| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from unittest.mock import patch |
| 5 | |
| 6 | import pytest |
| 7 | from openai_harmony import Author, Message, Role, TextContent |
| 8 | |
| 9 | from vllm.entrypoints.openai.responses.context import ( |
| 10 | HarmonyContext, |
| 11 | SimpleContext, |
| 12 | TurnMetrics, |
| 13 | ) |
| 14 | from vllm.outputs import CompletionOutput, RequestOutput |
| 15 | from vllm.parser.harmony import ChunkResult, HarmonyParser, Segment |
| 16 | |
| 17 | |
| 18 | def create_mock_request_output( |
| 19 | prompt_token_ids=None, |
| 20 | output_token_ids=None, |
| 21 | num_cached_tokens=0, |
| 22 | finished=True, |
| 23 | ): |
| 24 | """Helper function to create a mock RequestOutput object for testing.""" |
| 25 | outputs = [] |
| 26 | token_ids = output_token_ids if output_token_ids is not None else [] |
| 27 | outputs = [ |
| 28 | CompletionOutput( |
| 29 | index=0, |
| 30 | text="Test output", |
| 31 | token_ids=token_ids, |
| 32 | cumulative_logprob=0.0, |
| 33 | logprobs=None, |
| 34 | finish_reason=None, |
| 35 | stop_reason=None, |
| 36 | ) |
| 37 | ] |
| 38 | |
| 39 | return RequestOutput( |
| 40 | request_id="test-id", |
| 41 | prompt="Test prompt", |
| 42 | prompt_token_ids=prompt_token_ids, |
| 43 | prompt_logprobs=None, |
| 44 | outputs=outputs, |
| 45 | finished=finished, |
| 46 | num_cached_tokens=num_cached_tokens, |
| 47 | ) |
| 48 | |
| 49 | |
| 50 | async def generate_mock_outputs( |
| 51 | num_turns, prompt_token_counts, output_token_counts, cached_token_counts=None |
| 52 | ): |
| 53 | """Generate a sequence of mock RequestOutput objects to simulate multiple |
| 54 | turns.""" |
| 55 | if cached_token_counts is None: |
| 56 | cached_token_counts = [0] * num_turns |
| 57 | |
| 58 | for i in range(num_turns): |
| 59 | # Create mock prompt token IDs and output token IDs |
| 60 | prompt_token_ids = list(range(1, prompt_token_counts[i] + 1)) |
| 61 | output_token_ids = list(range(1, output_token_counts[i] + 1)) |
| 62 | |
| 63 | # Create and yield the RequestOutput |
| 64 | yield create_mock_request_output( |
| 65 | prompt_token_ids=prompt_token_ids, |
| 66 | output_token_ids=output_token_ids, |
| 67 | num_cached_tokens=cached_token_counts[i], |
| 68 | ) |
| 69 | |
| 70 | |
| 71 | class FakeHarmonyParser(HarmonyParser): |
| 72 | def __init__(self): |
| 73 | # Skip HarmonyParser initialization and script outputs directly. |
| 74 | self._reasoning_parser = None |
| 75 | self._tool_parser = None |
| 76 | self._chunk_results: list[ChunkResult] = [] |
| 77 | self._flush_results: list[list[Segment]] = [] |
| 78 | self.processed_chunks: list[list[int]] = [] |
| 79 | |
| 80 | def enqueue_chunk_result( |
| 81 | self, |
| 82 | segments: list[Segment] | None = None, |
| 83 | reasoning_token_count: int = 0, |
| 84 | ) -> None: |
| 85 | self._chunk_results.append( |
| 86 | ChunkResult( |
| 87 | segments=[] if segments is None else segments, |
| 88 | reasoning_token_count=reasoning_token_count, |
| 89 | ) |
| 90 | ) |
| 91 | |
| 92 | def enqueue_flush_result(self, segment: list[Segment]) -> None: |
| 93 | self._flush_results.append(segment) |
| 94 | |
| 95 | def process_chunk(self, token_ids) -> ChunkResult: |
| 96 | self.processed_chunks.append(list(token_ids)) |
| 97 | if self._chunk_results: |
| 98 | return self._chunk_results.pop(0) |
| 99 | return ChunkResult(segments=[], reasoning_token_count=0) |
| 100 | |
| 101 | def flush(self) -> list[Segment]: |
| 102 | if self._flush_results: |
| 103 | return self._flush_results.pop(0) |
| 104 | return [] |
| 105 | |
| 106 | |
| 107 | def make_harmony_context( |
| 108 | messages=None, available_tools=None, function_tool_names=None |
| 109 | ) -> tuple[HarmonyContext, FakeHarmonyParser]: |
| 110 | fake_parser = FakeHarmonyParser() |
| 111 | context = HarmonyContext( |
| 112 | messages=[] if messages is None else messages, |
| 113 | available_tools=[] if available_tools is None else available_tools, |
| 114 | function_tool_names=function_tool_names, |
| 115 | response_parser=fake_parser, |
| 116 | ) |
| 117 | return context, fake_parser |
| 118 | |
| 119 | |
| 120 | def test_single_turn_token_counting(): |
| 121 | """Test token counting behavior for a single turn.""" |
| 122 | # Create a context |
