sglang_scheduler.rs 23.9 KB
Newer Older
1
use std::convert::TryFrom;
2
3
4
5
use std::pin::Pin;
use std::sync::atomic::{AtomicBool, Ordering};
use std::sync::Arc;
use std::task::{Context, Poll};
6
use std::time::Duration;
7
use tonic::{transport::Channel, Request, Streaming};
8
use tracing::{debug, warn};
9

10
11
12
13
use crate::protocols::chat::ChatCompletionRequest;
use crate::protocols::common::{ResponseFormat, StringOrArray, ToolChoice, ToolChoiceValue};
use crate::protocols::generate::GenerateRequest;
use crate::protocols::sampling_params::SamplingParams as GenerateSamplingParams;
14

15
16
17
18
19
20
21
22
// Include the generated protobuf code
pub mod proto {
    tonic::include_proto!("sglang.grpc.scheduler");
}

// The generated module structure depends on the package name in the .proto file
// package sglang.grpc.scheduler; generates a nested module structure

23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
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
95
96
97
98
99
100
101
102
103
104
105
106
107
108
/// A smart wrapper around Streaming<GenerateResponse> that automatically
/// sends abort when dropped (e.g., due to client disconnection or early termination).
///
/// This leverages Rust's RAII pattern to ensure cleanup happens automatically,
/// regardless of how the stream is dropped (panic, early return, client disconnect, etc.).
pub struct AbortOnDropStream {
    inner: Streaming<proto::GenerateResponse>,
    request_id: String,
    client: SglangSchedulerClient,
    aborted: Arc<AtomicBool>,
}

impl AbortOnDropStream {
    /// Create a new auto-aborting stream wrapper
    pub fn new(
        stream: Streaming<proto::GenerateResponse>,
        request_id: String,
        client: SglangSchedulerClient,
    ) -> Self {
        debug!("Created AbortOnDropStream for request {}", request_id);
        Self {
            inner: stream,
            request_id,
            client,
            aborted: Arc::new(AtomicBool::new(false)),
        }
    }

    /// Manually mark the request as completed to prevent abort on drop.
    /// Call this when the request completes successfully to avoid unnecessary abort RPC.
    pub fn mark_completed(&self) {
        // Use Release ordering to ensure that this write is visible to other threads
        // that use Acquire on the same atomic variable
        self.aborted.store(true, Ordering::Release);
        debug!("Request {} marked as completed", self.request_id);
    }
}

impl Drop for AbortOnDropStream {
    fn drop(&mut self) {
        // Atomically check and set the aborted flag using compare_exchange.
        // If compare_exchange fails, it means the flag was already true (from mark_completed),
        // so we don't need to send abort. AcqRel is used for success to synchronize with
        // mark_completed's Release, and Acquire for failure to see writes from mark_completed.
        if self
            .aborted
            .compare_exchange(false, true, Ordering::AcqRel, Ordering::Acquire)
            .is_err()
        {
            return;
        }

        let client = self.client.clone();
        let request_id = self.request_id.clone();

        // Spawn a background task to send abort (since Drop is sync but abort_request is async)
        tokio::spawn(async move {
            debug!(
                "Stream dropped without completion for request {}, sending abort",
                request_id
            );
            // Clone request_id for the error message since abort_request takes ownership
            let request_id_for_log = request_id.clone();
            if let Err(e) = client
                .abort_request(request_id, "Stream dropped".to_string())
                .await
            {
                warn!(
                    "Failed to send abort on drop for request {}: {}",
                    request_id_for_log, e
                );
            }
        });
    }
}

// Implement Stream trait to make AbortOnDropStream work like the original Streaming
impl futures::Stream for AbortOnDropStream {
    type Item = Result<proto::GenerateResponse, tonic::Status>;

    fn poll_next(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Option<Self::Item>> {
        // Delegate to the inner stream
        Pin::new(&mut self.inner).poll_next(cx)
    }
}

