| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import contextlib |
| 4 | import os |
| 5 | import random |
| 6 | |
| 7 | import pytest |
| 8 | import torch |
| 9 | from transformers import AutoTokenizer |
| 10 | from utils import ( |
| 11 | BACKENDS, |
| 12 | TEST_MODEL, |
| 13 | _extract_step_logprobs, |
| 14 | _random_prompt, |
| 15 | skip_if_not_cuda, |
| 16 | skip_unsupported, |
| 17 | ) |
| 18 | |
| 19 | import vllm.envs as envs |
| 20 | from vllm import LLM, SamplingParams |
| 21 | |
| 22 | |
| 23 | @skip_unsupported |
| 24 | @pytest.mark.flaky(reruns=3) |
| 25 | @pytest.mark.timeout(1000) |
| 26 | @pytest.mark.parametrize( |
| 27 | "backend", |
| 28 | BACKENDS, |
| 29 | ) |
| 30 | @pytest.mark.parametrize("rms_norm_impl", ["default", "vllm_c"]) |
| 31 | def test_v1_generation_is_deterministic_across_batch_sizes_with_needle( |
| 32 | backend, |
| 33 | rms_norm_impl, |
| 34 | ): |
| 35 | """Ensures that the same request (the 'needle' prompt) yields identical output |
| 36 | whether run alone (bs=1) or mixed into a larger batch (e.g., bs=64), |
| 37 | using the high-level v1 LLM() API only (no manual batching). |
| 38 | |
| 39 | Strategy: |
| 40 | - Create a single LLM engine configured for the larger batch limit (N). |
| 41 | - Compute a baseline output for the needle prompt when it is run alone. |
| 42 | - For many trials, generate a mixed batch (size N) where the needle appears |
| 43 | at a random position among random filler prompts using the same engine. |
| 44 | - Track how many trials match vs mismatch, and report totals at the end. |
| 45 | The test fails if any mismatches occur, but we still dump pass/fail |
| 46 | counts. |
| 47 | |
| 48 | Notes: |
| 49 | - Use seeded stochastic sampling with a fixed seed to test determinism. |
| 50 | - Outputs are intentionally longer and sampled at higher temperature/top_p |
| 51 | to produce a more random-sounding phrase, yet remain deterministic by |
| 52 | seed. |
| 53 | - Keep max_tokens and max_model_len bounded for speed and memory use. |
| 54 | |
| 55 | """ |
| 56 | seed = int(os.getenv("VLLM_TEST_SEED", "12345")) |
| 57 | random.seed(seed) |
| 58 | |
| 59 | attention_config = {"backend": backend} |
| 60 | # Force the C++ RMSNorm implementation so we actually exercise the |
| 61 | # num_tokens-dependent block-size branches. |
| 62 | kernel_config = None |
| 63 | if rms_norm_impl == "vllm_c": |
| 64 | kernel_config = { |
| 65 | "ir_op_priority": { |
| 66 | "rms_norm": ["vllm_c"], |
| 67 | "fused_add_rms_norm": ["vllm_c"], |
| 68 | } |
| 69 | } |
| 70 | # Allow overrides from environment (useful for CI tuning) |
| 71 | # "facebook/opt-125m" is too small, doesn't reliably test determinism |
| 72 | model = TEST_MODEL |
| 73 | num_trials = int(os.getenv("VLLM_NEEDLE_TRIALS", "5")) |
| 74 | max_batch_size = int(os.getenv("VLLM_NEEDLE_BATCH_SIZE", "128")) |
| 75 | min_random_prompt = int(os.getenv("VLLM_MIN_PROMPT", "1024")) |
| 76 | max_random_prompt = int(os.getenv("VLLM_MAX_PROMPT", "2048")) |
| 77 | assert max_batch_size >= 2, "Batch size should be >= 2 to mix needle." |
| 78 | |
| 79 | # Keep GPU memory usage low to avoid startup allocation failures. |
| 80 | gpu_mem_util = float(os.getenv("VLLM_GPU_MEMORY_UTILIZATION", "0.5")) |
| 81 | max_model_len = int(os.getenv("VLLM_MAX_MODEL_LEN", "5120")) |
| 82 | |
| 83 | # Sampling parameters: longer outputs with a more random-sounding |
| 84 | # continuation,but still deterministic due to fixed seed. |
| 85 | temperature = float(os.getenv("VLLM_NEEDLE_TEMPERATURE", "0.0")) |
| 86 | top_p = float(os.getenv("VLLM_NEEDLE_TOP_P", "0.95")) |
| 87 | max_tokens = int(os.getenv("VLLM_NEEDLE_MAX_TOKENS", "128")) |
| 88 | |
| 89 | sampling = SamplingParams( |
| 90 | temperature=temperature, |
| 91 | top_p=top_p, |
| 92 | max_tokens=max_tokens, |
| 93 | seed=20240919, |
| 94 | ) |
| 95 | |
| 96 | needle_prompt = "There once was a " |
| 97 | |
| 98 | llm = None |
| 99 | try: |
| 100 | llm = LLM_with_max_seqs( |
| 101 | model=model, |
| 102 | max_num_seqs=max_batch_size, |
| 103 | gpu_memory_utilization=gpu_mem_util, |
| 104 | max_model_len=max_model_len, |
| 105 | attention_config=attention_config, |
| 106 | kernel_config=kernel_config, |
| 107 | ) |
| 108 | |
| 109 | # Baseline generation for the needle prompt alone. |
| 110 | baseline_out = llm.generate([needle_prompt], sampling) |
| 111 | assert len(baseline_out) == 1 |
| 112 | assert len(baseline_out[0].outputs) >= 1 |
| 113 | baseline_text = baseline_out[0].outputs[0].text |
| 114 | |
| 115 | mismatches = 0 |
| 116 | |
| 117 | for trial in range(num_trials): |
| 118 | # Create a batch of size `max_batch_size` and insert the needle at |
| 119 | # a random index |
| 120 | prompts: list[str] = [] |
| 121 | batch_size = random.randint(max_batch_size // 2, max_batch_size) |
