|
| 1 | +"""TRL TrainerCallback implementations for telemetry and Weave tracing. |
| 2 | +
|
| 3 | +Provides callbacks that integrate with TRL's GRPOTrainer to automatically |
| 4 | +track training events via our telemetry system and optionally log to Weave. |
| 5 | +
|
| 6 | +Usage:: |
| 7 | +
|
| 8 | + from trl import GRPOConfig, GRPOTrainer |
| 9 | + from openadapt_evals.integrations.trl_callbacks import TelemetryCallback |
| 10 | +
|
| 11 | + trainer = GRPOTrainer( |
| 12 | + model=model, |
| 13 | + args=config, |
| 14 | + callbacks=[TelemetryCallback(model_name="Qwen/Qwen2.5-VL-7B-Instruct")], |
| 15 | + ... |
| 16 | + ) |
| 17 | + trainer.train() |
| 18 | +""" |
| 19 | + |
| 20 | +from __future__ import annotations |
| 21 | + |
| 22 | +import logging |
| 23 | +import time |
| 24 | +from typing import Any |
| 25 | + |
| 26 | +logger = logging.getLogger(__name__) |
| 27 | + |
| 28 | + |
| 29 | +class TelemetryCallback: |
| 30 | + """TRL TrainerCallback that emits telemetry events for training lifecycle. |
| 31 | +
|
| 32 | + Calls our telemetry functions at key training milestones: |
| 33 | + - ``on_train_begin`` -> ``track_training_run(phase="start", ...)`` |
| 34 | + - ``on_step_end`` -> ``track_training_step(step=..., reward_mean=..., loss=...)`` |
| 35 | + - ``on_save`` -> ``track_checkpoint_saved(step=...)`` |
| 36 | + - ``on_train_end`` -> ``track_training_run(phase="completed", ...)`` |
| 37 | +
|
| 38 | + Inherits from ``transformers.TrainerCallback`` at runtime so it can be |
| 39 | + passed directly to TRL's ``GRPOTrainer(callbacks=[...])``. |
| 40 | +
|
| 41 | + All telemetry calls are wrapped in try/except so failures never interrupt |
| 42 | + training. |
| 43 | + """ |
| 44 | + |
| 45 | + def __init__( |
| 46 | + self, |
| 47 | + model_name: str | None = None, |
| 48 | + task_count: int | None = None, |
| 49 | + constrained_decoding: bool = False, |
| 50 | + vision_loss_mode: str | None = None, |
| 51 | + ) -> None: |
| 52 | + self.model_name = model_name |
| 53 | + self.task_count = task_count |
| 54 | + self.constrained_decoding = constrained_decoding |
| 55 | + self.vision_loss_mode = vision_loss_mode |
| 56 | + self._train_start_time: float | None = None |
| 57 | + |
| 58 | + def on_train_begin( |
| 59 | + self, |
| 60 | + args: Any, |
| 61 | + state: Any, |
| 62 | + control: Any, |
| 63 | + **kwargs: Any, |
| 64 | + ) -> None: |
| 65 | + """Called at the start of training.""" |
| 66 | + self._train_start_time = time.time() |
| 67 | + try: |
| 68 | + from openadapt_evals.telemetry import track_training_run |
| 69 | + |
| 70 | + num_steps = getattr(args, "max_steps", None) |
| 71 | + num_rollouts = getattr(args, "num_generations", None) |
| 72 | + track_training_run( |
| 73 | + phase="start", |
| 74 | + model_name=self.model_name, |
| 75 | + num_steps=num_steps, |
| 76 | + num_rollouts_per_step=num_rollouts, |
| 77 | + task_count=self.task_count, |
| 78 | + constrained_decoding=self.constrained_decoding, |
| 79 | + vision_loss_mode=self.vision_loss_mode, |
| 80 | + ) |
| 81 | + logger.debug("Telemetry: training_run start emitted") |
| 82 | + except Exception as exc: |
| 83 | + logger.debug("Telemetry on_train_begin failed: %s", exc) |
| 84 | + |
| 85 | + def on_step_end( |
| 86 | + self, |
| 87 | + args: Any, |
| 88 | + state: Any, |
| 89 | + control: Any, |
| 90 | + **kwargs: Any, |
