| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import contextlib |
| 5 | import enum |
| 6 | import json |
| 7 | |
| 8 | import torch |
| 9 | |
| 10 | from vllm.config import VllmConfig |
| 11 | from vllm.logger import init_logger |
| 12 | from vllm.v1.core.sched.output import SchedulerOutput |
| 13 | from vllm.v1.metrics.stats import SchedulerStats |
| 14 | from vllm.version import __version__ as VLLM_VERSION |
| 15 | |
| 16 | logger = init_logger(__name__) |
| 17 | |
| 18 | |
| 19 | def prepare_object_to_dump(obj) -> str: |
| 20 | if isinstance(obj, str): |
| 21 | return f"'{obj}'" # Double quotes |
| 22 | elif isinstance(obj, dict): |
| 23 | dict_str = ", ".join( |
| 24 | {f"{str(k)}: {prepare_object_to_dump(v)}" for k, v in obj.items()} |
| 25 | ) |
| 26 | return f"{{{dict_str}}}" |
| 27 | elif isinstance(obj, list): |
| 28 | return f"[{', '.join([prepare_object_to_dump(v) for v in obj])}]" |
| 29 | elif isinstance(obj, set): |
| 30 | return f"[{', '.join([prepare_object_to_dump(v) for v in list(obj)])}]" |
| 31 | # return [prepare_object_to_dump(v) for v in list(obj)] |
| 32 | elif isinstance(obj, tuple): |
| 33 | return f"[{', '.join([prepare_object_to_dump(v) for v in obj])}]" |
| 34 | elif isinstance(obj, enum.Enum): |
| 35 | return repr(obj) |
| 36 | elif isinstance(obj, torch.Tensor): |
| 37 | # We only print the 'draft' of the tensor to not expose sensitive data |
| 38 | # and to get some metadata in case of CUDA runtime crashed |
| 39 | return f"Tensor(shape={obj.shape}, device={obj.device},dtype={obj.dtype})" |
| 40 | elif hasattr(obj, "anon_repr"): |
| 41 | return obj.anon_repr() |
| 42 | elif hasattr(obj, "__dict__"): |
| 43 | items = obj.__dict__.items() |
| 44 | dict_str = ", ".join( |
| 45 | [f"{str(k)}={prepare_object_to_dump(v)}" for k, v in items] |
| 46 | ) |
| 47 | return f"{type(obj).__name__}({dict_str})" |
| 48 | else: |
| 49 | # Hacky way to make sure we can serialize the object in JSON format |
| 50 | try: |
| 51 | return json.dumps(obj) |
| 52 | except (TypeError, OverflowError): |
| 53 | return repr(obj) |
| 54 | |
| 55 | |
| 56 | def dump_engine_exception( |
| 57 | config: VllmConfig, |
| 58 | scheduler_output: SchedulerOutput, |
| 59 | scheduler_stats: SchedulerStats | None, |
| 60 | ): |
| 61 | # NOTE: ensure we can log extra info without risking raises |
| 62 | # unexpected errors during logging |
| 63 | with contextlib.suppress(Exception): |
| 64 | _dump_engine_exception(config, scheduler_output, scheduler_stats) |
| 65 | |
| 66 | |
| 67 | def _dump_engine_exception( |
| 68 | config: VllmConfig, |
| 69 | scheduler_output: SchedulerOutput, |
| 70 | scheduler_stats: SchedulerStats | None, |
| 71 | ): |
| 72 | logger.error( |
| 73 | "Dumping input data for V1 LLM engine (v%s) with config: %s, ", |
| 74 | VLLM_VERSION, |
| 75 | config, |
| 76 | ) |
| 77 | try: |
| 78 | dump_obj = prepare_object_to_dump(scheduler_output) |
| 79 | logger.error("Dumping scheduler output for model execution: %s", dump_obj) |
| 80 | if scheduler_stats: |
| 81 | logger.error("Dumping scheduler stats: %s", scheduler_stats) |
| 82 | except Exception: |
| 83 | logger.exception("Error preparing object to dump") |