| 122 | needle_pos = random.randint(0, batch_size - 1) |
| 123 | for i in range(batch_size): |
| 124 | if i == needle_pos: |
| 125 | prompts.append(needle_prompt) |
| 126 | else: |
| 127 | prompts.append(_random_prompt(min_random_prompt, max_random_prompt)) |
| 128 | |
| 129 | # Generate with the same engine but in a larger batch. |
| 130 | outputs = llm.generate(prompts, sampling) |
| 131 | # Find the needle output by position |
| 132 | needle_output = outputs[needle_pos] |
| 133 | assert needle_output.prompt == needle_prompt |
| 134 | assert len(needle_output.outputs) >= 1 |
| 135 | text = needle_output.outputs[0].text |
| 136 | |
| 137 | if text != baseline_text: |
| 138 | print(f"{text}\n\n== Not the same as ==\n\n{baseline_text}\n\n") |
| 139 | mismatches += 1 |
| 140 | |
| 141 | passes = num_trials - mismatches |
| 142 | # Dump how many passed vs failed |
| 143 | print( |
| 144 | f"[determinism] total={num_trials}, passed={passes}, " |
| 145 | f"failed={mismatches}, max_batch_size={max_batch_size}" |
| 146 | ) |
| 147 | |
| 148 | if mismatches > 0: |
| 149 | pytest.fail( |
| 150 | f"Nondeterministic outputs detected: {mismatches} failed out " |
| 151 | f"of {num_trials} trials (max_batch_size={max_batch_size})." |
| 152 | ) |
| 153 | |
| 154 | finally: |
| 155 | # Ensure engines are shutdown to free GPU/VRAM across test sessions |
| 156 | if llm is not None: |
| 157 | with contextlib.suppress(Exception): |
| 158 | llm.llm_engine.engine_core.shutdown() |
| 159 | |
| 160 | |
| 161 | @skip_unsupported |
| 162 | @pytest.mark.parametrize( |
| 163 | "backend", |
| 164 | BACKENDS, |
| 165 | ) |
| 166 | @pytest.mark.parametrize( |
| 167 | "block_m,block_n", |
| 168 | [(16, 16), (8, 16)], |
| 169 | ) |
| 170 | def test_logprobs_bitwise_batch_invariance_bs1_vs_bsN( |
| 171 | backend, |
| 172 | block_m, |
| 173 | block_n, |
| 174 | ): |
| 175 | seed = int(os.getenv("VLLM_TEST_SEED", "12345")) |
| 176 | random.seed(seed) |
| 177 | tp_size = int(os.getenv("VLLM_TEST_TP_SIZE", "1")) |
| 178 | |
| 179 | llm = LLM( |
| 180 | model=TEST_MODEL, |
| 181 | tensor_parallel_size=tp_size, |
| 182 | max_num_seqs=128, |
| 183 | max_model_len=8192, |
| 184 | dtype="auto", # not everything is supported |
| 185 | gpu_memory_utilization=0.9, |
| 186 | attention_config={ |
| 187 | "backend": backend, |
| 188 | "flex_attn_block_m": block_m, |
| 189 | "flex_attn_block_n": block_n, |
| 190 | }, |
| 191 | ) |
| 192 | |
| 193 | # Use more realistic prompts for better token generation |
| 194 | prompts = [_random_prompt(10, 50) for _ in range(32)] |
| 195 | |
| 196 | # Chunked prefill is covered by |
| 197 | # test_logprobs_bitwise_batch_invariance_ragged_chunked_prefill below. |
| 198 | |
| 199 | sp = SamplingParams( |
| 200 | temperature=0.6, |
| 201 | top_p=1.0, |
| 202 | max_tokens=16, |
| 203 | seed=1234, |
| 204 | logprobs=5, |
| 205 | ) |
| 206 | |
| 207 | # BS=1: run prompts individually and collect logprobs per step. |
| 208 | print("\n" + "=" * 80) |
| 209 | print("STARTING BS=1 RUNS (each prompt individually)") |
| 210 | print("=" * 80 + "\n") |
| 211 | |
| 212 | bs1_logprobs_per_prompt = [] |
| 213 | bs1_tokens_per_prompt = [] |
| 214 | for idx, p in enumerate(prompts): |
| 215 | print(f"\n[BS=1] Running prompt {idx}/{len(prompts)} - Preview: {p[:80]}...") |
| 216 | outs = llm.generate([p], sp, use_tqdm=False) |
| 217 | assert len(outs) == 1 |
| 218 | step_logprobs, token_ids = _extract_step_logprobs(outs[0]) |
| 219 | if step_logprobs is None: |
| 220 | pytest.skip( |
| 221 | "Logits are not available on RequestOutput; " |
| 222 | "enable logprobs return to run this test." |
| 223 | ) |
| 224 | bs1_logprobs_per_prompt.append(step_logprobs) |
| 225 | bs1_tokens_per_prompt.append(token_ids) |
| 226 | print(f"[BS=1] Prompt {idx} generated tokens: {token_ids}") |
| 227 | |
| 228 | # BS=N: run prompts in a batch and collect logprobs per step for each |
| 229 | # prompt. |
| 230 | print("\n" + "=" * 80) |
| 231 | print(f"STARTING BS={len(prompts)} RUN (all prompts batched)") |
| 232 | print("=" * 80 + "\n") |
| 233 | |
| 234 | outs_batched = llm.generate(prompts, sp, use_tqdm=False) |
| 235 | assert len(outs_batched) == len(prompts) |
| 236 | bsN_logprobs_per_prompt = [] |
| 237 | bsN_tokens_per_prompt = [] |
| 238 | |
| 239 | print(f"\n[BS={len(prompts)}] Processing batched outputs...") |
| 240 | for idx, o in enumerate(outs_batched): |
| 241 | tokens = o.outputs[0].token_ids if o.outputs else "N/A" |
| 242 | print(f"[BS={len(prompts)}] Prompt {idx} generated tokens: {tokens}") |
| 243 | step_logprobs, token_ids = _extract_step_logprobs(o) |
| 244 | if step_logprobs is None: |
| 245 | pytest.skip( |
| 246 | "Logits are not available on RequestOutput; " |
| 247 | "enable logprobs return to run this test." |
