| 1 | // SPDX-License-Identifier: Apache-2.0 |
| 2 | // SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | //! Roundtrip tests for the real chat-template and output-processor pairing. |
| 5 | //! |
| 6 | //! The invariant under test is that a structured assistant message rendered as history can be |
| 7 | //! parsed from the generated assistant completion and then rendered back to the exact same |
| 8 | //! assistant completion. |
| 9 | //! |
| 10 | //! The tool-call fixture also snapshots the output grammar built by the initialized parser for each |
| 11 | //! tool-choice variant, with the rendered completion as its generation, as one file per model under |
| 12 | //! `tests/grammar_replay/`. `grammar_replay.rs` replays these cases through the real XGrammar. |
| 13 | |
| 14 | use std::collections::{BTreeMap, BTreeSet}; |
| 15 | use std::fmt::Debug; |
| 16 | use std::ops::Range; |
| 17 | use std::path::PathBuf; |
| 18 | use std::pin::Pin; |
| 19 | use std::sync::Arc; |
| 20 | |
| 21 | use anyhow::{Context as _, Result, bail, ensure}; |
| 22 | use futures::{Stream, StreamExt as _, stream}; |
| 23 | use serde::ser::Error as _; |
| 24 | use serde::{Serialize, Serializer}; |
| 25 | use serde_json_fmt::JsonFormat as JsonFmt; |
| 26 | use serial_test::file_serial; |
| 27 | use vllm_chat::{ |
| 28 | AssistantContentBlock, AssistantMessage, AssistantMessageExt as _, AssistantToolCall, |
| 29 | ChatEvent, ChatMessage, ChatRequest, ChatRole, ChatTool, ChatToolChoice, FinishReason, |
| 30 | GenerationPromptMode, LoadModelBackendsOptions, NewChatOutputProcessorOptions, ParserSelection, |
| 31 | RendererSelection, ToolStrictLevel, load_model_backends, |
| 32 | }; |
| 33 | use vllm_parser::output_grammar::BuiltOutputGrammar; |
| 34 | use vllm_text::{DecodedTextEvent, Finished, Prompt}; |
| 35 | use vllm_tokenizer::Tokenizer; |
| 36 | |
| 37 | const TEXT_COMPLETION_CHUNK_CHARS: usize = 7; |
| 38 | const TOKEN_COMPLETION_CHUNK_TOKENS: usize = 1; |
| 39 | |
| 40 | /// One model/parser configuration used to run the fixed roundtrip fixtures. |
| 41 | #[derive(Clone)] |
| 42 | struct RoundtripCase { |
| 43 | /// Hugging Face model id or local checkpoint resolved through the production backend loader. |
| 44 | model_id: &'static str, |
| 45 | /// Final assistant-history suffix rendered by the chat template but not |
| 46 | /// generated by the model body consumed by the output processor. |
| 47 | // TODO: we should adopt `ContinueFinalAssistant` mode to naturally handle this. |
| 48 | assistant_stop_suffix: &'static str, |
| 49 | /// Tool parser selection used by the output processor. |
| 50 | tool_call_parser: ParserSelection, |
| 51 | /// Reasoning parser selection used by the output processor. |
| 52 | reasoning_parser: ParserSelection, |
| 53 | /// How this model's chat template handles thinking mode. |
| 54 | thinking_behavior: ThinkingBehavior, |
| 55 | /// JSON formatting expected after this model's template has materialized |
| 56 | /// tool-call arguments. |
| 57 | json_fmt: JsonFmt, |
| 58 | /// Whether the template renders tool-call argument object keys in sorted order. |
| 59 | sort_json_keys: bool, |
| 60 | } |
| 61 | |
| 62 | #[derive(Clone, Copy)] |
| 63 | enum ThinkingBehavior { |
| 64 | /// The chat template accepts explicit thinking on/off kwargs, and uses |
| 65 | /// `default` when the request does not specify either kwarg. |
| 66 | Toggleable { default: bool }, |
| 67 | /// The chat template always behaves as `value` for this fixture. |
| 68 | Always { value: bool }, |
| 69 | /// The chat template selects thinking mode through `reasoning_effort`. |
| 70 | ReasoningEffort { |
| 71 | default: bool, |
| 72 | enabled: &'static str, |
| 73 | disabled: &'static str, |
| 74 | }, |
| 75 | } |
| 76 | |
| 77 | impl ThinkingBehavior { |
| 78 | fn default(self) -> bool { |
| 79 | match self { |
| 80 | Self::Toggleable { default } => default, |
| 81 | Self::Always { value } => value, |
| 82 | Self::ReasoningEffort { default, .. } => default, |
| 83 | } |
| 84 | } |
| 85 | |
| 86 | fn fixtures(self) -> Vec<Option<bool>> { |
| 87 | match self { |
| 88 | Self::Toggleable { .. } => vec![ |
| 89 | Some(true), // explicitly enable thinking |
| 90 | Some(false), // explicitly disable thinking |
| 91 | None, // use default template behavior |
| 92 | ], |
| 93 | Self::Always { value } => vec![ |
| 94 | Some(value), // explicitly request the supported thinking behavior |
| 95 | None, // use default template behavior |
| 96 | ], |
| 97 | Self::ReasoningEffort { .. } => vec![ |
| 98 | Some(true), // explicitly enable thinking |
| 99 | Some(false), // explicitly disable thinking |
| 100 | None, // use default template behavior |
| 101 | ], |
| 102 | } |
| 103 | } |
| 104 | |
| 105 | fn apply(self, request: &mut ChatRequest, thinking: Option<bool>) { |
| 106 | let Some(thinking) = thinking else { |
| 107 | return; |
| 108 | }; |
| 109 | |
| 110 | match self { |
| 111 | Self::Toggleable { .. } | Self::Always { .. } => { |
| 112 | for key in ["thinking", "enable_thinking"] { |
| 113 | request.chat_options.template_kwargs.insert(key.to_string(), thinking.into()); |
| 114 | } |
| 115 | } |
| 116 | Self::ReasoningEffort { |
| 117 | enabled, disabled, .. |
| 118 | } => { |
| 119 | request.chat_options.template_kwargs.insert( |
| 120 | "reasoning_effort".to_string(), |
| 121 | if thinking { enabled } else { disabled }.into(), |
| 122 | ); |
| 123 | } |
| 124 | } |
| 125 | } |
| 126 | } |
| 127 | |
| 128 | /// Tool-choice variant of the tool-call fixture. |
| 129 | #[derive(Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Serialize)] |
| 130 | #[serde(rename_all = "snake_case")] |
| 131 | enum ToolChoiceVariant { |
| 132 | /// `auto`, with content before the tool calls. |
| 133 | Auto, |
| 134 | /// `required`, with tool calls only: builder grammars admit no content |
| 135 | /// before the calls. |
| 136 | Required, |
| 137 | /// Named `get_weather`, with a single call to it and no content. |
| 138 | Named, |
| 139 | } |
| 140 | |
| 141 | impl ToolChoiceVariant { |
| 142 | const ALL: [Self; 3] = [Self::Auto, Self::Required, Self::Named]; |
| 143 | |
| 144 | fn tool_choice(self) -> ChatToolChoice { |
| 145 | match self { |
| 146 | Self::Auto => ChatToolChoice::Auto, |
| 147 | Self::Required => ChatToolChoice::Required, |
| 148 | Self::Named => ChatToolChoice::Function { |
| 149 | name: "get_weather".to_string(), |
| 150 | }, |
| 151 | } |
| 152 | } |
| 153 | } |
| 154 | |
| 155 | impl RoundtripCase { |
| 156 | /// Qwen3 XML tool-call format with `qwen3` reasoning tags. |
| 157 | fn qwen3() -> Self { |
| 158 | Self { |
| 159 | model_id: "Qwen/Qwen3-0.6B", |