| 123 | context, _ = make_harmony_context() |
| 124 | |
| 125 | # Create a mock RequestOutput with specific token counts |
| 126 | mock_output = create_mock_request_output( |
| 127 | prompt_token_ids=[1, 2, 3, 4, 5], # 5 prompt tokens |
| 128 | output_token_ids=[6, 7, 8], # 3 output tokens |
| 129 | num_cached_tokens=2, # 2 cached tokens |
| 130 | ) |
| 131 | |
| 132 | # Append the output to the context |
| 133 | context.append_output(mock_output) |
| 134 | |
| 135 | # Verify the token counts |
| 136 | assert context.num_prompt_tokens == 5 |
| 137 | assert context.num_output_tokens == 3 |
| 138 | assert context.num_cached_tokens == 2 |
| 139 | assert context.num_tool_output_tokens == 0 # No tool tokens in first turn |
| 140 | |
| 141 | # Verify internal state tracking |
| 142 | assert not context.is_first_turn |
| 143 | assert len(context.all_turn_metrics) == 1 |
| 144 | previous_turn = context.all_turn_metrics[0] |
| 145 | assert previous_turn.input_tokens == 5 |
| 146 | assert previous_turn.output_tokens == 3 |
| 147 | assert previous_turn.cached_input_tokens == 2 |
| 148 | assert previous_turn.tool_output_tokens == 0 |
| 149 | |
| 150 | |
| 151 | @pytest.mark.asyncio |
| 152 | async def test_multi_turn_token_counting(): |
| 153 | """Test token counting behavior across multiple turns with tool output.""" |
| 154 | # Create a context |
| 155 | context, _ = make_harmony_context(available_tools=["browser"]) |
| 156 | |
| 157 | # Simulate a conversation with 3 turns |
| 158 | # Turn 1: prefill 5, decode 3, tool 7 |
| 159 | # Turn 2: prefill 15, cached 5, decode 4, tool 1 |
| 160 | # Turn 3: prefill 20, cached 15, decode 5 |
| 161 | prompt_token_counts = [5, 15, 20] |
| 162 | output_token_counts = [3, 4, 5] |
| 163 | cached_token_counts = [0, 5, 15] |
| 164 | mock_generator = generate_mock_outputs( |
| 165 | 3, prompt_token_counts, output_token_counts, cached_token_counts |
| 166 | ) |
| 167 | |
| 168 | # First turn - initial prompt and response |
| 169 | mock_output1 = await anext(mock_generator) |
| 170 | context.append_output(mock_output1) |
| 171 | |
| 172 | # At this point, we should have 5 prompt tokens and 3 output tokens |
| 173 | assert context.num_prompt_tokens == 5 |
| 174 | assert context.num_output_tokens == 3 |
| 175 | assert context.num_tool_output_tokens == 0 |
| 176 | |
| 177 | # Second turn - after tool output |
| 178 | mock_output2 = await anext(mock_generator) |
| 179 | context.append_output(mock_output2) |
| 180 | # Current prompt tokens (15) - last_turn_input_tokens (5) - |
| 181 | # last_turn_output_tokens (3) = 7 |
| 182 | expected_tool_output = 7 |
| 183 | |
| 184 | assert context.num_prompt_tokens == 5 + 15 |
| 185 | assert context.num_output_tokens == 3 + 4 |
| 186 | assert context.num_tool_output_tokens == expected_tool_output |
| 187 | assert context.num_cached_tokens == 5 |
| 188 | |
| 189 | # Third turn - final response |
| 190 | mock_output3 = await anext(mock_generator) |
| 191 | context.append_output(mock_output3) |
| 192 | # Additional tool output tokens from third turn: |
| 193 | # Current prompt (20) - last_turn_input_tokens (15) - |
| 194 | # last_turn_output_tokens (4) = 1 |
| 195 | expected_tool_output = 7 + 1 |
| 196 | |
| 197 | assert context.num_prompt_tokens == 5 + 15 + 20 |
| 198 | assert context.num_output_tokens == 3 + 4 + 5 |
| 199 | assert context.num_tool_output_tokens == expected_tool_output |
| 200 | assert context.num_cached_tokens == 5 + 15 |
| 201 | |
| 202 | # Validate all turn metrics |
| 203 | assert len(context.all_turn_metrics) == 3 |
| 204 | for i, turn in enumerate(context.all_turn_metrics): |
| 205 | assert turn.input_tokens == prompt_token_counts[i] |
| 206 | assert turn.output_tokens == output_token_counts[i] |
| 207 | assert turn.cached_input_tokens == cached_token_counts[i] |
| 208 | assert context.all_turn_metrics[1].tool_output_tokens == 7 |
| 209 | assert context.all_turn_metrics[2].tool_output_tokens == 1 |
| 210 | |
| 211 | |
| 212 | def test_empty_output_tokens(): |
| 213 | """Test behavior when RequestOutput has empty output tokens.""" |
| 214 | context, _ = make_harmony_context() |
| 215 | |
| 216 | # Create a RequestOutput with empty output tokens |
| 217 | mock_output = create_mock_request_output( |
| 218 | prompt_token_ids=[1, 2, 3], # 3 prompt tokens |