| 248 | ) |
| 249 | bsN_logprobs_per_prompt.append(step_logprobs) |
| 250 | bsN_tokens_per_prompt.append(token_ids) |
| 251 | |
| 252 | # Compare step-by-step logprobs for each prompt between BS=1 and BS=N runs. |
| 253 | failed_prompts = [] |
| 254 | for i, (logprobs_bs1, logprobs_bsN, tokens_bs1, tokens_bsN) in enumerate( |
| 255 | zip( |
| 256 | bs1_logprobs_per_prompt, |
| 257 | bsN_logprobs_per_prompt, |
| 258 | bs1_tokens_per_prompt, |
| 259 | bsN_tokens_per_prompt, |
| 260 | ) |
| 261 | ): |
| 262 | if len(logprobs_bs1) != len(logprobs_bsN): |
| 263 | reason = ( |
| 264 | f"Different number of steps: {len(logprobs_bs1)} (BS=1) " |
| 265 | f"vs {len(logprobs_bsN)} (BS=N)" |
| 266 | ) |
| 267 | failed_prompts.append( |
| 268 | { |
| 269 | "prompt_idx": i, |
| 270 | "step": "all", |
| 271 | "reason": reason, |
| 272 | "prompt_preview": prompts[i][:100], |
| 273 | "bs1_tokens": tokens_bs1, |
| 274 | "bsN_tokens": tokens_bsN, |
| 275 | } |
| 276 | ) |
| 277 | continue |
| 278 | |
| 279 | # Check if tokens match first |
| 280 | if tokens_bs1 != tokens_bsN: |
| 281 | failed_prompts.append( |
| 282 | { |
| 283 | "prompt_idx": i, |
| 284 | "step": "sampling", |
| 285 | "reason": "Different tokens sampled", |
| 286 | "prompt_preview": prompts[i][:100], |
| 287 | "bs1_tokens": tokens_bs1, |
| 288 | "bsN_tokens": tokens_bsN, |
| 289 | "bs1_all_logprobs": [ |
| 290 | logprobs_bs1[s].tolist() for s in range(len(logprobs_bs1)) |
| 291 | ], |
| 292 | "bsN_all_logprobs": [ |
| 293 | logprobs_bsN[s].tolist() for s in range(len(logprobs_bsN)) |
| 294 | ], |
| 295 | } |
| 296 | ) |
| 297 | continue |
| 298 | |
| 299 | for t, (a, b) in enumerate(zip(logprobs_bs1, logprobs_bsN)): |
| 300 | if a.shape != b.shape: |
| 301 | failed_prompts.append( |
| 302 | { |
| 303 | "prompt_idx": i, |
| 304 | "step": t, |
| 305 | "reason": f"Shape mismatch: {a.shape} vs {b.shape}", |
| 306 | "prompt_preview": prompts[i][:100], |
| 307 | "bs1_tokens": tokens_bs1, |
| 308 | "bsN_tokens": tokens_bsN, |
| 309 | } |
| 310 | ) |
| 311 | break |
| 312 | |
| 313 | if not torch.equal(a, b): |
| 314 | max_diff = torch.abs(a - b).max().item() |
| 315 | # Print which token failed |
| 316 | print(f"\n[DIVERGENCE] Prompt {i}, Token {t}: max_diff={max_diff:.6e}") |
| 317 | bs1_tok = tokens_bs1[t] if t < len(tokens_bs1) else "N/A" |
| 318 | bsN_tok = tokens_bsN[t] if t < len(tokens_bsN) else "N/A" |
| 319 | print(f" Token IDs: bs1={bs1_tok}, bsN={bsN_tok}") |
| 320 | print(f" BS=1 logprob: {a.tolist()}") |
| 321 | print(f" BS=N logprob: {b.tolist()}") |
| 322 | failed_prompts.append( |
| 323 | { |
| 324 | "prompt_idx": i, |
| 325 | "step": t, |
| 326 | "reason": f"Bitwise mismatch (max_diff={max_diff:.6e})", |
| 327 | "prompt_preview": prompts[i][:100], |
| 328 | "bs1_tokens": tokens_bs1, |
| 329 | "bsN_tokens": tokens_bsN, |
| 330 | "bs1_all_logprobs": [ |
| 331 | logprobs_bs1[s].tolist() for s in range(len(logprobs_bs1)) |
| 332 | ], |
| 333 | "bsN_all_logprobs": [ |
| 334 | logprobs_bsN[s].tolist() for s in range(len(logprobs_bsN)) |
| 335 | ], |
| 336 | } |
| 337 | ) |
| 338 | break |
| 339 | |
| 340 | # Print summary of all failures |
| 341 | if failed_prompts: |
| 342 | print(f"\n{'=' * 80}") |
| 343 | fail_msg = ( |
| 344 | f"BATCH INVARIANCE FAILURES: {len(failed_prompts)}/" |
| 345 | f"{len(prompts)} prompts failed" |
| 346 | ) |
| 347 | print(fail_msg) |
| 348 | print(f"{'=' * 80}") |
| 349 | for fail in failed_prompts: |
| 350 | print(f"\nPrompt {fail['prompt_idx']} (step {fail['step']}):") |
| 351 | print(f" Reason: {fail['reason']}") |
| 352 | print(f" Preview: {fail['prompt_preview']}...") |
| 353 | |
| 354 | # Always show the tokens |
| 355 | if "bs1_tokens" in fail: |
| 356 | print(f" BS=1 tokens: {fail['bs1_tokens']}") |
| 357 | if "bsN_tokens" in fail: |
| 358 | print(f" BS=N tokens: {fail['bsN_tokens']}") |
| 359 | |
| 360 | if "bs1_all_logprobs" in fail: |
| 361 | print(f" BS=1 logprobs for all {len(fail['bs1_all_logprobs'])} steps:") |
| 362 | for step_idx, logprobs in enumerate(fail["bs1_all_logprobs"]): |
| 363 | print(f" Step {step_idx}: {logprobs}") |
| 364 | print(f" BS=N logprobs for all {len(fail['bsN_all_logprobs'])} steps:") |
| 365 | for step_idx, logprobs in enumerate(fail["bsN_all_logprobs"]): |
| 366 | print(f" Step {step_idx}: {logprobs}") |
| 367 | print(f"{'=' * 80}\n") |
| 368 | |
| 369 | # Fail the test with summary |
| 370 | msg = ( |
| 371 | f"Batch invariance violated in {len(failed_prompts)}/" |
| 372 | f"{len(prompts)} prompts. See output above for details." |
| 373 | ) |
| 374 | pytest.fail(msg) |
| 375 | |
| 376 | |
| 377 | @skip_unsupported |
| 378 | @pytest.mark.parametrize( |
| 379 | "backend", |
| 380 | BACKENDS, |
| 381 | ) |
| 382 | def test_logprobs_bitwise_batch_invariance_ragged_chunked_prefill(backend): |
