| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | # Adapted from |
| 4 | # https://github.com/vllm-project/vllm/blob/main/vllm/entrypoints/openai/chat_completion/serving.py |
| 5 | |
| 6 | """Anthropic Messages API serving handler.""" |
| 7 | |
| 8 | import json |
| 9 | import logging |
| 10 | import time |
| 11 | import uuid |
| 12 | from collections.abc import AsyncGenerator |
| 13 | from typing import Any, Literal, get_args |
| 14 | |
| 15 | import jinja2 |
| 16 | from fastapi import Request |
| 17 | |
| 18 | from vllm.engine.protocol import EngineClient |
| 19 | from vllm.entrypoints.anthropic.protocol import ( |
| 20 | AnthropicContentBlock, |
| 21 | AnthropicContextManagement, |
| 22 | AnthropicCountTokensRequest, |
| 23 | AnthropicCountTokensResponse, |
| 24 | AnthropicDelta, |
| 25 | AnthropicDisabledThinkingEffort, |
| 26 | AnthropicDisabledThinkingEffortOption, |
| 27 | AnthropicEffort, |
| 28 | AnthropicError, |
| 29 | AnthropicMessagesRequest, |
| 30 | AnthropicMessagesResponse, |
| 31 | AnthropicOutputConfig, |
| 32 | AnthropicStreamEvent, |
| 33 | AnthropicThinkingConfig, |
| 34 | AnthropicTool, |
| 35 | AnthropicToolChangeDefinition, |
| 36 | AnthropicUsage, |
| 37 | ) |
| 38 | from vllm.entrypoints.chat_utils import ChatTemplateContentFormatOption |
| 39 | from vllm.entrypoints.generate.base.protocol import ( |
| 40 | JsonSchemaResponseFormat, |
| 41 | ResponseFormat, |
| 42 | StreamOptions, |
| 43 | ) |
| 44 | from vllm.entrypoints.openai.chat_completion.protocol import ( |
| 45 | ChatCompletionNamedToolChoiceParam, |
| 46 | ChatCompletionRequest, |
| 47 | ChatCompletionResponse, |
| 48 | ChatCompletionStreamResponse, |
| 49 | ChatCompletionToolsParam, |
| 50 | ) |
| 51 | from vllm.entrypoints.openai.chat_completion.serving import OpenAIServingChat |
| 52 | from vllm.entrypoints.openai.models.serving import OpenAIServingModels |
| 53 | from vllm.entrypoints.serve.engine.protocol import ErrorResponse, UsageInfo |
| 54 | from vllm.entrypoints.serve.exception_handling.utils import sanitize_message |
| 55 | from vllm.entrypoints.serve.utils.request_logger import RequestLogger |
| 56 | from vllm.logger import init_logger |
| 57 | from vllm.renderers.hf import HfRenderer, resolve_chat_template |
| 58 | from vllm.renderers.online_renderer import OnlineRenderer |
| 59 | |
| 60 | logger = init_logger(__name__) |
| 61 | |
| 62 | |
| 63 | def _build_anthropic_usage( |
| 64 | usage: UsageInfo | None, |
| 65 | ) -> AnthropicUsage: |
| 66 | """Build an AnthropicUsage from UsageInfo. |
| 67 | |
| 68 | Anthropic defines ``total_input == input_tokens + cache_read + |
| 69 | cache_creation``. vLLM's ``prompt_tokens`` is the total, so |
| 70 | ``input_tokens = prompt_tokens - cache_read - cache_creation``. |
| 71 | |
| 72 | Cache fields are taken from ``UsageInfo.prompt_tokens_details``. |
| 73 | When cache info is absent (e.g. ``--enable-prompt-tokens-details`` |
| 74 | off, or a streaming chunk that hasn't carried it yet), cache fields |
| 75 | are left **unset** so ``exclude_unset=True`` serialization omits them |
| 76 | entirely. |
| 77 | |
| 78 | ``completion_tokens`` follows ``UsageInfo`` and may be ``None`` on |
| 79 | intermediate stream chunks; we coerce to ``0`` for the wire format. |
| 80 | """ |
| 81 | kwargs = {} |
| 82 | if usage is None: |
| 83 | kwargs["input_tokens"] = 0 |
| 84 | kwargs["output_tokens"] = 0 |
| 85 | else: |
| 86 | kwargs["output_tokens"] = usage.completion_tokens or 0 |
| 87 | input_tokens = usage.prompt_tokens |
| 88 | |
| 89 | if (details := usage.prompt_tokens_details) is not None: |
| 90 | if (cache_read := details.cached_tokens) is not None: |
| 91 | input_tokens -= cache_read |
| 92 | kwargs["cache_read_input_tokens"] = cache_read |
| 93 | |
| 94 | if (cache_creation := details.created_cache_tokens) is not None: |
| 95 | input_tokens -= cache_creation |
| 96 | kwargs["cache_creation_input_tokens"] = cache_creation |
| 97 | |
| 98 | kwargs["input_tokens"] = max(0, input_tokens) |
| 99 | return AnthropicUsage(**kwargs) |
| 100 | |
| 101 | |
| 102 | def _to_anthropic_stop_reason( |
| 103 | finish_reason: str | None, |
| 104 | stop_reason: int | str | None, |
| 105 | tool_used: bool, |
| 106 | tool_complete: bool, |
| 107 | ) -> tuple[ |
| 108 | Literal["end_turn", "max_tokens", "stop_sequence", "tool_use"] | None, |
| 109 | str | None, |
| 110 | ]: |
| 111 | """Map an OpenAI finish reason to Anthropic ``(stop_reason, stop_sequence)``. |
| 112 | |
| 113 | Anthropic reports ``tool_use`` whenever the model invoked a tool. The chat |
| 114 | layer reports ``finish_reason="stop"`` for a named (forced) tool call, so |
| 115 | ``tool_used`` records whether a ``tool_use`` block was produced. |
| 116 | ``max_tokens`` always wins. A matched stop string wins unless the tool |
| 117 | calls are complete (``tool_complete``): a call it cut short must not run. |
| 118 | """ |
| 119 | if finish_reason == "length": |
| 120 | return "max_tokens", None |
| 121 | if finish_reason not in ("stop", "tool_calls"): |
| 122 | return None, None |
| 123 | if tool_complete: |
| 124 | return "tool_use", None |
| 125 | # vLLM reports the matched stop string in stop_reason (a str); |
| 126 | # an int stop-token-id or None (natural EOS) is not a stop sequence. |
| 127 | if isinstance(stop_reason, str): |
| 128 | return "stop_sequence", stop_reason |
| 129 | if finish_reason == "tool_calls" or tool_used: |
| 130 | return "tool_use", None |
| 131 | return "end_turn", None |
| 132 | |
| 133 | |
| 134 | def _is_json(text: str) -> bool: |
| 135 | try: |
| 136 | json.loads(text) |
| 137 | except json.JSONDecodeError: |
| 138 | return False |
| 139 | return True |
| 140 | |
| 141 | |
| 142 | def wrap_data_with_event(data: str, event: str): |
| 143 | return f"event: {event}\ndata: {data}\n\n" |
| 144 | |
| 145 | |
| 146 | class AnthropicServingMessages(OpenAIServingChat): |
| 147 | """Handler for Anthropic Messages API requests.""" |
| 148 | |
| 149 | def __init__( |
| 150 | self, |
| 151 | engine_client: EngineClient, |
| 152 | models: OpenAIServingModels, |
| 153 | response_role: str, |
| 154 | *, |
| 155 | online_renderer: "OnlineRenderer", |