109
/// gRPC client for SGLang scheduler
110
#[derive(Clone)]
111
112
113
114
115
116
pub struct SglangSchedulerClient {
    client: proto::sglang_scheduler_client::SglangSchedulerClient<Channel>,
}

impl SglangSchedulerClient {
    /// Create a new client and connect to the scheduler
117
    pub async fn connect(endpoint: &str) -> Result<Self, Box<dyn std::error::Error + Send + Sync>> {
118
119
        debug!("Connecting to SGLang scheduler at {}", endpoint);

120
121
122
        // Convert grpc:// to http:// for tonic
        let http_endpoint = if let Some(addr) = endpoint.strip_prefix("grpc://") {
            format!("http://{}", addr)
Chang Su's avatar
Chang Su committed
123
124
125
126
127
        } else {
            endpoint.to_string()
        };

        let channel = Channel::from_shared(http_endpoint)?
128
            .timeout(Duration::from_secs(600)) // 10 minute timeout for connection
129
130
131
132
133
134
135
136
            .http2_keep_alive_interval(Duration::from_secs(30))
            .keep_alive_timeout(Duration::from_secs(10))
            .keep_alive_while_idle(true)
            .tcp_keepalive(Some(Duration::from_secs(60)))
            .tcp_nodelay(true)
            .http2_adaptive_window(true)
            .initial_stream_window_size(Some(16 * 1024 * 1024)) // 16MB
            .initial_connection_window_size(Some(32 * 1024 * 1024)) // 32MB
137
138
139
140
141
142
143
144
            .connect()
            .await?;

        let client = proto::sglang_scheduler_client::SglangSchedulerClient::new(channel);

        Ok(Self { client })
    }

145
146
147
148
149
150
    /// Submit a generation request (returns auto-aborting streaming response)
    ///
    /// The returned stream automatically sends an abort request when dropped,
    /// ensuring proper cleanup even if the HTTP client disconnects or an error occurs.
    /// Call `mark_completed()` on the stream after successful completion to prevent
    /// unnecessary abort RPCs.
151
    pub async fn generate(
152
        &self,
153
        req: proto::GenerateRequest,
154
155
    ) -> Result<AbortOnDropStream, Box<dyn std::error::Error + Send + Sync>> {
        let request_id = req.request_id.clone();
156
        let mut client = self.client.clone();
157
        let request = Request::new(req);
158
        let response = client.generate(request).await?;
159
160
161
162
163
164

        Ok(AbortOnDropStream::new(
            response.into_inner(),
            request_id,
            self.clone(),
        ))
165
166
167
168
    }

    /// Perform health check
    pub async fn health_check(
169
        &self,
170
    ) -> Result<proto::HealthCheckResponse, Box<dyn std::error::Error + Send + Sync>> {
Chang Su's avatar
Chang Su committed
171
        debug!("Sending health check request");
172
173
        // HealthCheckRequest is now empty - server generates its own health check internally
        let request = Request::new(proto::HealthCheckRequest {});
174

175
176
        let mut client = self.client.clone();
        let response = client.health_check(request).await?;
Chang Su's avatar
Chang Su committed
177
        debug!("Health check response received");
178
179
180
181
182
        Ok(response.into_inner())
    }

    /// Abort a request
    pub async fn abort_request(
183
        &self,
184
185
        request_id: String,
        reason: String,
186
    ) -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
187
188
189
190
191
192
193
194
        debug!(
            "Sending abort request for {} (reason: {})",
            request_id, reason
        );
        let request = Request::new(proto::AbortRequest {
            request_id: request_id.clone(),
            reason,
        });
195

196
        let mut client = self.client.clone();
197
198
199
200
201
202
203
        let response = client.abort(request).await?;
        debug!(
            "Abort response for {}: success={}, message={}",
            request_id,
            response.get_ref().success,
            response.get_ref().message
        );
204
205
        Ok(())
    }
206

207
208
    /// Get model information
    pub async fn get_model_info(
209
        &self,
210
211
212
213
    ) -> Result<proto::GetModelInfoResponse, Box<dyn std::error::Error + Send + Sync>> {
        debug!("Requesting model info");
        let request = Request::new(proto::GetModelInfoRequest {});

214
215
        let mut client = self.client.clone();
        let response = client.get_model_info(request).await?;
216
217
218
219
220
221
        debug!("Model info response received");
        Ok(response.into_inner())
    }