| 383 | """Batch invariance must hold when a prefill is split across chunks.""" |
| 384 | random.seed(int(os.getenv("VLLM_TEST_SEED", "12345"))) |
| 385 | |
| 386 | prompts = ( |
| 387 | [_random_prompt(10, 50) for _ in range(8)] |
| 388 | + [_random_prompt(300, 450) for _ in range(4)] |
| 389 | + [_random_prompt(800, 1200) for _ in range(2)] |
| 390 | ) |
| 391 | random.shuffle(prompts) |
| 392 | |
| 393 | # _random_prompt takes a word target, not a token count, and the ratio |
| 394 | # depends on the tokenizer, so derive the budget rather than hardcode it: |
| 395 | # half the longest prompt splits that prefill whatever TEST_MODEL is, while |
| 396 | # the shorter prompts are still co-scheduled whole. Floor is max_num_seqs. |
| 397 | tokenizer = AutoTokenizer.from_pretrained(TEST_MODEL) |
| 398 | prompt_lens = [len(tokenizer(p).input_ids) for p in prompts] |
| 399 | max_num_batched_tokens = max(len(prompts), max(prompt_lens) // 2) |
| 400 | assert max_num_batched_tokens < max(prompt_lens), ( |
| 401 | f"a {max_num_batched_tokens}-token budget does not split the longest " |
| 402 | f"prompt ({max(prompt_lens)} tokens), so nothing would be chunked" |
| 403 | ) |
| 404 | |
| 405 | sp = SamplingParams( |
| 406 | temperature=0.6, top_p=1.0, max_tokens=16, seed=1234, logprobs=5 |
| 407 | ) |
| 408 | llm = LLM_with_max_seqs( |
| 409 | model=TEST_MODEL, |
| 410 | max_num_seqs=len(prompts), |
| 411 | gpu_memory_utilization=float(os.getenv("VLLM_GPU_MEMORY_UTILIZATION", "0.5")), |
| 412 | max_model_len=4096, |
| 413 | attention_config={"backend": backend}, |
| 414 | max_num_batched_tokens=max_num_batched_tokens, |
| 415 | ) |
| 416 | |
| 417 | try: |
| 418 | bs1 = [] |
| 419 | for p in prompts: |
| 420 | out = llm.generate([p], sp, use_tqdm=False)[0] |
| 421 | bs1.append(_extract_step_logprobs(out)) |
| 422 | if any(logprobs is None for logprobs, _ in bs1): |
| 423 | pytest.skip("Logprobs are not available on RequestOutput.") |
| 424 | |
| 425 | outs_batched = llm.generate(prompts, sp, use_tqdm=False) |
| 426 | assert len(outs_batched) == len(prompts) |
| 427 | |
| 428 | mismatches = [] |
| 429 | for i, out in enumerate(outs_batched): |
| 430 | logprobs, tokens = _extract_step_logprobs(out) |
| 431 | bs1_logprobs, bs1_tokens = bs1[i] |
| 432 | n = prompt_lens[i] |
| 433 | if tokens != bs1_tokens: |
| 434 | mismatches.append(f"prompt {i} ({n} tokens): tokens differ") |
| 435 | elif not torch.equal(bs1_logprobs, logprobs): |
| 436 | delta = torch.max(torch.abs(bs1_logprobs - logprobs)).item() |
| 437 | mismatches.append( |
| 438 | f"prompt {i} ({n} tokens): logprobs differ, max|delta|={delta:.3e}" |
| 439 | ) |
| 440 | |
| 441 | assert not mismatches, ( |
| 442 | "Batch invariance violated under chunked prefill " |
| 443 | f"(max_num_batched_tokens={max_num_batched_tokens}): " |
| 444 | + "; ".join(mismatches) |
| 445 | ) |
| 446 | finally: |
| 447 | with contextlib.suppress(Exception): |
| 448 | llm.llm_engine.engine_core.shutdown() |
| 449 | |
| 450 | |
| 451 | @skip_unsupported |
| 452 | @pytest.mark.parametrize( |
| 453 | "backend", |
| 454 | BACKENDS, |
| 455 | ) |
| 456 | def test_simple_generation(backend): |
| 457 | """Simple test that runs the model with a basic prompt and prints the output. |
| 458 | Useful for quick smoke testing and debugging. |
| 459 | """ |
| 460 | model = TEST_MODEL |
| 461 | |
| 462 | llm = LLM( |
| 463 | model=model, |
| 464 | max_num_seqs=1, |
| 465 | tensor_parallel_size=int(os.getenv("VLLM_TP_SIZE", "1")), |
| 466 | gpu_memory_utilization=0.9, |
| 467 | max_model_len=2048, |
| 468 | dtype="auto", |
| 469 | enable_prefix_caching=False, |
| 470 | attention_config={"backend": backend}, |
| 471 | ) |
| 472 | |
| 473 | prompt = "the capital of france is" |
| 474 | sampling_params = SamplingParams( |
| 475 | temperature=0.0, |
| 476 | max_tokens=20, |
| 477 | ) |
| 478 | |
| 479 | print(f"\n{'=' * 80}") |
| 480 | print("Running simple generation test") |
| 481 | print(f"Prompt: '{prompt}'") |
| 482 | print(f"{'=' * 80}\n") |
| 483 | |
| 484 | try: |
| 485 | outputs = llm.generate([prompt], sampling_params) |
| 486 | |
| 487 | assert len(outputs) == 1 |
| 488 | output_text = outputs[0].outputs[0].text |
| 489 | |
| 490 | print(f"Output: '{output_text}'") |
| 491 | print(f"\n{'=' * 80}") |
| 492 | print(f"Full completion: '{prompt}{output_text}'") |
| 493 | print(f"{'=' * 80}\n") |
| 494 | |
| 495 | finally: |
| 496 | with contextlib.suppress(Exception): |
| 497 | llm.llm_engine.engine_core.shutdown() |
| 498 | |
| 499 | |
| 500 | @skip_unsupported |
| 501 | @pytest.mark.parametrize( |
| 502 | "backend", |
| 503 | BACKENDS, |
| 504 | ) |
| 505 | def test_logprobs_without_batch_invariance_should_fail( |
| 506 | backend, monkeypatch: pytest.MonkeyPatch |