| 160 | assistant_stop_suffix: "<|im_end|>\n", |
| 161 | tool_call_parser: ParserSelection::Auto, |
| 162 | reasoning_parser: ParserSelection::Auto, |
| 163 | thinking_behavior: ThinkingBehavior::Toggleable { default: true }, |
| 164 | json_fmt: spaced_json_fmt(), |
| 165 | sort_json_keys: false, |
| 166 | } |
| 167 | } |
| 168 | |
| 169 | /// Qwen3.5 coder-style JSON tool-call format with `qwen3` reasoning tags. |
| 170 | fn qwen35() -> Self { |
| 171 | Self { |
| 172 | model_id: "Qwen/Qwen3.5-4B", |
| 173 | assistant_stop_suffix: "<|im_end|>\n", |
| 174 | tool_call_parser: ParserSelection::Auto, |
| 175 | reasoning_parser: ParserSelection::Auto, |
| 176 | thinking_behavior: ThinkingBehavior::Toggleable { default: true }, |
| 177 | json_fmt: compact_json_fmt(), |
| 178 | sort_json_keys: false, |
| 179 | } |
| 180 | } |
| 181 | |
| 182 | /// MiMo V2.6 uses bare reasoning and Qwen Coder tool tags. |
| 183 | fn mimo_v26() -> Self { |
| 184 | Self { |
| 185 | model_id: "XiaomiMiMo/MiMo-V2.6-Flash-RL", |
| 186 | assistant_stop_suffix: "<|im_end|>", |
| 187 | tool_call_parser: ParserSelection::Auto, |
| 188 | reasoning_parser: ParserSelection::Auto, |
| 189 | thinking_behavior: ThinkingBehavior::Toggleable { default: true }, |
| 190 | json_fmt: compact_json_fmt(), |
| 191 | sort_json_keys: false, |
| 192 | } |
| 193 | } |
| 194 | |
| 195 | /// MiniMax M2.5 XML invoke format with `<think>` reasoning tags. |
| 196 | fn minimax_m25() -> Self { |
| 197 | Self { |
| 198 | model_id: "MiniMaxAI/MiniMax-M2.5", |
| 199 | assistant_stop_suffix: "[e~[\n", |
| 200 | tool_call_parser: ParserSelection::Auto, |
| 201 | reasoning_parser: ParserSelection::Auto, |
| 202 | thinking_behavior: ThinkingBehavior::Always { value: true }, |
| 203 | json_fmt: compact_json_fmt(), |
| 204 | sort_json_keys: false, |
| 205 | } |
| 206 | } |
| 207 | |
| 208 | /// MiniMax M3 invoke format with `<mm:think>` reasoning tags. |
| 209 | fn minimax_m3() -> Self { |
| 210 | Self { |
| 211 | model_id: "MiniMaxAI/MiniMax-M3", |
| 212 | assistant_stop_suffix: "[e~[\n", |
| 213 | tool_call_parser: ParserSelection::Auto, |
| 214 | reasoning_parser: ParserSelection::Auto, |
| 215 | thinking_behavior: ThinkingBehavior::Always { value: true }, |
| 216 | json_fmt: compact_json_fmt(), |
| 217 | sort_json_keys: false, |
| 218 | } |
| 219 | } |
| 220 | |
| 221 | /// DeepSeek V4 DSML tool-call format. |
| 222 | fn deepseek_v4() -> Self { |
| 223 | Self { |
| 224 | model_id: "deepseek-ai/DeepSeek-V4-Flash", |
| 225 | assistant_stop_suffix: "<｜end▁of▁sentence｜>", |
| 226 | tool_call_parser: ParserSelection::Auto, |
| 227 | reasoning_parser: ParserSelection::Auto, |
| 228 | thinking_behavior: ThinkingBehavior::Toggleable { default: false }, |
| 229 | json_fmt: compact_json_fmt(), |
| 230 | sort_json_keys: false, |
| 231 | } |
| 232 | } |
| 233 | |
| 234 | /// DeepSeek V4.1 Flash DSML tool-call format. |
| 235 | fn deepseek_v41() -> Self { |
| 236 | Self { |
| 237 | model_id: "deepseek-ai/DeepSeek-V4.1-Flash", |
| 238 | assistant_stop_suffix: "<｜end▁of▁sentence｜>", |
| 239 | tool_call_parser: ParserSelection::Auto, |
| 240 | reasoning_parser: ParserSelection::Auto, |
| 241 | thinking_behavior: ThinkingBehavior::Toggleable { default: true }, |
| 242 | json_fmt: compact_json_fmt(), |
| 243 | sort_json_keys: false, |
| 244 | } |
| 245 | } |
| 246 | |
| 247 | /// DeepSeek V3.2 DSML tool-call format. |
| 248 | fn deepseek_v32() -> Self { |
| 249 | Self { |
| 250 | model_id: "deepseek-ai/DeepSeek-V3.2-Exp", |
| 251 | assistant_stop_suffix: "<｜end▁of▁sentence｜>", |
| 252 | tool_call_parser: ParserSelection::Auto, |
| 253 | reasoning_parser: ParserSelection::Auto, |
| 254 | thinking_behavior: ThinkingBehavior::Toggleable { default: false }, |
| 255 | json_fmt: compact_json_fmt(), |
| 256 | sort_json_keys: false, |
| 257 | } |
| 258 | } |
| 259 | |
| 260 | /// GLM-4.5 XML-like argument format with `<think>` reasoning tags. |
| 261 | fn glm45() -> Self { |
| 262 | Self { |
| 263 | model_id: "zai-org/GLM-4.5", |
| 264 | assistant_stop_suffix: "", |
| 265 | tool_call_parser: ParserSelection::Auto, |
| 266 | reasoning_parser: ParserSelection::Auto, |
| 267 | thinking_behavior: ThinkingBehavior::Toggleable { default: true }, |
| 268 | json_fmt: compact_json_fmt(), |
| 269 | sort_json_keys: false, |
| 270 | } |
| 271 | } |
| 272 | |
| 273 | /// GLM-4.7 XML-like argument format with `<think>` reasoning tags. |
| 274 | fn glm47() -> Self { |
| 275 | Self { |
| 276 | model_id: "zai-org/GLM-4.7-Flash", |
| 277 | assistant_stop_suffix: "", |
| 278 | tool_call_parser: ParserSelection::Auto, |
| 279 | reasoning_parser: ParserSelection::Auto, |
| 280 | thinking_behavior: ThinkingBehavior::Toggleable { default: true }, |
| 281 | json_fmt: compact_json_fmt(), |
| 282 | sort_json_keys: false, |
| 283 | } |
| 284 | } |
| 285 | |
| 286 | /// GLM-5.2 XML-like argument format with `<think>` reasoning tags. |
| 287 | fn glm52() -> Self { |
| 288 | Self { |
| 289 | model_id: "zai-org/GLM-5.2-FP8", |
| 290 | assistant_stop_suffix: "", |
| 291 | tool_call_parser: ParserSelection::Auto, |
| 292 | reasoning_parser: ParserSelection::Auto, |
| 293 | thinking_behavior: ThinkingBehavior::Toggleable { default: true }, |
| 294 | json_fmt: compact_json_fmt(), |
| 295 | sort_json_keys: false, |
| 296 | } |
| 297 | } |
| 298 | |
| 299 | /// Gemma4 channel reasoning with custom function-call arguments. |
| 300 | fn gemma4() -> Self { |
| 301 | Self { |
| 302 | model_id: "google/gemma-4-E4B-it", |
| 303 | assistant_stop_suffix: "<|tool_response>", |
| 304 | tool_call_parser: ParserSelection::Auto, |
| 305 | reasoning_parser: ParserSelection::Auto, |
| 306 | thinking_behavior: ThinkingBehavior::Always { value: true }, |
| 307 | json_fmt: compact_json_fmt(), |
| 308 | sort_json_keys: true, |
| 309 | } |
| 310 | } |
| 311 | |
| 312 | /// Gemma4 through the checkpoint's `response_template` (`hf` parser). |
| 313 | fn gemma4_hf() -> Self { |
| 314 | Self { |
| 315 | tool_call_parser: ParserSelection::Explicit("hf".to_string()), |
| 316 | reasoning_parser: ParserSelection::Explicit("hf".to_string()), |
| 317 | ..Self::gemma4() |
| 318 | } |
| 319 | } |
| 320 | |
| 321 | /// Kimi K2.5 tool-call format with `<think>` reasoning tags. |
| 322 | #[allow(dead_code)] |
| 323 | fn kimi_k25() -> Self { |
| 324 | Self { |
| 325 | model_id: "moonshotai/Kimi-K2.5", |
| 326 | assistant_stop_suffix: "<|im_end|>", |
| 327 | tool_call_parser: ParserSelection::Auto, |
| 328 | reasoning_parser: ParserSelection::Auto, |
| 329 | thinking_behavior: ThinkingBehavior::Toggleable { default: true }, |
| 330 | json_fmt: spaced_json_fmt(), |
