| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Unit tests for ``vllm/renderers/cohere.py``. |
| 4 | |
| 5 | The tests focus on the pure-Python helpers that produce the render-config |
| 6 | dicts passed to ``cohere_melody.render_cmd3`` / ``render_cmd4``. We also |
| 7 | include a class-level instantiation + async-non-blocking test that |
| 8 | mirrors the analogous ``test_mistral.py`` pattern, exercising the |
| 9 | :class:`CohereRenderer` end-to-end with mocked ``model_config`` / |
| 10 | tokenizer / melody bindings. |
| 11 | """ |
| 12 | |
| 13 | import asyncio |
| 14 | import json |
| 15 | import time |
| 16 | from dataclasses import dataclass |
| 17 | from typing import Any |
| 18 | from unittest.mock import Mock |
| 19 | |
| 20 | import pytest |
| 21 | |
| 22 | from vllm.renderers import ChatParams |
| 23 | from vllm.renderers.cohere import ( |
| 24 | CohereRenderer, |
| 25 | MelodyContentType, |
| 26 | _build_render_config, |
| 27 | _content_blocks, |
| 28 | _conversation_to_melody_messages, |
| 29 | _document_to_melody, |
| 30 | _normalize_tool_call, |
| 31 | _role_to_melody, |
| 32 | _tool_to_melody, |
| 33 | _v2_tool_content_to_melody_block, |
| 34 | ) |
| 35 | from vllm.tokenizers.hf import HfTokenizer |
| 36 | |
| 37 | # ====================================================================== |
| 38 | # _role_to_melody |
| 39 | # ====================================================================== |
| 40 | |
| 41 | |
| 42 | class TestRoleToMelody: |
| 43 | def test_assistant_maps_to_chatbot(self): |
| 44 | # melody's templates use the legacy Cohere ``chatbot`` role name. |
| 45 | assert _role_to_melody("assistant") == "chatbot" |
| 46 | |
| 47 | def test_developer_aliases_to_system(self): |
| 48 | # OpenAI's ``developer`` role is documented as high-priority |
| 49 | # instructions; map it onto the ``system`` slot rather than |
| 50 | # letting the templates drop it on the floor. |
| 51 | assert _role_to_melody("developer") == "system" |
| 52 | |
| 53 | @pytest.mark.parametrize("role", ["user", "system", "tool", "chatbot"]) |
| 54 | def test_recognized_roles_passthrough(self, role): |
| 55 | assert _role_to_melody(role) == role |
| 56 | |
| 57 | @pytest.mark.parametrize( |
| 58 | "role,expected", |
| 59 | [ |
| 60 | ("ASSISTANT", "chatbot"), |
| 61 | ("Developer", "system"), |
| 62 | ("User", "user"), |
| 63 | ("SYSTEM", "system"), |
| 64 | ], |
| 65 | ) |
| 66 | def test_role_normalization_is_case_insensitive(self, role, expected): |
| 67 | # cmd3 / cmd4 templates lowercase the role before matching, so |
| 68 | # accept any casing the caller provides. |
| 69 | assert _role_to_melody(role) == expected |
| 70 | |
| 71 | @pytest.mark.parametrize("role", ["function", "moderator", "", "anything"]) |
| 72 | def test_unknown_roles_raise(self, role): |
| 73 | # Silently dropping unknown roles produces malformed prompts |
| 74 | # (the templates' role chain has no else branch). |
| 75 | with pytest.raises(ValueError, match="Unsupported message role"): |
| 76 | _role_to_melody(role) |
| 77 | |
| 78 | def test_non_string_role_rejected(self): |
| 79 | # The function is typed ``role: str`` and the implementation |
| 80 | # relies on Python's attribute lookup (``role.lower()``) to |
| 81 | # reject non-strings — any exception type is acceptable as long |
| 82 | # as we don't silently produce a malformed prompt. |
| 83 | with pytest.raises((AttributeError, TypeError, ValueError)): |
| 84 | _role_to_melody(None) |
| 85 | |
| 86 | |
| 87 | # ====================================================================== |
| 88 | # _normalize_tool_call |
| 89 | # ====================================================================== |
| 90 | |
| 91 | |
| 92 | class TestNormalizeToolCall: |
| 93 | def test_openai_dict_with_dict_arguments_json_encoded(self): |
| 94 | # melody expects ``parameters`` as a JSON-encoded string even when |
| 95 | # OpenAI delivers an already-parsed dict. |
| 96 | out = _normalize_tool_call( |
| 97 | { |
| 98 | "id": "c1", |
| 99 | "type": "function", |
| 100 | "function": {"name": "f", "arguments": {"a": 1}}, |
| 101 | } |
| 102 | ) |
| 103 | assert out == {"id": "c1", "name": "f", "parameters": '{"a": 1}'} |
| 104 | |
| 105 | def test_openai_dict_with_string_arguments_preserved(self): |
| 106 | out = _normalize_tool_call( |
| 107 | { |
| 108 | "id": "c1", |
| 109 | "type": "function", |
| 110 | "function": {"name": "f", "arguments": '{"a":1}'}, |
| 111 | } |
| 112 | ) |
| 113 | assert out["parameters"] == '{"a":1}' |
| 114 | |
| 115 | def test_flat_dict_without_function_wrapper(self): |
| 116 | out = _normalize_tool_call({"id": "c1", "name": "f", "arguments": '{"k": 1}'}) |
| 117 | # Falls back to top-level ``name`` / ``arguments``. |
| 118 | assert out == {"id": "c1", "name": "f", "parameters": '{"k": 1}'} |
| 119 | |
| 120 | def test_missing_id_becomes_empty_string(self): |
| 121 | out = _normalize_tool_call({"function": {"name": "f", "arguments": "{}"}}) |
| 122 | assert out["id"] == "" |
| 123 | |
| 124 | def test_pydantic_model_dump_supported(self): |
| 125 | class _Fake: |
| 126 | def model_dump(self): |
| 127 | return { |
| 128 | "id": "c1", |
| 129 | "type": "function", |
| 130 | "function": {"name": "f", "arguments": "{}"}, |
| 131 | } |
| 132 | |
| 133 | out = _normalize_tool_call(_Fake()) |
| 134 | assert out == {"id": "c1", "name": "f", "parameters": "{}"} |
| 135 | |
| 136 | def test_invalid_type_rejected(self): |
| 137 | with pytest.raises(TypeError, match="Unexpected tool_call value"): |
| 138 | _normalize_tool_call(42) |
| 139 | |
| 140 | |
| 141 | # ====================================================================== |
| 142 | # _content_blocks |
| 143 | # ====================================================================== |
| 144 | |
| 145 | |
| 146 | class TestContentBlocks: |
| 147 | def test_none_returns_empty_list(self): |
| 148 | assert _content_blocks(None) == [] |
| 149 | |
| 150 | def test_string_wrapped_in_text_block(self): |
| 151 | out = _content_blocks("hi") |
| 152 | assert out == [{"type": MelodyContentType.TEXT, "text": "hi"}] |
| 153 | |
| 154 | def test_string_item_in_list_wrapped(self): |
| 155 | out = _content_blocks(["a", "b"]) |
| 156 | assert out == [ |
| 157 | {"type": MelodyContentType.TEXT, "text": "a"}, |
| 158 | {"type": MelodyContentType.TEXT, "text": "b"}, |
| 159 | ] |
| 160 | |
| 161 | @pytest.mark.parametrize( |
| 162 | "part_type", |
| 163 | ["text", "input_text", "output_text", "refusal"], |