| 219 | output_token_ids=[], # Empty output tokens list |
| 220 | num_cached_tokens=1, |
| 221 | ) |
| 222 | |
| 223 | context.append_output(mock_output) |
| 224 | |
| 225 | # Should handle empty outputs gracefully |
| 226 | assert context.num_prompt_tokens == 3 |
| 227 | assert context.num_output_tokens == 0 # No output tokens |
| 228 | assert context.num_cached_tokens == 1 |
| 229 | assert context.num_tool_output_tokens == 0 |
| 230 | |
| 231 | |
| 232 | def test_missing_prompt_token_ids(): |
| 233 | """Test behavior when RequestOutput has None prompt_token_ids.""" |
| 234 | context, _ = make_harmony_context() |
| 235 | |
| 236 | mock_output = create_mock_request_output( |
| 237 | prompt_token_ids=None, # No prompt token IDs |
| 238 | output_token_ids=[1, 2], # 2 output tokens |
| 239 | num_cached_tokens=0, |
| 240 | ) |
| 241 | |
| 242 | # Logger.error will be called, but we don't need to check for warnings |
| 243 | # here Just ensure it doesn't raise an exception |
| 244 | context.append_output(mock_output) |
| 245 | |
| 246 | # Should handle missing prompt tokens gracefully |
| 247 | assert context.num_prompt_tokens == 0 |
| 248 | assert context.num_output_tokens == 2 |
| 249 | assert context.num_cached_tokens == 0 |
| 250 | assert context.num_tool_output_tokens == 0 |
| 251 | |
| 252 | |
| 253 | def test_reasoning_tokens_counting(): |
| 254 | """Test that reasoning tokens are counted correctly.""" |
| 255 | context, parser = make_harmony_context() |
| 256 | parser.enqueue_chunk_result(reasoning_token_count=4) |
| 257 | |
| 258 | mock_output = create_mock_request_output( |
| 259 | prompt_token_ids=[1, 2, 3], |
| 260 | output_token_ids=[4, 5, 6, 7], # 4 tokens, all in reasoning |
| 261 | num_cached_tokens=0, |
| 262 | ) |
| 263 | |
| 264 | context.append_output(mock_output) |
| 265 | |
| 266 | # All output tokens should be counted as reasoning |
| 267 | assert context.num_reasoning_tokens == 4 |
| 268 | assert context.num_output_tokens == 4 |
| 269 | |
| 270 | |
| 271 | def test_preamble_tokens_not_counted_as_reasoning(): |
| 272 | """Preambles (commentary with no recipient) are visible user text, |
| 273 | not hidden reasoning. They must NOT inflate num_reasoning_tokens.""" |
| 274 | context, parser = make_harmony_context() |
| 275 | parser.enqueue_chunk_result(reasoning_token_count=0) |
| 276 | |
| 277 | mock_output = create_mock_request_output( |
| 278 | prompt_token_ids=[1, 2, 3], |
| 279 | output_token_ids=[4, 5, 6], |
| 280 | num_cached_tokens=0, |
| 281 | ) |
| 282 | context.append_output(mock_output) |
| 283 | |
| 284 | assert context.num_reasoning_tokens == 0 |
| 285 | assert context.num_output_tokens == 3 |
| 286 | |
| 287 | |
| 288 | def test_commentary_with_recipient_counted_as_reasoning(): |
| 289 | """Commentary directed at a tool (recipient != None) is hidden from |
| 290 | the user, so it should still count as reasoning tokens.""" |
| 291 | context, parser = make_harmony_context() |
| 292 | parser.enqueue_chunk_result(reasoning_token_count=3) |
| 293 | |
| 294 | mock_output = create_mock_request_output( |
| 295 | prompt_token_ids=[1, 2, 3], |
| 296 | output_token_ids=[4, 5, 6], |
| 297 | num_cached_tokens=0, |
| 298 | ) |
| 299 | context.append_output(mock_output) |
| 300 | |
| 301 | assert context.num_reasoning_tokens == 3 |
| 302 | assert context.num_output_tokens == 3 |
| 303 | |
| 304 | |
| 305 | def test_zero_tokens_edge_case(): |
| 306 | """Test behavior with all zero token counts.""" |
| 307 | context, _ = make_harmony_context() |
| 308 | |
| 309 | # Create a request with empty lists (not None) for both prompt and |
| 310 | # output tokens |
| 311 | mock_output = create_mock_request_output( |
| 312 | prompt_token_ids=[], # Empty prompt tokens |
| 313 | output_token_ids=[], # Empty output tokens |
| 314 | num_cached_tokens=0, |
| 315 | ) |
| 316 | |
| 317 | context.append_output(mock_output) |
| 318 | |
| 319 | # All counts should be zero |
| 320 | assert context.num_prompt_tokens == 0 |
| 321 | assert context.num_output_tokens == 0 |
| 322 | assert context.num_cached_tokens == 0 |
| 323 | assert context.num_tool_output_tokens == 0 |
| 324 | assert context.num_reasoning_tokens == 0 |
| 325 | |
| 326 | |
| 327 | @pytest.mark.asyncio |