| 331 | sort_json_keys: false, |
| 332 | } |
| 333 | } |
| 334 | |
| 335 | /// Kimi K3 XTML tool/reasoning channels (native renderer + unified parser). |
| 336 | /// |
| 337 | /// Needs HF tokenizer files under `HF_HOME` (`tiktoken.model` + |
| 338 | /// `tokenizer_config.json`). Weights are not required for this text-level |
| 339 | /// roundtrip. |
| 340 | fn kimi_k3() -> Self { |
| 341 | Self { |
| 342 | model_id: "moonshotai/Kimi-K3", |
| 343 | assistant_stop_suffix: "<|end_of_msg|>", |
| 344 | tool_call_parser: ParserSelection::Auto, |
| 345 | reasoning_parser: ParserSelection::Auto, |
| 346 | thinking_behavior: ThinkingBehavior::Toggleable { default: true }, |
| 347 | json_fmt: compact_json_fmt(), |
| 348 | sort_json_keys: false, |
| 349 | } |
| 350 | } |
| 351 | |
| 352 | /// HY3 suffixed reasoning/tool markers discovered from tokenizer added vocab. |
| 353 | fn hy_v3() -> Self { |
| 354 | Self { |
| 355 | model_id: "tencent/Hy3", |
| 356 | assistant_stop_suffix: "<｜hy_eos:opensource｜>", |
| 357 | tool_call_parser: ParserSelection::Auto, |
| 358 | reasoning_parser: ParserSelection::Auto, |
| 359 | thinking_behavior: ThinkingBehavior::ReasoningEffort { |
| 360 | default: false, |
| 361 | enabled: "high", |
| 362 | disabled: "no_think", |
| 363 | }, |
| 364 | json_fmt: compact_json_fmt(), |
| 365 | sort_json_keys: false, |
| 366 | } |
| 367 | } |
| 368 | |
| 369 | /// SeedOSS with `<seed:think>` / `</seed:think>` reasoning tags. |
| 370 | fn seed_oss() -> Self { |
| 371 | Self { |
| 372 | model_id: "ByteDance-Seed/Seed-OSS-36B-Instruct", |
| 373 | assistant_stop_suffix: "<seed:eos>", |
| 374 | tool_call_parser: ParserSelection::Auto, |
| 375 | reasoning_parser: ParserSelection::Auto, |
| 376 | thinking_behavior: ThinkingBehavior::Always { value: true }, |
| 377 | json_fmt: compact_json_fmt(), |
| 378 | sort_json_keys: false, |
| 379 | } |
| 380 | } |
| 381 | |
| 382 | /// Step-3.5 with `<think>` / `</think>` reasoning tags and newline trimming. |
| 383 | fn step3p5() -> Self { |
| 384 | Self { |
| 385 | model_id: "stepfun-ai/Step-3.5-Flash", |
| 386 | assistant_stop_suffix: "<|im_end|>\n", |
| 387 | tool_call_parser: ParserSelection::Auto, |
| 388 | reasoning_parser: ParserSelection::Auto, |
| 389 | thinking_behavior: ThinkingBehavior::Always { value: true }, |
| 390 | json_fmt: compact_json_fmt(), |
| 391 | sort_json_keys: false, |
| 392 | } |
| 393 | } |
| 394 | |
| 395 | /// Nemotron V3 with `<think>` / `</think>` reasoning tags. |
| 396 | fn nemotron_v3() -> Self { |
| 397 | Self { |
| 398 | model_id: "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16", |
| 399 | assistant_stop_suffix: "<|im_end|>\n", |
| 400 | tool_call_parser: ParserSelection::Auto, |
| 401 | reasoning_parser: ParserSelection::Auto, |
| 402 | thinking_behavior: ThinkingBehavior::Always { value: true }, |
| 403 | json_fmt: compact_json_fmt(), |
| 404 | sort_json_keys: false, |
| 405 | } |
| 406 | } |
| 407 | |
| 408 | /// GPT-OSS Harmony token-id renderer and native Harmony output processor. |
| 409 | fn gpt_oss() -> Self { |
| 410 | Self { |
| 411 | model_id: "openai/gpt-oss-20b", |
| 412 | assistant_stop_suffix: "", |
| 413 | tool_call_parser: ParserSelection::Auto, |
| 414 | reasoning_parser: ParserSelection::Auto, |
| 415 | thinking_behavior: ThinkingBehavior::Always { value: true }, |
| 416 | json_fmt: compact_json_fmt(), |
| 417 | sort_json_keys: false, |
| 418 | } |
| 419 | } |
| 420 | |
| 421 | /// Inkling typed content blocks with native token-id rendering. |
| 422 | fn inkling() -> Self { |
| 423 | Self { |
| 424 | model_id: "thinkingmachines/Inkling", |
| 425 | assistant_stop_suffix: "<|content_model_end_sampling|>", |
| 426 | tool_call_parser: ParserSelection::Auto, |
| 427 | reasoning_parser: ParserSelection::Auto, |
| 428 | thinking_behavior: ThinkingBehavior::Always { value: true }, |
| 429 | json_fmt: compact_json_fmt(), |
| 430 | sort_json_keys: true, |
| 431 | } |
| 432 | } |
| 433 | } |
| 434 | |
| 435 | macro_rules! roundtrip_tests { |
| 436 | ($($case:ident => $(#[$case_attr:meta])* [$($fixture:ident),* $(,)?]),+ $(,)?) => { |
| 437 | paste::paste! { |
| 438 | $( |
| 439 | #[tokio::test] |
| 440 | #[file_serial([<hf_ $case>])] |
| 441 | $(#[$case_attr])* |
| 442 | async fn [<roundtrip_ $case>]() -> Result<()> { |
| 443 | let case = RoundtripCase::$case(); |
| 444 | let backends = load_roundtrip_backends(&case).await?; |
| 445 | $( |
| 446 | [<run_roundtrip_ $fixture>](&case, &backends).await?; |
| 447 | )* |
| 448 | Ok(()) |
| 449 | } |
| 450 | )+ |
| 451 | } |
| 452 | }; |
| 453 | } |
| 454 | |
| 455 | roundtrip_tests! { |
| 456 | qwen3 => [reasoning_and_content, tool_call_mix], |
| 457 | qwen35 => [reasoning_and_content, tool_call_mix], |
| 458 | mimo_v26 => [reasoning_and_content, tool_call_mix], |
| 459 | minimax_m25 => [reasoning_and_content, tool_call_mix], |
| 460 | minimax_m3 => [reasoning_and_content, tool_call_mix], |
| 461 | deepseek_v4 => [reasoning_and_content, tool_call_mix], |
| 462 | deepseek_v41 => [reasoning_and_content, tool_call_mix], |
| 463 | deepseek_v32 => [tool_call_mix], |
| 464 | glm45 => [reasoning_and_content, tool_call_mix], |
| 465 | glm47 => [reasoning_and_content, tool_call_mix], |
| 466 | glm52 => [reasoning_and_content, tool_call_mix], |
| 467 | seed_oss => [reasoning_and_content, tool_call_mix], |
| 468 | step3p5 => [reasoning_and_content], |
| 469 | nemotron_v3 => [reasoning_and_content], |
| 470 | gemma4 => [tool_call_mix], // Gemma4 strips reasoning in history if there's no tool call |
| 471 | gemma4_hf => [tool_call_mix], |
| 472 | kimi_k25 => [tool_call_mix], // Kimi K2.5 strips reasoning in history |
| 473 | // K3 drops plain-assistant reasoning in history; tool-call turns keep it. |
| 474 | kimi_k3 => [tool_call_mix], |
| 475 | // HY3's final plain-assistant history omits EOS; tool-call history keeps it |
| 476 | // and exercises both stages of the tokenizer-derived unified parser. |
| 477 | hy_v3 => [tool_call_mix], |
| 478 | gpt_oss => [tool_call_mix], // Harmony strips reasoning in history if there's no tool call |
| 479 | inkling => [reasoning_and_content, tool_call_mix], |
| 480 | } |
| 481 | |
| 482 | /// Run the fixed reasoning+content fixture for one model/parser case. |
| 483 | async fn run_roundtrip_reasoning_and_content( |
| 484 | case: &RoundtripCase, |
| 485 | backends: &vllm_chat::LoadedModelBackends, |
| 486 | ) -> Result<()> { |
| 487 | for thinking in case.thinking_behavior.fixtures() { |
| 488 | run_roundtrip_reasoning_and_content_inner(case, backends, thinking).await?; |
| 489 | } |
| 490 | Ok(()) |
| 491 | } |
| 492 | |