| 156 | request_logger: RequestLogger | None, |
| 157 | chat_template: str | None, |
| 158 | chat_template_content_format: ChatTemplateContentFormatOption, |
| 159 | return_tokens_as_token_ids: bool = False, |
| 160 | reasoning_parser: str = "", |
| 161 | enable_auto_tools: bool = False, |
| 162 | tool_parser: str | None = None, |
| 163 | tool_strict_level: str = "auto", |
| 164 | enable_prompt_tokens_details: bool = False, |
| 165 | enable_force_include_usage: bool = False, |
| 166 | default_chat_template_kwargs: dict[str, Any] | None = None, |
| 167 | disabled_thinking_effort: AnthropicDisabledThinkingEffortOption = "auto", |
| 168 | ): |
| 169 | super().__init__( |
| 170 | engine_client=engine_client, |
| 171 | models=models, |
| 172 | response_role=response_role, |
| 173 | online_renderer=online_renderer, |
| 174 | request_logger=request_logger, |
| 175 | chat_template=chat_template, |
| 176 | chat_template_content_format=chat_template_content_format, |
| 177 | return_tokens_as_token_ids=return_tokens_as_token_ids, |
| 178 | reasoning_parser=reasoning_parser, |
| 179 | enable_auto_tools=enable_auto_tools, |
| 180 | tool_strict_level=tool_strict_level, |
| 181 | tool_parser=tool_parser, |
| 182 | enable_prompt_tokens_details=enable_prompt_tokens_details, |
| 183 | enable_force_include_usage=enable_force_include_usage, |
| 184 | default_chat_template_kwargs=default_chat_template_kwargs, |
| 185 | ) |
| 186 | self._merge_inline_system = self._should_merge_inline_system(online_renderer) |
| 187 | # Resolved lazily from the renderer when "auto". |
| 188 | self._disabled_thinking_effort: AnthropicDisabledThinkingEffort | None = ( |
| 189 | None if disabled_thinking_effort == "auto" else disabled_thinking_effort |
| 190 | ) |
| 191 | |
| 192 | async def _get_disabled_thinking_effort(self) -> AnthropicDisabledThinkingEffort: |
| 193 | if self._disabled_thinking_effort is None: |
| 194 | self._disabled_thinking_effort = ( |
| 195 | await self._probe_disabled_thinking_effort() |
| 196 | ) |
| 197 | logger.info( |
| 198 | "Anthropic thinking.type=disabled maps to reasoning_effort=%r", |
| 199 | self._disabled_thinking_effort, |
| 200 | ) |
| 201 | return self._disabled_thinking_effort |
| 202 | |
| 203 | async def _probe_disabled_thinking_effort(self) -> AnthropicDisabledThinkingEffort: |
| 204 | """Use ``low`` if the renderer rejects or ignores ``none``. |
| 205 | |
| 206 | ``none`` is ignored when it renders the same prompt as a thinking |
| 207 | effort, e.g. GLM-5.3 treats unknown efforts as ``max``. |
| 208 | """ |
| 209 | none_prompt = await self._render_probe_prompt("none") |
| 210 | if none_prompt is None: |
| 211 | return "low" |
| 212 | for effort in get_args(AnthropicEffort): |
| 213 | if await self._render_probe_prompt(effort) == none_prompt: |
| 214 | return "low" |
| 215 | return "none" |
| 216 | |
| 217 | async def _render_probe_prompt( |
| 218 | self, effort: AnthropicDisabledThinkingEffort |
| 219 | ) -> tuple[list[int] | None, str | None] | None: |
| 220 | request = ChatCompletionRequest( |
| 221 | messages=[{"role": "user", "content": "Hi"}], |
| 222 | reasoning_effort=effort, |
| 223 | ) |
| 224 | try: |
| 225 | result = await self.online_renderer.render_chat(request) |
| 226 | except Exception: |
| 227 | logger.debug("Rendering reasoning_effort=%r failed", effort, exc_info=True) |
| 228 | return None |
| 229 | if isinstance(result, ErrorResponse): |
| 230 | return None |
| 231 | _, (engine_input,) = result |
| 232 | components = self._extract_prompt_components(engine_input) |
| 233 | return components.token_ids, components.text |
| 234 | |
| 235 | @classmethod |
| 236 | def _should_merge_inline_system(cls, online_renderer: OnlineRenderer) -> bool: |
| 237 | """Probe the chat templates the renderer applies, not the CLI override. |
| 238 | |
| 239 | Both the tool and non-tool templates are checked since they may differ. |
| 240 | """ |
| 241 | renderer = online_renderer.renderer |
| 242 | tool_variants: tuple[list[dict[str, Any]] | None, ...] = (None, []) |
| 243 | if not isinstance(renderer, HfRenderer) or renderer.tokenizer is None: |
| 244 | merge = cls._detect_merge_inline_system(online_renderer.chat_template) |
| 245 | else: |
| 246 | merge = any( |
| 247 | cls._detect_merge_inline_system( |
| 248 | resolve_chat_template( |
| 249 | renderer.tokenizer, |
| 250 | online_renderer.chat_template, |
| 251 | tools, |
| 252 | model_config=online_renderer.model_config, |
| 253 | ) |
| 254 | ) |
| 255 | for tools in tool_variants |
| 256 | ) |
| 257 | if merge: |
| 258 | logger.warning_once( |
| 259 | "The chat template requires system-first ordering, so inline " |
| 260 | "system messages in /v1/messages requests (e.g. Claude Code's " |
| 261 | "per-turn reminders) are merged into the leading system prompt. " |
| 262 | "Each new one changes the prompt prefix, so the rest of the " |
| 263 | "conversation misses the prefix cache. Pass a --chat-template " |
| 264 | "that accepts non-leading system messages to avoid this." |
| 265 | ) |
| 266 | return merge |
| 267 | |
| 268 | @staticmethod |
| 269 | def _detect_merge_inline_system(chat_template: str | None) -> bool: |
| 270 | """Auto-detect whether the chat template requires system-first ordering. |
| 271 | |
| 272 | Renders a [system, user, system, user] conversation against the |
| 273 | template; if it raises (e.g. Qwen's ``loop.first`` guard), the |
| 274 | model needs inline system messages merged into the leading block. |
| 275 | """ |
| 276 | if not chat_template: |
| 277 | return True |
| 278 | try: |
| 279 | env = jinja2.sandbox.ImmutableSandboxedEnvironment( |
| 280 | trim_blocks=True, |
| 281 | lstrip_blocks=True, |
| 282 | extensions=[jinja2.ext.loopcontrols], |
| 283 | ) |
| 284 | env.from_string(chat_template).render( |
| 285 | messages=[ |
| 286 | {"role": "system", "content": "t"}, |
| 287 | {"role": "user", "content": "t"}, |
| 288 | {"role": "system", "content": "t"}, |
| 289 | {"role": "user", "content": "t"}, |
| 290 | ], |
| 291 | add_generation_prompt=False, |
| 292 | ) |
| 293 | return False |
| 294 | except jinja2.TemplateError: |