| 328 | async def test_single_turn_no_tool_output(): |
| 329 | """Test that first turn never generates tool output tokens.""" |
| 330 | context, _ = make_harmony_context(available_tools=["browser"]) |
| 331 | |
| 332 | # Even with large prompt in first turn, no tool tokens should be counted |
| 333 | mock_output = create_mock_request_output( |
| 334 | prompt_token_ids=list(range(100)), # 100 tokens |
| 335 | output_token_ids=[1, 2, 3], |
| 336 | num_cached_tokens=0, |
| 337 | ) |
| 338 | |
| 339 | context.append_output(mock_output) |
| 340 | |
| 341 | # First turn should never have tool output tokens |
| 342 | assert context.num_tool_output_tokens == 0 |
| 343 | assert context.is_first_turn is False # Should be updated after first turn |
| 344 | |
| 345 | |
| 346 | @pytest.mark.asyncio |
| 347 | async def test_negative_tool_tokens_edge_case(): |
| 348 | """Test edge case where calculation could result in negative tool |
| 349 | tokens. We should log an error and clamp the value to 0.""" |
| 350 | # Use patch to check if logger.error was called |
| 351 | with patch("vllm.entrypoints.openai.responses.context.logger.error") as mock_log: |
| 352 | context, _ = make_harmony_context(available_tools=["browser"]) |
| 353 | |
| 354 | # First turn |
| 355 | mock_output1 = create_mock_request_output( |
| 356 | prompt_token_ids=list(range(10)), # 10 tokens |
| 357 | output_token_ids=[1, 2, 3, 4, 5], # 5 tokens |
| 358 | ) |
| 359 | context.append_output(mock_output1) |
| 360 | |
| 361 | # Second turn with fewer new tokens than previous output |
| 362 | # This could happen in edge cases with aggressive caching |
| 363 | mock_output2 = create_mock_request_output( |
| 364 | prompt_token_ids=list(range(12)), # 12 tokens (only 2 new) |
| 365 | output_token_ids=[6, 7], # 2 tokens |
| 366 | ) |
| 367 | context.append_output(mock_output2) |
| 368 | |
| 369 | # Calculated negative tool tokens (12 - 10 - 5 = -3) should be clamped |
| 370 | # to 0 and an error should be logged |
| 371 | assert context.num_tool_output_tokens == 0 |
| 372 | assert context.num_prompt_tokens == 10 + 12 |
| 373 | assert context.num_output_tokens == 5 + 2 |
| 374 | |
| 375 | # Verify the error was logged properly |
| 376 | mock_log.assert_called_once() |
| 377 | |
| 378 | # Extract the actual log message and arguments from the call |
| 379 | args, _ = mock_log.call_args |
| 380 | log_message = args[0] |
| 381 | |
| 382 | # Check for key parts of the message |
| 383 | assert "Negative tool output tokens calculated" in log_message |
| 384 | assert "-3" in str(args) # Check that -3 is in the arguments |
| 385 | |
| 386 | |
| 387 | @pytest.mark.asyncio |
| 388 | async def test_streaming_multi_turn_token_counting(): |
| 389 | """Test token counting for streaming multi-turn conversations. |
| 390 | |
| 391 | This test focuses on how HarmonyContext counts tokens in a |
| 392 | multi-turn conversation with streaming (token-by-token) outputs and |
| 393 | message boundaries. |
| 394 | """ |
| 395 | # Create a streaming context |
| 396 | context, parser = make_harmony_context(available_tools=["browser"]) |
| 397 | |
| 398 | num_prompt_tokens = [3, 8, 13] |
| 399 | num_output_tokens = [3, 3, 2] |
| 400 | num_cached_tokens = [0, 3, 8] |
| 401 | |
| 402 | # Simulate three turns of conversation: |
| 403 | # Turn 1: stream tokens one by one, then finish the message |
| 404 | # Turn 2: new prompt, stream more tokens with a reasoning segment |
| 405 | # Turn 3: new prompt with tool output and cached tokens |
| 406 | |
| 407 | # First turn: 3 tokens streamed one by one |
| 408 | # First token of first turn |
| 409 | context.append_output( |
| 410 | create_mock_request_output( |
| 411 | prompt_token_ids=[1, 2, 3], # 3 prompt tokens |
| 412 | output_token_ids=[101], # Single token |
| 413 | num_cached_tokens=num_cached_tokens[0], |
| 414 | finished=False, # Not end of message yet |
| 415 | ) |
| 416 | ) |
| 417 | |
| 418 | # Second token of first turn |
| 419 | context.append_output( |
| 420 | create_mock_request_output( |
| 421 | output_token_ids=[102], |
| 422 | finished=False, |
| 423 | ) |
| 424 | ) |
| 425 | |
| 426 | # Last token of first turn (finished=True signals end of message) |
| 427 | context.append_output( |
| 428 | create_mock_request_output( |