| 493 | async fn run_roundtrip_reasoning_and_content_inner( |
| 494 | case: &RoundtripCase, |
| 495 | backends: &vllm_chat::LoadedModelBackends, |
| 496 | thinking: Option<bool>, |
| 497 | ) -> Result<()> { |
| 498 | let request = roundtrip_request( |
| 499 | "roundtrip-reasoning-content", |
| 500 | vec![ChatMessage::text(ChatRole::User, "What is 2 + 2?")], |
| 501 | Vec::new(), |
| 502 | None, |
| 503 | thinking, |
| 504 | case.thinking_behavior, |
| 505 | ); |
| 506 | let expected_reasoning = "Need compute 2 + 2 directly."; |
| 507 | let expected_text = "The answer is 4."; |
| 508 | let effective_thinking = thinking.unwrap_or(case.thinking_behavior.default()); |
| 509 | |
| 510 | let assistant = { |
| 511 | let mut content = Vec::new(); |
| 512 | if effective_thinking { |
| 513 | content.push(AssistantContentBlock::Reasoning { |
| 514 | text: expected_reasoning.to_string(), |
| 515 | }); |
| 516 | } |
| 517 | content.push(AssistantContentBlock::Text { |
| 518 | text: expected_text.to_string(), |
| 519 | }); |
| 520 | AssistantMessage { content } |
| 521 | }; |
| 522 | let result = run_roundtrip(case, backends, &request, assistant).await?; |
| 523 | |
| 524 | assert_eq!( |
| 525 | result.parsed_message.reasoning().as_deref(), |
| 526 | effective_thinking.then_some(expected_reasoning), |
| 527 | "parsed message: {:#?}", |
| 528 | result.parsed_message |
| 529 | ); |
| 530 | assert_eq!(result.parsed_message.text(), expected_text); |
| 531 | assert_eq!(result.parsed_message.tool_calls().count(), 0); |
| 532 | |
| 533 | assert_eq!( |
| 534 | result.rerendered_closed_completion, |
| 535 | result.closed_completion |
| 536 | ); |
| 537 | |
| 538 | Ok(()) |
| 539 | } |
| 540 | |
| 541 | /// Run the fixed reasoning+multiple-tools fixture for one model/parser case, |
| 542 | /// once per tool-choice variant. |
| 543 | async fn run_roundtrip_tool_call_mix( |
| 544 | case: &RoundtripCase, |
| 545 | backends: &vllm_chat::LoadedModelBackends, |
| 546 | ) -> Result<()> { |
| 547 | let mut results = Vec::new(); |
| 548 | for variant in ToolChoiceVariant::ALL { |
| 549 | let result = run_roundtrip_tool_call_mix_inner(case, backends, variant).await?; |
| 550 | results.push((variant, result)); |
| 551 | } |
| 552 | check_grammar_replay_file(case, backends, results) |
| 553 | } |
| 554 | |
| 555 | async fn run_roundtrip_tool_call_mix_inner( |
| 556 | case: &RoundtripCase, |
| 557 | backends: &vllm_chat::LoadedModelBackends, |
| 558 | variant: ToolChoiceVariant, |
| 559 | ) -> Result<RoundtripResult> { |
| 560 | let request = roundtrip_request( |
| 561 | "roundtrip-reasoning-tools", |
| 562 | vec![ChatMessage::text( |
| 563 | ChatRole::User, |
| 564 | "Check Shanghai weather and add 1.0 plus 2.", |
| 565 | )], |
| 566 | test_tools(), |
| 567 | Some(variant.tool_choice()), |
| 568 | Some(true), // always enable thinking in this fixture |
| 569 | case.thinking_behavior, |
| 570 | ); |
| 571 | let expected_reasoning = "Need call the weather and add tools."; |
| 572 | let (expected_text, tool_call_count) = match variant { |
| 573 | ToolChoiceVariant::Auto => (Some("I will call the tools."), 2), |
| 574 | ToolChoiceVariant::Required => (None, 2), |
| 575 | ToolChoiceVariant::Named => (None, 1), |
| 576 | }; |
| 577 | |
| 578 | let mut content = vec![AssistantContentBlock::Reasoning { |
| 579 | text: expected_reasoning.to_string(), |
| 580 | }]; |
| 581 | content.extend(expected_text.map(|text| AssistantContentBlock::Text { |
| 582 | text: text.to_string(), |
| 583 | })); |
| 584 | let tool_calls = [ |
| 585 | AssistantContentBlock::ToolCall(AssistantToolCall { |
| 586 | id: "functions.get_weather:0".to_string(), |
| 587 | name: "get_weather".to_string(), |
| 588 | arguments: r#"{"location":"Shanghai"}"#.to_string(), |
| 589 | }), |
| 590 | AssistantContentBlock::ToolCall(AssistantToolCall { |
| 591 | id: "functions.add:1".to_string(), |
| 592 | name: "add".to_string(), |
| 593 | // Intentionally use a non-lexical order of keys to verify text-level |
| 594 | // fidelity of the roundtrip where JSON formatting remains stable. The |
| 595 | // `items` key also exercises templates that call `arguments.items()`. |
| 596 | arguments: r#"{"y":1.0,"x":2,"items":["left","right"]}"#.to_string(), |
| 597 | }), |
| 598 | ]; |
| 599 | content.extend(tool_calls.into_iter().take(tool_call_count)); |
| 600 | let result = run_roundtrip(case, backends, &request, AssistantMessage { content }).await?; |
| 601 | |
| 602 | assert_eq!( |
| 603 | result.parsed_message.reasoning().as_deref(), |
| 604 | Some(expected_reasoning), |
| 605 | "parsed message: {:#?}", |
| 606 | result.parsed_message |
| 607 | ); |
| 608 | assert_eq!( |
| 609 | result.parsed_message.text(), |
| 610 | expected_text.unwrap_or_default() |
| 611 | ); |
| 612 | |
| 613 | let tool_calls = result.parsed_message.tool_calls().collect::<Vec<_>>(); |
| 614 | assert_eq!( |
| 615 | tool_calls.len(), |
| 616 | tool_call_count, |
| 617 | "parsed message: {:#?}", |
| 618 | result.parsed_message |
| 619 | ); |
| 620 | assert_eq!(tool_calls[0].name, "get_weather"); |
| 621 | assert_eq!( |
| 622 | tool_calls[0].arguments, |
| 623 | expected_arguments(case, r#"{"location": "Shanghai"}"#)?, |
| 624 | ); |
| 625 | if let Some(add) = tool_calls.get(1) { |
| 626 | assert_eq!(add.name, "add"); |
| 627 | assert_eq!( |
| 628 | add.arguments, |
| 629 | expected_arguments(case, r#"{"y": 1.0, "x": 2, "items": ["left", "right"]}"#)?, |
| 630 | ); |
| 631 | } |
| 632 | |
| 633 | assert_eq!( |
| 634 | result.rerendered_closed_completion, |
| 635 | result.closed_completion |
| 636 | ); |
| 637 | |
| 638 | Ok(result) |
| 639 | } |
| 640 | |
| 641 | /// Compact JSON argument formatting used by JSON-native parsers/renderers. |
| 642 | fn compact_json_fmt() -> JsonFmt { |
| 643 | JsonFmt::new() |
| 644 | } |
| 645 | |
| 646 | /// Python `json.dumps`-style compact formatting with a space after commas and |
| 647 | /// colons. |
| 648 | fn spaced_json_fmt() -> JsonFmt { |
| 649 | JsonFmt::new() |
| 650 | .comma(", ") |
| 651 | .expect("literal comma separator is valid JSON") |
| 652 | .colon(": ") |
| 653 | .expect("literal colon separator is valid JSON") |
| 654 | } |
| 655 | |
| 656 | /// Parse and format expected tool-call arguments from raw JSON text. |
| 657 | /// Pass in a raw JSON string instead of a structured value to ensure the exact precision and |
| 658 | /// formatting of numbers are preserved. |
| 659 | fn expected_arguments(case: &RoundtripCase, raw_json: &str) -> Result<String> { |
| 660 | let mut value: serde_json::Value = |
| 661 | serde_json::from_str(raw_json).context("invalid expected tool-call arguments")?; |
| 662 | if case.sort_json_keys { |