    /// Get server information
    pub async fn get_server_info(
222
        &self,
223
224
225
226
    ) -> Result<proto::GetServerInfoResponse, Box<dyn std::error::Error + Send + Sync>> {
        debug!("Requesting server info");
        let request = Request::new(proto::GetServerInfoRequest {});

227
228
        let mut client = self.client.clone();
        let response = client.get_server_info(request).await?;
229
230
231
232
        debug!("Server info response received");
        Ok(response.into_inner())
    }

233
234
235
236
237
238
    /// Build a single SGLang GenerateRequest from OpenAI ChatCompletionRequest
    pub fn build_generate_request(
        &self,
        request_id: String,
        body: &ChatCompletionRequest,
        processed_text: String,
239
        token_ids: Vec<u32>,
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
        multimodal_inputs: Option<proto::MultimodalInputs>,
        tool_call_constraint: Option<(String, String)>, // (constraint_type, constraint_value)
    ) -> Result<proto::GenerateRequest, String> {
        // Build sampling params
        let sampling_params = self.build_grpc_sampling_params(body, tool_call_constraint)?;

        let grpc_request = proto::GenerateRequest {
            request_id,
            tokenized: Some(proto::TokenizedInput {
                original_text: processed_text,
                input_ids: token_ids,
            }),
            mm_inputs: multimodal_inputs,
            sampling_params: Some(sampling_params),
            return_logprob: body.logprobs,
            logprob_start_len: -1,
            top_logprobs_num: body.top_logprobs.unwrap_or(0) as i32,
            return_hidden_states: body.return_hidden_states,
258
            stream: body.stream,
259
260
261
262
263
264
            ..Default::default()
        };

        Ok(grpc_request)
    }

265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
    /// Build a basic GenerateRequest from the SGLang spec GenerateRequest
    pub fn build_plain_generate_request(
        &self,
        request_id: String,
        body: &GenerateRequest,
        original_text: Option<String>,
        token_ids: Vec<u32>,
    ) -> Result<proto::GenerateRequest, String> {
        let sampling_params =
            Self::build_sampling_params_from_plain(body.sampling_params.as_ref())?;

        let grpc_request = proto::GenerateRequest {
            request_id,
            tokenized: Some(proto::TokenizedInput {
                original_text: original_text.unwrap_or_default(),
                input_ids: token_ids,
            }),
            sampling_params: Some(sampling_params),
283
284
285
286
            return_logprob: body.return_logprob.unwrap_or(false),
            logprob_start_len: body.logprob_start_len.unwrap_or(-1),
            top_logprobs_num: body.top_logprobs_num.unwrap_or(0),
            token_ids_logprob: body.token_ids_logprob.clone().unwrap_or_default(),
287
288
            return_hidden_states: body.return_hidden_states,
            stream: body.stream,
289
            log_metrics: body.log_metrics,
290
291
292
293
294
295
            ..Default::default()
        };

        Ok(grpc_request)
    }

296
297
298
299
300
301
302
303
    /// Build gRPC SamplingParams from OpenAI request
    fn build_grpc_sampling_params(
        &self,
        request: &ChatCompletionRequest,
        tool_call_constraint: Option<(String, String)>,
    ) -> Result<proto::SamplingParams, String> {
        let stop_sequences = self.extract_stop_strings(request);

304
        let max_new_tokens = request.max_completion_tokens.map(|v| v as i32);
305
306
307
308

        // Handle skip_special_tokens: set to false if tools are present and tool_choice is not "none"
        let skip_special_tokens = if request.tools.is_some() {
            match &request.tool_choice {
309
                Some(ToolChoice::Value(ToolChoiceValue::None)) => request.skip_special_tokens,
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
                Some(_) => false, // tool_choice is not "none"
                None => false, // TODO: this assumes tool_choice defaults to "auto" when tools present
            }
        } else {
            request.skip_special_tokens
        };