| 429 | output_token_ids=[103], |
| 430 | finished=True, # End of message |
| 431 | ) |
| 432 | ) |
| 433 | |
| 434 | # Check token counts after first turn |
| 435 | assert context.num_prompt_tokens == 3 # Initial prompt tokens |
| 436 | assert context.num_output_tokens == 3 # Three output tokens |
| 437 | assert context.num_cached_tokens == 0 |
| 438 | assert context.num_tool_output_tokens == 0 # No tool output in first turn |
| 439 | assert context.first_tok_of_message is True # Ready for next message |
| 440 | |
| 441 | # First token of second turn |
| 442 | parser.enqueue_chunk_result(reasoning_token_count=1) |
| 443 | context.append_output( |
| 444 | create_mock_request_output( |
| 445 | prompt_token_ids=[ |
| 446 | 1, |
| 447 | 2, |
| 448 | 3, |
| 449 | 101, |
| 450 | 102, |
| 451 | 103, |
| 452 | 4, |
| 453 | 5, |
| 454 | ], # 8 tokens (includes previous) |
| 455 | output_token_ids=[201], |
| 456 | num_cached_tokens=num_cached_tokens[1], # Some tokens cached |
| 457 | finished=False, |
| 458 | ) |
| 459 | ) |
| 460 | |
| 461 | # More tokens in reasoning channel |
| 462 | parser.enqueue_chunk_result(reasoning_token_count=1) |
| 463 | context.append_output( |
| 464 | create_mock_request_output( |
| 465 | output_token_ids=[202], |
| 466 | finished=False, |
| 467 | ) |
| 468 | ) |
| 469 | |
| 470 | parser.enqueue_chunk_result(reasoning_token_count=1) |
| 471 | context.append_output( |
| 472 | create_mock_request_output( |
| 473 | output_token_ids=[203], |
| 474 | finished=True, # End of reasoning message |
| 475 | ) |
| 476 | ) |
| 477 | |
| 478 | # Check counts after second turn (reasoning message) |
| 479 | assert context.num_prompt_tokens == 3 + 8 # Initial + second prompt |
| 480 | assert context.num_output_tokens == 3 + 3 # First turn + second turn |
| 481 | assert context.num_reasoning_tokens == 3 # All tokens in analysis channel |
| 482 | assert context.num_cached_tokens == 3 # Cached tokens from second turn |
| 483 | |
| 484 | # Formula: this turn prompt tokens - last turn prompt - last turn output |
| 485 | expected_tool_tokens = 8 - 3 - 3 # = 2 |
| 486 | assert context.num_tool_output_tokens == expected_tool_tokens |
| 487 | |
| 488 | # Third turn (with more cached tokens) |
| 489 | context.append_output( |
| 490 | create_mock_request_output( |
| 491 | prompt_token_ids=[ |
| 492 | 1, |
| 493 | 2, |
| 494 | 3, |
| 495 | 101, |
| 496 | 102, |
| 497 | 103, |
| 498 | 4, |
| 499 | 5, |
| 500 | 201, |
| 501 | 202, |
| 502 | 203, |
| 503 | 6, |
| 504 | 7, |
| 505 | ], # 13 tokens |
| 506 | output_token_ids=[301], |
| 507 | num_cached_tokens=num_cached_tokens[2], # More cached tokens |
| 508 | finished=False, |
| 509 | ) |
| 510 | ) |
| 511 | |
| 512 | context.append_output( |
| 513 | create_mock_request_output( |
| 514 | output_token_ids=[302], |
| 515 | finished=True, |
| 516 | ) |
| 517 | ) |
| 518 | |
| 519 | # Final token counts check |
| 520 | assert context.num_prompt_tokens == sum(num_prompt_tokens) # All prompts |
| 521 | assert context.num_output_tokens == sum(num_output_tokens) # All outputs |
| 522 | assert context.num_reasoning_tokens == 3 # Unchanged from second turn |
| 523 | assert context.num_cached_tokens == sum( |
| 524 | num_cached_tokens |
| 525 | ) # Accumulated cached tokens |
| 526 | |
| 527 | # Additional tool tokens from third turn |
| 528 | # Formula: this turn prompt - last turn prompt - last turn output |
| 529 | additional_tool_tokens = 13 - 8 - 3 # = 2 |
| 530 | assert ( |
| 531 | context.num_tool_output_tokens == expected_tool_tokens + additional_tool_tokens |
| 532 | ) |
| 533 | |
| 534 | # Validate all turn metrics |
| 535 | assert len(context.all_turn_metrics) == 3 |
| 536 | for i, turn in enumerate(context.all_turn_metrics): |
| 537 | assert turn.input_tokens == num_prompt_tokens[i] |
| 538 | assert turn.output_tokens == num_output_tokens[i] |
| 539 | assert turn.cached_input_tokens == num_cached_tokens[i] |
| 540 | assert context.all_turn_metrics[1].tool_output_tokens == 2 |
| 541 | assert context.all_turn_metrics[2].tool_output_tokens == 2 |
| 542 | |
| 543 | |
| 544 | @pytest.mark.asyncio |
| 545 | async def test_streaming_message_synchronization(): |
| 546 | """Completed messages from append-local and flush segments sync into context.""" |