| 663 | sort_json_value(&mut value); |
| 664 | } |
| 665 | |
| 666 | case.json_fmt |
| 667 | .format_to_string(&value) |
| 668 | .context("failed to format expected tool-call arguments") |
| 669 | } |
| 670 | |
| 671 | /// Sort JSON object keys recursively to match templates that render mappings with `dictsort`. |
| 672 | fn sort_json_value(value: &mut serde_json::Value) { |
| 673 | match value { |
| 674 | serde_json::Value::Object(map) => { |
| 675 | for value in map.values_mut() { |
| 676 | sort_json_value(value); |
| 677 | } |
| 678 | |
| 679 | let mut entries = std::mem::take(map).into_iter().collect::<Vec<_>>(); |
| 680 | entries.sort_by(|(left, _), (right, _)| left.cmp(right)); |
| 681 | map.extend(entries); |
| 682 | } |
| 683 | serde_json::Value::Array(values) => { |
| 684 | for value in values { |
| 685 | sort_json_value(value); |
| 686 | } |
| 687 | } |
| 688 | _ => {} |
| 689 | } |
| 690 | } |
| 691 | |
| 692 | /// Load the real model chat/text backend for one roundtrip case. |
| 693 | async fn load_roundtrip_backends(case: &RoundtripCase) -> Result<vllm_chat::LoadedModelBackends> { |
| 694 | load_model_backends( |
| 695 | case.model_id, |
| 696 | LoadModelBackendsOptions { |
| 697 | renderer: RendererSelection::Auto, |
| 698 | ..Default::default() |
| 699 | }, |
| 700 | ) |
| 701 | .await |
| 702 | .with_context(|| format!("failed to load HF model files for {}", case.model_id)) |
| 703 | } |
| 704 | |
| 705 | /// Roundtrip artifacts needed for semantic and exact-text assertions. |
| 706 | struct RoundtripResult { |
| 707 | /// Final assistant message reconstructed by the output processor. |
| 708 | parsed_message: AssistantMessage, |
| 709 | /// Assistant-completion suffix cut from rendering the expected assistant as |
| 710 | /// history. |
| 711 | closed_completion: Prompt, |
| 712 | /// Assistant-completion suffix cut after rendering the parsed assistant |
| 713 | /// back as history. |
| 714 | rerendered_closed_completion: Prompt, |
| 715 | /// Output grammar built by the same initialized parser that parsed the |
| 716 | /// completion. |
| 717 | output_grammar: Option<BuiltOutputGrammar>, |
| 718 | } |
| 719 | |
| 720 | /// Render, parse, and rerender one assistant turn through the production |
| 721 | /// renderer/output-processor boundary. |
| 722 | async fn run_roundtrip( |
| 723 | case: &RoundtripCase, |
| 724 | backends: &vllm_chat::LoadedModelBackends, |
| 725 | request: &ChatRequest, |
| 726 | assistant: AssistantMessage, |
| 727 | ) -> Result<RoundtripResult> { |
| 728 | let renderer = backends.chat_backend.chat_renderer(); |
| 729 | let rendered = render_closed_completion(renderer.as_ref(), request, &assistant)?; |
| 730 | |
| 731 | let ParsedCompletion { |
| 732 | message: parsed_message, |
| 733 | output_grammar, |
| 734 | } = parse_completion(case, backends, request, &rendered).await?; |
| 735 | let rerendered = render_closed_completion(renderer.as_ref(), request, &parsed_message)?; |
| 736 | |
| 737 | Ok(RoundtripResult { |
| 738 | parsed_message, |
| 739 | closed_completion: rendered.completion, |
| 740 | rerendered_closed_completion: rerendered.completion, |
| 741 | output_grammar, |
| 742 | }) |
| 743 | } |
| 744 | |
| 745 | /// Rendered prompt/completion artifacts at the renderer boundary. |
| 746 | struct RenderedTurn { |
| 747 | prompt: Prompt, |
| 748 | completion: Prompt, |
| 749 | } |
| 750 | |
| 751 | /// Render `history` as a production prompt and `history + assistant` as closed |
| 752 | /// history, then return the production prompt and the assistant completion. |
| 753 | fn render_closed_completion( |
| 754 | renderer: &dyn vllm_chat::ChatRenderer, |
| 755 | base_request: &ChatRequest, |
| 756 | assistant: &AssistantMessage, |
| 757 | ) -> Result<RenderedTurn> { |
| 758 | let mut prompt_request = base_request.clone(); |
| 759 | prompt_request.chat_options.generation_prompt_mode = GenerationPromptMode::StartNewAssistant; |
| 760 | let prompt = renderer.render(&prompt_request).context("failed to render prompt")?.prompt; |
| 761 | |
| 762 | let mut full_request = base_request.clone(); |
| 763 | full_request.chat_options.generation_prompt_mode = GenerationPromptMode::NoGenerationPrompt; |
| 764 | full_request.messages.push(ChatMessage::from(assistant.clone())); |
| 765 | let full = renderer.render(&full_request).context("failed to render full prompt")?.prompt; |
| 766 | |
| 767 | let completion = match (&prompt, full) { |
| 768 | (Prompt::Text(prompt), Prompt::Text(full)) => { |
| 769 | let range = completion_range(prompt.as_bytes(), full.as_bytes()) |
| 770 | .with_context(|| format!("prompt: {prompt:?}\nfull: {full:?}"))?; |
| 771 | let completion = full.get(range).context("completion splits a UTF-8 character")?; |
| 772 | Prompt::Text(completion.to_string()) |
| 773 | } |
| 774 | (Prompt::TokenIds(prompt), Prompt::TokenIds(full)) => { |
| 775 | let range = completion_range(prompt, &full) |
| 776 | .with_context(|| format!("prompt: {prompt:?}\nfull: {full:?}"))?; |
| 777 | Prompt::TokenIds(full[range].to_vec()) |
| 778 | } |
| 779 | (prompt, full) => bail!("prompt kind changed between renders: {prompt:?} vs {full:?}"), |
| 780 | }; |
| 781 | |
| 782 | Ok(RenderedTurn { prompt, completion }) |
| 783 | } |
| 784 | |
| 785 | /// Locate the assistant completion in `full`, the history rendered with the |
| 786 | /// assistant message, given `prompt`, the same history rendered for generation. |
| 787 | /// |
| 788 | /// Some templates render request-scoped control messages after the whole |
| 789 | /// history, such as Kimi K3's `tool_choice=required` message. With history `A`, |
| 790 | /// control messages `C`, generation prompt `G`, and completion `B`, the renders |
| 791 | /// are `prompt = A + C + G` and `full = A + G + B + C`. `C` is empty for most |
| 792 | /// templates, which makes `prompt` a prefix of `full`. |
| 793 | fn completion_range<T: PartialEq + Debug>(prompt: &[T], full: &[T]) -> Result<Range<usize>> { |
| 794 | if full.starts_with(prompt) { |
| 795 | return Ok(prompt.len()..full.len()); |
| 796 | } |
| 797 | ensure!( |
| 798 | full.len() >= prompt.len(), |
| 799 | "full prompt is shorter than production prompt" |
| 800 | ); |
| 801 | |
| 802 | // `A` is a common prefix of both renders; try every split of the remaining |
| 803 | // `prompt` into `C + G` and require exactly one to fit `full`. |
| 804 | let common_prefix_len = prompt.iter().zip(full).take_while(|(p, f)| p == f).count(); |
| 805 | let mut control_lens = Vec::new(); |
| 806 | for history_len in 0..=common_prefix_len { |
| 807 | let rest = &prompt[history_len..]; |
