OpenAI compliance"
                        );

                        Some((annotated_usage, inner))
                    } else {
                        // stream closed
                        inner.finished = true; // Mark as finished
                        None
                    }
Biswa Panda's avatar
Biswa Panda committed
574
575
                }
            }
Ryan Olson's avatar
Ryan Olson committed
576
        })
Biswa Panda's avatar
Biswa Panda committed
577
    }
578
579

    /// Transform engine embedding output stream to OpenAI embedding response stream
Ryan Olson's avatar
Ryan Olson committed
580
581
    pub fn transform_embedding_postprocessor_stream<S>(
        stream: S,
582
        original_request: NvCreateEmbeddingRequest,
Ryan Olson's avatar
Ryan Olson committed
583
584
585
586
587
    ) -> impl Stream<Item = Annotated<NvCreateEmbeddingResponse>> + Send
    where
        S: Stream<Item = Annotated<EmbeddingsEngineOutput>> + Send + 'static,
    {
        stream.map(move |output| {
588
589
            output.map_data(|engine_output| {
                // Convert engine output to OpenAI response format
590
                let embeddings: Vec<dynamo_async_openai::types::Embedding> = engine_output
591
592
593
                    .embeddings
                    .into_iter()
                    .enumerate()
594
                    .map(|(index, embedding)| dynamo_async_openai::types::Embedding {
595
596
597
598
599
600
601
                        index: index as u32,
                        object: "embedding".to_string(),
                        embedding: embedding.into_iter().map(|f| f as f32).collect(),
                    })
                    .collect();

                let response = NvCreateEmbeddingResponse {
602
                    inner: dynamo_async_openai::types::CreateEmbeddingResponse {
603
604
605
                        object: "list".to_string(),
                        model: original_request.inner.model.clone(),
                        data: embeddings,
606
                        usage: dynamo_async_openai::types::EmbeddingUsage {
607
608
609
610
611
612
613
614
                            prompt_tokens: engine_output.prompt_tokens,
                            total_tokens: engine_output.total_tokens,
                        },
                    },
                };

                Ok(response)
            })
Ryan Olson's avatar
Ryan Olson committed
615
        })
616
617
    }

Ryan Olson's avatar
Ryan Olson committed
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
    /// Determine if we should apply the tool calling jail based on configuration
    /// Returns Ok(true) if jail should be applied, Ok(false) if not, or Err if invalid config
    pub fn should_apply_tool_jail(
        tool_call_parser: Option<&String>,
        tool_choice: Option<&ChatCompletionToolChoiceOption>,
        has_tools: bool,
    ) -> std::result::Result<bool, Error> {
        match (tool_call_parser, tool_choice, has_tools) {
            // No parser but tools requested - error cases
            (None, Some(ChatCompletionToolChoiceOption::Required), true) => {
                tracing::warn!(
                    "Tool choice 'required' specified but no tool parser configured; proceeding without jailing"
                );
                Ok(false)
            }
            (None, Some(ChatCompletionToolChoiceOption::Auto), true) => {
                tracing::warn!(
                    "Tool choice 'auto' specified but no tool parser configured; proceeding without jailing"
                );
                Ok(false)
            }
            (None, Some(ChatCompletionToolChoiceOption::Named(_)), _) => {
                tracing::warn!(
                    "Named tool choice specified but no tool parser configured; proceeding without jailing"
                );
                Ok(false)
644
645
            }

Ryan Olson's avatar
Ryan Olson committed
646
647
648
            // Parser exists and tools might be called
            (Some(_), Some(ChatCompletionToolChoiceOption::None), _) => {
                Ok(false) // Explicitly disabled
649
            }
Ryan Olson's avatar
Ryan Olson committed
650
651
            (Some(_), Some(_), true) => Ok(true), // Any other tool_choice with tools
            (Some(_), None, true) => Ok(true),    // Default behavior when tools present
652

Ryan Olson's avatar
Ryan Olson committed
653
654
655
656
            // No tools or no parser
            _ => Ok(false),
        }
    }
657

Ryan Olson's avatar
Ryan Olson committed
658
659
660
661
662
663
664
665
666
667
668
669
670
    /// Apply tool calling jail to the stream if needed
    pub fn apply_tool_calling_jail<S>(
        tool_call_parser: String,
        stream: S,
    ) -> impl Stream<Item = Annotated<NvCreateChatCompletionStreamResponse>> + Send
    where
        S: Stream<Item = Annotated<NvCreateChatCompletionStreamResponse>> + Send + 'static,
    {
        let jail = JailedStream::builder()
            .tool_call_parser(tool_call_parser)
            .build();
        jail.apply(stream)
    }
Biswa Panda's avatar
Biswa Panda committed
671
672
673
674
675
676
677
678
679
680
}