        Ok(proto::SamplingParams {
            temperature: request.temperature.unwrap_or(1.0),
            top_p: request.top_p.unwrap_or(1.0),
            top_k: request.top_k.unwrap_or(-1),
            min_p: request.min_p.unwrap_or(0.0),
            frequency_penalty: request.frequency_penalty.unwrap_or(0.0),
            presence_penalty: request.presence_penalty.unwrap_or(0.0),
            repetition_penalty: request.repetition_penalty.unwrap_or(1.0),
            max_new_tokens,
            stop: stop_sequences,
            stop_token_ids: request.stop_token_ids.clone().unwrap_or_default(),
            skip_special_tokens,
329
            spaces_between_special_tokens: true, // Default from Python SamplingParams
330
331
            ignore_eos: request.ignore_eos,
            no_stop_trim: request.no_stop_trim,
332
333
334
335
336
337
338
339
340
            n: request.n.unwrap_or(1) as i32,
            constraint: self.build_constraint(request, tool_call_constraint)?,
            ..Default::default()
        })
    }

    /// Extract stop strings from request
    fn extract_stop_strings(&self, request: &ChatCompletionRequest) -> Vec<String> {
        match &request.stop {
341
342
            Some(StringOrArray::String(s)) => vec![s.clone()],
            Some(StringOrArray::Array(arr)) => arr.clone(),
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
            None => vec![],
        }
    }

    /// Build constraint for structured generation
    fn build_constraint(
        &self,
        request: &ChatCompletionRequest,
        tool_call_constraint: Option<(String, String)>,
    ) -> Result<Option<proto::sampling_params::Constraint>, String> {
        let mut constraints = Vec::new();

        if let Some(ResponseFormat::JsonSchema { json_schema }) = &request.response_format {
            let schema_str = serde_json::to_string(&json_schema.schema)
                .map_err(|e| format!("Failed to serialize JSON schema: {}", e))?;
            constraints.push(proto::sampling_params::Constraint::JsonSchema(schema_str));
        }

        if let Some(ebnf) = &request.ebnf {
            constraints.push(proto::sampling_params::Constraint::EbnfGrammar(
                ebnf.clone(),
            ));
        }

        if let Some(regex) = &request.regex {
            constraints.push(proto::sampling_params::Constraint::Regex(regex.clone()));
        }

        // Handle tool call constraint
        if let Some((constraint_type, constraint_value)) = tool_call_constraint {
            if !constraints.is_empty() {
                return Err("Constrained decoding is not compatible with tool calls.".to_string());
            }
            let tool_constraint = match constraint_type.as_str() {
                "structural_tag" => {
                    proto::sampling_params::Constraint::StructuralTag(constraint_value)
                }
                "json_schema" => proto::sampling_params::Constraint::JsonSchema(constraint_value),
                "ebnf" => proto::sampling_params::Constraint::EbnfGrammar(constraint_value),
                "regex" => proto::sampling_params::Constraint::Regex(constraint_value),
                _ => return Err(format!("Unknown constraint type: {}", constraint_type)),
            };
            constraints.push(tool_constraint);
        }

        match constraints.len() {
            0 => Ok(None),
            1 => Ok(constraints.pop()),
            _ => Err("Multiple constraints are not allowed.".to_string()),
        }
    }
394
395
396
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
423
424
425
426
427
428

    fn build_single_constraint_from_plain(
        params: &GenerateSamplingParams,
    ) -> Result<Option<proto::sampling_params::Constraint>, String> {
        let mut constraints = Vec::new();
        if let Some(json_schema) = &params.json_schema {
            constraints.push(proto::sampling_params::Constraint::JsonSchema(
                json_schema.clone(),
            ));
        }
        if let Some(regex) = &params.regex {
            constraints.push(proto::sampling_params::Constraint::Regex(regex.clone()));
        }
        if let Some(ebnf) = &params.ebnf {
            constraints.push(proto::sampling_params::Constraint::EbnfGrammar(
                ebnf.clone(),
            ));
        }

        match constraints.len() {
            0 => Ok(None),
            1 => Ok(constraints.pop()),
            _ => Err("Multiple structured constraints are not allowed".to_string()),
        }
    }

    fn build_sampling_params_from_plain(
        params: Option<&GenerateSamplingParams>,
    ) -> Result<proto::SamplingParams, String> {
        let mut sampling = proto::SamplingParams {
            temperature: 1.0,
            top_p: 1.0,
            top_k: -1,
            repetition_penalty: 1.0,
            n: 1,
429
430
            skip_special_tokens: true,
            spaces_between_special_tokens: true,
431
432
433
434
435
436
437
438
439
440
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
            ..Default::default()
        };

        let Some(p) = params else {
            return Ok(sampling);
        };