| 808 | for control_len in 1..=rest.len() { |
| 809 | let (control, generation) = rest.split_at(control_len); |
| 810 | if full[history_len..].starts_with(generation) && full.ends_with(control) { |
| 811 | control_lens.push(control_len); |
| 812 | } |
| 813 | } |
| 814 | } |
| 815 | let [control_len] = control_lens[..] else { |
| 816 | bail!( |
| 817 | "full prompt must be `A + G + B + C` for production prompt `A + C + G`; found {} splits", |
| 818 | control_lens.len() |
| 819 | ); |
| 820 | }; |
| 821 | |
| 822 | Ok(prompt.len() - control_len..full.len() - control_len) |
| 823 | } |
| 824 | |
| 825 | #[test] |
| 826 | fn completion_range_skips_trailing_control_messages() { |
| 827 | let history = "<user>hi</user>"; |
| 828 | let control = "<system>call tools</system>"; |
| 829 | let generation = "<assistant>"; |
| 830 | let completion = "call()</assistant>"; |
| 831 | |
| 832 | let prompt = format!("{history}{generation}"); |
| 833 | let full = format!("{history}{generation}{completion}"); |
| 834 | let range = completion_range(prompt.as_bytes(), full.as_bytes()).unwrap(); |
| 835 | assert_eq!(&full[range], completion); |
| 836 | |
| 837 | let prompt = format!("{history}{control}{generation}"); |
| 838 | let full = format!("{history}{generation}{completion}{control}"); |
| 839 | let range = completion_range(prompt.as_bytes(), full.as_bytes()).unwrap(); |
| 840 | assert_eq!(&full[range], completion); |
| 841 | } |
| 842 | |
| 843 | /// Output of one production output processor over a rendered completion. |
| 844 | struct ParsedCompletion { |
| 845 | /// Terminal assistant message. |
| 846 | message: AssistantMessage, |
| 847 | /// Output grammar built by the processor after initialization. |
| 848 | output_grammar: Option<BuiltOutputGrammar>, |
| 849 | } |
| 850 | |
| 851 | /// Feed one rendered assistant completion body into the real output processor |
| 852 | /// and collect its terminal assistant message, along with the output grammar |
| 853 | /// the initialized processor builds for the request. |
| 854 | async fn parse_completion( |
| 855 | case: &RoundtripCase, |
| 856 | backends: &vllm_chat::LoadedModelBackends, |
| 857 | base_request: &ChatRequest, |
| 858 | rendered: &RenderedTurn, |
| 859 | ) -> Result<ParsedCompletion> { |
| 860 | let tokenizer = backends.text_backend.tokenizer(); |
| 861 | let prompt_token_ids = match &rendered.prompt { |
| 862 | Prompt::Text(prompt) => tokenizer |
| 863 | .encode(prompt, base_request.add_special_tokens) |
| 864 | .context("failed to encode rendered prompt")?, |
| 865 | Prompt::TokenIds(token_ids) => token_ids.clone(), |
| 866 | }; |
| 867 | |
| 868 | let mut request = base_request.clone(); |
| 869 | let mut processor = backends.chat_backend.new_chat_output_processor( |
| 870 | &mut request, |
| 871 | NewChatOutputProcessorOptions { |
| 872 | // Pin every tool schema so that builder-backed parsers build an |
| 873 | // output grammar for replay. The strict level does not affect parsing. |
| 874 | tool_strict_level: ToolStrictLevel::Parameter, |
| 875 | tool_call_parser: &case.tool_call_parser, |
| 876 | reasoning_parser: &case.reasoning_parser, |
| 877 | }, |
| 878 | )?; |
| 879 | processor.initialize(&prompt_token_ids)?; |
| 880 | let output_grammar = processor.build_output_grammar()?; |
| 881 | |
| 882 | let decoded = decoded_completion_stream( |
| 883 | tokenizer.as_ref(), |
| 884 | prompt_token_ids, |
| 885 | &rendered.completion, |
| 886 | case.assistant_stop_suffix, |
| 887 | )?; |
| 888 | let mut events = processor.process(decoded)?; |
| 889 | |
| 890 | while let Some(event) = events.next().await { |
| 891 | if let ChatEvent::Done { message, .. } = event? { |
| 892 | return Ok(ParsedCompletion { |
| 893 | message, |
| 894 | output_grammar, |
| 895 | }); |
| 896 | } |
| 897 | } |
| 898 | |
| 899 | bail!("output processor finished without a Done event") |
| 900 | } |
| 901 | |
| 902 | /// Build a decoded-text stream from an already-rendered completion body. |
| 903 | /// |
| 904 | /// The first event carries real prompt token ids so reasoning parsers can |
| 905 | /// initialize from the same prompt boundary production uses. Completion text is |
| 906 | /// split into small chunks to exercise streaming parser state across marker |
| 907 | /// and JSON boundaries. |
| 908 | fn decoded_completion_stream( |
| 909 | tokenizer: &dyn Tokenizer, |
| 910 | prompt_token_ids: Vec<u32>, |
| 911 | completion: &Prompt, |
| 912 | assistant_stop_suffix: &str, |
| 913 | ) -> Result<Pin<Box<dyn Stream<Item = vllm_chat::Result<DecodedTextEvent>> + Send>>> { |
| 914 | let mut events = vec![DecodedTextEvent::Start { |
| 915 | prompt_token_ids: Arc::from(prompt_token_ids.clone().into_boxed_slice()), |
| 916 | prompt_logprobs: None, |
| 917 | }]; |
| 918 | |
| 919 | let chunks = match completion_body(tokenizer, completion, assistant_stop_suffix)? { |
| 920 | CompletionBody::Text(body) => { |
| 921 | // The engine returns token ids, so attribute the rendered text |
| 922 | // through the production decoder, then split it into small char |
| 923 | // chunks to exercise marker splits. |
| 924 | let token_ids = |
| 925 | tokenizer.encode(body, false).context("failed to encode text completion")?; |
| 926 | let mut rest = vllm_text::DecodedText::default(); |
| 927 | for chunk in incremental_decode_chunks( |
| 928 | tokenizer, |
| 929 | &prompt_token_ids, |
| 930 | &token_ids, |
| 931 | TOKEN_COMPLETION_CHUNK_TOKENS, |
| 932 | )? { |
| 933 | rest.append(chunk.delta); |
| 934 | } |
| 935 | ensure!( |
| 936 | rest.text == body, |
| 937 | "text completion must roundtrip through the tokenizer\nbody: {body:?}\ndecoded: {:?}", |
| 938 | rest.text |
| 939 | ); |
| 940 | split_by_chars(body, TEXT_COMPLETION_CHUNK_CHARS) |
| 941 | .into_iter() |
| 942 | .map(|delta| DecodedCompletionChunk { |
| 943 | delta: rest.drain_prefix(delta.len()), |
| 944 | token_ids: Vec::new(), // unused for text-level roundtrip cases |
| 945 | }) |
| 946 | .collect() |
| 947 | } |
| 948 | CompletionBody::TokenIds(body) => incremental_decode_chunks( |
| 949 | tokenizer, |
| 950 | &prompt_token_ids, |
| 951 | body, |
| 952 | TOKEN_COMPLETION_CHUNK_TOKENS, |
| 953 | )?, |
| 954 | }; |
| 955 | |
| 956 | if chunks.is_empty() { |
| 957 | events.push({ |
| 958 | DecodedTextEvent::TextDelta { |
| 959 | decoded: vllm_text::DecodedText::default(), |
| 960 | sampled: vllm_text::SampledDelta::default(), |
| 961 | finished: Some(Box::new(Finished { |
| 962 | usage: Default::default(), |
| 963 | finish_reason: FinishReason::stop_eos(), |
| 964 | kv_transfer_params: None, |
| 965 | ec_transfer_params: None, |
| 966 | sampling_mask: None, |
| 967 | })), |