| 164 | ) |
| 165 | def test_text_variants_normalized(self, part_type): |
| 166 | out = _content_blocks([{"type": part_type, "text": "hello"}]) |
| 167 | assert out == [{"type": MelodyContentType.TEXT, "text": "hello"}] |
| 168 | |
| 169 | def test_thinking_block(self): |
| 170 | out = _content_blocks([{"type": "thinking", "thinking": "thoughts"}]) |
| 171 | assert out == [{"type": MelodyContentType.THINKING, "thinking": "thoughts"}] |
| 172 | |
| 173 | def test_image_block_with_default_placeholder(self): |
| 174 | out = _content_blocks([{"type": "image"}]) |
| 175 | assert out == [ |
| 176 | { |
| 177 | "type": MelodyContentType.IMAGE, |
| 178 | "image": {"template_placeholder": "<image>"}, |
| 179 | } |
| 180 | ] |
| 181 | |
| 182 | def test_image_block_custom_placeholder(self): |
| 183 | out = _content_blocks([{"type": "image", "template_placeholder": "[[IMG]]"}]) |
| 184 | assert out[0]["image"]["template_placeholder"] == "[[IMG]]" |
| 185 | |
| 186 | def test_document_block_dict_passthrough(self): |
| 187 | out = _content_blocks( |
| 188 | [{"type": "document", "document": {"data": {"text": "doc"}}}] |
| 189 | ) |
| 190 | assert out == [ |
| 191 | { |
| 192 | "type": MelodyContentType.DOCUMENT, |
| 193 | "document": {"data": {"text": "doc"}}, |
| 194 | } |
| 195 | ] |
| 196 | |
| 197 | def test_document_block_with_non_dict_falls_back_to_json_text(self): |
| 198 | out = _content_blocks([{"type": "document", "document": "raw string doc"}]) |
| 199 | assert out[0]["type"] == MelodyContentType.TEXT |
| 200 | # JSON-encoded for safety since melody expects a structured doc. |
| 201 | assert out[0]["text"] == json.dumps("raw string doc") |
| 202 | |
| 203 | def test_tool_reference_emitted_as_text(self): |
| 204 | out = _content_blocks([{"type": "tool_reference", "name": "calc"}]) |
| 205 | assert out == [{"type": MelodyContentType.TEXT, "text": "calc"}] |
| 206 | |
| 207 | def test_unknown_block_type_fallback_to_text(self): |
| 208 | # Unknown block type with a string value is wrapped in a text block. |
| 209 | out = _content_blocks([{"type": "custom", "custom": "value"}]) |
| 210 | assert out == [{"type": MelodyContentType.TEXT, "text": "value"}] |
| 211 | |
| 212 | def test_unknown_block_type_dict_value_json_encoded(self): |
| 213 | out = _content_blocks([{"type": "custom", "custom": {"k": 1}}]) |
| 214 | assert out == [{"type": MelodyContentType.TEXT, "text": json.dumps({"k": 1})}] |
| 215 | |
| 216 | def test_non_string_non_dict_part_rejected(self): |
| 217 | with pytest.raises(TypeError, match="Unexpected content part"): |
| 218 | _content_blocks([42]) # type: ignore[list-item] |
| 219 | |
| 220 | |
| 221 | # ====================================================================== |
| 222 | # _document_to_melody |
| 223 | # ====================================================================== |
| 224 | |
| 225 | |
| 226 | class TestDocumentToMelody: |
| 227 | def test_string_wrapped_in_text_dict(self): |
| 228 | assert _document_to_melody("hello") == {"text": "hello"} |
| 229 | |
| 230 | def test_pure_dict_passthrough(self): |
| 231 | inp = {"text": "x", "id": "d1"} |
| 232 | out = _document_to_melody(inp) |
| 233 | assert out == {"text": "x", "id": "d1"} |
| 234 | # Must be a defensive copy so caller-side mutations of the |
| 235 | # returned dict don't leak back into the input. |
| 236 | out["new_key"] = "value" |
| 237 | assert "new_key" not in inp |
| 238 | |
| 239 | def test_data_wrapper_flattened(self): |
| 240 | # Cohere v2 documents use ``{id, data: {...}}``; melody expects |
| 241 | # the flat shape with ``id`` merged into the payload. |
| 242 | out = _document_to_melody({"id": "d1", "data": {"text": "hello", "title": "t"}}) |
| 243 | assert out == {"id": "d1", "text": "hello", "title": "t"} |
| 244 | |
| 245 | def test_data_wrapper_preserves_inner_id(self): |
| 246 | # If the inner ``data`` already has an ``id``, it wins. |
| 247 | out = _document_to_melody({"id": "outer", "data": {"id": "inner", "text": "x"}}) |
| 248 | assert out["id"] == "inner" |
| 249 | |
| 250 | def test_invalid_type_rejected(self): |
| 251 | with pytest.raises(TypeError, match="Unsupported document type"): |
| 252 | _document_to_melody(42) |
| 253 | |
| 254 | |
| 255 | # ====================================================================== |
| 256 | # _tool_to_melody |
| 257 | # ====================================================================== |
| 258 | |
| 259 | |
| 260 | class TestToolToMelody: |
| 261 | def test_openai_wrapper(self): |
| 262 | out = _tool_to_melody( |
| 263 | { |
| 264 | "type": "function", |
| 265 | "function": { |
| 266 | "name": "calc", |
| 267 | "description": "calculate", |
| 268 | "parameters": {"type": "object"}, |
| 269 | }, |
| 270 | } |
| 271 | ) |
| 272 | assert out == { |
| 273 | "name": "calc", |
| 274 | "description": "calculate", |
| 275 | "parameters": {"type": "object"}, |
| 276 | } |
| 277 | |
| 278 | def test_flat_dict(self): |
| 279 | out = _tool_to_melody({"name": "calc", "description": "d", "parameters": {}}) |
| 280 | assert out["name"] == "calc" |
| 281 | assert out["parameters"] == {} |
| 282 | |
| 283 | def test_pydantic_like_model_dump(self): |
| 284 | class _Fake: |
| 285 | def model_dump(self): |
| 286 | return { |
| 287 | "type": "function", |
| 288 | "function": { |
| 289 | "name": "calc", |
| 290 | "description": "x", |
| 291 | "parameters": {}, |
| 292 | }, |
| 293 | } |
| 294 | |
| 295 | out = _tool_to_melody(_Fake()) |
| 296 | assert out["name"] == "calc" |
| 297 | |
| 298 | def test_missing_description_becomes_empty(self): |
| 299 | out = _tool_to_melody({"name": "calc"}) |
| 300 | assert out["description"] == "" |
| 301 | assert out["parameters"] == {} |
| 302 | |
| 303 | def test_invalid_type_rejected(self): |
| 304 | with pytest.raises(TypeError, match="Unsupported tool type"): |
| 305 | _tool_to_melody(42) |
| 306 | |
| 307 | |
| 308 | # ====================================================================== |
| 309 | # _v2_tool_content_to_melody_block |
| 310 | # ====================================================================== |
| 311 | |
| 312 | |
| 313 | class TestV2ToolContentToMelodyBlock: |
| 314 | """Pins the v2 → melody mapping applied to every tool-message |
| 315 | content block. The serving layer forwards the original v2 blocks |
| 316 | verbatim under :data:`TOOL_MESSAGE_V2_CONTENT_KEY`; the renderer |
| 317 | owns this mapping so cmd3 / cmd4 render each result field as its |