| 295 | return True |
| 296 | |
| 297 | @staticmethod |
| 298 | def _convert_image_source_to_url(source: dict[str, Any]) -> str: |
| 299 | """Convert an Anthropic image source to an OpenAI-compatible URL. |
| 300 | |
| 301 | Anthropic supports two image source types: |
| 302 | - base64: {"type": "base64", "media_type": "image/jpeg", "data": "..."} |
| 303 | - url: {"type": "url", "url": "https://..."} |
| 304 | |
| 305 | For base64 sources, this constructs a proper data URI that |
| 306 | downstream processors (e.g. vLLM's media connector) can handle. |
| 307 | """ |
| 308 | source_type = source.get("type") |
| 309 | if source_type == "url": |
| 310 | return source.get("url", "") |
| 311 | # Default to base64 processing if type is "base64" |
| 312 | # or missing, ensuring a proper data URI is always |
| 313 | # constructed for non-URL sources. |
| 314 | media_type = source.get("media_type", "image/jpeg") |
| 315 | data = source.get("data", "") |
| 316 | return f"data:{media_type};base64,{data}" |
| 317 | |
| 318 | @classmethod |
| 319 | def to_chat_completion_request( |
| 320 | cls, |
| 321 | anthropic_request: AnthropicMessagesRequest | AnthropicCountTokensRequest, |
| 322 | *, |
| 323 | merge_inline_system: bool = False, |
| 324 | disabled_thinking_effort: AnthropicDisabledThinkingEffort = "none", |
| 325 | ) -> ChatCompletionRequest: |
| 326 | """Convert Anthropic message format to OpenAI format.""" |
| 327 | openai_messages: list[dict[str, Any]] = [] |
| 328 | |
| 329 | cls._convert_system_message( |
| 330 | anthropic_request, |
| 331 | openai_messages, |
| 332 | merge_inline_system=merge_inline_system, |
| 333 | ) |
| 334 | cls._convert_messages( |
| 335 | anthropic_request.messages, |
| 336 | openai_messages, |
| 337 | merge_inline_system=merge_inline_system, |
| 338 | ) |
| 339 | req = cls._build_base_request(anthropic_request, openai_messages) |
| 340 | cls._handle_streaming_options(req, anthropic_request) |
| 341 | cls._handle_output_config(req, anthropic_request) |
| 342 | cls._handle_thinking(req, anthropic_request, disabled_thinking_effort) |
| 343 | cls._convert_tool_choice(anthropic_request, req) |
| 344 | cls._convert_tools(anthropic_request, req) |
| 345 | return req |
| 346 | |
| 347 | @classmethod |
| 348 | def _convert_system_message( |
| 349 | cls, |
| 350 | anthropic_request: AnthropicMessagesRequest | AnthropicCountTokensRequest, |
| 351 | openai_messages: list[dict[str, Any]], |
| 352 | *, |
| 353 | merge_inline_system: bool = False, |
| 354 | ) -> None: |
| 355 | """Convert Anthropic system message to OpenAI format.""" |
| 356 | system_parts: list[str] = [] |
| 357 | |
| 358 | # Top-level system field |
| 359 | if anthropic_request.system: |
| 360 | if isinstance(anthropic_request.system, str): |
| 361 | system_parts.append(anthropic_request.system) |
| 362 | else: |
| 363 | for block in anthropic_request.system: |
| 364 | if block.type == "text" and block.text: |
| 365 | # Strip Claude Code's attribution header which contains |
| 366 | # a per-request hash that defeats prefix caching. |
| 367 | if block.text.startswith("x-anthropic-billing-header"): |
| 368 | continue |
| 369 | system_parts.append(block.text) |
| 370 | |
| 371 | # When the template requires system-first ordering, extract inline |
| 372 | # system messages from the messages array and merge them into the |
| 373 | # top-level block so the template doesn't reject them. |
| 374 | if merge_inline_system: |
| 375 | for msg in anthropic_request.messages: |
| 376 | if msg.role != "system": |
| 377 | continue |
| 378 | text = cls._extract_system_text(msg) |
| 379 | if text: |
| 380 | system_parts.append(text) |
| 381 | |
| 382 | if system_parts: |
| 383 | openai_messages.append({"role": "system", "content": "".join(system_parts)}) |
| 384 | |
| 385 | @classmethod |
| 386 | def _extract_system_text(cls, msg) -> str | None: |
| 387 | """Extract text from a system message, stripping billing headers.""" |
| 388 | if isinstance(msg.content, str): |
| 389 | text = msg.content |
| 390 | if text.startswith("x-anthropic-billing-header"): |
| 391 | return None |
| 392 | return text |
| 393 | parts: list[str] = [] |
| 394 | for block in msg.content: |
| 395 | if block.type == "text" and block.text: |
| 396 | if block.text.startswith("x-anthropic-billing-header"): |
| 397 | continue |
| 398 | parts.append(block.text) |
| 399 | return "".join(parts) if parts else None |
| 400 | |
| 401 | @classmethod |
| 402 | def _convert_messages( |
| 403 | cls, |
| 404 | messages: list, |
| 405 | openai_messages: list[dict[str, Any]], |
| 406 | *, |
| 407 | merge_inline_system: bool = False, |
| 408 | ) -> None: |
| 409 | """Convert Anthropic messages to OpenAI format.""" |
| 410 | for msg in messages: |
| 411 | # Handle system messages in-place: extract text, strip billing |
| 412 | # headers, and only emit if there is real content. This avoids |
| 413 | # going through _convert_block / _convert_message_content which |
| 414 | # doesn't strip billing headers and may produce messages with |
| 415 | # no "content" key. |
| 416 | if msg.role == "system": |
| 417 | if merge_inline_system: |
| 418 | continue # already merged into top-level by _convert_system_message |
| 419 | text = cls._extract_system_text(msg) |
| 420 | if text: |
| 421 | openai_messages.append({"role": "system", "content": text}) |
| 422 | continue |
| 423 | |
| 424 | openai_msg: dict[str, Any] = {"role": msg.role} # type: ignore |
| 425 | |
| 426 | if isinstance(msg.content, str): |
| 427 | openai_msg["content"] = msg.content |
| 428 | else: |
| 429 | cls._convert_message_content(msg, openai_msg, openai_messages) |
| 430 | |
| 431 | if not (msg.role == "user" and "content" not in openai_msg): |
| 432 | openai_messages.append(openai_msg) |
| 433 | |
| 434 | @classmethod |
| 435 | def _convert_message_content( |
| 436 | cls, |
| 437 | msg, |
| 438 | openai_msg: dict[str, Any], |
| 439 | openai_messages: list[dict[str, Any]], |
| 440 | ) -> None: |
| 441 | """Convert complex message content blocks.""" |
| 442 | content_parts: list[dict[str, Any]] = [] |
| 443 | tool_calls: list[dict[str, Any]] = [] |