| 547 | # Create a streaming context with some initial messages |
| 548 | initial_messages = [ |
| 549 | Message( |
| 550 | author=Author(role=Role.USER, name="user"), |
| 551 | content=[TextContent(text="Hello")], |
| 552 | recipient=Role.ASSISTANT, |
| 553 | ) |
| 554 | ] |
| 555 | context, parser = make_harmony_context(messages=initial_messages) |
| 556 | |
| 557 | # Verify initial state |
| 558 | assert len(context._messages) == 1 |
| 559 | assert context.num_init_messages == 1 |
| 560 | |
| 561 | response_text = "First response" |
| 562 | message = Message( |
| 563 | author=Author(role=Role.ASSISTANT, name="assistant"), |
| 564 | content=[TextContent(text=response_text)], |
| 565 | recipient=Role.USER, |
| 566 | ) |
| 567 | parser.enqueue_chunk_result( |
| 568 | segments=[ |
| 569 | Segment( |
| 570 | channel="commentary", |
| 571 | recipient=None, |
| 572 | delta="", |
| 573 | completed_message=message, |
| 574 | ) |
| 575 | ] |
| 576 | ) |
| 577 | |
| 578 | # This should sync the completed message from the latest append |
| 579 | context.append_output( |
| 580 | create_mock_request_output( |
| 581 | prompt_token_ids=[1, 2, 3], output_token_ids=[101], finished=False |
| 582 | ) |
| 583 | ) |
| 584 | |
| 585 | # Verify that messages were synchronized |
| 586 | assert len(context._messages) == 2 |
| 587 | |
| 588 | # Verify the new messages were added correctly |
| 589 | assert context._messages[1].content[0].text == response_text |
| 590 | |
| 591 | messages_minus_init = len(context._messages) - context.num_init_messages |
| 592 | assert messages_minus_init == 1 |
| 593 | |
| 594 | response_text = "Second response" |
| 595 | message = Message( |
| 596 | author=Author(role=Role.ASSISTANT, name="assistant"), |
| 597 | content=[TextContent(text=response_text)], |
| 598 | recipient=Role.USER, |
| 599 | ) |
| 600 | flush_segments = [ |
| 601 | Segment( |
| 602 | channel="final", |
| 603 | recipient=None, |
| 604 | delta=response_text, |
| 605 | completed_message=None, |
| 606 | ), |
| 607 | Segment( |
| 608 | channel="final", |
| 609 | recipient=None, |
| 610 | delta="", |
| 611 | completed_message=message, |
| 612 | ), |
| 613 | ] |
| 614 | parser.enqueue_flush_result(flush_segments) |
| 615 | |
| 616 | # Create another output to trigger synchronization via flush() |
| 617 | context.append_output( |
| 618 | create_mock_request_output( |
| 619 | prompt_token_ids=[1, 2, 3], output_token_ids=[102], finished=True |
| 620 | ) |
| 621 | ) |
| 622 | |
| 623 | # Verify the flushed response was added, num_init_messages is still 1 |
| 624 | assert len(context._messages) == 3 |
| 625 | assert context.num_init_messages == 1 |
| 626 | assert context._messages[2].content[0].text == response_text |
| 627 | assert context.last_append_flush_status is True |
| 628 | assert len(context.last_append_segments) == 2 |
| 629 | assert context.last_append_segments[-2].delta == response_text |
| 630 | assert context.last_append_segments[-1].completed_message is message |
| 631 | |
| 632 | |
| 633 | def test_turn_metrics_copy_and_reset(): |
| 634 | """Test TurnMetrics copy and reset methods work correctly.""" |
| 635 | # Create a TurnMetrics with specific values |
| 636 | original_metrics = TurnMetrics( |
| 637 | input_tokens=10, |
| 638 | output_tokens=20, |
| 639 | cached_input_tokens=5, |
| 640 | tool_output_tokens=3, |
| 641 | ) |
| 642 | |
| 643 | # Test copy functionality |
| 644 | copied_metrics = original_metrics.copy() |
| 645 | |
| 646 | # Verify copy has same values |
| 647 | assert copied_metrics.input_tokens == 10 |
| 648 | assert copied_metrics.output_tokens == 20 |
| 649 | assert copied_metrics.cached_input_tokens == 5 |
| 650 | assert copied_metrics.tool_output_tokens == 3 |
| 651 | |
| 652 | # Verify they are separate objects |
| 653 | assert copied_metrics is not original_metrics |
| 654 | |
| 655 | # Modify copy to ensure independence |
| 656 | copied_metrics.input_tokens = 999 |
| 657 | assert original_metrics.input_tokens == 10 # Original unchanged |
| 658 | assert copied_metrics.input_tokens == 999 |
| 659 | |
| 660 | # Test reset functionality |
| 661 | original_metrics.reset() |
| 662 | |
| 663 | # Verify all fields are reset to zero |
| 664 | assert original_metrics.input_tokens == 0 |