| 318 | own entry rather than a single stringified-JSON blob. |
| 319 | """ |
| 320 | |
| 321 | def test_document_block_unwraps_data_into_melody_document(self): |
| 322 | # ``document.data`` becomes the melody ``document`` payload |
| 323 | # directly. Keeping the wrapping ``data`` key would produce a |
| 324 | # redundant ``{"data": {...}}`` nesting at render time. |
| 325 | # ``document.id`` is dropped here on purpose: outbound |
| 326 | # citation binding tracks IDs through the separate |
| 327 | # ``POSITION_TO_SOURCE_KEY`` map. |
| 328 | result = _v2_tool_content_to_melody_block( |
| 329 | { |
| 330 | "type": "document", |
| 331 | "document": { |
| 332 | "id": "d1", |
| 333 | "data": {"result": "The second tallest mountain is K2."}, |
| 334 | }, |
| 335 | } |
| 336 | ) |
| 337 | assert result == { |
| 338 | "type": "document", |
| 339 | "document": {"result": "The second tallest mountain is K2."}, |
| 340 | } |
| 341 | |
| 342 | def test_document_block_with_missing_data_yields_empty_payload(self): |
| 343 | # A document block whose ``data`` isn't a mapping (e.g. |
| 344 | # missing entirely) still produces a document block so |
| 345 | # per-slot numbering stays aligned with the outbound citation |
| 346 | # walk in ``CohereServingChatV2._walk_citable_positions``. |
| 347 | result = _v2_tool_content_to_melody_block( |
| 348 | {"type": "document", "document": {"id": "d1"}} |
| 349 | ) |
| 350 | assert result == {"type": "document", "document": {}} |
| 351 | |
| 352 | def test_json_parseable_text_becomes_document_dict(self): |
| 353 | # A ``text`` block whose payload parses to a JSON object lands |
| 354 | # directly under ``document`` (no ``content`` wrapper) so |
| 355 | # cmd3/cmd4 render each field individually. |
| 356 | result = _v2_tool_content_to_melody_block( |
| 357 | {"type": "text", "text": '{"temperature": "20C"}'} |
| 358 | ) |
| 359 | assert result == {"type": "document", "document": {"temperature": "20C"}} |
| 360 | |
| 361 | def test_non_json_text_wrapped_under_content_key(self): |
| 362 | # Plain-text tool output gets wrapped under a ``content`` key |
| 363 | # so cmd3/cmd4 templates have a stable field to look up. |
| 364 | result = _v2_tool_content_to_melody_block( |
| 365 | {"type": "text", "text": "K2 is 8611m."} |
| 366 | ) |
| 367 | assert result == {"type": "document", "document": {"content": "K2 is 8611m."}} |
| 368 | |
| 369 | def test_json_text_that_is_not_a_dict_is_treated_as_plain_text(self): |
| 370 | # A ``text`` block whose payload parses as a JSON *array* or |
| 371 | # *scalar* isn't a valid melody document (which must be a |
| 372 | # mapping). Fall back to the ``{"content": <text>}`` wrapper so |
| 373 | # the raw text still reaches the prompt. |
| 374 | result = _v2_tool_content_to_melody_block({"type": "text", "text": "[1, 2, 3]"}) |
| 375 | assert result == {"type": "document", "document": {"content": "[1, 2, 3]"}} |
| 376 | |
| 377 | def test_empty_text_still_produces_content_block(self): |
| 378 | # Even an empty text block must produce an entry so melody's |
| 379 | # per-content-block numbering stays aligned with the |
| 380 | # citation-side walk in ``_walk_citable_positions``. |
| 381 | result = _v2_tool_content_to_melody_block({"type": "text", "text": ""}) |
| 382 | assert result == {"type": "document", "document": {"content": ""}} |
| 383 | |
| 384 | def test_missing_text_field_treated_as_empty(self): |
| 385 | # ``model_dump(exclude_none=True)`` on a ``TextToolContent`` |
| 386 | # with an explicitly empty text still emits ``text: ""``, but |
| 387 | # guard against a future v2 shape that omits it entirely. |
| 388 | result = _v2_tool_content_to_melody_block({"type": "text"}) |
| 389 | assert result == {"type": "document", "document": {"content": ""}} |
| 390 | |
| 391 | def test_unknown_block_type_returns_none(self): |
| 392 | # Callers use ``None`` to know they should skip the slot; a |
| 393 | # future v2 addition shouldn't silently render as an empty |
| 394 | # document. |
| 395 | assert _v2_tool_content_to_melody_block({"type": "audio"}) is None |
| 396 | assert _v2_tool_content_to_melody_block({"not": "a block"}) is None |
| 397 | |
| 398 | def test_document_block_with_non_dict_document_returns_none(self): |
| 399 | # Malformed ``document`` field -- e.g. someone forwarding a |
| 400 | # raw string rather than the ``{id, data}`` wrapper -- is |
| 401 | # unusable, so return ``None`` so the caller skips it. |
| 402 | assert ( |
| 403 | _v2_tool_content_to_melody_block( |
| 404 | {"type": "document", "document": "not a dict"} |
| 405 | ) |
| 406 | is None |
| 407 | ) |
| 408 | |
| 409 | |
| 410 | # ====================================================================== |
| 411 | # _conversation_to_melody_messages |
| 412 | # ====================================================================== |
| 413 | |
| 414 | |
| 415 | class TestConversationToMelody: |
| 416 | def test_basic_user_assistant_pair(self): |
| 417 | conv = [ |
| 418 | {"role": "user", "content": "hi"}, |
| 419 | {"role": "assistant", "content": "hello"}, |
| 420 | ] |
| 421 | out = _conversation_to_melody_messages(conv) |
| 422 | assert out == [ |
| 423 | { |
| 424 | "role": "user", |
| 425 | "content": [{"type": MelodyContentType.TEXT, "text": "hi"}], |
| 426 | "tool_calls": [], |
| 427 | }, |
| 428 | { |
| 429 | "role": "chatbot", |
| 430 | "content": [{"type": MelodyContentType.TEXT, "text": "hello"}], |
| 431 | "tool_calls": [], |
| 432 | }, |
| 433 | ] |
| 434 | |
| 435 | def test_assistant_reasoning_prepended_as_thinking_block(self): |
| 436 | # ``reasoning`` (or ``reasoning_content``) is prepended as a |
| 437 | # ``thinking`` block on assistant turns, preserving multi-turn |
| 438 | # chain-of-thought across the rendered prompt. |
| 439 | conv = [ |
| 440 | { |
| 441 | "role": "assistant", |
| 442 | "content": "answer", |
| 443 | "reasoning": "thoughts", |
| 444 | } |
| 445 | ] |
| 446 | out = _conversation_to_melody_messages(conv) |
| 447 | assert out[0]["content"] == [ |
| 448 | {"type": MelodyContentType.THINKING, "thinking": "thoughts"}, |
| 449 | {"type": MelodyContentType.TEXT, "text": "answer"}, |
| 450 | ] |
| 451 | |
| 452 | def test_assistant_reasoning_content_alias_accepted(self): |
| 453 | conv = [ |
| 454 | { |
| 455 | "role": "assistant", |
| 456 | "content": "answer", |
| 457 | "reasoning_content": "thoughts", |
| 458 | } |
| 459 | ] |
| 460 | out = _conversation_to_melody_messages(conv) |
| 461 | assert out[0]["content"][0] == { |
| 462 | "type": MelodyContentType.THINKING, |