| 444 | reasoning_parts: list[str] = [] |
| 445 | |
| 446 | for block in msg.content: |
| 447 | cls._convert_block( |
| 448 | block, |
| 449 | msg.role, |
| 450 | content_parts, |
| 451 | tool_calls, |
| 452 | reasoning_parts, |
| 453 | openai_messages, |
| 454 | ) |
| 455 | |
| 456 | if reasoning_parts: |
| 457 | openai_msg["reasoning"] = "".join(reasoning_parts) |
| 458 | |
| 459 | if tool_calls: |
| 460 | openai_msg["tool_calls"] = tool_calls # type: ignore |
| 461 | |
| 462 | if content_parts: |
| 463 | if len(content_parts) == 1 and content_parts[0]["type"] == "text": |
| 464 | openai_msg["content"] = content_parts[0]["text"] |
| 465 | else: |
| 466 | openai_msg["content"] = content_parts # type: ignore |
| 467 | elif not tool_calls and not reasoning_parts: |
| 468 | return |
| 469 | |
| 470 | @classmethod |
| 471 | def _convert_block( |
| 472 | cls, |
| 473 | block, |
| 474 | role: str, |
| 475 | content_parts: list[dict[str, Any]], |
| 476 | tool_calls: list[dict[str, Any]], |
| 477 | reasoning_parts: list[str], |
| 478 | openai_messages: list[dict[str, Any]], |
| 479 | ) -> None: |
| 480 | """Convert individual content block.""" |
| 481 | if block.type == "text" and block.text: |
| 482 | content_parts.append({"type": "text", "text": block.text}) |
| 483 | elif block.type == "image" and block.source: |
| 484 | image_url = cls._convert_image_source_to_url(block.source) |
| 485 | content_parts.append({"type": "image_url", "image_url": {"url": image_url}}) |
| 486 | elif block.type == "thinking" and block.thinking is not None: |
| 487 | reasoning_parts.append(block.thinking) |
| 488 | elif block.type == "redacted_thinking": |
| 489 | # Redacted thinking blocks contain safety-filtered reasoning. |
| 490 | # We skip them as the content is opaque (base64 'data' field), |
| 491 | # but accepting the block prevents a validation error when the |
| 492 | # client echoes back the full assistant message. |
| 493 | pass |
| 494 | elif block.type == "tool_use": |
| 495 | cls._convert_tool_use_block(block, tool_calls) |
| 496 | elif block.type == "tool_result": |
| 497 | cls._convert_tool_result_block(block, role, openai_messages, content_parts) |
| 498 | elif block.type == "tool_reference": |
| 499 | # Tool references are expanded during tool_result processing |
| 500 | # when they appear inside tool_result content. |
| 501 | pass |
| 502 | |
| 503 | @classmethod |
| 504 | def _convert_tool_use_block(cls, block, tool_calls: list[dict[str, Any]]) -> None: |
| 505 | """Convert tool_use block to OpenAI function call format.""" |
| 506 | tool_call = { |
| 507 | "id": block.id or f"call_{int(time.time())}", |
| 508 | "type": "function", |
| 509 | "function": { |
| 510 | "name": block.name or "", |
| 511 | "arguments": json.dumps(block.input or {}), |
| 512 | }, |
| 513 | } |
| 514 | tool_calls.append(tool_call) |
| 515 | |
| 516 | @classmethod |
| 517 | def _convert_tool_result_block( |
| 518 | cls, |
| 519 | block, |
| 520 | role: str, |
| 521 | openai_messages: list[dict[str, Any]], |
| 522 | content_parts: list[dict[str, Any]], |
| 523 | ) -> None: |
| 524 | """Convert tool_result block to OpenAI format.""" |
| 525 | if role == "user": |
| 526 | cls._convert_user_tool_result(block, openai_messages) |
| 527 | else: |
| 528 | tool_result_text = str(block.content) if block.content else "" |
| 529 | content_parts.append( |
| 530 | {"type": "text", "text": f"Tool result: {tool_result_text}"} |
| 531 | ) |
| 532 | |
| 533 | @classmethod |
| 534 | def _convert_user_tool_result( |
| 535 | cls, block, openai_messages: list[dict[str, Any]] |
| 536 | ) -> None: |
| 537 | """Convert user tool_result with text and image support.""" |
| 538 | tool_text = "" |
| 539 | tool_image_urls: list[str] = [] |
| 540 | tool_reference: list[dict[str, Any]] = [] |
| 541 | |
| 542 | if isinstance(block.content, str): |
| 543 | tool_text = block.content |
| 544 | elif isinstance(block.content, list): |
| 545 | text_parts: list[str] = [] |
| 546 | for item in block.content: |
| 547 | if not isinstance(item, dict): |
| 548 | continue |
| 549 | item_type = item.get("type") |
| 550 | if item_type == "text": |
| 551 | text_parts.append(item.get("text", "")) |
| 552 | elif item_type == "image": |
| 553 | source = item.get("source", {}) |
| 554 | url = cls._convert_image_source_to_url(source) |
| 555 | if url: |
| 556 | tool_image_urls.append(url) |
| 557 | elif item_type == "tool_reference": |
| 558 | ref_name = item.get("tool_name") or item.get("name") |
| 559 | if ref_name: |
| 560 | tool_reference.append( |
| 561 | {"type": "tool_reference", "name": ref_name} |
| 562 | ) |
| 563 | tool_text = "\n".join(text_parts) |
| 564 | |
| 565 | openai_messages.append( |
| 566 | { |
| 567 | "role": "tool", |
| 568 | "tool_call_id": block.tool_use_id or "", |
| 569 | "content": tool_text or "", |
| 570 | } |
| 571 | ) |
| 572 | |
| 573 | if tool_image_urls: |
| 574 | openai_messages.append( |
| 575 | { |
| 576 | "role": "user", |
| 577 | "content": [ # type: ignore[dict-item] |
| 578 | {"type": "image_url", "image_url": {"url": img}} |
| 579 | for img in tool_image_urls |
| 580 | ], |
| 581 | } |
| 582 | ) |
| 583 | |
| 584 | if tool_reference: |
| 585 | openai_messages.append( |
| 586 | { |
| 587 | "role": "tool", |
| 588 | "tool_call_id": block.tool_use_id or "", |
| 589 | "content": tool_reference, # type: ignore[dict-item] |
| 590 | } |
| 591 | ) |
| 592 | |
| 593 | @classmethod |
| 594 | def _build_base_request( |
| 595 | cls, |
| 596 | anthropic_request: AnthropicMessagesRequest | AnthropicCountTokensRequest, |
| 597 | openai_messages: list[dict[str, Any]], |
| 598 | ) -> ChatCompletionRequest: |
| 599 | """Build base ChatCompletionRequest.""" |
| 600 | if isinstance(anthropic_request, AnthropicCountTokensRequest): |
| 601 | return ChatCompletionRequest( |
| 602 | model=anthropic_request.model, |
| 603 | messages=openai_messages, |
| 604 | chat_template_kwargs=anthropic_request.chat_template_kwargs, |
| 605 | ) |
| 606 | |
| 607 | return ChatCompletionRequest( |
| 608 | model=anthropic_request.model, |
| 609 | messages=openai_messages, |
| 610 | max_tokens=anthropic_request.max_tokens, |