| 507 | ): |
| 508 | """This test is the inverse of test_logprobs_bitwise_batch_invariance_bs1_vs_bsN. |
| 509 | It DISABLES batch invariance mode and expects to see non-deterministic behavior |
| 510 | between BS=1 and BS=N runs. This demonstrates that batch invariance is actually |
| 511 | doing something useful. |
| 512 | |
| 513 | The test will PASS if we detect differences (proving batch invariance matters). |
| 514 | The test will FAIL if everything matches (suggesting batch invariance isn't needed). |
| 515 | """ |
| 516 | # CRITICAL: Disable batch invariance for this test |
| 517 | monkeypatch.setenv("VLLM_BATCH_INVARIANT", "0") |
| 518 | monkeypatch.setattr(envs, "VLLM_BATCH_INVARIANT", False) |
| 519 | seed = int(os.getenv("VLLM_TEST_SEED", "12345")) |
| 520 | random.seed(seed) |
| 521 | tp_size = int(os.getenv("VLLM_TEST_TP_SIZE", "1")) |
| 522 | |
| 523 | print(f"\n{'=' * 80}") |
| 524 | print("BATCH INVARIANCE DISABLED: Expecting non-deterministic behavior") |
| 525 | print(f"{'=' * 80}\n") |
| 526 | |
| 527 | llm = LLM( |
| 528 | model=TEST_MODEL, |
| 529 | tensor_parallel_size=tp_size, |
| 530 | max_num_seqs=32, |
| 531 | max_model_len=8192, |
| 532 | dtype="auto", |
| 533 | attention_config={"backend": backend}, |
| 534 | ) |
| 535 | |
| 536 | # build ragged prompts to change shapes significantly across BS=1 vs BS=N |
| 537 | long_min = int(os.getenv("VLLM_MIN_PROMPT", "768")) |
| 538 | long_max = int(os.getenv("VLLM_MAX_PROMPT", "2048")) |
| 539 | prompts: list[str] = [] |
| 540 | options = [ |
| 541 | (max(long_min, 1536), max(long_max, 3072)), # very long |
| 542 | (max(1024, long_min), max(2048, long_max)), # long |
| 543 | (256, 512), # mid |
| 544 | (10, 20), # short |
| 545 | ] |
| 546 | |
| 547 | for _ in range(32): |
| 548 | lo, hi = random.choice(options) |
| 549 | prompts.append(_random_prompt(lo, hi)) |
| 550 | |
| 551 | sp = SamplingParams( |
| 552 | temperature=0.6, |
| 553 | top_p=1.0, |
| 554 | max_tokens=8, |
| 555 | seed=1234, |
| 556 | logprobs=5, |
| 557 | ) |
| 558 | |
| 559 | # BS=1: run prompts individually and collect logprobs per step. |
| 560 | print("\n" + "=" * 80) |
| 561 | print("STARTING BS=1 RUNS (each prompt individually)") |
| 562 | print("=" * 80 + "\n") |
| 563 | |
| 564 | bs1_logprobs_per_prompt = [] |
| 565 | bs1_tokens_per_prompt = [] |
| 566 | for idx, p in enumerate(prompts): |
| 567 | print(f"\n[BS=1] Running prompt {idx}/{len(prompts)} - Preview: {p[:80]}...") |
| 568 | outs = llm.generate([p], sp, use_tqdm=False) |
| 569 | assert len(outs) == 1 |
| 570 | step_logprobs, token_ids = _extract_step_logprobs(outs[0]) |
| 571 | if step_logprobs is None: |
| 572 | pytest.skip( |
| 573 | "Logits are not available on RequestOutput; " |
| 574 | "enable logprobs return to run this test." |
| 575 | ) |
| 576 | bs1_logprobs_per_prompt.append(step_logprobs) |
| 577 | bs1_tokens_per_prompt.append(token_ids) |
| 578 | print(f"[BS=1] Prompt {idx} generated tokens: {token_ids}") |
| 579 | |
| 580 | # BS=N: run prompts in a batch and collect logprobs per step for each prompt. |
| 581 | print("\n" + "=" * 80) |
| 582 | print(f"STARTING BS={len(prompts)} RUN (all prompts batched)") |
| 583 | print("=" * 80 + "\n") |
| 584 | |
| 585 | outs_batched = llm.generate(prompts, sp, use_tqdm=False) |
| 586 | assert len(outs_batched) == len(prompts) |
| 587 | bsN_logprobs_per_prompt = [] |
| 588 | bsN_tokens_per_prompt = [] |
| 589 | |
| 590 | print(f"\n[BS={len(prompts)}] Processing batched outputs...") |
| 591 | for idx, o in enumerate(outs_batched): |
| 592 | tokens = o.outputs[0].token_ids if o.outputs else "N/A" |
| 593 | print(f"[BS={len(prompts)}] Prompt {idx} generated tokens: {tokens}") |
| 594 | step_logprobs, token_ids = _extract_step_logprobs(o) |
| 595 | if step_logprobs is None: |
| 596 | pytest.skip( |
| 597 | "Logits are not available on RequestOutput; " |
| 598 | "enable logprobs return to run this test." |
| 599 | ) |
| 600 | bsN_logprobs_per_prompt.append(step_logprobs) |
| 601 | bsN_tokens_per_prompt.append(token_ids) |
| 602 | |
| 603 | # Compare step-by-step logprobs for each prompt between BS=1 and BS=N runs. |
| 604 | differences_found = [] |
| 605 | for i, (logprobs_bs1, logprobs_bsN, tokens_bs1, tokens_bsN) in enumerate( |
| 606 | zip( |
| 607 | bs1_logprobs_per_prompt, |
| 608 | bsN_logprobs_per_prompt, |
| 609 | bs1_tokens_per_prompt, |
| 610 | bsN_tokens_per_prompt, |
| 611 | ) |
| 612 | ): |
| 613 | if len(logprobs_bs1) != len(logprobs_bsN): |
| 614 | reason = ( |
| 615 | f"Different number of steps: {len(logprobs_bs1)} (BS=1) " |
| 616 | f"vs {len(logprobs_bsN)} (BS=N)" |
| 617 | ) |
| 618 | differences_found.append( |
| 619 | { |
| 620 | "prompt_idx": i, |
| 621 | "step": "all", |
| 622 | "reason": reason, |
| 623 | "prompt_preview": prompts[i][:100], |