| 463 | "thinking": "thoughts", |
| 464 | } |
| 465 | |
| 466 | def test_user_reasoning_ignored(self): |
| 467 | # Only assistant turns get reasoning-as-thinking lifting; user |
| 468 | # turns with a ``reasoning`` key (which shouldn't happen in |
| 469 | # practice) must not produce a phantom thinking block. |
| 470 | conv = [ |
| 471 | { |
| 472 | "role": "user", |
| 473 | "content": "hi", |
| 474 | "reasoning": "should be ignored", |
| 475 | } |
| 476 | ] |
| 477 | out = _conversation_to_melody_messages(conv) |
| 478 | assert out[0]["content"] == [{"type": MelodyContentType.TEXT, "text": "hi"}] |
| 479 | |
| 480 | def test_tool_calls_normalized(self): |
| 481 | conv = [ |
| 482 | { |
| 483 | "role": "assistant", |
| 484 | "content": "", |
| 485 | "tool_calls": [ |
| 486 | { |
| 487 | "id": "c1", |
| 488 | "type": "function", |
| 489 | "function": {"name": "f", "arguments": '{"a":1}'}, |
| 490 | } |
| 491 | ], |
| 492 | } |
| 493 | ] |
| 494 | out = _conversation_to_melody_messages(conv) |
| 495 | assert out[0]["tool_calls"] == [ |
| 496 | {"id": "c1", "name": "f", "parameters": '{"a":1}'} |
| 497 | ] |
| 498 | |
| 499 | def test_tool_call_id_preserved_on_tool_role(self): |
| 500 | conv = [ |
| 501 | { |
| 502 | "role": "tool", |
| 503 | "content": "result", |
| 504 | "tool_call_id": "c1", |
| 505 | } |
| 506 | ] |
| 507 | out = _conversation_to_melody_messages(conv) |
| 508 | assert out[0]["tool_call_id"] == "c1" |
| 509 | |
| 510 | def test_messages_citations_attached_by_index(self): |
| 511 | # ``messages_citations`` is a dict keyed by message index. Only |
| 512 | # the message at the matching index should receive the |
| 513 | # citations; other messages must be unaffected. |
| 514 | conv = [ |
| 515 | {"role": "user", "content": "q"}, |
| 516 | {"role": "assistant", "content": "a"}, |
| 517 | ] |
| 518 | citations = { |
| 519 | 1: [ |
| 520 | { |
| 521 | "start_index": 0, |
| 522 | "end_index": 1, |
| 523 | "text": "a", |
| 524 | "sources": [ |
| 525 | { |
| 526 | "tool_call_index": 0, |
| 527 | "tool_result_indices": [0], |
| 528 | } |
| 529 | ], |
| 530 | "is_thinking": False, |
| 531 | } |
| 532 | ] |
| 533 | } |
| 534 | out = _conversation_to_melody_messages(conv, citations) |
| 535 | assert "citations" not in out[0] |
| 536 | assert out[1]["citations"] == citations[1] |
| 537 | |
| 538 | def test_messages_citations_none_is_a_no_op(self): |
| 539 | conv = [{"role": "assistant", "content": "a"}] |
| 540 | out = _conversation_to_melody_messages(conv, None) |
| 541 | assert "citations" not in out[0] |
| 542 | |
| 543 | def test_messages_citations_missing_index_is_a_no_op(self): |
| 544 | # A ``messages_citations`` dict whose key doesn't hit any |
| 545 | # message must not attach anything (and must not raise). |
| 546 | conv = [{"role": "assistant", "content": "a"}] |
| 547 | out = _conversation_to_melody_messages(conv, {5: [{"anything": 1}]}) |
| 548 | assert "citations" not in out[0] |
| 549 | |
| 550 | def test_tool_message_v2_content_replaces_default_content(self): |
| 551 | # When ``tool_message_v2_content`` has an entry for a message |
| 552 | # index, the renderer maps those v2 blocks through |
| 553 | # :func:`_v2_tool_content_to_melody_block` and uses the result |
| 554 | # instead of the block list ``_content_blocks`` would derive |
| 555 | # from ``msg["content"]``. |
| 556 | conv = [ |
| 557 | {"role": "user", "content": "q"}, |
| 558 | { |
| 559 | "role": "tool", |
| 560 | "tool_call_id": "c1", |
| 561 | # ``_convert_tool_message`` on the serving side flattens |
| 562 | # the original ``document`` block to this text part. |
| 563 | # The renderer should ignore this flattened content |
| 564 | # because the forwarded map carries the structural |
| 565 | # truth. |
| 566 | "content": '[{"type":"text","text":"{\\"result\\":\\"K2\\"}"}]', |
| 567 | }, |
| 568 | ] |
| 569 | v2_content = { |
| 570 | 1: [ |
| 571 | { |
| 572 | "type": "document", |
| 573 | "document": {"id": "d1", "data": {"result": "K2"}}, |
| 574 | } |
| 575 | ] |
| 576 | } |
| 577 | out = _conversation_to_melody_messages( |
| 578 | conv, |
| 579 | tool_message_v2_content=v2_content, |
| 580 | ) |
| 581 | # User message (no forwarded entry) still goes through the |
| 582 | # default ``_content_blocks`` path. |
| 583 | assert out[0]["content"] == [{"type": "text", "text": "q"}] |
| 584 | # Tool message picks up the mapped melody document block. |
| 585 | assert out[1]["content"] == [{"type": "document", "document": {"result": "K2"}}] |
| 586 | assert out[1]["tool_call_id"] == "c1" |
| 587 | |
| 588 | def test_tool_message_v2_content_skips_unrecognized_blocks(self): |
| 589 | # ``_v2_tool_content_to_melody_block`` returns ``None`` for |
| 590 | # shapes it doesn't recognise; the renderer must skip those |
| 591 | # rather than let a ``None`` reach melody. |
| 592 | conv = [{"role": "tool", "tool_call_id": "c", "content": "ignored"}] |
| 593 | v2_content = { |
| 594 | 0: [ |
| 595 | {"type": "text", "text": "keep me"}, |
| 596 | {"type": "mystery", "unexpected": True}, |
| 597 | ] |
| 598 | } |
| 599 | out = _conversation_to_melody_messages( |
| 600 | conv, |
| 601 | tool_message_v2_content=v2_content, |
| 602 | ) |
| 603 | assert out[0]["content"] == [ |
| 604 | {"type": "document", "document": {"content": "keep me"}} |
| 605 | ] |
| 606 | |
| 607 | def test_tool_message_v2_content_none_falls_back_to_default(self): |
| 608 | conv = [{"role": "tool", "tool_call_id": "c", "content": "plain text"}] |
| 609 | out = _conversation_to_melody_messages( |
| 610 | conv, |
| 611 | tool_message_v2_content=None, |
| 612 | ) |
| 613 | assert out[0]["content"] == [{"type": "text", "text": "plain text"}] |
| 614 | |
| 615 | |
| 616 | # ====================================================================== |
| 617 | # _build_render_config |
| 618 | # ====================================================================== |
| 619 | |
| 620 | |
| 621 | class TestBuildRenderConfig: |
| 622 | def _conv(self): |
| 623 | return [{"role": "user", "content": "hi"}] |
| 624 | |
| 625 | def test_default_format_is_cmd4(self): |
| 626 | # Bare kwargs -> cmd4 (the current Command A+ prompt format). |
| 627 | # Mirrors ``_DEFAULT_FORMAT`` in ``vllm/renderers/cohere.py`` and |
| 628 | # the ``--cohere-format`` CLI default. |
| 629 | fmt, cfg = _build_render_config(self._conv(), {}) |
| 630 | assert fmt == "cmd4" |
| 631 | assert cfg["use_jinja"] is True |
| 632 | assert isinstance(cfg["messages"], list) |