| 611 | max_completion_tokens=anthropic_request.max_tokens, |
| 612 | stop=anthropic_request.stop_sequences, |
| 613 | temperature=anthropic_request.temperature, |
| 614 | top_p=anthropic_request.top_p, |
| 615 | top_k=anthropic_request.top_k, |
| 616 | cache_salt=anthropic_request.cache_salt, |
| 617 | kv_transfer_params=anthropic_request.kv_transfer_params, |
| 618 | ec_transfer_params=anthropic_request.ec_transfer_params, |
| 619 | vllm_xargs=anthropic_request.vllm_xargs, |
| 620 | chat_template_kwargs=anthropic_request.chat_template_kwargs, |
| 621 | ) |
| 622 | |
| 623 | @classmethod |
| 624 | def _handle_thinking( |
| 625 | cls, |
| 626 | req: ChatCompletionRequest, |
| 627 | anthropic_request: AnthropicMessagesRequest | AnthropicCountTokensRequest, |
| 628 | disabled_thinking_effort: AnthropicDisabledThinkingEffort = "none", |
| 629 | ) -> None: |
| 630 | """Handle extended-thinking configuration. |
| 631 | |
| 632 | ``display`` is intentionally ignored: suppressing reasoning would mark |
| 633 | it ended for structured outputs and drop it from multi-turn history. |
| 634 | """ |
| 635 | if isinstance(anthropic_request, AnthropicCountTokensRequest): |
| 636 | return |
| 637 | thinking: AnthropicThinkingConfig | None = anthropic_request.thinking |
| 638 | if thinking is None: |
| 639 | return |
| 640 | |
| 641 | if thinking.type == "disabled": |
| 642 | # "none" clears enable_thinking for templates that honor it; models |
| 643 | # that cannot disable thinking are configured with a low effort. |
| 644 | req.reasoning_effort = disabled_thinking_effort |
| 645 | elif thinking.type == "enabled": |
| 646 | req.thinking_token_budget = thinking.budget_tokens |
| 647 | # "adaptive" pins nothing: the model chooses depth beneath the ceiling |
| 648 | # already set from output_config.effort. |
| 649 | |
| 650 | @classmethod |
| 651 | def _handle_output_config( |
| 652 | cls, |
| 653 | req: ChatCompletionRequest, |
| 654 | anthropic_request: AnthropicMessagesRequest | AnthropicCountTokensRequest, |
| 655 | ) -> None: |
| 656 | """Handle output configuration such as output format and effort.""" |
| 657 | if isinstance(anthropic_request, AnthropicCountTokensRequest): |
| 658 | return |
| 659 | output_config: AnthropicOutputConfig | None = anthropic_request.output_config |
| 660 | if output_config and output_config.format and output_config.format.json_schema: |
| 661 | req.response_format = ResponseFormat( |
| 662 | type=output_config.format.type, |
| 663 | json_schema=JsonSchemaResponseFormat( |
| 664 | schema=output_config.format.json_schema, |
| 665 | name=output_config.format.type, |
| 666 | ), |
| 667 | ) |
| 668 | if output_config and output_config.effort is not None: |
| 669 | req.reasoning_effort = output_config.effort |
| 670 | |
| 671 | @classmethod |
| 672 | def _handle_streaming_options( |
| 673 | cls, |
| 674 | req: ChatCompletionRequest, |
| 675 | anthropic_request: AnthropicMessagesRequest | AnthropicCountTokensRequest, |
| 676 | ) -> None: |
| 677 | """Handle streaming configuration.""" |
| 678 | if isinstance(anthropic_request, AnthropicCountTokensRequest): |
| 679 | return |
| 680 | if anthropic_request.stream: |
| 681 | req.stream = anthropic_request.stream |
| 682 | req.stream_options = StreamOptions.model_validate( |
| 683 | {"include_usage": True, "continuous_usage_stats": True} |
| 684 | ) |
| 685 | |
| 686 | @classmethod |
| 687 | def _convert_tool_choice( |
| 688 | cls, |
| 689 | anthropic_request: AnthropicMessagesRequest | AnthropicCountTokensRequest, |
| 690 | req: ChatCompletionRequest, |
| 691 | ) -> None: |
| 692 | """Convert Anthropic tool_choice to OpenAI format.""" |
| 693 | if anthropic_request.tool_choice is None: |
| 694 | req.tool_choice = None |
| 695 | return |
| 696 | |
| 697 | req.parallel_tool_calls = ( |
| 698 | not anthropic_request.tool_choice.disable_parallel_tool_use |
| 699 | ) |
| 700 | tool_choice_type = anthropic_request.tool_choice.type |
| 701 | if tool_choice_type == "auto": |
| 702 | req.tool_choice = "auto" |
| 703 | elif tool_choice_type == "any": |
| 704 | req.tool_choice = "required" |
| 705 | elif tool_choice_type == "none": |
| 706 | req.tool_choice = "none" |
| 707 | elif tool_choice_type == "tool": |
| 708 | req.tool_choice = ChatCompletionNamedToolChoiceParam.model_validate( |
| 709 | { |
| 710 | "type": "function", |
| 711 | "function": {"name": anthropic_request.tool_choice.name}, |
| 712 | } |
| 713 | ) |
| 714 | |
| 715 | @classmethod |
| 716 | def _resolve_tools( |
| 717 | cls, |
| 718 | anthropic_request: AnthropicMessagesRequest | AnthropicCountTokensRequest, |
| 719 | ) -> list[AnthropicTool] | None: |
| 720 | """Apply tool_addition/tool_removal blocks in ``messages`` to ``tools``. |
| 721 | |
| 722 | Chat templates see a single tool list, so this resolves the tools in |
| 723 | effect at the end of the conversation: a tool added by reference is no |
| 724 | longer deferred, a tool added by value is declared, and a removed tool |
| 725 | is dropped. |
| 726 | """ |
| 727 | changes = [ |
| 728 | block |
| 729 | for msg in anthropic_request.messages |
| 730 | if not isinstance(msg.content, str) |
| 731 | for block in msg.content |
| 732 | if block.type in ("tool_addition", "tool_removal") |
| 733 | ] |
| 734 | if not changes: |
| 735 | return anthropic_request.tools |
| 736 | |
| 737 | tools = {tool.name: tool for tool in anthropic_request.tools or []} |
| 738 | removed: set[str] = set() |
| 739 | for block in changes: |
| 740 | change = block.tool |
| 741 | assert change is not None |
| 742 | if isinstance(change, AnthropicToolChangeDefinition): |
| 743 | name = change.definition.name |
| 744 | tools[name] = change.definition |
| 745 | else: |
| 746 | name = change.name |
| 747 | if name not in tools: |
| 748 | raise ValueError( |
| 749 | f"{block.type} references tool {name!r}, " |
| 750 | "which is not declared in tools" |
| 751 | ) |
| 752 | if block.type == "tool_addition": |
| 753 | tools[name] = tools[name].model_copy(update={"defer_loading": None}) |
| 754 | if block.type == "tool_removal": |
| 755 | removed.add(name) |
| 756 | else: |
| 757 | removed.discard(name) |