| 91 | + ) -> None: |
| 92 | + """Called at the end of each training step.""" |
| 93 | + try: |
| 94 | + from openadapt_evals.telemetry import track_training_step |
| 95 | + |
| 96 | + # TRL logs metrics in state.log_history |
| 97 | + reward_mean = None |
| 98 | + loss = None |
| 99 | + if state.log_history: |
| 100 | + last_log = state.log_history[-1] |
| 101 | + reward_mean = last_log.get("reward", last_log.get("reward_mean")) |
| 102 | + loss = last_log.get("loss") |
| 103 | + |
| 104 | + track_training_step( |
| 105 | + step=state.global_step, |
| 106 | + reward_mean=reward_mean, |
| 107 | + loss=loss, |
| 108 | + ) |
| 109 | + except Exception as exc: |
| 110 | + logger.debug("Telemetry on_step_end failed: %s", exc) |
| 111 | + |
| 112 | + def on_save( |
| 113 | + self, |
| 114 | + args: Any, |
| 115 | + state: Any, |
| 116 | + control: Any, |
| 117 | + **kwargs: Any, |
| 118 | + ) -> None: |
| 119 | + """Called when a checkpoint is saved.""" |
| 120 | + try: |
| 121 | + from openadapt_evals.telemetry import track_checkpoint_saved |
| 122 | + |
| 123 | + track_checkpoint_saved(step=state.global_step) |
| 124 | + logger.debug("Telemetry: checkpoint_saved at step %d", state.global_step) |
| 125 | + except Exception as exc: |
| 126 | + logger.debug("Telemetry on_save failed: %s", exc) |
| 127 | + |
| 128 | + def on_train_end( |
| 129 | + self, |
| 130 | + args: Any, |
| 131 | + state: Any, |
| 132 | + control: Any, |
| 133 | + **kwargs: Any, |
| 134 | + ) -> None: |
| 135 | + """Called at the end of training.""" |
| 136 | + try: |
| 137 | + from openadapt_evals.telemetry import track_training_run |
| 138 | + |
| 139 | + duration = None |
| 140 | + if self._train_start_time is not None: |
| 141 | + duration = time.time() - self._train_start_time |
| 142 | + |
| 143 | + # Extract final reward from last log entry |
| 144 | + reward_mean = None |
| 145 | + loss = None |
| 146 | + if state.log_history: |
| 147 | + last_log = state.log_history[-1] |
| 148 | + reward_mean = last_log.get("reward", last_log.get("reward_mean")) |
| 149 | + loss = last_log.get("loss") |
| 150 | + |
| 151 | + track_training_run( |
| 152 | + phase="completed", |
| 153 | + model_name=self.model_name, |
| 154 | + num_steps=state.global_step, |
| 155 | + duration_seconds=duration, |
| 156 | + reward_mean=reward_mean, |
| 157 | + loss=loss, |
| 158 | + ) |
| 159 | + logger.debug("Telemetry: training_run completed emitted") |
| 160 | + except Exception as exc: |
| 161 | + logger.debug("Telemetry on_train_end failed: %s", exc) |
| 162 | + |
| 163 | + |
| 164 | +# Register as a TrainerCallback subclass at import time so TRL recognizes it. |
| 165 | +# If transformers is not installed, the class still works as a plain object |
| 166 | +# (the callback methods are called by name, not by inheritance check in recent |
| 167 | +# TRL versions). |
| 168 | +try: |
| 169 | + from transformers import TrainerCallback as _TrainerCallback |
| 170 | + |
| 171 | + # Dynamically add TrainerCallback as a base class |
| 172 | + TelemetryCallback.__bases__ = (_TrainerCallback,) + TelemetryCallback.__bases__ |
| 173 | +except ImportError: |
| 174 | + logger.debug( |
| 175 | + "transformers not installed; TelemetryCallback will work as a " |
| 176 | + "duck-typed callback but won't inherit from TrainerCallback" |
| 177 | + ) |
0 commit comments