| 633 | # No additional_template_fields when no extra kwargs are set. |
| 634 | assert "additional_template_fields" not in cfg |
| 635 | |
| 636 | def test_explicit_cmd3(self): |
| 637 | fmt, cfg = _build_render_config(self._conv(), {"cohere_format": "cmd3"}) |
| 638 | assert fmt == "cmd3" |
| 639 | |
| 640 | def test_invalid_format_raises(self): |
| 641 | with pytest.raises(ValueError, match="Invalid cohere_format"): |
| 642 | _build_render_config(self._conv(), {"cohere_format": "cmd5"}) |
| 643 | |
| 644 | def test_documents_converted(self): |
| 645 | _, cfg = _build_render_config( |
| 646 | self._conv(), |
| 647 | { |
| 648 | "documents": [ |
| 649 | "doc text", |
| 650 | {"id": "d1", "data": {"text": "wrapped"}}, |
| 651 | ] |
| 652 | }, |
| 653 | ) |
| 654 | assert cfg["documents"] == [ |
| 655 | {"text": "doc text"}, |
| 656 | {"id": "d1", "text": "wrapped"}, |
| 657 | ] |
| 658 | |
| 659 | def test_available_tools_take_precedence_over_tools(self): |
| 660 | _, cfg = _build_render_config( |
| 661 | self._conv(), |
| 662 | { |
| 663 | "tools": [{"type": "function", "function": {"name": "from_tools"}}], |
| 664 | "available_tools": [ |
| 665 | {"type": "function", "function": {"name": "preferred"}} |
| 666 | ], |
| 667 | }, |
| 668 | ) |
| 669 | names = [t["name"] for t in cfg["available_tools"]] |
| 670 | assert names == ["preferred"] |
| 671 | |
| 672 | def test_tools_used_when_no_available_tools(self): |
| 673 | _, cfg = _build_render_config( |
| 674 | self._conv(), |
| 675 | {"tools": [{"type": "function", "function": {"name": "from_tools"}}]}, |
| 676 | ) |
| 677 | assert [t["name"] for t in cfg["available_tools"]] == ["from_tools"] |
| 678 | |
| 679 | @pytest.mark.parametrize("value", ["enabled", "disabled"]) |
| 680 | def test_reasoning_type_direct(self, value): |
| 681 | _, cfg = _build_render_config(self._conv(), {"reasoning_type": value}) |
| 682 | assert cfg["reasoning_type"] == value |
| 683 | |
| 684 | def test_thinking_dict_shorthand_resolves_reasoning_type(self): |
| 685 | _, cfg = _build_render_config(self._conv(), {"thinking": {"type": "enabled"}}) |
| 686 | assert cfg["reasoning_type"] == "enabled" |
| 687 | |
| 688 | def test_thinking_shorthand_ignores_unknown_type(self): |
| 689 | _, cfg = _build_render_config(self._conv(), {"thinking": {"type": "auto"}}) |
| 690 | assert "reasoning_type" not in cfg |
| 691 | |
| 692 | def test_dev_instruction_forwarded(self): |
| 693 | _, cfg = _build_render_config(self._conv(), {"dev_instruction": "be brief"}) |
| 694 | assert cfg["dev_instruction"] == "be brief" |
| 695 | |
| 696 | def test_response_format_json_object_sets_json_mode(self): |
| 697 | _, cfg = _build_render_config( |
| 698 | self._conv(), {"response_format": {"type": "json_object"}} |
| 699 | ) |
| 700 | assert cfg["json_mode"] is True |
| 701 | assert "json_schema" not in cfg |
| 702 | |
| 703 | def test_response_format_json_schema_sets_json_schema(self): |
| 704 | schema = {"type": "object"} |
| 705 | _, cfg = _build_render_config( |
| 706 | self._conv(), |
| 707 | {"response_format": {"type": "json_schema", "schema": schema}}, |
| 708 | ) |
| 709 | # JSON-encoded for melody (string-only schema field). |
| 710 | assert cfg["json_schema"] == json.dumps(schema) |
| 711 | |
| 712 | def test_response_format_nested_json_schema_unwrapped(self): |
| 713 | # When the SDK shape is ``{type: json_schema, schema: {schema: |
| 714 | # {...}}}``, the inner ``schema`` value is used. |
| 715 | inner = {"type": "object"} |
| 716 | _, cfg = _build_render_config( |
| 717 | self._conv(), |
| 718 | { |
| 719 | "response_format": { |
| 720 | "type": "json_schema", |
| 721 | "schema": {"schema": inner}, |
| 722 | } |
| 723 | }, |
| 724 | ) |
| 725 | assert cfg["json_schema"] == json.dumps(inner) |
| 726 | |
| 727 | def test_json_schema_kwarg_direct(self): |
| 728 | # Caller can also pass ``json_schema`` directly, both as dict and |
| 729 | # as a pre-stringified value. |
| 730 | _, cfg = _build_render_config(self._conv(), {"json_schema": {"a": 1}}) |
| 731 | assert cfg["json_schema"] == '{"a": 1}' |
| 732 | _, cfg = _build_render_config( |
| 733 | self._conv(), {"json_schema": "raw-string-schema"} |
| 734 | ) |
| 735 | assert cfg["json_schema"] == "raw-string-schema" |
| 736 | |
| 737 | def test_json_mode_kwarg_overrides(self): |
| 738 | _, cfg = _build_render_config(self._conv(), {"json_mode": True}) |
| 739 | assert cfg["json_mode"] is True |
| 740 | |
| 741 | def test_cmd3_safety_mode_lowercased(self): |
| 742 | _, cfg = _build_render_config( |
| 743 | self._conv(), |
| 744 | {"cohere_format": "cmd3", "safety_mode": "CONTEXTUAL"}, |
| 745 | ) |
| 746 | assert cfg["safety_mode"] == "contextual" |
| 747 | |
| 748 | def test_cmd3_citation_quality_direct(self): |
| 749 | _, cfg = _build_render_config( |
| 750 | self._conv(), |
| 751 | {"cohere_format": "cmd3", "citation_quality": "ACCURATE"}, |
| 752 | ) |
| 753 | assert cfg["citation_quality"] == "accurate" |
| 754 | |
| 755 | def test_cmd3_citation_quality_derived_from_citation_options(self): |
| 756 | # When ``citation_quality`` is unset, ``citation_options.mode`` is |
| 757 | # collapsed to on/off so cmd3's binary toggle has a value. |
| 758 | _, cfg = _build_render_config( |
| 759 | self._conv(), |
| 760 | {"cohere_format": "cmd3", "citation_options": {"mode": "accurate"}}, |
| 761 | ) |
| 762 | assert cfg["citation_quality"] == "on" |
| 763 | |
| 764 | _, cfg = _build_render_config( |
| 765 | self._conv(), |
| 766 | {"cohere_format": "cmd3", "citation_options": {"mode": "off"}}, |
| 767 | ) |
| 768 | assert cfg["citation_quality"] == "off" |
| 769 | |
| 770 | def test_cmd3_skip_preamble_forwarded(self): |
| 771 | _, cfg = _build_render_config( |
| 772 | self._conv(), |
| 773 | {"cohere_format": "cmd3", "skip_preamble": True}, |
| 774 | ) |
| 775 | assert cfg["skip_preamble"] is True |
| 776 | |
| 777 | def test_cmd3_no_grounding_field(self): |
| 778 | # cmd3 should never emit a cmd4-only ``grounding`` field. |
| 779 | _, cfg = _build_render_config( |
| 780 | self._conv(), |
| 781 | {"cohere_format": "cmd3", "grounding": "fast"}, |
| 782 | ) |
| 783 | assert "grounding" not in cfg |
| 784 | |
| 785 | @pytest.mark.parametrize( |
| 786 | "raw,expected", |
| 787 | [ |
| 788 | ("FAST", "enabled"), |
| 789 | ("ACCURATE", "enabled"), |
| 790 | ("OFF", "disabled"), |
| 791 | ("enabled", "enabled"), |
| 792 | ("disabled", "disabled"), |
| 793 | ("unknown", "unknown"), |
| 794 | ], |
| 795 | ) |