| 624 | "bs1_tokens": tokens_bs1, |
| 625 | "bsN_tokens": tokens_bsN, |
| 626 | } |
| 627 | ) |
| 628 | continue |
| 629 | |
| 630 | # Check if tokens match first |
| 631 | if tokens_bs1 != tokens_bsN: |
| 632 | differences_found.append( |
| 633 | { |
| 634 | "prompt_idx": i, |
| 635 | "step": "sampling", |
| 636 | "reason": "Different tokens sampled", |
| 637 | "prompt_preview": prompts[i][:100], |
| 638 | "bs1_tokens": tokens_bs1, |
| 639 | "bsN_tokens": tokens_bsN, |
| 640 | } |
| 641 | ) |
| 642 | continue |
| 643 | |
| 644 | for t, (a, b) in enumerate(zip(logprobs_bs1, logprobs_bsN)): |
| 645 | if a.shape != b.shape: |
| 646 | differences_found.append( |
| 647 | { |
| 648 | "prompt_idx": i, |
| 649 | "step": t, |
| 650 | "reason": f"Shape mismatch: {a.shape} vs {b.shape}", |
| 651 | "prompt_preview": prompts[i][:100], |
| 652 | "bs1_tokens": tokens_bs1, |
| 653 | "bsN_tokens": tokens_bsN, |
| 654 | } |
| 655 | ) |
| 656 | break |
| 657 | |
| 658 | if not torch.equal(a, b): |
| 659 | max_diff = torch.abs(a - b).max().item() |
| 660 | print( |
| 661 | f"\n[EXPECTED DIVERGENCE FOUND] Prompt {i}, " |
| 662 | f"Token {t}: max_diff={max_diff:.6e}" |
| 663 | ) |
| 664 | bs1_tok = tokens_bs1[t] if t < len(tokens_bs1) else "N/A" |
| 665 | bsN_tok = tokens_bsN[t] if t < len(tokens_bsN) else "N/A" |
| 666 | print(f" Token IDs: bs1={bs1_tok}, bsN={bsN_tok}") |
| 667 | print(f" BS=1 logprob: {a.tolist()}") |
| 668 | print(f" BS=N logprob: {b.tolist()}") |
| 669 | differences_found.append( |
| 670 | { |
| 671 | "prompt_idx": i, |
| 672 | "step": t, |
| 673 | "reason": f"Bitwise mismatch (max_diff={max_diff:.6e})", |
| 674 | "prompt_preview": prompts[i][:100], |
| 675 | "bs1_tokens": tokens_bs1, |
| 676 | "bsN_tokens": tokens_bsN, |
| 677 | } |
| 678 | ) |
| 679 | break |
| 680 | |
| 681 | # Print summary |
| 682 | print(f"\n{'=' * 80}") |
| 683 | if differences_found: |
| 684 | success_msg = ( |
| 685 | f"✓ SUCCESS: Batch invariance is doing something! " |
| 686 | f"Found {len(differences_found)}/{len(prompts)} prompts " |
| 687 | f"with differences when batch invariance was DISABLED." |
| 688 | ) |
| 689 | print(success_msg) |
| 690 | print(f"{'=' * 80}") |
| 691 | for diff in differences_found: |
| 692 | print(f"\nPrompt {diff['prompt_idx']} (step {diff['step']}):") |
| 693 | print(f" Reason: {diff['reason']}") |
| 694 | print(f" Preview: {diff['prompt_preview']}...") |
| 695 | if "bs1_tokens" in diff: |
| 696 | print(f" BS=1 tokens: {diff['bs1_tokens']}") |
| 697 | if "bsN_tokens" in diff: |
| 698 | print(f" BS=N tokens: {diff['bsN_tokens']}") |
| 699 | print(f"{'=' * 80}\n") |
| 700 | # Test PASSES because we found differences (batch invariance matters!) |
| 701 | return |
| 702 | else: |
| 703 | # Test FAILS because everything matched even without batch invariance |
| 704 | fail_msg = ( |
| 705 | f"✗ UNEXPECTED: All {len(prompts)} prompts matched " |
| 706 | f"between BS=1 and BS=N even with batch invariance DISABLED. " |
| 707 | f"This suggests batch invariance might not be necessary, " |
| 708 | f"or the test needs more sensitive prompts." |
| 709 | ) |
| 710 | print(fail_msg) |
| 711 | print(f"{'=' * 80}\n") |
| 712 | pytest.fail(fail_msg) |
| 713 | |
| 714 | |
| 715 | @skip_if_not_cuda |
| 716 | @pytest.mark.parametrize("backend", ["FLASH_ATTN"]) |
| 717 | def test_decode_logprobs_match_prefill_logprobs( |
| 718 | backend, |
| 719 | ): |
| 720 | """Test that verifies decode logprobs match prefill logprobs. |
| 721 | |
| 722 | For each decoded token at position i: |
| 723 | 1. Run decode to generate N tokens and collect their logprobs |
| 724 | 2. For each position i in [0, N): |
| 725 | - Take prefix = prompt + tokens[0:i] |
| 726 | - Run prefill(prefix + tokens[i]) to get logprob of tokens[i] |
| 727 | - Verify prefill logprob matches decode logprob bitwise |
| 728 | |
| 729 | This ensures that the logprobs from decode are consistent with what |
| 730 | we would get if we ran prefill on each prefix. |
| 731 | """ |
| 732 | seed = int(os.getenv("VLLM_TEST_SEED", "12345")) |
| 733 | random.seed(seed) |
| 734 | tp_size = int(os.getenv("VLLM_TEST_TP_SIZE", "1")) |
| 735 | |
| 736 | llm = LLM( |
| 737 | model=TEST_MODEL, |
| 738 | tensor_parallel_size=tp_size, |
| 739 | max_num_seqs=32, |
| 740 | max_model_len=8192, |
| 741 | dtype="auto", |
| 742 | attention_config={"backend": backend}, |
| 743 | ) |
| 744 | |
| 745 | # Use a few test prompts |
| 746 | num_test_prompts = int(os.getenv("VLLM_DECODE_PREFILL_NUM_PROMPTS", "4")) |
| 747 | prompts = [_random_prompt(10, 50) for _ in range(num_test_prompts)] |
| 748 | |
| 749 | # Generate longer sequences to test multiple decode steps |
| 750 | max_tokens = int(os.getenv("VLLM_DECODE_PREFILL_MAX_TOKENS", "16")) |
| 751 | |
| 752 | sp = SamplingParams( |