| 665 | assert original_metrics.output_tokens == 0 |
| 666 | assert original_metrics.cached_input_tokens == 0 |
| 667 | assert original_metrics.tool_output_tokens == 0 |
| 668 | |
| 669 | # Verify copied metrics are unaffected by reset |
| 670 | assert copied_metrics.input_tokens == 999 |
| 671 | assert copied_metrics.output_tokens == 20 |
| 672 | assert copied_metrics.cached_input_tokens == 5 |
| 673 | assert copied_metrics.tool_output_tokens == 3 |
| 674 | |
| 675 | |
| 676 | # ==================== SimpleContext Tests ==================== |
| 677 | |
| 678 | |
| 679 | def create_simple_context_output( |
| 680 | text="", |
| 681 | token_ids=None, |
| 682 | prompt="Test prompt", |
| 683 | prompt_token_ids=None, |
| 684 | num_cached_tokens=0, |
| 685 | num_cache_creation_tokens=None, |
| 686 | logprobs=None, |
| 687 | finished=True, |
| 688 | ): |
| 689 | """Helper to create a RequestOutput with customizable text for |
| 690 | SimpleContext tests.""" |
| 691 | if token_ids is None: |
| 692 | token_ids = [] |
| 693 | return RequestOutput( |
| 694 | request_id="test-id", |
| 695 | prompt=prompt, |
| 696 | prompt_token_ids=prompt_token_ids, |
| 697 | prompt_logprobs=None, |
| 698 | outputs=[ |
| 699 | CompletionOutput( |
| 700 | index=0, |
| 701 | text=text, |
| 702 | token_ids=token_ids, |
| 703 | cumulative_logprob=0.0, |
| 704 | logprobs=logprobs, |
| 705 | finish_reason=None, |
| 706 | stop_reason=None, |
| 707 | ) |
| 708 | ], |
| 709 | finished=finished, |
| 710 | num_cached_tokens=num_cached_tokens, |
| 711 | num_cache_creation_tokens=num_cache_creation_tokens, |
| 712 | ) |
| 713 | |
| 714 | |
| 715 | def test_simple_context_output_messages_empty(): |
| 716 | """output_messages should be empty before any output is appended.""" |
| 717 | context = SimpleContext() |
| 718 | assert context.output_messages == [] |
| 719 | |
| 720 | |
| 721 | def test_simple_context_output_messages_single_call(): |
| 722 | """Non-streaming: single append_output produces a single output message.""" |
| 723 | context = SimpleContext() |
| 724 | output = create_simple_context_output( |
| 725 | text="Hello world", |
| 726 | token_ids=[10, 20, 30], |
| 727 | prompt_token_ids=[1, 2, 3], |
| 728 | ) |
| 729 | context.append_output(output) |
| 730 | |
| 731 | messages = context.output_messages |
| 732 | assert len(messages) == 1 |
| 733 | assert messages[0].message == "Hello world" |
| 734 | assert messages[0].tokens == [10, 20, 30] |
| 735 | assert messages[0].type == "raw_message_tokens" |
| 736 | |
| 737 | |
| 738 | def test_simple_context_output_messages_streaming_consolidation(): |
| 739 | """Streaming: multiple append_output calls consolidate into one message.""" |
| 740 | context = SimpleContext() |
| 741 | |
| 742 | # Simulate 3 streaming deltas |
| 743 | context.append_output( |
| 744 | create_simple_context_output( |
| 745 | text="Hello", |
| 746 | token_ids=[10], |
| 747 | prompt_token_ids=[1, 2, 3], |
| 748 | ) |
| 749 | ) |
| 750 | context.append_output( |
| 751 | create_simple_context_output( |
| 752 | text=" world", |
| 753 | token_ids=[20], |
| 754 | prompt_token_ids=[1, 2, 3], |
| 755 | ) |
| 756 | ) |
| 757 | context.append_output( |
| 758 | create_simple_context_output( |
| 759 | text="!", |
| 760 | token_ids=[30], |
| 761 | prompt_token_ids=[1, 2, 3], |
| 762 | ) |
| 763 | ) |
| 764 | |
| 765 | messages = context.output_messages |
| 766 | assert len(messages) == 1 |
| 767 | assert messages[0].message == "Hello world!" |
| 768 | assert messages[0].tokens == [10, 20, 30] |
| 769 | |
| 770 | |
| 771 | def test_simple_context_output_messages_many_deltas(): |
| 772 | """Streaming with many small deltas still produces a single message.""" |
| 773 | context = SimpleContext() |
| 774 | |
| 775 | words = ["The", " quick", " brown", " fox", " jumps"] |
| 776 | for i, word in enumerate(words): |
| 777 | context.append_output( |
| 778 | create_simple_context_output( |
| 779 | text=word, |
| 780 | token_ids=[100 + i], |
| 781 | prompt_token_ids=[1, 2], |
| 782 | ) |
| 783 | ) |
| 784 | |
| 785 | messages = context.output_messages |
| 786 | assert len(messages) == 1 |
| 787 | assert messages[0].message == "The quick brown fox jumps" |
| 788 | assert messages[0].tokens == [100, 101, 102, 103, 104] |
| 789 | |
| 790 | |