// for pals, we do not want to add the generation prompt to the formatted prompt
// we also need to know if the template support this add_generation_prompt bool
// any prompt template that does not support this should return an error
// oob - we should update any prompt template that does not support this to support it

#[async_trait]
impl
    Operator<
681
        SingleIn<NvCreateChatCompletionRequest>,
682
        ManyOut<Annotated<NvCreateChatCompletionStreamResponse>>,
683
        SingleIn<PreprocessedRequest>,
Biswa Panda's avatar
Biswa Panda committed
684
685
686
687
688
        ManyOut<Annotated<BackendOutput>>,
    > for OpenAIPreprocessor
{
    async fn generate(
        &self,
689
        request: SingleIn<NvCreateChatCompletionRequest>,
Biswa Panda's avatar
Biswa Panda committed
690
        next: Arc<
691
            dyn AsyncEngine<SingleIn<PreprocessedRequest>, ManyOut<Annotated<BackendOutput>>, Error>,
Biswa Panda's avatar
Biswa Panda committed
692
        >,
693
    ) -> Result<ManyOut<Annotated<NvCreateChatCompletionStreamResponse>>, Error> {
Biswa Panda's avatar
Biswa Panda committed
694
695
696
697
        // unpack the request
        let (request, context) = request.into_parts();

        // create a response generator
698
        let response_generator = request.response_generator(context.id().to_string());
Ryan Olson's avatar
Ryan Olson committed
699
700
701
702

        // convert the chat completion request to a common completion request
        let (common_request, annotations) = self.preprocess_request(&request)?;

Biswa Panda's avatar
Biswa Panda committed
703
704
        let mut response_generator = Box::new(response_generator);

705
706
        // set the runtime configuration
        response_generator.set_reasoning_parser(self.runtime_config.clone());
Biswa Panda's avatar
Biswa Panda committed
707
708

        // update isl
Paul Hendricks's avatar
Paul Hendricks committed
709
        response_generator.update_isl(common_request.token_ids.len() as u32);
Biswa Panda's avatar
Biswa Panda committed
710
711
712
713
714

        // repack the common completion request
        let common_request = context.map(|_| common_request);

        // create a stream of annotations this will be prepend to the response stream
715
        let annotations: Vec<Annotated<NvCreateChatCompletionStreamResponse>> = annotations
Biswa Panda's avatar
Biswa Panda committed
716
717
718
719
720
721
722
723
            .into_iter()
            .flat_map(|(k, v)| Annotated::from_annotation(k, &v))
            .collect();
        let annotations_stream = stream::iter(annotations);

        // forward the common completion request to the next operator
        let response_stream = next.generate(common_request).await?;

Ryan Olson's avatar
Ryan Olson committed
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
        // Extract context once
        let context = response_stream.context();

        // transform the postprocessor stream (no boxing yet)
        let stream = Self::transform_postprocessor_stream(
            response_stream,
            response_generator,
            context.clone(),
        );

        // Check if tools are present and if we should apply jail
        let has_tools =
            request.inner.tools.is_some() && !request.inner.tools.as_ref().unwrap().is_empty();

        // Context was already extracted above from response_stream
Biswa Panda's avatar
Biswa Panda committed
739

Ryan Olson's avatar
Ryan Olson committed
740
741
742
743
744
745
746
747
748
749
750
751
752
753
        // Determine if we should apply jail (do this before moving request)
        let should_jail = Self::should_apply_tool_jail(
            self.tool_call_parser.as_ref(),
            request.inner.tool_choice.as_ref(),
            has_tools,
        )?;

        // Apply jail conditionally
        let stream: Pin<Box<dyn Stream<Item = _> + Send>> = if should_jail {
            if let Some(parser) = self.tool_call_parser.clone() {
                Box::pin(Self::apply_tool_calling_jail(parser, stream))
            } else {
                Box::pin(stream) // Should not happen due to should_jail check
            }
754
        } else {
Ryan Olson's avatar
Ryan Olson committed
755
            Box::pin(stream)
756
        };
Biswa Panda's avatar
Biswa Panda committed
757
758
759
        // prepend the annotations to the response stream
        let stream = annotations_stream.chain(stream);

Ryan Olson's avatar
Ryan Olson committed
760
        // return the response stream - single boxing at the end
Biswa Panda's avatar
Biswa Panda committed
761
762
763
764
765
766
767
        Ok(ResponseStream::new(Box::pin(stream), context))
    }
}