| 753 | temperature=0.0, # Greedy for determinism |
| 754 | max_tokens=max_tokens, |
| 755 | logprobs=5, |
| 756 | ) |
| 757 | |
| 758 | print("\n" + "=" * 80) |
| 759 | print("STEP 1: Running decode to generate tokens and collect logprobs") |
| 760 | print("=" * 80 + "\n") |
| 761 | |
| 762 | # Step 1: Run decode and collect logprobs |
| 763 | decode_outputs = llm.generate(prompts, sp, use_tqdm=False) |
| 764 | |
| 765 | failed_comparisons = [] |
| 766 | |
| 767 | for prompt_idx, (prompt, decode_output) in enumerate(zip(prompts, decode_outputs)): |
| 768 | print(f"\n[Prompt {prompt_idx}] Testing: {prompt[:80]}...") |
| 769 | |
| 770 | # Extract decode logprobs and tokens |
| 771 | decode_logprobs, token_ids = _extract_step_logprobs(decode_output) |
| 772 | if decode_logprobs is None: |
| 773 | pytest.skip( |
| 774 | "Logprobs are not available on RequestOutput; " |
| 775 | "enable logprobs return to run this test." |
| 776 | ) |
| 777 | |
| 778 | print(f"[Prompt {prompt_idx}] Generated {len(token_ids)} tokens: {token_ids}") |
| 779 | print(f"[Prompt {prompt_idx}] Decode logprobs: {decode_logprobs.tolist()}") |
| 780 | |
| 781 | # Step 2: For each token position, run prefill and compare |
| 782 | print(f"\n[Prompt {prompt_idx}] Verifying each token via prefill...") |
| 783 | |
| 784 | for token_idx in range(len(token_ids)): |
| 785 | # Construct the prefix up to (but not including) this token |
| 786 | current_token = token_ids[token_idx] |
| 787 | |
| 788 | # We need to detokenize to get the text prefix |
| 789 | # For this, we'll use the tokenizer from the LLM |
| 790 | # However, the LLM API doesn't expose tokenizer easily, so we'll |
| 791 | # construct the prefix by decoding from the original prompt |
| 792 | |
| 793 | # Get text up to this point by using the output text |
| 794 | # This is approximate but should work for verification |
| 795 | if token_idx == 0: |
| 796 | prefix_prompt = prompt |
| 797 | else: |
| 798 | # No per-token text, so regenerate with max_tokens = token_idx |
| 799 | # to get the prefix text |
| 800 | prefix_sp = SamplingParams( |
| 801 | temperature=0.0, |
| 802 | max_tokens=token_idx, |
| 803 | logprobs=1, |
| 804 | ) |
| 805 | prefix_output = llm.generate([prompt], prefix_sp, use_tqdm=False)[0] |
| 806 | prefix_prompt = prompt + prefix_output.outputs[0].text |
| 807 | |
| 808 | # Now run prefill with max_tokens=1 to get the logprob of the next token |
| 809 | prefill_sp = SamplingParams( |
| 810 | temperature=0.0, |
| 811 | max_tokens=1, |
| 812 | logprobs=5, |
| 813 | ) |
| 814 | |
| 815 | print( |
| 816 | f" [Token {token_idx}] Running prefill for prefix " |
| 817 | f"(len={len(prefix_prompt)})..." |
| 818 | ) |
| 819 | prefill_output = llm.generate([prefix_prompt], prefill_sp, use_tqdm=False)[ |
| 820 | 0 |
| 821 | ] |
| 822 | prefill_logprobs, prefill_token_ids = _extract_step_logprobs(prefill_output) |
| 823 | |
| 824 | if prefill_logprobs is None: |
| 825 | print(f" [Token {token_idx}] Warning: No prefill logprobs available") |
| 826 | continue |
| 827 | |
| 828 | # The first token from prefill should match the current token |
| 829 | prefill_token = prefill_token_ids[0] |
| 830 | prefill_logprob = prefill_logprobs[0].item() |
| 831 | decode_logprob = decode_logprobs[token_idx].item() |
| 832 | |
| 833 | print( |
| 834 | f" [Token {token_idx}] Decode token: {current_token}, " |
| 835 | f"logprob: {decode_logprob:.8f}" |
| 836 | ) |
| 837 | print( |
| 838 | f" [Token {token_idx}] Prefill token: {prefill_token}, " |
| 839 | f"logprob: {prefill_logprob:.8f}" |
| 840 | ) |
| 841 | |
| 842 | # Check if tokens match |
| 843 | if current_token != prefill_token: |
| 844 | failed_comparisons.append( |
| 845 | { |
| 846 | "prompt_idx": prompt_idx, |
| 847 | "token_idx": token_idx, |
| 848 | "reason": "Token mismatch", |
| 849 | "decode_token": current_token, |
| 850 | "prefill_token": prefill_token, |
| 851 | "decode_logprob": decode_logprob, |
| 852 | "prefill_logprob": prefill_logprob, |
| 853 | "prompt_text": prompt[:100], |
| 854 | "prefix_text": prefix_prompt[:100], |
| 855 | } |
| 856 | ) |
| 857 | print(f" [Token {token_idx}] ✗ TOKEN MISMATCH!") |
| 858 | continue |
| 859 | |
| 860 | # Check if logprobs match bitwise |
| 861 | if decode_logprob != prefill_logprob: |
| 862 | diff = abs(decode_logprob - prefill_logprob) |
| 863 | failed_comparisons.append( |
| 864 | { |
| 865 | "prompt_idx": prompt_idx, |
| 866 | "token_idx": token_idx, |
| 867 | "reason": "Logprob mismatch", |
| 868 | "decode_token": current_token, |
| 869 | "prefill_token": prefill_token, |
| 870 | "decode_logprob": decode_logprob, |
| 871 | "prefill_logprob": prefill_logprob, |
| 872 | "diff": diff, |
| 873 | "prompt_text": prompt[:100], |
| 874 | "prefix_text": prefix_prompt[:100], |