| 758 | return [tool for name, tool in tools.items() if name not in removed] |
| 759 | |
| 760 | @classmethod |
| 761 | def _convert_tools( |
| 762 | cls, |
| 763 | anthropic_request: AnthropicMessagesRequest | AnthropicCountTokensRequest, |
| 764 | req: ChatCompletionRequest, |
| 765 | ) -> None: |
| 766 | """Convert Anthropic tools to OpenAI format.""" |
| 767 | anthropic_tools = cls._resolve_tools(anthropic_request) |
| 768 | if anthropic_tools is None: |
| 769 | return |
| 770 | |
| 771 | tools = [] |
| 772 | for tool in anthropic_tools: |
| 773 | tools.append( |
| 774 | ChatCompletionToolsParam.model_validate( |
| 775 | { |
| 776 | "type": "function", |
| 777 | "function": { |
| 778 | "name": tool.name, |
| 779 | "description": tool.description, |
| 780 | "parameters": tool.input_schema, |
| 781 | "strict": tool.strict, |
| 782 | "defer_loading": tool.defer_loading, |
| 783 | }, |
| 784 | } |
| 785 | ) |
| 786 | ) |
| 787 | |
| 788 | if req.tool_choice is None: |
| 789 | req.tool_choice = "auto" |
| 790 | req.tools = tools |
| 791 | |
| 792 | async def create_messages( |
| 793 | self, |
| 794 | request: AnthropicMessagesRequest, |
| 795 | raw_request: Request | None = None, |
| 796 | ) -> AsyncGenerator[str, None] | AnthropicMessagesResponse | ErrorResponse: |
| 797 | """Messages API similar to Anthropic's API. |
| 798 | |
| 799 | See https://docs.anthropic.com/en/api/messages |
| 800 | for the API specification. This API mimics the Anthropic messages API. |
| 801 | """ |
| 802 | if logger.isEnabledFor(logging.DEBUG): |
| 803 | logger.debug("Received messages request %s", request.model_dump_json()) |
| 804 | disabled_thinking_effort: AnthropicDisabledThinkingEffort = "none" |
| 805 | if request.thinking is not None and request.thinking.type == "disabled": |
| 806 | disabled_thinking_effort = await self._get_disabled_thinking_effort() |
| 807 | chat_req = self.to_chat_completion_request( |
| 808 | request, |
| 809 | merge_inline_system=self._merge_inline_system, |
| 810 | disabled_thinking_effort=disabled_thinking_effort, |
| 811 | ) |
| 812 | if logger.isEnabledFor(logging.DEBUG): |
| 813 | logger.debug("Convert to OpenAI request %s", chat_req.model_dump_json()) |
| 814 | generator = await self.create_chat_completion(chat_req, raw_request) |
| 815 | |
| 816 | if isinstance(generator, ErrorResponse): |
| 817 | return generator |
| 818 | |
| 819 | elif isinstance(generator, ChatCompletionResponse): |
| 820 | return self.messages_full_converter(generator) |
| 821 | |
| 822 | return self.message_stream_converter(generator) |
| 823 | |
| 824 | def messages_full_converter( |
| 825 | self, |
| 826 | generator: ChatCompletionResponse, |
| 827 | ) -> AnthropicMessagesResponse: |
| 828 | result = AnthropicMessagesResponse( |
| 829 | id=generator.id, |
| 830 | content=[], |
| 831 | model=generator.model, |
| 832 | usage=_build_anthropic_usage(generator.usage), |
| 833 | kv_transfer_params=generator.kv_transfer_params, |
| 834 | ec_transfer_params=generator.ec_transfer_params, |
| 835 | ) |
| 836 | choice = generator.choices[0] |
| 837 | result.stop_reason, result.stop_sequence = _to_anthropic_stop_reason( |
| 838 | choice.finish_reason, |
| 839 | choice.stop_reason, |
| 840 | tool_used=bool(choice.message.tool_calls), |
| 841 | # Their arguments are json-decoded below, so parsed calls are complete. |
| 842 | tool_complete=bool(choice.message.tool_calls), |
| 843 | ) |
| 844 | |
| 845 | content: list[AnthropicContentBlock] = [] |
| 846 | if choice.message.reasoning: |
| 847 | content.append( |
| 848 | AnthropicContentBlock( |
| 849 | type="thinking", |
| 850 | thinking=choice.message.reasoning, |
| 851 | signature=uuid.uuid4().hex, |
| 852 | ) |
| 853 | ) |
| 854 | if choice.message.content: |
| 855 | content.append( |
| 856 | AnthropicContentBlock( |
| 857 | type="text", |
| 858 | text=choice.message.content, |
| 859 | ) |
| 860 | ) |
| 861 | |
| 862 | for tool_call in choice.message.tool_calls: |
| 863 | anthropic_tool_call = AnthropicContentBlock( |
| 864 | type="tool_use", |
| 865 | id=tool_call.id, |
| 866 | name=tool_call.function.name, |
| 867 | input=json.loads(tool_call.function.arguments), |
| 868 | ) |
| 869 | content += [anthropic_tool_call] |
| 870 | |
| 871 | # Anthropic's canonical shape for an empty completion is a single |
| 872 | # empty text block, not []. Some strict clients assume content[0] |
| 873 | # exists, so emit one here. |
| 874 | if not content: |
| 875 | content.append(AnthropicContentBlock(type="text", text="")) |
| 876 | |
| 877 | result.content = content |
| 878 | |
| 879 | return result |
| 880 | |
| 881 | async def message_stream_converter( |
| 882 | self, |
| 883 | generator: AsyncGenerator[str, None], |
| 884 | ) -> AsyncGenerator[str, None]: |
| 885 | try: |
| 886 | |
| 887 | class _ActiveBlockState: |
| 888 | def __init__(self) -> None: |
| 889 | self.content_block_index = 0 |
| 890 | self.block_type: str | None = None |
| 891 | self.block_index: int | None = None |
| 892 | self.block_signature: str | None = None |
| 893 | self.signature_emitted: bool = False |
| 894 | self.tool_use_id: str | None = None |
| 895 | self.pending_content: list[str] = [] |
| 896 | |
| 897 | def reset(self) -> None: |
| 898 | self.block_type = None |
| 899 | self.block_index = None |
| 900 | self.block_signature = None |
| 901 | self.signature_emitted = False |
| 902 | self.tool_use_id = None |
| 903 | self.pending_content.clear() |
| 904 | |
| 905 | def start(self, block: AnthropicContentBlock) -> None: |
| 906 | self.block_type = block.type |
| 907 | self.block_index = self.content_block_index |
| 908 | if block.type == "thinking": |
| 909 | self.block_signature = uuid.uuid4().hex |
| 910 | self.signature_emitted = False |
| 911 | self.tool_use_id = None |
| 912 | elif block.type == "tool_use": |
| 913 | self.block_signature = None |
| 914 | self.signature_emitted = True |
| 915 | self.tool_use_id = block.id |
| 916 | else: |
| 917 | self.block_signature = None |
| 918 | self.signature_emitted = True |
| 919 | self.tool_use_id = None |
| 920 | |
| 921 | first_item = True |
| 922 | finish_reason = None |