| 796 | def test_cmd4_grounding_direct(self, raw, expected): |
| 797 | # melody's cmd4 only accepts ``unknown``/``enabled``/``disabled``, |
| 798 | # so the renderer normalizes any of the v2-facing values into |
| 799 | # that vocab. |
| 800 | _, cfg = _build_render_config( |
| 801 | self._conv(), |
| 802 | {"cohere_format": "cmd4", "grounding": raw}, |
| 803 | ) |
| 804 | assert cfg["grounding"] == expected |
| 805 | |
| 806 | @pytest.mark.parametrize( |
| 807 | "mode,expected", |
| 808 | [ |
| 809 | ("ACCURATE", "enabled"), |
| 810 | ("FAST", "enabled"), |
| 811 | ("OFF", "disabled"), |
| 812 | ], |
| 813 | ) |
| 814 | def test_cmd4_grounding_from_citation_options_mode(self, mode, expected): |
| 815 | _, cfg = _build_render_config( |
| 816 | self._conv(), |
| 817 | { |
| 818 | "cohere_format": "cmd4", |
| 819 | "citation_options": {"mode": mode}, |
| 820 | }, |
| 821 | ) |
| 822 | assert cfg["grounding"] == expected |
| 823 | |
| 824 | def test_cmd4_grounding_rejects_unknown_value(self): |
| 825 | with pytest.raises(ValueError, match="Unrecognized cmd4 grounding"): |
| 826 | _build_render_config( |
| 827 | self._conv(), |
| 828 | {"cohere_format": "cmd4", "grounding": "foobar"}, |
| 829 | ) |
| 830 | |
| 831 | def test_cmd4_platform_instruction(self): |
| 832 | _, cfg = _build_render_config( |
| 833 | self._conv(), |
| 834 | { |
| 835 | "cohere_format": "cmd4", |
| 836 | "platform_instruction": "do this", |
| 837 | }, |
| 838 | ) |
| 839 | assert cfg["platform_instruction"] == "do this" |
| 840 | |
| 841 | def test_cmd4_no_safety_mode_field(self): |
| 842 | # cmd4 should never carry cmd3-only ``safety_mode``/``citation_quality``. |
| 843 | _, cfg = _build_render_config( |
| 844 | self._conv(), |
| 845 | { |
| 846 | "cohere_format": "cmd4", |
| 847 | "safety_mode": "contextual", |
| 848 | "citation_quality": "on", |
| 849 | }, |
| 850 | ) |
| 851 | assert "safety_mode" not in cfg |
| 852 | assert "citation_quality" not in cfg |
| 853 | |
| 854 | def test_extra_kwargs_become_additional_template_fields(self): |
| 855 | # Anything not in the renderer's consumed-keys set is forwarded |
| 856 | # verbatim under ``additional_template_fields`` so jinja templates |
| 857 | # can resolve ``{{ var }}`` directly. |
| 858 | _, cfg = _build_render_config( |
| 859 | self._conv(), |
| 860 | { |
| 861 | "reasoning_effort": "low", |
| 862 | "my_var": "x", |
| 863 | "documents": ["doc"], # consumed, must NOT leak through |
| 864 | }, |
| 865 | ) |
| 866 | extras = cfg["additional_template_fields"] |
| 867 | assert extras == {"reasoning_effort": "low", "my_var": "x"} |
| 868 | # Sanity: the consumed key still produced its dedicated config slot. |
| 869 | assert cfg["documents"] == [{"text": "doc"}] |
| 870 | |
| 871 | def test_template_id_passthrough(self): |
| 872 | # ``template_id`` is a safe selector for one of melody's built-in |
| 873 | # template variants (not raw source) and is still accepted via |
| 874 | # ``chat_template_kwargs``. |
| 875 | _, cfg = _build_render_config( |
| 876 | self._conv(), |
| 877 | {"template_id": "tpl1"}, |
| 878 | ) |
| 879 | assert cfg["template_id"] == "tpl1" |
| 880 | # use_jinja is always True, regardless of caller input. |
| 881 | assert cfg["use_jinja"] is True |
| 882 | |
| 883 | @pytest.mark.parametrize("cohere_only_key", ["template_jinja", "template"]) |
| 884 | def test_cohere_only_template_kwargs_are_rejected(self, cohere_only_key): |
| 885 | # ``chat_template_kwargs.template_jinja`` / ``.template`` are |
| 886 | # accepted at Cohere's own API surface but not at vLLM's -- raw |
| 887 | # template source in vLLM must flow through the standard |
| 888 | # ``chat_template`` request field so the |
| 889 | # ``--trust-request-chat-template`` guard applies uniformly. |
| 890 | # Silently dropping these keys would hide client misconfiguration, |
| 891 | # so we reject them loudly instead. |
| 892 | with pytest.raises( |
| 893 | ValueError, |
| 894 | match=f"chat_template_kwargs.{cohere_only_key!r}", |
| 895 | ): |
| 896 | _build_render_config( |
| 897 | self._conv(), |
| 898 | {cohere_only_key: "raw {{ jinja }}"}, |
| 899 | ) |
| 900 | |
| 901 | @pytest.mark.parametrize("cohere_only_key", ["template_jinja", "template"]) |
| 902 | def test_cohere_only_template_kwargs_none_is_tolerated(self, cohere_only_key): |
| 903 | # An explicit ``None`` (e.g. from ``.model_dump(exclude_none=False)`` |
| 904 | # on an optional field) is treated as absent -- we neither raise |
| 905 | # nor let it fall through as a Jinja variable. |
| 906 | _, cfg = _build_render_config( |
| 907 | self._conv(), |
| 908 | {cohere_only_key: None}, |
| 909 | ) |
| 910 | assert cohere_only_key not in cfg |
| 911 | assert "additional_template_fields" not in cfg |
| 912 | |
| 913 | def test_chat_template_arg_populates_template_jinja(self): |
| 914 | # The standard vLLM ``chat_template`` request field is the sole |
| 915 | # supported channel for raw template source and is forwarded to |
| 916 | # melody under its ``template_jinja`` config key. |
| 917 | _, cfg = _build_render_config( |
| 918 | self._conv(), |
| 919 | {}, |
| 920 | "raw {{ jinja }}", |
| 921 | ) |
| 922 | assert cfg["template_jinja"] == "raw {{ jinja }}" |
| 923 | assert cfg["use_jinja"] is True |
| 924 | |
| 925 | def test_chat_template_arg_none_leaves_template_jinja_unset(self): |
| 926 | _, cfg = _build_render_config( |
| 927 | self._conv(), |
| 928 | {}, |
| 929 | None, |
| 930 | ) |
| 931 | assert "template_jinja" not in cfg |
| 932 | |
| 933 | def test_tool_message_v2_content_key_picked_up_from_kwargs(self): |
| 934 | # ``_build_render_config`` must pull |
| 935 | # ``TOOL_MESSAGE_V2_CONTENT_KEY`` out of |
| 936 | # ``chat_template_kwargs`` and hand it to |
| 937 | # ``_conversation_to_melody_messages`` so the renderer's |
| 938 | # v2 → melody mapping runs over the original v2 blocks rather |
| 939 | # than the flattened string content on the OpenAI-shape tool |
| 940 | # message. |
| 941 | from vllm.renderers.cohere import TOOL_MESSAGE_V2_CONTENT_KEY |
| 942 | |
| 943 | conv = [ |
| 944 | {"role": "user", "content": "q"}, |
| 945 | {"role": "tool", "tool_call_id": "c1", "content": '{"r":"k"}'}, |
| 946 | ] |
| 947 | kwargs = { |
| 948 | TOOL_MESSAGE_V2_CONTENT_KEY: { |
| 949 | 1: [ |
| 950 | { |
| 951 | "type": "document", |
| 952 | "document": {"id": "d1", "data": {"r": "k"}}, |
| 953 | } |
| 954 | ] |
| 955 | } |
| 956 | } |
| 957 | _, cfg = _build_render_config(conv, kwargs) |
| 958 | tool_msg = cfg["messages"][1] |