        // Simple field mappings using a macro
        macro_rules! map_field {
            ($field:ident) => {
                if let Some(val) = p.$field {
                    sampling.$field = val;
                }
            };
        }

        map_field!(temperature);
        map_field!(top_p);
        map_field!(top_k);
        map_field!(frequency_penalty);
        map_field!(presence_penalty);
        map_field!(repetition_penalty);
        map_field!(min_p);
        map_field!(ignore_eos);
        map_field!(skip_special_tokens);
        map_field!(no_stop_trim);

        // Handle stop sequences
        if let Some(stop) = &p.stop {
            match stop {
                StringOrArray::String(s) => sampling.stop.push(s.clone()),
                StringOrArray::Array(arr) => sampling.stop.extend(arr.clone()),
            }
        }

        // Handle stop token IDs
        if let Some(stop_token_ids) = &p.stop_token_ids {
            sampling.stop_token_ids = stop_token_ids.clone();
        }

        // Handle max_new_tokens with conversion
        if let Some(max_new_tokens) = p.max_new_tokens {
            sampling.max_new_tokens =
                Some(i32::try_from(max_new_tokens).map_err(|_| {
                    "max_new_tokens must fit into a 32-bit signed integer".to_string()
                })?);
        }

479
480
481
482
        // Handle min_new_tokens with conversion
        if let Some(min_new_tokens) = p.min_new_tokens {
            sampling.min_new_tokens = i32::try_from(min_new_tokens)
                .map_err(|_| "min_new_tokens must fit into a 32-bit signed integer".to_string())?;
483
484
        }

485
486
487
488
489
490
        // Handle n with conversion
        if let Some(n) = p.n {
            sampling.n = i32::try_from(n)
                .map_err(|_| "n must fit into a 32-bit signed integer".to_string())?;
        }

491
492
493
494
495
        // Handle constraints (exactly one allowed)
        sampling.constraint = Self::build_single_constraint_from_plain(p)?;

        Ok(sampling)
    }
496
497
498
499
500
501
502
503
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn test_proto_types_compilation() {
504
505
        let _health_req = proto::HealthCheckRequest {};
        // HealthCheckRequest is now empty - no fields to test
506
507
508
509
510
511
    }

    #[test]
    fn test_generate_request_construction() {
        let sampling_params = proto::SamplingParams {
            temperature: 0.7,
512
            max_new_tokens: Some(128),
513
514
515
516
517
518
519
520
            top_p: 0.9,
            top_k: 50,
            stop: vec!["</s>".to_string()],
            ..Default::default()
        };

        let gen_req = proto::GenerateRequest {
            request_id: "test-req-123".to_string(),
521
522
523
524
            tokenized: Some(proto::TokenizedInput {
                original_text: "Hello world".to_string(),
                input_ids: vec![9906, 1917], // Mock token IDs for "Hello world"
            }),
525
526
527
528
529
530
531
532
            sampling_params: Some(sampling_params),
            return_logprob: true,
            logprob_start_len: 0,
            top_logprobs_num: 5,
            ..Default::default()
        };

        assert_eq!(gen_req.request_id, "test-req-123");
533
534
        if let Some(ref tokenized) = &gen_req.tokenized {
            assert_eq!(tokenized.original_text, "Hello world");
535
536
537
538
539
540
        }
        assert!(gen_req.return_logprob);
        assert_eq!(gen_req.top_logprobs_num, 5);

        let params = gen_req.sampling_params.unwrap();
        assert_eq!(params.temperature, 0.7);
541
        assert_eq!(params.max_new_tokens, Some(128));
542
543
544
545
546
        assert_eq!(params.stop, vec!["</s>"]);
    }

    #[test]
    fn test_health_check_request() {
547
548
        let _health_req = proto::HealthCheckRequest {};
        // HealthCheckRequest is now empty - server generates its own test internally
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
    }