| 923 | # Matched stop string, when generation stopped on one (a str); |
| 924 | # int stop-token-id / None are not stop sequences. |
| 925 | stop_sequence: int | str | None = None |
| 926 | tool_used = False |
| 927 | # Arguments of the latest tool_use block; only the last call can be |
| 928 | # cut short by a stop string. |
| 929 | last_tool_args: list[str] = [] |
| 930 | state = _ActiveBlockState() |
| 931 | # Map from tool call index to tool_use_id |
| 932 | tool_index_to_id: dict[int, str] = {} |
| 933 | |
| 934 | def stop_active_block(): |
| 935 | events: list[str] = [] |
| 936 | if state.block_type is None: |
| 937 | return events |
| 938 | if ( |
| 939 | state.block_type == "thinking" |
| 940 | and state.block_signature is not None |
| 941 | and not state.signature_emitted |
| 942 | ): |
| 943 | chunk = AnthropicStreamEvent( |
| 944 | index=state.block_index, |
| 945 | type="content_block_delta", |
| 946 | delta=AnthropicDelta( |
| 947 | type="signature_delta", |
| 948 | signature=state.block_signature, |
| 949 | ), |
| 950 | ) |
| 951 | data = chunk.model_dump_json(exclude_unset=True) |
| 952 | events.append(wrap_data_with_event(data, "content_block_delta")) |
| 953 | state.signature_emitted = True |
| 954 | stop_chunk = AnthropicStreamEvent( |
| 955 | index=state.block_index, |
| 956 | type="content_block_stop", |
| 957 | ) |
| 958 | data = stop_chunk.model_dump_json(exclude_unset=True) |
| 959 | events.append(wrap_data_with_event(data, "content_block_stop")) |
| 960 | state.reset() |
| 961 | state.content_block_index += 1 |
| 962 | return events |
| 963 | |
| 964 | def start_block(block: AnthropicContentBlock): |
| 965 | chunk = AnthropicStreamEvent( |
| 966 | index=state.content_block_index, |
| 967 | type="content_block_start", |
| 968 | content_block=block, |
| 969 | ) |
| 970 | data = chunk.model_dump_json(exclude_unset=True) |
| 971 | event = wrap_data_with_event(data, "content_block_start") |
| 972 | state.start(block) |
| 973 | return event |
| 974 | |
| 975 | def stop_and_flush() -> list[str]: |
| 976 | buffered = list(state.pending_content) |
| 977 | state.pending_content.clear() |
| 978 | events = stop_active_block() |
| 979 | if not buffered: |
| 980 | return events |
| 981 | text = "".join(buffered) |
| 982 | events.append(start_block(AnthropicContentBlock(type="text", text=""))) |
| 983 | pc_chunk = AnthropicStreamEvent( |
| 984 | index=state.block_index, |
| 985 | type="content_block_delta", |
| 986 | delta=AnthropicDelta(type="text_delta", text=text), |
| 987 | ) |
| 988 | pc_data = pc_chunk.model_dump_json(exclude_unset=True) |
| 989 | events.append(wrap_data_with_event(pc_data, "content_block_delta")) |
| 990 | events.extend(stop_active_block()) |
| 991 | return events |
| 992 | |
| 993 | async for item in generator: |
| 994 | if item.startswith("data:"): |
| 995 | data_str = item[5:].strip().rstrip("\n") |
| 996 | if data_str == "[DONE]": |
| 997 | for event in stop_and_flush(): |
| 998 | yield event |
| 999 | stop_message = AnthropicStreamEvent( |
| 1000 | type="message_stop", |
| 1001 | ) |
| 1002 | data = stop_message.model_dump_json( |
| 1003 | exclude_unset=True, exclude_none=True |
| 1004 | ) |
| 1005 | yield wrap_data_with_event(data, "message_stop") |
| 1006 | else: |
| 1007 | origin_chunk = ChatCompletionStreamResponse.model_validate_json( |
| 1008 | data_str |
| 1009 | ) |
| 1010 | |
| 1011 | if first_item: |
| 1012 | chunk = AnthropicStreamEvent( |
| 1013 | type="message_start", |
| 1014 | message=AnthropicMessagesResponse( |
| 1015 | id=origin_chunk.id, |
| 1016 | # Set explicitly: this event is serialized |
| 1017 | # with exclude_unset=True, which drops |
| 1018 | # default-valued fields, while strict |
| 1019 | # Anthropic SDK clients require |
| 1020 | # message.type/role (issue #45367). |
| 1021 | type="message", |
| 1022 | role="assistant", |
| 1023 | content=[], |
| 1024 | model=origin_chunk.model, |
| 1025 | stop_reason=None, |
| 1026 | stop_sequence=None, |
| 1027 | usage=_build_anthropic_usage(origin_chunk.usage), |
| 1028 | ), |
| 1029 | ) |
| 1030 | first_item = False |
| 1031 | data = chunk.model_dump_json(exclude_unset=True) |
| 1032 | yield wrap_data_with_event(data, "message_start") |
| 1033 | continue |
| 1034 | |
| 1035 | # last chunk including usage info |
| 1036 | if len(origin_chunk.choices) == 0: |
| 1037 | for event in stop_and_flush(): |
| 1038 | yield event |
| 1039 | anthropic_stop_reason, anthropic_stop_sequence = ( |
| 1040 | _to_anthropic_stop_reason( |
| 1041 | finish_reason or "stop", |
| 1042 | stop_sequence, |
| 1043 | tool_used=tool_used, |
| 1044 | tool_complete=( |
| 1045 | tool_used and _is_json("".join(last_tool_args)) |
| 1046 | ), |
| 1047 | ) |
| 1048 | ) |
| 1049 | stop_delta = AnthropicDelta( |
| 1050 | stop_reason=anthropic_stop_reason, |
| 1051 | # Set explicitly so exclude_unset=True keeps it. |
| 1052 | stop_sequence=anthropic_stop_sequence, |
| 1053 | ) |
| 1054 | chunk = AnthropicStreamEvent( |
| 1055 | type="message_delta", |
| 1056 | delta=stop_delta, |
| 1057 | usage=_build_anthropic_usage(origin_chunk.usage), |
| 1058 | ) |
| 1059 | data = chunk.model_dump_json(exclude_unset=True) |
| 1060 | yield wrap_data_with_event(data, "message_delta") |
| 1061 | continue |
| 1062 | |
| 1063 | if origin_chunk.choices[0].finish_reason is not None: |
| 1064 | finish_reason = origin_chunk.choices[0].finish_reason |
| 1065 | stop_sequence = origin_chunk.choices[0].stop_reason |
| 1066 | # continue |
| 1067 | |
| 1068 | # thinking / text content |
| 1069 | reasoning_delta = origin_chunk.choices[0].delta.reasoning |
| 1070 | if reasoning_delta is not None: |
| 1071 | if reasoning_delta == "": |
| 1072 | pass |
| 1073 | else: |
| 1074 | if state.block_type != "thinking": |
| 1075 | for event in stop_and_flush(): |
| 1076 | yield event |
| 1077 | start_event = start_block( |
| 1078 | AnthropicContentBlock( |
| 1079 | type="thinking", thinking="" |
| 1080 | ) |
| 1081 | ) |
| 1082 | yield start_event |
| 1083 | chunk = AnthropicStreamEvent( |
| 1084 | index=( |
| 1085 | state.block_index |
| 1086 | if state.block_index is not None |
| 1087 | else state.content_block_index |
