distributed fix

docs/telemetry.qmd

@@ -0,0 +1,59 @@
+---
+title: Telemetry
+description: A description of the opt-out telemetry implementation in Axolotl.
+---
+
+# Telemetry in Axolotl
+
+Axolotl implements anonymous telemetry to help maintainers understand how the library
+is used and where users encounter issues. This data helps prioritize features, optimize
+performance, and fix bugs.
+
+## Data Collection
+
+We collect:
+
+- System info: OS, Python version, Axolotl version, PyTorch version, Transformers
+version, etc.
+- Hardware info: CPU count, memory, GPU count and models
+- Runtime metrics: Training progress, memory usage, timing information
+- Usage patterns: Models (from a whitelist) and configurations used
+- Error tracking: Stack traces and error messages (sanitized to remove personal
+information)
+
+No personally identifiable information (PII) is collected.
+
+## Implementation
+
+Telemetry is implemented using PostHog and consists of:
+
+- `axolotl.telemetry.TelemetryManager`: A singleton class that initializes the
+telemetry system and provides methods for tracking events.
+- `axolotl.telemetry.errors.send_errors`: A decorator that captures exceptions and
+sends sanitized stack traces.
+- `axolotl.telemetry.runtime_metrics.RuntimeMetricsTracker`: A class that tracks
+runtime metrics during training.
+- `axolotl.telemetry.callbacks.TelemetryCallback`: A Trainer callback that sends
+runtime metrics telemetry.
+
+The telemetry system will block training startup for 15 seconds to ensure users are
+aware of data collection, unless telemetry is explicitly enabled or disabled.
+
+## Opt-Out Mechanism
+
+Telemetry is **enabled by default** on an opt-out basis. To disable it, set either:
+
+- `AXOLOTL_DO_NOT_TRACK=1` (Axolotl-specific)
+- `DO_NOT_TRACK=1` (Global standard; see https://consoledonottrack.com/)
+
+To acknowledge and explicitly enable telemetry (and remove the warning message), set:
+`AXOLOTL_DO_NOT_TRACK=0`.
+
+## Privacy
+
+- All path-like config information is automatically redacted from telemetry data
+- Model information is only collected for whitelisted organizations
+    - See `axolotl/telemetry/whitelist.yaml` for the set of whitelisted organizations
+- Each run generates a unique anonymous ID
+    - This allows us to link different telemetry events in a single same training run
+- Telemetry is only sent from the main process to avoid duplicate events

requirements.txt

@@ -63,3 +63,6 @@ torchao==0.7.0
 schedulefree==1.3.0
 
 axolotl-contribs-lgpl==0.0.3
+
+# telemetry
+posthog>=3.15.1

src/axolotl/cli/config.py

@@ -14,6 +14,8 @@ import yaml
 from transformers.utils import is_torch_bf16_gpu_available
 
 from axolotl.integrations.base import PluginManager
+from axolotl.telemetry.errors import send_errors
+from axolotl.telemetry.manager import TelemetryManager
 from axolotl.utils.comet_ import setup_comet_env_vars
 from axolotl.utils.config import (
     normalize_cfg_datasets,
@@ -27,6 +29,8 @@ from axolotl.utils.wandb_ import setup_wandb_env_vars
 
 LOG = logging.getLogger(__name__)
 
+TELEMETRY_MANAGER = TelemetryManager.get_instance()
+
 
 def check_remote_config(config: Union[str, Path]) -> Union[str, Path]:
     """
@@ -152,6 +156,7 @@ def prepare_plugins(cfg: DictDefault):
             plugin_manager.register(plugin_name)
 
 
+@send_errors
 def load_cfg(config: Union[str, Path] = Path("examples/"), **kwargs) -> DictDefault:
     """
     Loads the `axolotl` configuration stored at `config`, validates it, and performs
@@ -171,6 +176,7 @@ def load_cfg(config: Union[str, Path] = Path("examples/"), **kwargs) -> DictDefa
     # Load the config from the yaml file
     with open(config, encoding="utf-8") as file:
         cfg: DictDefault = DictDefault(yaml.safe_load(file))
+    TELEMETRY_MANAGER.send_event(event_type="config-loaded", properties=cfg)
 