| 875 | "decode_all_tokens": token_ids, |
| 876 | "decode_all_logprobs": decode_logprobs.tolist(), |
| 877 | } |
| 878 | ) |
| 879 | print(f" [Token {token_idx}] ✗ LOGPROB MISMATCH! diff={diff:.8e}") |
| 880 | else: |
| 881 | print(f" [Token {token_idx}] ✓ Match (bitwise equal)") |
| 882 | |
| 883 | # Print summary |
| 884 | print(f"\n{'=' * 80}") |
| 885 | if failed_comparisons: |
| 886 | print(f"DECODE-PREFILL MISMATCH: {len(failed_comparisons)} failures detected") |
| 887 | print(f"{'=' * 80}") |
| 888 | |
| 889 | # Group failures by prompt for better readability |
| 890 | failures_by_prompt: dict[int, list[dict]] = {} |
| 891 | for fail in failed_comparisons: |
| 892 | pid = fail["prompt_idx"] |
| 893 | if pid not in failures_by_prompt: |
| 894 | failures_by_prompt[pid] = [] |
| 895 | failures_by_prompt[pid].append(fail) |
| 896 | |
| 897 | for prompt_idx, failures in failures_by_prompt.items(): |
| 898 | print(f"\n{'=' * 80}") |
| 899 | print(f"PROMPT {prompt_idx}: {failures[0]['prompt_text']}...") |
| 900 | print(f"{'=' * 80}") |
| 901 | print(f"Total failures for this prompt: {len(failures)}") |
| 902 | |
| 903 | # Show where mismatches occur (which token positions) |
| 904 | mismatch_positions = [f["token_idx"] for f in failures] |
| 905 | print(f"Mismatch at token positions: {mismatch_positions}") |
| 906 | |
| 907 | # Show first few failures in detail |
| 908 | for i, fail in enumerate(failures[:5]): # Show first 5 failures per prompt |
| 909 | print(f"\n [Failure {i + 1}] Token position {fail['token_idx']}:") |
| 910 | print(f" Reason: {fail['reason']}") |
| 911 | print(f" Prefix text: '{fail['prefix_text']}...'") |
| 912 | print( |
| 913 | f" Decode: token={fail['decode_token']}, " |
| 914 | f"logprob={fail['decode_logprob']:.10f}" |
| 915 | ) |
| 916 | print( |
| 917 | f" Prefill: token={fail['prefill_token']}, " |
| 918 | f"logprob={fail['prefill_logprob']:.10f}" |
| 919 | ) |
| 920 | if "diff" in fail: |
| 921 | print(f" Difference: {fail['diff']:.10e}") |
| 922 | # Show in hex to see bitwise difference |
| 923 | import struct |
| 924 | |
| 925 | decode_hex = struct.pack("f", fail["decode_logprob"]).hex() |
| 926 | prefill_hex = struct.pack("f", fail["prefill_logprob"]).hex() |
| 927 | print(f" Decode logprob (hex): 0x{decode_hex}") |
| 928 | print(f" Prefill logprob (hex): 0x{prefill_hex}") |
| 929 | |
| 930 | # If we have all tokens/logprobs, show the context |
| 931 | if "decode_all_tokens" in fail and "decode_all_logprobs" in fail: |
| 932 | token_idx = fail["token_idx"] |
| 933 | all_tokens = fail["decode_all_tokens"] |
| 934 | all_logprobs = fail["decode_all_logprobs"] |
| 935 | |
| 936 | # Show context: 2 tokens before and after |
| 937 | start = max(0, token_idx - 2) |
| 938 | end = min(len(all_tokens), token_idx + 3) |
| 939 | |
| 940 | print(f" Context (tokens {start} to {end - 1}):") |
| 941 | for j in range(start, end): |
| 942 | marker = " <-- MISMATCH" if j == token_idx else "" |
| 943 | print( |
| 944 | f" [{j}] token={all_tokens[j]}, " |
| 945 | f"logprob={all_logprobs[j]:.8f}{marker}" |
| 946 | ) |
| 947 | |
| 948 | if len(failures) > 5: |
| 949 | print(f"\n ... and {len(failures) - 5} more failures for this prompt") |
| 950 | |
| 951 | print(f"\n{'=' * 80}\n") |
| 952 | |
| 953 | pytest.fail( |
| 954 | f"Decode logprobs do not match prefill logprobs: " |
| 955 | f"{len(failed_comparisons)} mismatches found." |
| 956 | ) |
| 957 | else: |
| 958 | print("✓ SUCCESS: All decode logprobs match prefill logprobs bitwise!") |
| 959 | print(f"{'=' * 80}\n") |
| 960 | |
| 961 | |
| 962 | def LLM_with_max_seqs( |
| 963 | model: str, |
| 964 | max_num_seqs: int, |
| 965 | gpu_memory_utilization: float, |
| 966 | max_model_len: int, |
| 967 | attention_config: dict | None = None, |
| 968 | kernel_config: dict | None = None, |
| 969 | max_num_batched_tokens: int | None = None, |
| 970 | ) -> LLM: |
| 971 | """Helper to construct an LLM with a specific max_num_seqs (batch-size limit) |
| 972 | using the high-level v1 LLM API, while constraining memory usage. |
| 973 | """ |
| 974 | extra_kwargs: dict = {} |
| 975 | if kernel_config is not None: |
| 976 | extra_kwargs["kernel_config"] = kernel_config |
| 977 | if max_num_batched_tokens is not None: |
| 978 | extra_kwargs["max_num_batched_tokens"] = max_num_batched_tokens |
| 979 | return LLM( |
| 980 | model=model, |
| 981 | max_num_seqs=max_num_seqs, |
| 982 | gpu_memory_utilization=gpu_memory_utilization, |
| 983 | max_model_len=max_model_len, |
| 984 | dtype="auto", |
| 985 | tensor_parallel_size=int(os.getenv("VLLM_TP_SIZE", "1")), |
| 986 | enable_prefix_caching=False, |
| 987 | attention_config=attention_config, |
| 988 | # Enable for MOE models |
| 989 | # enable_expert_parallel=True, |
| 990 | **extra_kwargs, |
| 991 | ) |