| 791 | def test_simple_context_input_messages(): |
| 792 | """input_messages is populated on the first append_output call.""" |
| 793 | context = SimpleContext() |
| 794 | assert context.input_messages == [] |
| 795 | |
| 796 | context.append_output( |
| 797 | create_simple_context_output( |
| 798 | text="Hi", |
| 799 | token_ids=[10], |
| 800 | prompt="My prompt text", |
| 801 | prompt_token_ids=[1, 2, 3], |
| 802 | ) |
| 803 | ) |
| 804 | |
| 805 | assert len(context.input_messages) == 1 |
| 806 | assert context.input_messages[0].message == "My prompt text" |
| 807 | assert context.input_messages[0].tokens == [1, 2, 3] |
| 808 | |
| 809 | # Second call should not add another input message |
| 810 | context.append_output( |
| 811 | create_simple_context_output( |
| 812 | text=" there", |
| 813 | token_ids=[20], |
| 814 | prompt="My prompt text", |
| 815 | prompt_token_ids=[1, 2, 3], |
| 816 | ) |
| 817 | ) |
| 818 | |
| 819 | assert len(context.input_messages) == 1 |
| 820 | |
| 821 | |
| 822 | def test_simple_context_token_counting(): |
| 823 | """Token counting accumulates across streaming deltas.""" |
| 824 | context = SimpleContext() |
| 825 | |
| 826 | context.append_output( |
| 827 | create_simple_context_output( |
| 828 | text="a", |
| 829 | token_ids=[10, 11], |
| 830 | prompt_token_ids=[1, 2, 3, 4, 5], |
| 831 | num_cached_tokens=2, |
| 832 | num_cache_creation_tokens=3, |
| 833 | ) |
| 834 | ) |
| 835 | context.append_output( |
| 836 | create_simple_context_output( |
| 837 | text="b", |
| 838 | token_ids=[12], |
| 839 | prompt_token_ids=[1, 2, 3, 4, 5], |
| 840 | num_cached_tokens=2, |
| 841 | num_cache_creation_tokens=3, |
| 842 | ) |
| 843 | ) |
| 844 | |
| 845 | assert context.num_prompt_tokens == 5 |
| 846 | assert context.num_output_tokens == 3 # 2 + 1 |
| 847 | assert context.num_cached_tokens == 2 |
| 848 | assert context.num_cache_creation_tokens == 3 |
| 849 | |
| 850 | |
| 851 | def test_simple_context_final_output(): |
| 852 | """final_output reconstructs accumulated text and token_ids.""" |
| 853 | context = SimpleContext() |
| 854 | |
| 855 | context.append_output( |
| 856 | create_simple_context_output( |
| 857 | text="foo", |
| 858 | token_ids=[1, 2], |
| 859 | prompt_token_ids=[10], |
| 860 | ) |
| 861 | ) |
| 862 | context.append_output( |
| 863 | create_simple_context_output( |
| 864 | text="bar", |
| 865 | token_ids=[3], |
| 866 | prompt_token_ids=[10], |
| 867 | ) |
| 868 | ) |
| 869 | |
| 870 | final = context.final_output |
| 871 | assert final is not None |
| 872 | assert final.outputs[0].text == "foobar" |
| 873 | assert final.outputs[0].token_ids == (1, 2, 3) |
| 874 | |
| 875 | |
| 876 | def test_simple_context_output_messages_empty_text_with_tokens(): |
| 877 | """output_messages should be returned when tokens exist even if text is |
| 878 | empty (e.g. special tokens).""" |
| 879 | context = SimpleContext() |
| 880 | context.append_output( |
| 881 | create_simple_context_output( |
| 882 | text="", |
| 883 | token_ids=[99], |
| 884 | prompt_token_ids=[1], |
| 885 | ) |
| 886 | ) |
| 887 | |
| 888 | messages = context.output_messages |
| 889 | assert len(messages) == 1 |
| 890 | assert messages[0].message == "" |
| 891 | assert messages[0].tokens == [99] |
| 892 | |
| 893 | |
| 894 | def test_simple_context_output_messages_no_mutation(): |
| 895 | """Each call to output_messages returns a fresh list; callers can't |
| 896 | corrupt internal state.""" |
| 897 | context = SimpleContext() |
| 898 | context.append_output( |
| 899 | create_simple_context_output( |
| 900 | text="hello", |
| 901 | token_ids=[1], |
| 902 | prompt_token_ids=[10], |
| 903 | ) |
| 904 | ) |
| 905 | |
| 906 | msgs1 = context.output_messages |
| 907 | msgs2 = context.output_messages |
| 908 | assert msgs1 is not msgs2 |
| 909 | assert msgs1[0].message == msgs2[0].message |
| 910 | |
| 911 | # Appending more output updates the property |
| 912 | context.append_output( |
| 913 | create_simple_context_output( |
| 914 | text=" world", |
| 915 | token_ids=[2], |
| 916 | prompt_token_ids=[10], |
| 917 | ) |
| 918 | ) |
| 919 | |
| 920 | msgs3 = context.output_messages |
| 921 | assert len(msgs3) == 1 |
| 922 | assert msgs3[0].message == "hello world" |
| 923 | assert msgs3[0].tokens == [1, 2] |