     # If there are any options passed in the cli, if it is something that seems valid
     # from the yaml, then overwrite the value
@@ -214,4 +220,6 @@ def load_cfg(config: Union[str, Path] = Path("examples/"), **kwargs) -> DictDefa
     setup_mlflow_env_vars(cfg)
     setup_comet_env_vars(cfg)
 
+    TELEMETRY_MANAGER.send_event(event_type="config-processed", properties=cfg)
+
     return cfg

src/axolotl/cli/inference.py

@@ -17,6 +17,7 @@ from axolotl.cli.args import InferenceCliArgs
 from axolotl.cli.art import print_axolotl_text_art
 from axolotl.cli.config import load_cfg
 from axolotl.cli.utils import load_model_and_tokenizer
+from axolotl.telemetry.errors import send_errors
 from axolotl.utils.chat_templates import (
     get_chat_template,
     get_chat_template_from_config,
@@ -42,6 +43,7 @@ def get_multi_line_input() -> str:
     return instruction
 
 
+@send_errors
 def do_inference(
     *,
     cfg: DictDefault,
@@ -135,6 +137,7 @@ def do_inference(
         print(tokenizer.decode(generated["sequences"].cpu().tolist()[0]))
 
 
+@send_errors
 def do_inference_gradio(
     *,
     cfg: DictDefault,

src/axolotl/cli/merge_lora.py

@@ -12,11 +12,13 @@ from axolotl.cli.args import TrainerCliArgs
 from axolotl.cli.art import print_axolotl_text_art
 from axolotl.cli.config import load_cfg
 from axolotl.cli.utils import load_model_and_tokenizer
+from axolotl.telemetry.errors import send_errors
 from axolotl.utils.dict import DictDefault
 
 LOG = logging.getLogger(__name__)
 
 
+@send_errors
 def do_merge_lora(*, cfg: DictDefault) -> None:
     """
     Calls `transformers`' `merge_and_unload` on the model given in the `axolotl` config

src/axolotl/cli/merge_sharded_fsdp_weights.py

@@ -27,6 +27,7 @@ from torch.distributed.checkpoint.format_utils import _EmptyStateDictLoadPlanner
 from axolotl.cli.args import TrainerCliArgs
 from axolotl.cli.art import print_axolotl_text_art
 from axolotl.cli.config import load_cfg
+from axolotl.telemetry.errors import send_errors
 
 LOG = logging.getLogger(__name__)
 
@@ -120,6 +121,7 @@ def _distributed_checkpoint_to_merged_weights(
     return save_path_
 
 
+@send_errors
 def merge_fsdp_weights(
     checkpoint_dir: str,
     output_path: str,

src/axolotl/cli/preprocess.py

@@ -18,12 +18,14 @@ from axolotl.cli.checks import check_accelerate_default_config, check_user_token
 from axolotl.cli.config import load_cfg
 from axolotl.common.const import DEFAULT_DATASET_PREPARED_PATH
 from axolotl.common.datasets import load_datasets, load_preference_datasets
+from axolotl.telemetry.errors import send_errors
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.trainer import disable_datasets_caching
 
 LOG = logging.getLogger(__name__)
 
 
+@send_errors
 def do_preprocess(cfg: DictDefault, cli_args: PreprocessCliArgs) -> None:
     """
     Preprocesses dataset specified in axolotl config.

src/axolotl/common/datasets.py

@@ -10,6 +10,7 @@ from datasets import Dataset
 
 import axolotl.monkeypatch.data.batch_dataset_fetcher  # pylint: disable=unused-import  # noqa: F401
 from axolotl.cli.args import PreprocessCliArgs, TrainerCliArgs
+from axolotl.telemetry.errors import send_errors
 from axolotl.utils.data import prepare_dataset
 from axolotl.utils.data.rl import load_prepare_preference_datasets
 from axolotl.utils.dict import DictDefault
@@ -44,6 +45,7 @@ def sample_dataset(dataset: Dataset, num_samples: int) -> Dataset:
     )
 
 
+@send_errors
 def load_datasets(
     *,
     cfg: DictDefault,
@@ -103,6 +105,7 @@ def load_datasets(
     )
 