| 959 | assert tool_msg["content"] == [{"type": "document", "document": {"r": "k"}}] |
| 960 | # And the key must be *consumed* -- not surfaced as a stray |
| 961 | # Jinja variable (``_RENDERER_CONSUMED_KEYS`` covers this but |
| 962 | # pin it here so a removal from that set surfaces immediately). |
| 963 | assert TOOL_MESSAGE_V2_CONTENT_KEY not in cfg.get( |
| 964 | "additional_template_fields", {} |
| 965 | ) |
| 966 | |
| 967 | |
| 968 | # ====================================================================== |
| 969 | # End-to-end async rendering (mirrors ``test_mistral.py``) |
| 970 | # ====================================================================== |
| 971 | # |
| 972 | # Verifies that the synchronous melody bindings run on the renderer's |
| 973 | # thread pool so the asyncio event loop stays responsive under |
| 974 | # concurrent load. Mirrors |
| 975 | # ``test_async_mistral_tokenizer_does_not_block_event_loop`` so future |
| 976 | # regressions in either path are caught uniformly. |
| 977 | |
| 978 | |
| 979 | @dataclass |
| 980 | class _MockHFConfig: |
| 981 | model_type: str = "any" |
| 982 | |
| 983 | |
| 984 | @dataclass |
| 985 | class _MockModelConfig: |
| 986 | runner_type = "generate" |
| 987 | model: str = "cohere-test" |
| 988 | tokenizer: str = "cohere-test" |
| 989 | trust_remote_code: bool = False |
| 990 | max_model_len: int = 100 |
| 991 | tokenizer_revision = None |
| 992 | tokenizer_mode = "cohere" |
| 993 | hf_config = _MockHFConfig() |
| 994 | hf_text_config = _MockHFConfig() |
| 995 | encoder_config: dict[str, Any] | None = None |
| 996 | enable_prompt_embeds: bool = True |
| 997 | skip_tokenizer_init: bool = True |
| 998 | is_encoder_decoder: bool = False |
| 999 | is_multimodal_model: bool = False |
| 1000 | supports_multimodal_inputs: bool = False |
| 1001 | renderer_num_workers: int = 1 |
| 1002 | |
| 1003 | |
| 1004 | @dataclass |
| 1005 | class _MockParallelConfig: |
| 1006 | _api_process_rank: int = 0 |
| 1007 | |
| 1008 | |
| 1009 | @dataclass |
| 1010 | class _MockVllmConfig: |
| 1011 | model_config: _MockModelConfig |
| 1012 | parallel_config: _MockParallelConfig |
| 1013 | |
| 1014 | |
| 1015 | @pytest.mark.asyncio |
| 1016 | async def test_async_cohere_renderer_does_not_block_event_loop(): |
| 1017 | expected_prompt = "MOCK_RENDERED_PROMPT" |
| 1018 | |
| 1019 | def slow_render(*_a, **_kw): |
| 1020 | time.sleep(2) |
| 1021 | return expected_prompt |
| 1022 | |
| 1023 | mock_tokenizer = Mock(spec=HfTokenizer) |
| 1024 | renderer = CohereRenderer( |
| 1025 | _MockVllmConfig(_MockModelConfig(), _MockParallelConfig()), |
| 1026 | tokenizer=mock_tokenizer, |
| 1027 | ) |
| 1028 | |
| 1029 | # Replace the (already-imported) ``cohere_melody`` bindings with a |
| 1030 | # blocking mock. ``_render`` reads ``self._melody`` at call time, so |
| 1031 | # this works even though ``_render_async`` was bound at __init__. |
| 1032 | fake_melody = Mock() |
| 1033 | fake_melody.render_cmd3 = slow_render |
| 1034 | fake_melody.render_cmd4 = slow_render |
| 1035 | renderer._melody = fake_melody |
| 1036 | |
| 1037 | task = renderer.render_messages_async([], ChatParams()) |
| 1038 | |
| 1039 | # Ensure the event loop is not blocked while the (blocking) render |
| 1040 | # call is in flight on the thread pool. |
| 1041 | blocked_count = 0 |
| 1042 | for _ in range(20): # ~2 seconds at 0.1s slices |
| 1043 | start = time.perf_counter() |
| 1044 | await asyncio.sleep(0) |
| 1045 | elapsed = time.perf_counter() - start |
| 1046 | if elapsed >= 0.5: |
| 1047 | blocked_count += 1 |
| 1048 | await asyncio.sleep(0.1) |
| 1049 | |
| 1050 | _, prompt = await task |
| 1051 | assert prompt["prompt"] == expected_prompt, "Mocked blocking render was not called" |
| 1052 | assert blocked_count == 0, "Event loop blocked during rendering" |
| 1053 | |
| 1054 | |
| 1055 | # ====================================================================== |
| 1056 | # End-to-end: request-level citations -> rendered prompt markup |
| 1057 | # ====================================================================== |
| 1058 | # |
| 1059 | # Ties the whole pipeline together: a Cohere v2 request carrying an |
| 1060 | # assistant message with citations must produce a rendered prompt that |
| 1061 | # contains melody's inline ``<co>...</co: <id>>`` markup around the |
| 1062 | # cited span. Preserving this invariant is the whole point of the |
| 1063 | # ``_messages_citations`` chat_template_kwargs entry. |
| 1064 | |
| 1065 | |
| 1066 | class TestRequestCitationsReachRenderedPrompt: |
| 1067 | """End-to-end verification that assistant-message citations on the |
| 1068 | request survive the OpenAI-shape round-trip and land in the melody- |
| 1069 | rendered prompt as inline ``<co>...</co>`` markers. |
| 1070 | |
| 1071 | The chain covered here: |
| 1072 | |
| 1073 | CohereChatV2Request |
| 1074 | -> CohereServingChatV2._convert_v2_to_chat_completion |
| 1075 | -> ChatCompletionRequest.chat_template_kwargs["_messages_citations"] |
| 1076 | -> _build_render_config (reads the entry) |
| 1077 | -> cohere_melody.render_cmd4 (renders <co>...</co>) |
| 1078 | |
| 1079 | A regression that drops the citations or mangles the melody |
| 1080 | ``FilterCitation`` payload would remove the citation markers from |
| 1081 | the output, so a text-in / text-out assertion is enough. |
| 1082 | """ |
| 1083 | |
| 1084 | @staticmethod |
| 1085 | def _melody(): |
| 1086 | # Import locally so missing optional deps skip this test class |
| 1087 | # rather than failing collection. Two guards: |
| 1088 | # * ``cohere_melody`` -- the Rust binding this test drives. |
| 1089 | # * ``cohere`` -- required transitively by |
| 1090 | # ``vllm.entrypoints.cohere.{protocol,serving}`` (both |
| 1091 | # unconditionally ``from cohere.types import ...`` at module |
| 1092 | # scope) which every test in this class imports locally. |
| 1093 | pytest.importorskip("cohere") |
| 1094 | pytest.importorskip("cohere_melody") |
| 1095 | import cohere_melody |
| 1096 | |
| 1097 | return cohere_melody |
| 1098 | |
| 1099 | @staticmethod |
| 1100 | def _openai_msgs_to_conversation( |
| 1101 | openai_messages: list[dict[str, Any]], |
| 1102 | ) -> list[dict[str, Any]]: |
| 1103 | """Cheap stand-in for ``parse_chat_messages`` on text-only inputs. |
| 1104 | |
| 1105 | ``_convert_v2_to_chat_completion`` emits OpenAI-shape assistant |
| 1106 | dicts with ``content`` as a plain string. The real chat-utils |