#[async_trait]
impl
    Operator<
768
        SingleIn<NvCreateCompletionRequest>,
769
        ManyOut<Annotated<NvCreateCompletionResponse>>,
770
        SingleIn<PreprocessedRequest>,
Biswa Panda's avatar
Biswa Panda committed
771
772
773
774
775
        ManyOut<Annotated<BackendOutput>>,
    > for OpenAIPreprocessor
{
    async fn generate(
        &self,
776
        request: SingleIn<NvCreateCompletionRequest>,
Biswa Panda's avatar
Biswa Panda committed
777
        next: Arc<
778
            dyn AsyncEngine<SingleIn<PreprocessedRequest>, ManyOut<Annotated<BackendOutput>>, Error>,
Biswa Panda's avatar
Biswa Panda committed
779
        >,
780
    ) -> Result<ManyOut<Annotated<NvCreateCompletionResponse>>, Error> {
Biswa Panda's avatar
Biswa Panda committed
781
782
783
784
        // unpack the request
        let (request, context) = request.into_parts();

        // create a response generator
785
        let response_generator = request.response_generator(context.id().to_string());
Biswa Panda's avatar
Biswa Panda committed
786
787
        let mut response_generator = Box::new(response_generator);
        // convert the chat completion request to a common completion request
788
789
790
        let mut builder = self.builder(&request)?;
        let annotations = self.gather_tokens(&request, &mut builder, None)?;
        let common_request = builder.build()?;
Biswa Panda's avatar
Biswa Panda committed
791
792

        // update isl
793
        response_generator.update_isl(common_request.token_ids.len() as u32);
Biswa Panda's avatar
Biswa Panda committed
794
795
796
797
798

        // repack the common completion request
        let common_request = context.map(|_| common_request);

        // create a stream of annotations this will be prepend to the response stream
799
        let annotations: Vec<Annotated<NvCreateCompletionResponse>> = annotations
Biswa Panda's avatar
Biswa Panda committed
800
801
802
803
804
805
806
807
            .into_iter()
            .flat_map(|(k, v)| Annotated::from_annotation(k, &v))
            .collect();
        let annotations_stream = stream::iter(annotations);

        // forward the common completion request to the next operator
        let response_stream = next.generate(common_request).await?;

Ryan Olson's avatar
Ryan Olson committed
808
809
810
        // Extract context once
        let context = response_stream.context();

Biswa Panda's avatar
Biswa Panda committed
811
        // transform the postprocessor stream
Ryan Olson's avatar
Ryan Olson committed
812
813
814
815
816
        let stream = Self::transform_postprocessor_stream(
            response_stream,
            response_generator,
            context.clone(),
        );
Biswa Panda's avatar
Biswa Panda committed
817
818
819
820
821
822
823
824

        // prepend the annotations to the response stream
        let stream = annotations_stream.chain(stream);

        // return the response stream
        Ok(ResponseStream::new(Box::pin(stream), context))
    }
}
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839

#[async_trait]
impl
    Operator<
        SingleIn<NvCreateEmbeddingRequest>,
        ManyOut<Annotated<NvCreateEmbeddingResponse>>,
        SingleIn<PreprocessedEmbeddingRequest>,
        ManyOut<Annotated<EmbeddingsEngineOutput>>,
    > for OpenAIPreprocessor
{
    async fn generate(
        &self,
        request: SingleIn<NvCreateEmbeddingRequest>,
        next: Arc<
            dyn AsyncEngine<
840
841
842
843
                    SingleIn<PreprocessedEmbeddingRequest>,
                    ManyOut<Annotated<EmbeddingsEngineOutput>>,
                    Error,
                >,
844
845
846
847
848
849
850
851
852
853
854
855
856
        >,
    ) -> Result<ManyOut<Annotated<NvCreateEmbeddingResponse>>, Error> {
        // Unpack request
        let (request, context) = request.into_parts();

        // Preprocess the embedding request
        let (preprocessed_request, annotations) =
            self.preprocess_embedding_request(&request).await?;

        // Forward to next stage
        let preprocessed_request = context.map(|_| preprocessed_request);
        let response_stream = next.generate(preprocessed_request).await?;

Ryan Olson's avatar
Ryan Olson committed
857
858
859
        // Extract context once
        let context = response_stream.context();

860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
        // Transform response stream back to OpenAI format
        let stream = Self::transform_embedding_postprocessor_stream(response_stream, request);

        // Prepend annotations
        let annotations_stream = stream::iter(
            annotations
                .into_iter()
                .flat_map(|(k, v)| Annotated::from_annotation(k, &v))
                .collect::<Vec<_>>(),
        );

        let combined_stream = annotations_stream.chain(stream);
        Ok(ResponseStream::new(Box::pin(combined_stream), context))
    }
}
Ryan Olson's avatar
Ryan Olson committed
875
876

// Note: tests for jailing and parser detection live in `lib/llm/tests/test_jail.rs`