| 1088 | ), |
| 1089 | type="content_block_delta", |
| 1090 | delta=AnthropicDelta( |
| 1091 | type="thinking_delta", |
| 1092 | thinking=reasoning_delta, |
| 1093 | ), |
| 1094 | ) |
| 1095 | data = chunk.model_dump_json(exclude_unset=True) |
| 1096 | yield wrap_data_with_event(data, "content_block_delta") |
| 1097 | |
| 1098 | if origin_chunk.choices[0].delta.content is not None: |
| 1099 | if origin_chunk.choices[0].delta.content == "": |
| 1100 | pass |
| 1101 | elif state.block_type == "tool_use": |
| 1102 | state.pending_content.append( |
| 1103 | origin_chunk.choices[0].delta.content |
| 1104 | ) |
| 1105 | else: |
| 1106 | if state.block_type != "text": |
| 1107 | for event in stop_and_flush(): |
| 1108 | yield event |
| 1109 | start_event = start_block( |
| 1110 | AnthropicContentBlock(type="text", text="") |
| 1111 | ) |
| 1112 | yield start_event |
| 1113 | chunk = AnthropicStreamEvent( |
| 1114 | index=( |
| 1115 | state.block_index |
| 1116 | if state.block_index is not None |
| 1117 | else state.content_block_index |
| 1118 | ), |
| 1119 | type="content_block_delta", |
| 1120 | delta=AnthropicDelta( |
| 1121 | type="text_delta", |
| 1122 | text=origin_chunk.choices[0].delta.content, |
| 1123 | ), |
| 1124 | ) |
| 1125 | data = chunk.model_dump_json(exclude_unset=True) |
| 1126 | yield wrap_data_with_event(data, "content_block_delta") |
| 1127 | |
| 1128 | # tool calls - process all tool calls in the delta |
| 1129 | if len(origin_chunk.choices[0].delta.tool_calls) > 0: |
| 1130 | for tool_call in origin_chunk.choices[0].delta.tool_calls: |
| 1131 | if tool_call.id is not None: |
| 1132 | # Update mapping for incremental updates |
| 1133 | tool_index_to_id[tool_call.index] = tool_call.id |
| 1134 | # Only create new block if different tool call |
| 1135 | # AND has a name |
| 1136 | tool_name = ( |
| 1137 | tool_call.function.name |
| 1138 | if tool_call.function |
| 1139 | else None |
| 1140 | ) |
| 1141 | if ( |
| 1142 | state.tool_use_id != tool_call.id |
| 1143 | and tool_name is not None |
| 1144 | ): |
| 1145 | for event in stop_and_flush(): |
| 1146 | yield event |
| 1147 | tool_used = True |
| 1148 | last_tool_args = [] |
| 1149 | start_event = start_block( |
| 1150 | AnthropicContentBlock( |
| 1151 | type="tool_use", |
| 1152 | id=tool_call.id, |
| 1153 | name=tool_name, |
| 1154 | input={}, |
| 1155 | ) |
| 1156 | ) |
| 1157 | yield start_event |
| 1158 | # Handle initial arguments if present |
| 1159 | if ( |
| 1160 | tool_call.function |
| 1161 | and tool_call.function.arguments |
| 1162 | and state.tool_use_id == tool_call.id |
| 1163 | ): |
| 1164 | last_tool_args.append( |
| 1165 | tool_call.function.arguments |
| 1166 | ) |
| 1167 | chunk = AnthropicStreamEvent( |
| 1168 | index=( |
| 1169 | state.block_index |
| 1170 | if state.block_index is not None |
| 1171 | else state.content_block_index |
| 1172 | ), |
| 1173 | type="content_block_delta", |
| 1174 | delta=AnthropicDelta( |
| 1175 | type="input_json_delta", |
| 1176 | partial_json=tool_call.function.arguments, |
| 1177 | ), |
| 1178 | ) |
| 1179 | data = chunk.model_dump_json(exclude_unset=True) |
| 1180 | yield wrap_data_with_event( |
| 1181 | data, "content_block_delta" |
| 1182 | ) |
| 1183 | else: |
| 1184 | # Incremental update - use index to find tool_use_id |
| 1185 | tool_use_id = tool_index_to_id.get(tool_call.index) |
| 1186 | if ( |
| 1187 | tool_use_id is not None |
| 1188 | and tool_call.function |
| 1189 | and tool_call.function.arguments |
| 1190 | and state.tool_use_id == tool_use_id |
| 1191 | ): |
| 1192 | last_tool_args.append( |
| 1193 | tool_call.function.arguments |
| 1194 | ) |
| 1195 | chunk = AnthropicStreamEvent( |
| 1196 | index=( |
| 1197 | state.block_index |
| 1198 | if state.block_index is not None |
| 1199 | else state.content_block_index |
| 1200 | ), |
| 1201 | type="content_block_delta", |
| 1202 | delta=AnthropicDelta( |
| 1203 | type="input_json_delta", |
| 1204 | partial_json=tool_call.function.arguments, |
| 1205 | ), |
| 1206 | ) |
| 1207 | data = chunk.model_dump_json(exclude_unset=True) |
| 1208 | yield wrap_data_with_event( |
| 1209 | data, "content_block_delta" |
| 1210 | ) |
| 1211 | continue |
| 1212 | else: |
| 1213 | error_response = AnthropicStreamEvent( |
| 1214 | type="error", |
| 1215 | error=AnthropicError( |
| 1216 | type="internal_error", |
| 1217 | message="Invalid data format received", |
| 1218 | ), |
| 1219 | ) |
| 1220 | data = error_response.model_dump_json(exclude_unset=True) |
| 1221 | yield wrap_data_with_event(data, "error") |
| 1222 | |
| 1223 | except Exception as e: |
| 1224 | logger.exception("Error in message stream converter.") |
| 1225 | error_response = AnthropicStreamEvent( |
| 1226 | type="error", |
| 1227 | error=AnthropicError( |
| 1228 | type="internal_error", message=sanitize_message(str(e)) |
| 1229 | ), |
| 1230 | ) |
| 1231 | data = error_response.model_dump_json(exclude_unset=True) |
| 1232 | yield wrap_data_with_event(data, "error") |
| 1233 | |
| 1234 | async def count_tokens( |
| 1235 | self, |
| 1236 | request: AnthropicCountTokensRequest, |
| 1237 | raw_request: Request | None = None, |
| 1238 | ) -> AnthropicCountTokensResponse | ErrorResponse: |
| 1239 | """Implements Anthropic's messages.count_tokens endpoint.""" |
| 1240 | chat_req = self.to_chat_completion_request( |
| 1241 | request, |
| 1242 | merge_inline_system=self._merge_inline_system, |
| 1243 | ) |
| 1244 | result = await self.render_chat_request(chat_req) |
| 1245 | if isinstance(result, ErrorResponse): |
| 1246 | return result |
| 1247 | |
| 1248 | _, engine_inputs = result |
| 1249 | |
| 1250 | input_tokens = sum( # type: ignore |
| 1251 | len(engine_input["prompt_token_ids"]) # type: ignore[typeddict-item, misc] |
| 1252 | for engine_input in engine_inputs |
| 1253 | if "prompt_token_ids" in engine_input |
| 1254 | ) |
| 1255 | |
| 1256 | response = AnthropicCountTokensResponse( |
| 1257 | input_tokens=input_tokens, |
| 1258 | context_management=AnthropicContextManagement( |
| 1259 | original_input_tokens=input_tokens |
| 1260 | ), |
| 1261 | ) |
| 1262 | |
| 1263 | return response |