| 1107 | pipeline calls ``parse_chat_messages`` (which requires a full |
| 1108 | model config / mm tracker), but for text-only content it's |
| 1109 | essentially a passthrough -- we normalize ``content`` to the |
| 1110 | list-of-parts shape ``_conversation_to_melody_messages`` expects |
| 1111 | and preserve every other key the renderer reads. |
| 1112 | """ |
| 1113 | conv: list[dict[str, Any]] = [] |
| 1114 | for m in openai_messages: |
| 1115 | entry: dict[str, Any] = dict(m) |
| 1116 | content = entry.get("content") |
| 1117 | if isinstance(content, str): |
| 1118 | entry["content"] = [{"type": "text", "text": content}] |
| 1119 | elif content is None: |
| 1120 | entry["content"] = [] |
| 1121 | conv.append(entry) |
| 1122 | return conv |
| 1123 | |
| 1124 | def test_document_citation_survives_to_prompt(self): |
| 1125 | # Prevent this test from being collected/run when the melody |
| 1126 | # extension isn't importable. |
| 1127 | melody = self._melody() |
| 1128 | |
| 1129 | from vllm.entrypoints.cohere.protocol import CohereChatV2Request |
| 1130 | from vllm.entrypoints.cohere.serving import CohereServingChatV2 |
| 1131 | from vllm.renderers.cohere import _build_render_config |
| 1132 | |
| 1133 | request = CohereChatV2Request( |
| 1134 | model="m", |
| 1135 | messages=[ |
| 1136 | {"role": "user", "content": "Who wrote Hamlet?"}, |
| 1137 | { |
| 1138 | "role": "assistant", |
| 1139 | "content": "Shakespeare wrote it around 1600.", |
| 1140 | "citations": [ |
| 1141 | { |
| 1142 | "start": 0, |
| 1143 | "end": 11, # "Shakespeare" |
| 1144 | "text": "Shakespeare", |
| 1145 | "sources": [{"type": "document", "id": "doc_shakespeare"}], |
| 1146 | "type": "TEXT_CONTENT", |
| 1147 | } |
| 1148 | ], |
| 1149 | }, |
| 1150 | {"role": "user", "content": "and what year exactly?"}, |
| 1151 | ], |
| 1152 | documents=[ |
| 1153 | { |
| 1154 | "id": "doc_shakespeare", |
| 1155 | "data": {"text": "Hamlet was written by Shakespeare c. 1600."}, |
| 1156 | } |
| 1157 | ], |
| 1158 | ) |
| 1159 | |
| 1160 | # Step 1: v2 -> ChatCompletionRequest. Citations must land in |
| 1161 | # ``chat_template_kwargs["_messages_citations"]``. |
| 1162 | chat_req = CohereServingChatV2._convert_v2_to_chat_completion(request) |
| 1163 | assert chat_req.chat_template_kwargs is not None |
| 1164 | assert "_messages_citations" in chat_req.chat_template_kwargs |
| 1165 | |
| 1166 | # Step 2: build the melody render config. The renderer helper |
| 1167 | # folds the per-message citations onto the melody message |
| 1168 | # dicts. |
| 1169 | conversation = self._openai_msgs_to_conversation(chat_req.messages) |
| 1170 | fmt, config = _build_render_config(conversation, chat_req.chat_template_kwargs) |
| 1171 | |
| 1172 | assistant_msg = config["messages"][1] |
| 1173 | assert assistant_msg["role"] == "chatbot" |
| 1174 | assert "citations" in assistant_msg, ( |
| 1175 | "citations were not attached to the melody assistant message dict" |
| 1176 | ) |
| 1177 | |
| 1178 | # Step 3: hand the config to melody and check the rendered |
| 1179 | # prompt actually contains inline citation markup around the |
| 1180 | # cited span. |
| 1181 | # |
| 1182 | # The exact id is deterministic for this input. Melody builds |
| 1183 | # ``</co: <tool_call_index>:[<tool_result_indices>]>`` where |
| 1184 | # ``tool_call_index=0`` is the reserved bucket for the top-level |
| 1185 | # ``documents`` array and ``tool_result_indices`` are positions |
| 1186 | # inside it (see ``PromptRenderIds`` in melody/src/templating/ |
| 1187 | # util.rs). ``doc_shakespeare`` sits at position 0 in the |
| 1188 | # request's ``documents`` list, so we expect ``0:[0]``. Two |
| 1189 | # historical regressions this pins: |
| 1190 | # * ``0:[]`` -- the source id was never resolved to an index |
| 1191 | # (documents didn't flow through) and melody had nothing to |
| 1192 | # anchor the marker on. |
| 1193 | # * ``1:[0]`` -- the citation was routed through the wrong |
| 1194 | # tool-call bucket while documents were present, so it |
| 1195 | # pointed at the wrong prompt slot. |
| 1196 | # |
| 1197 | # Note the same rendered prompt also contains an example |
| 1198 | # ``<co>span</co: 0:[1,2],1:[0]>`` marker baked into melody's |
| 1199 | # system-prompt boilerplate (placeholder text ``"span"``); the |
| 1200 | # substring below is specific enough to only match the marker |
| 1201 | # around the cited text. |
| 1202 | if fmt == "cmd4": |
| 1203 | rendered = melody.render_cmd4(config) |
| 1204 | else: |
| 1205 | rendered = melody.render_cmd3(config) |
| 1206 | |
| 1207 | assert "<co>Shakespeare</co: 0:[0]>" in rendered, ( |
| 1208 | f"expected inline citation markup around the cited span; " |
| 1209 | f"tail of rendered prompt: {rendered[-400:]!r}" |
| 1210 | ) |
| 1211 | |
| 1212 | # And the cited document's text itself must be in the prompt -- |
| 1213 | # otherwise the model would have no way to satisfy the citation. |
| 1214 | assert "Hamlet" in rendered |
| 1215 | |
| 1216 | def test_no_markup_when_no_citations(self): |
| 1217 | # Control: the same request shape without any citations must |
| 1218 | # NOT contain ``<co>`` anywhere in the rendered prompt. Guards |
| 1219 | # against a false-positive where melody injects citation |
| 1220 | # markers regardless of what we passed in. |
| 1221 | melody = self._melody() |
| 1222 | |
| 1223 | from vllm.entrypoints.cohere.protocol import CohereChatV2Request |
| 1224 | from vllm.entrypoints.cohere.serving import CohereServingChatV2 |
| 1225 | from vllm.renderers.cohere import _build_render_config |
| 1226 | |
| 1227 | request = CohereChatV2Request( |
| 1228 | model="m", |
| 1229 | messages=[ |
| 1230 | {"role": "user", "content": "Who wrote Hamlet?"}, |
| 1231 | { |
| 1232 | "role": "assistant", |
| 1233 | "content": "Shakespeare wrote it around 1600.", |
| 1234 | }, |
| 1235 | ], |
| 1236 | ) |
| 1237 | |
| 1238 | chat_req = CohereServingChatV2._convert_v2_to_chat_completion(request) |
| 1239 | assert (chat_req.chat_template_kwargs or {}).get("_messages_citations") is None |
| 1240 | |
| 1241 | conversation = self._openai_msgs_to_conversation(chat_req.messages) |
| 1242 | fmt, config = _build_render_config( |
| 1243 | conversation, chat_req.chat_template_kwargs or {} |
| 1244 | ) |
| 1245 | |
| 1246 | if fmt == "cmd4": |
| 1247 | rendered = melody.render_cmd4(config) |
| 1248 | else: |
| 1249 | rendered = melody.render_cmd3(config) |
| 1250 | |
| 1251 | assert "<co>" not in rendered |
| 1252 | assert "</co:" not in rendered |