step_train/src/utils/tb_writer.py

"""TensorBoard logging wrapper (platform-aware, fail-soft, rank-0-only).

Platform contract (用户平台要求): when TensorBoard is enabled on the task-creation
page the platform injects $TENSORBOARD_LOG_PATH and expects logs written there:

    log_dir = os.getenv('TENSORBOARD_LOG_PATH')
    writer = SummaryWriter(log_dir=log_dir)
    writer.add_scalar('Loss/train', train_loss, step)

This wrapper adds three robustness properties on top of the raw SummaryWriter:
  1. fail-soft import — if `tensorboard` isn't installed (e.g. CPU dev box) every
     call is a silent no-op, so training code never needs `try/except` around logging.
  2. rank-aware — only the main process writes (avoids 8 ranks clobbering one event
     file); non-main ranks get a no-op writer.
  3. path resolution — $TENSORBOARD_LOG_PATH first, else a config/explicit dir, else
     a sensible default under the run's output_dir; parent dirs are created.

Usage:
    from src.utils.tb_writer import TBWriter
    tb = TBWriter(fallback_dir=os.path.join(output_dir, "tb"))
    tb.add_scalar("Loss/train", loss, step)
    ...
    tb.close()
"""

import os
from typing import Dict, Optional

from src.utils.logging_utils import setup_logger

logger = setup_logger(__name__)


def resolve_log_dir(fallback_dir: Optional[str] = None) -> Optional[str]:
    """$TENSORBOARD_LOG_PATH (platform) > fallback_dir > None. (single-dir, legacy)."""
    env = os.getenv("TENSORBOARD_LOG_PATH")
    if env:
        return env
    return fallback_dir


def resolve_log_dirs(fallback_dir: Optional[str] = None) -> list:
    """Return ALL target dirs to write to (deduped, order-preserving).

    🔴 DUAL-WRITE: we write to BOTH the platform-injected $TENSORBOARD_LOG_PATH (for the
    live hosted dashboard) AND fallback_dir (a local, checkpoint-adjacent path) at once.
    The platform path (e.g. /mnt/tensorboard_logs) is typically wiped after the job ends,
    so the local copy under <output_dir>/tb survives for post-hoc 复盘. If only one is
    available we just write that one; if they coincide we write once.
    """
    dirs = []
    env = os.getenv("TENSORBOARD_LOG_PATH")
    if env:
        dirs.append(env)
    if fallback_dir and fallback_dir not in dirs:
        dirs.append(fallback_dir)
    return dirs


class TBWriter:
    """Thin SummaryWriter facade. No-op when disabled / tensorboard missing / non-main.

    Fans every write out to one OR MORE SummaryWriters (dual-write: platform dir +
    local persistent dir) — see resolve_log_dirs."""

    def __init__(
        self,
        fallback_dir: Optional[str] = None,
        enabled: bool = True,
        is_main: bool = True,
    ):
        self.writers = []        # list[SummaryWriter]
        self.log_dirs = []       # list[str], parallel to writers
        if not enabled or not is_main:
            return

        log_dirs = resolve_log_dirs(fallback_dir)
        if not log_dirs:
            logger.info("[TB] no $TENSORBOARD_LOG_PATH and no fallback dir → disabled")
            return
        try:
            from torch.utils.tensorboard import SummaryWriter
        except Exception as e:  # tensorboard not installed
            logger.warning(f"[TB] tensorboard unavailable ({type(e).__name__}) → logging disabled")
            return

        for d in log_dirs:
            try:
                os.makedirs(d, exist_ok=True)
                self.writers.append(SummaryWriter(log_dir=d))
                self.log_dirs.append(d)
            except Exception as e:  # a bad dir shouldn't kill the others / training
                logger.warning(f"[TB] cannot open {d} ({type(e).__name__}: {e}) → skipped")
        if self.log_dirs:
            logger.info(f"[TB] logging to {len(self.log_dirs)} dir(s): {self.log_dirs}")

    @property
    def enabled(self) -> bool:
        return bool(self.writers)

    @property
    def log_dir(self):
        """Back-compat: first (primary) dir, or None."""
        return self.log_dirs[0] if self.log_dirs else None

    def add_scalar(self, tag: str, value, step: int):
        if not self.writers or value is None:
            return
        for w in self.writers:
            try:
                w.add_scalar(tag, float(value), step)
            except Exception as e:  # never let logging crash training
                logger.debug(f"[TB] add_scalar({tag}) failed: {e}")

    def add_scalars(self, prefix: str, values: Dict[str, float], step: int):
        """Log each value as a separate `{prefix}/{key}` scalar.

        We deliberately avoid SummaryWriter.add_scalars (it spawns per-tag event
        subdirs that are awkward on hosted dashboards); flat `prefix/key` tags render
        cleanly and group under one section.
        """
        if not self.writers:
            return
        for k, v in values.items():
            self.add_scalar(f"{prefix}/{k}", v, step)

    def flush(self):
        for w in self.writers:
            w.flush()

    def close(self):
        for w in self.writers:
            w.flush()
            w.close()
        self.writers = []