| 968 | } |
| 969 | }); |
| 970 | } else { |
| 971 | let last_index = chunks.len() - 1; |
| 972 | for (index, chunk) in chunks.into_iter().enumerate() { |
| 973 | let finished = (index == last_index).then(|| Finished { |
| 974 | usage: Default::default(), |
| 975 | finish_reason: FinishReason::stop_eos(), |
| 976 | kv_transfer_params: None, |
| 977 | ec_transfer_params: None, |
| 978 | sampling_mask: None, |
| 979 | }); |
| 980 | events.push(DecodedTextEvent::TextDelta { |
| 981 | decoded: chunk.delta, |
| 982 | sampled: vllm_text::SampledDelta { |
| 983 | token_ids: chunk.token_ids, |
| 984 | logprobs: None, |
| 985 | }, |
| 986 | finished: finished.map(Box::new), |
| 987 | }); |
| 988 | } |
| 989 | } |
| 990 | |
| 991 | Ok(stream::iter(events).map(Ok).boxed()) |
| 992 | } |
| 993 | |
| 994 | /// Generated part of a closed assistant completion, without the rendered stop |
| 995 | /// suffix. |
| 996 | enum CompletionBody<'a> { |
| 997 | Text(&'a str), |
| 998 | TokenIds(&'a [u32]), |
| 999 | } |
| 1000 | |
| 1001 | /// Strip the rendered stop suffix from a closed assistant completion. |
| 1002 | fn completion_body<'a>( |
| 1003 | tokenizer: &dyn Tokenizer, |
| 1004 | completion: &'a Prompt, |
| 1005 | assistant_stop_suffix: &str, |
| 1006 | ) -> Result<CompletionBody<'a>> { |
| 1007 | match completion { |
| 1008 | Prompt::Text(text) => { |
| 1009 | let body = text.strip_suffix(assistant_stop_suffix).with_context(|| { |
| 1010 | format!( |
| 1011 | "closed assistant completion did not end with {:?}: {:?}", |
| 1012 | assistant_stop_suffix, text |
| 1013 | ) |
| 1014 | })?; |
| 1015 | Ok(CompletionBody::Text(body)) |
| 1016 | } |
| 1017 | Prompt::TokenIds(token_ids) => { |
| 1018 | let body = if assistant_stop_suffix.is_empty() { |
| 1019 | token_ids.as_slice() |
| 1020 | } else { |
| 1021 | let stop_token_ids = tokenizer |
| 1022 | .encode(assistant_stop_suffix, false) |
| 1023 | .context("failed to encode token-id completion stop suffix")?; |
| 1024 | token_ids.strip_suffix(stop_token_ids.as_slice()).with_context(|| { |
| 1025 | format!( |
| 1026 | "token-id completion did not end with {:?}: {:?}", |
| 1027 | assistant_stop_suffix, token_ids |
| 1028 | ) |
| 1029 | })? |
| 1030 | }; |
| 1031 | Ok(CompletionBody::TokenIds(body)) |
| 1032 | } |
| 1033 | } |
| 1034 | } |
| 1035 | |
| 1036 | /// One decoded completion chunk fed into the output processor. |
| 1037 | struct DecodedCompletionChunk { |
| 1038 | /// Attributed for token-id completions, as the production decoder emits it. |
| 1039 | delta: vllm_text::DecodedText, |
| 1040 | token_ids: Vec<u32>, |
| 1041 | } |
| 1042 | |
| 1043 | /// Split text into chunks containing at most `chunk_chars` Unicode scalar |
| 1044 | /// values. |
| 1045 | fn split_by_chars(text: &str, chunk_chars: usize) -> Vec<String> { |
| 1046 | let mut chunks = Vec::new(); |
| 1047 | let mut start = 0; |
| 1048 | let mut count = 0; |
| 1049 | |
| 1050 | for (index, _) in text.char_indices() { |
| 1051 | if count == chunk_chars { |
| 1052 | chunks.push(text[start..index].to_string()); |
| 1053 | start = index; |
| 1054 | count = 0; |
| 1055 | } |
| 1056 | count += 1; |
| 1057 | } |
| 1058 | |
| 1059 | if start < text.len() { |
| 1060 | chunks.push(text[start..].to_string()); |
| 1061 | } |
| 1062 | |
| 1063 | chunks |
| 1064 | } |
| 1065 | |
| 1066 | /// Split token ids into chunks containing at most `chunk_size` ids. |
| 1067 | fn split_by_count(token_ids: &[u32], chunk_size: usize) -> Vec<Vec<u32>> { |
| 1068 | token_ids.chunks(chunk_size).map(<[u32]>::to_vec).collect() |
| 1069 | } |
| 1070 | |
| 1071 | /// Decode token ids incrementally using the production tokenizer stream. |
| 1072 | fn incremental_decode_chunks( |
| 1073 | tokenizer: &dyn Tokenizer, |
| 1074 | prompt_token_ids: &[u32], |
| 1075 | token_ids: &[u32], |
| 1076 | chunk_size: usize, |
| 1077 | ) -> Result<Vec<DecodedCompletionChunk>> { |
| 1078 | let mut decoder = tokenizer.create_decode_stream(prompt_token_ids, false, 0); |
| 1079 | let mut chunks = Vec::new(); |
| 1080 | for chunk_token_ids in split_by_count(token_ids, chunk_size) { |
| 1081 | let mut delta = vllm_text::DecodedText::default(); |
| 1082 | for token_id in chunk_token_ids.iter().copied() { |
| 1083 | decoder.push_token(token_id)?; |
| 1084 | while let Some(chunk) = decoder.next_chunk() { |
| 1085 | delta.append(chunk); |
| 1086 | } |
| 1087 | } |
| 1088 | chunks.push(DecodedCompletionChunk { |
| 1089 | delta, |
| 1090 | token_ids: chunk_token_ids, |
| 1091 | }); |
| 1092 | } |
| 1093 | |
| 1094 | let (last_chunk, _) = decoder.flush(None)?; |
| 1095 | if let Some(last_chunk) = last_chunk { |
| 1096 | if let Some(delta) = chunks.last_mut() { |
| 1097 | delta.delta.append(last_chunk); |
| 1098 | } else { |
| 1099 | chunks.push(DecodedCompletionChunk { |
| 1100 | delta: last_chunk, |
| 1101 | token_ids: Vec::new(), |
| 1102 | }); |
| 1103 | } |
| 1104 | } |
| 1105 | |
| 1106 | Ok(chunks) |
| 1107 | } |
| 1108 | |
| 1109 | /// Grammar replay cases of one model, replayed through the real XGrammar by |
| 1110 | /// `grammar_replay.py`. |
| 1111 | #[derive(Serialize)] |
| 1112 | struct GrammarReplayFile { |
| 1113 | /// Hugging Face model id whose tokenizer the replay loads. |
| 1114 | model_id: &'static str, |
| 1115 | /// Model vocabulary size, which sizes the XGrammar token mask. |
| 1116 | vocab_size: usize, |
| 1117 | /// Request stop set, passed to the matcher as `override_stop_tokens`. |
| 1118 | #[serde(serialize_with = "serialize_one_line")] |
| 1119 | all_stop_token_ids: BTreeSet<u32>, |
| 1120 | /// Cases keyed by tool-choice variant. |
| 1121 | cases: BTreeMap<ToolChoiceVariant, GrammarReplayCase>, |
| 1122 | } |
| 1123 | |
| 1124 | /// One parser-built output grammar with the generation it must accept. |
| 1125 | #[derive(Serialize)] |
| 1126 | struct GrammarReplayCase { |
| 1127 | /// Parser-built grammar; only token-zero grammars are replayed from the |
| 1128 | /// first generated token. |
| 1129 | grammar: BuiltOutputGrammar, |
| 1130 | /// Generated completion text, without the stop token. |
| 1131 | completion: String, |
| 1132 | /// Generated token IDs presented to the matcher, ending with a stop token. |
| 1133 | #[serde(serialize_with = "serialize_one_line")] |
| 1134 | generation_token_ids: Vec<u32>, |
| 1135 | } |
| 1136 | |
| 1137 | /// Serialize a token ID list on one line inside the pretty-printed file. |
| 1138 | fn serialize_one_line<S: Serializer>( |
| 1139 | token_ids: &impl Serialize, |
| 1140 | serializer: S, |
| 1141 | ) -> std::result::Result<S::Ok, S::Error> { |
| 1142 | let json = spaced_json_fmt().format_to_string(token_ids).map_err(S::Error::custom)?; |