    #[test]
    fn test_abort_request_construction() {
        let abort_req = proto::AbortRequest {
            request_id: "req-456".to_string(),
            reason: "User canceled".to_string(),
        };
        assert_eq!(abort_req.request_id, "req-456");
        assert_eq!(abort_req.reason, "User canceled");
    }

    #[test]
    fn test_sampling_params_defaults() {
        let params = proto::SamplingParams::default();
564
        // Numeric fields have proto defaults (0)
565
566
567
        assert_eq!(params.temperature, 0.0);
        assert_eq!(params.top_p, 0.0);
        assert_eq!(params.top_k, 0);
568
569
570
571
572
573
574
575
576
577
578
579
580
581
        assert_eq!(params.repetition_penalty, 0.0);
        assert_eq!(params.n, 0);
        // Bool fields have proto defaults (false)
        assert!(!params.skip_special_tokens);
        assert!(!params.spaces_between_special_tokens);
        assert!(!params.ignore_eos);
        assert!(!params.no_stop_trim);
        // Optional int fields should be None
        assert_eq!(params.max_new_tokens, None);
        assert_eq!(params.stream_interval, None);
        // Other non-optional fields
        assert_eq!(params.min_p, 0.0);
        assert_eq!(params.frequency_penalty, 0.0);
        assert_eq!(params.presence_penalty, 0.0);
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
        assert!(params.stop.is_empty());
    }

    #[test]
    fn test_multimodal_inputs() {
        let mm_inputs = proto::MultimodalInputs {
            image_urls: vec!["http://example.com/image.jpg".to_string()],
            video_urls: vec![],
            audio_urls: vec![],
            image_data: vec![],
            video_data: vec![],
            audio_data: vec![],
            modalities: vec!["image".to_string()],
            ..Default::default()
        };

        assert_eq!(mm_inputs.image_urls.len(), 1);
        assert_eq!(mm_inputs.image_urls[0], "http://example.com/image.jpg");
        assert_eq!(mm_inputs.modalities[0], "image");
    }

603
    // TODO: SessionParams not in current proto - skip test
604
605
606
607
608

    #[test]
    fn test_embed_request() {
        let embed_req = proto::EmbedRequest {
            request_id: "embed-req-202".to_string(),
609
610
611
612
            tokenized: Some(proto::TokenizedInput {
                original_text: "This is a test sentence for embedding".to_string(),
                input_ids: vec![2028, 374, 264, 1296, 11914, 369, 28537], // Mock token IDs
            }),
613
614
615
616
617
618
            log_metrics: true,
            data_parallel_rank: 0,
            ..Default::default()
        };

        assert_eq!(embed_req.request_id, "embed-req-202");
619
620
621
622
623
        if let Some(ref tokenized) = &embed_req.tokenized {
            assert_eq!(
                tokenized.original_text,
                "This is a test sentence for embedding"
            );
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
        }
        assert!(embed_req.log_metrics);
        assert_eq!(embed_req.data_parallel_rank, 0);
    }

    #[tokio::test]
    async fn test_client_connect_invalid_endpoint() {
        let result = SglangSchedulerClient::connect("invalid://endpoint").await;
        assert!(result.is_err());
    }

    #[test]
    fn test_tokenized_input() {
        let tokenized = proto::TokenizedInput {
            original_text: "Hello world".to_string(),
            input_ids: vec![1, 15043, 1917, 2],
        };

        assert_eq!(tokenized.original_text, "Hello world");
        assert_eq!(tokenized.input_ids, vec![1, 15043, 1917, 2]);
    }

    #[test]
    fn test_generate_stream_chunk() {
        let chunk = proto::GenerateStreamChunk {
649
            token_ids: vec![1234, 5678],
650
651
652
653
654
655
            prompt_tokens: 5,
            completion_tokens: 2,
            cached_tokens: 3,
            ..Default::default()
        };

656
        assert_eq!(chunk.token_ids, vec![1234, 5678]);
657
658
659
660
661
        assert_eq!(chunk.prompt_tokens, 5);
        assert_eq!(chunk.completion_tokens, 2);
        assert_eq!(chunk.cached_tokens, 3);
    }

662
    // TODO: ModelInfo not in current proto - skip test
663
}