| 1143 | serde_json::value::RawValue::from_string(json) |
| 1144 | .map_err(S::Error::custom)? |
| 1145 | .serialize(serializer) |
| 1146 | } |
| 1147 | |
| 1148 | /// Directory of checked-in grammar replay files. |
| 1149 | fn grammar_replay_dir() -> PathBuf { |
| 1150 | PathBuf::from(env!("CARGO_MANIFEST_DIR")).join("tests/grammar_replay") |
| 1151 | } |
| 1152 | |
| 1153 | /// Snapshot the output grammars of one model's roundtrips, keyed by variant, |
| 1154 | /// as its grammar replay file, or check that no stale file remains when the |
| 1155 | /// parser builds none. |
| 1156 | fn check_grammar_replay_file( |
| 1157 | case: &RoundtripCase, |
| 1158 | backends: &vllm_chat::LoadedModelBackends, |
| 1159 | results: Vec<(ToolChoiceVariant, RoundtripResult)>, |
| 1160 | ) -> Result<()> { |
| 1161 | let model_name = case.model_id.rsplit_once('/').map_or(case.model_id, |(_, name)| name); |
| 1162 | let path = grammar_replay_dir().join(format!("{model_name}.json")); |
| 1163 | let outline_path = path.with_extension("txt"); |
| 1164 | let results = results |
| 1165 | .into_iter() |
| 1166 | .filter_map(|(variant, result)| { |
| 1167 | Some((variant, result.output_grammar?, result.closed_completion)) |
| 1168 | }) |
| 1169 | .collect::<Vec<_>>(); |
| 1170 | if results.is_empty() { |
| 1171 | for path in [&path, &outline_path] { |
| 1172 | ensure!( |
| 1173 | !path.exists(), |
| 1174 | "{} builds no output grammar, but grammar replay file {path:?} exists", |
| 1175 | case.model_id |
| 1176 | ); |
| 1177 | } |
| 1178 | return Ok(()); |
| 1179 | } |
| 1180 | |
| 1181 | let tokenizer = backends.text_backend.tokenizer(); |
| 1182 | let hints = backends.text_backend.sampling_hints()?; |
| 1183 | // Lowered `_all_stop_token_ids` of a request without user stop tokens. |
| 1184 | let mut all_stop_token_ids = hints.extra_eos_token_ids; |
| 1185 | all_stop_token_ids.extend(hints.primary_eos_token_id); |
| 1186 | |
| 1187 | // The generation ends with the first token of the rendered stop suffix, or |
| 1188 | // with the primary EOS for templates that render none. |
| 1189 | let stop_token_id = if case.assistant_stop_suffix.is_empty() { |
| 1190 | hints.primary_eos_token_id.context("model has no primary EOS token")? |
| 1191 | } else { |
| 1192 | tokenizer.encode(case.assistant_stop_suffix, false)?[0] |
| 1193 | }; |
| 1194 | ensure!( |
| 1195 | all_stop_token_ids.contains(&stop_token_id), |
| 1196 | "stop token {stop_token_id} of {:?} is not in the stop set {all_stop_token_ids:?}", |
| 1197 | case.assistant_stop_suffix |
| 1198 | ); |
| 1199 | |
| 1200 | let mut cases = BTreeMap::new(); |
| 1201 | for (variant, grammar, closed_completion) in results { |
| 1202 | let (completion, mut generation_token_ids) = match completion_body( |
| 1203 | tokenizer.as_ref(), |
| 1204 | &closed_completion, |
| 1205 | case.assistant_stop_suffix, |
| 1206 | )? { |
| 1207 | CompletionBody::Text(body) => (body.to_string(), tokenizer.encode(body, false)?), |
| 1208 | CompletionBody::TokenIds(body) => (tokenizer.decode(body, false)?, body.to_vec()), |
| 1209 | }; |
| 1210 | generation_token_ids.push(stop_token_id); |
| 1211 | cases.insert( |
| 1212 | variant, |
| 1213 | GrammarReplayCase { |
| 1214 | grammar, |
| 1215 | completion, |
| 1216 | generation_token_ids, |
| 1217 | }, |
| 1218 | ); |
| 1219 | } |
| 1220 | |
| 1221 | // A readable outline of the same grammars, for review only. |
| 1222 | let mut outline = String::new(); |
| 1223 | for (variant, case) in &cases { |
| 1224 | if !outline.is_empty() { |
| 1225 | outline.push('\n'); |
| 1226 | } |
| 1227 | let variant = serde_json::to_value(variant)?; |
| 1228 | let coverage = serde_json::to_value(case.grammar.coverage)?; |
| 1229 | outline.push_str(&format!( |
| 1230 | "# {} ({})\n", |
| 1231 | variant.as_str().context("variant name")?, |
| 1232 | coverage.as_str().context("coverage name")?, |
| 1233 | )); |
| 1234 | outline.push_str(&vllm_parser::output_grammar::test_utils::outline( |
| 1235 | &case.grammar.format, |
| 1236 | )); |
| 1237 | } |
| 1238 | expect_test::expect_file![outline_path].assert_eq(&outline); |
| 1239 | |
| 1240 | let file = GrammarReplayFile { |
| 1241 | model_id: case.model_id, |
| 1242 | vocab_size: backends.text_backend.model_vocab_size(), |
| 1243 | all_stop_token_ids, |
| 1244 | cases, |
| 1245 | }; |
| 1246 | let mut json = serde_json::to_string_pretty(&file)?; |
| 1247 | json.push('\n'); |
| 1248 | expect_test::expect_file![path].assert_eq(&json); |
| 1249 | |
| 1250 | Ok(()) |
| 1251 | } |
| 1252 | |
| 1253 | /// Build a chat request fixture with parser-enabling tool-choice semantics. |
| 1254 | fn roundtrip_request( |
| 1255 | request_id: impl Into<String>, |
| 1256 | messages: Vec<ChatMessage>, |
| 1257 | tools: Vec<ChatTool>, |
| 1258 | tool_choice: Option<ChatToolChoice>, |
| 1259 | thinking: Option<bool>, |
| 1260 | thinking_behavior: ThinkingBehavior, |
| 1261 | ) -> ChatRequest { |
| 1262 | let tool_context = vllm_chat::ResolvedToolContext::new(&messages, tools, tool_choice, true) |
| 1263 | .expect("tool context should resolve"); |
| 1264 | let mut request = ChatRequest { |
| 1265 | request_id: request_id.into(), |
| 1266 | messages, |
| 1267 | tool_context, |
| 1268 | ..ChatRequest::for_test() |
| 1269 | }; |
| 1270 | |
| 1271 | // Explicitly enable or disable thinking using the controls understood by |
| 1272 | // this model's template. If unspecified, use the template default. |
| 1273 | thinking_behavior.apply(&mut request, thinking); |
| 1274 | |
| 1275 | request |
| 1276 | } |
| 1277 | |
| 1278 | /// Return the function tools used by the multiple-tool-call fixture. |
| 1279 | fn test_tools() -> Vec<ChatTool> { |
| 1280 | vec![ |
| 1281 | ChatTool { |
| 1282 | name: "get_weather".to_string(), |
| 1283 | description: Some("Get weather for a location".to_string()), |
| 1284 | parameters: serde_json::json!({ |
| 1285 | "type": "object", |
| 1286 | "properties": { |
| 1287 | "location": { "type": "string" } |
| 1288 | }, |
| 1289 | "required": ["location"] |
| 1290 | }), |
| 1291 | strict: None, |
| 1292 | }, |
| 1293 | ChatTool { |
| 1294 | name: "add".to_string(), |
| 1295 | description: Some("Add two integers".to_string()), |
| 1296 | parameters: serde_json::json!({ |
| 1297 | "type": "object", |
| 1298 | "properties": { |
| 1299 | "y": { "type": "number" }, |
| 1300 | "x": { "type": "number" }, |
| 1301 | "items": { |
| 1302 | "type": "array", |
| 1303 | "items": { "type": "string" } |
| 1304 | } |
| 1305 | }, |
| 1306 | "required": ["y", "x", "items"] |
| 1307 | }), |
| 1308 | strict: None, |
| 1309 | }, |
| 1310 | ] |
| 1311 | } |