 
+@send_errors
 def load_preference_datasets(
     *,
     cfg: DictDefault,

src/axolotl/core/trainer_builder.py

@@ -61,6 +61,8 @@ from axolotl.core.training_args import (
 from axolotl.integrations.base import PluginManager
 from axolotl.monkeypatch.multipack import SUPPORTED_MULTIPACK_MODEL_TYPES
 from axolotl.monkeypatch.relora import ReLoRACallback
+from axolotl.telemetry.callbacks import TelemetryCallback
+from axolotl.telemetry.manager import TelemetryManager
 from axolotl.utils import is_comet_available, is_mlflow_available
 from axolotl.utils.callbacks import (
     EvalFirstStepCallback,
@@ -176,10 +178,8 @@ class TrainerBuilderBase(abc.ABC):
                 SaveAxolotlConfigtoMlflowCallback,
             )
 
-            callbacks.extend(
-                [
-                    SaveAxolotlConfigtoMlflowCallback(self.cfg.axolotl_config_path),
-                ]
+            callbacks.append(
+                SaveAxolotlConfigtoMlflowCallback(self.cfg.axolotl_config_path)
             )
         if self.cfg.use_comet and is_comet_available():
             from axolotl.utils.callbacks.comet_ import SaveAxolotlConfigtoCometCallback
@@ -188,6 +188,10 @@ class TrainerBuilderBase(abc.ABC):
                 SaveAxolotlConfigtoCometCallback(self.cfg.axolotl_config_path)
             )
 
+        telemetry_manager = TelemetryManager.get_instance()
+        if telemetry_manager.enabled:
+            callbacks.append(TelemetryCallback())
+
         return callbacks
 
     def get_post_trainer_create_callbacks(self, trainer):

src/axolotl/evaluate.py

@@ -10,6 +10,7 @@ import torch
 from accelerate.logging import get_logger
 
 from axolotl.logging_config import configure_logging
+from axolotl.telemetry.errors import send_errors
 from axolotl.train import TrainDatasetMeta
 from axolotl.utils import set_pytorch_cuda_alloc_conf
 from axolotl.utils.dict import DictDefault
@@ -61,6 +62,7 @@ def evaluate_dataset(
     return metrics
 
 
+@send_errors
 def evaluate(*, cfg: DictDefault, dataset_meta: TrainDatasetMeta) -> Dict[str, float]:
     """
     Evaluate a model on training and validation datasets

src/axolotl/telemetry/callbacks.py

@@ -0,0 +1,164 @@
+"""Trainer callbacks for reporting runtime metrics at regular intervals."""
+
+import logging
+import time
+
+from transformers import (
+    TrainerCallback,
+    TrainerControl,
+    TrainerState,
+    TrainingArguments,
+)
+
+from axolotl.telemetry.manager import TelemetryManager
+from axolotl.telemetry.runtime_metrics import RuntimeMetricsTracker
+
+LOG = logging.getLogger(__name__)
+
+TIME_SINCE_LAST = 30
+
+
+class TelemetryCallback(TrainerCallback):
+    """
+    Trainer callback for tracking and reporting runtime metrics.
+
+    This callback tracks training progress, runtime, and memory usage,
+    sending telemetry at configurable intervals.
+    """
+
+    report_interval_steps: int = 100
+
+    def __init__(self):
+        """Initialize the metrics callback."""
+        self.tracker = RuntimeMetricsTracker()
+        self.telemetry_manager = TelemetryManager.get_instance()
+        self.current_epoch = -1
+        self.start_time = time.time()
+        self.last_report_time = None
+        self.last_report_step = 0
+
+    def on_train_begin(
+        self,
+        args: TrainingArguments,
+        state: TrainerState,  # pylint: disable=unused-argument
+        control: TrainerControl,  # pylint: disable=unused-argument
+        **kwargs,  # pylint: disable=unused-argument
+    ):
+        """Handle training start."""
+        self.telemetry_manager.send_event(event_type="train-started")
+
+    def on_train_end(
+        self,
+        args: TrainingArguments,  # pylint: disable=unused-argument
+        state: TrainerState,
+        control: TrainerControl,  # pylint: disable=unused-argument
+        **kwargs,  # pylint: disable=unused-argument
+    ):
+        """Handle training end."""
+        # Send training completion event
+        self.telemetry_manager.send_event(
+            event_type="train-ended",
+            properties={
+                "loss": state.log_history[-1].get("loss", 0)
+                if state.log_history
+                else None,
+                "learning_rate": state.log_history[-1].get("learning_rate", 0)
+                if state.log_history
+                else None,
+            }
+            | self.tracker.metrics.to_dict(),
+        )
+
+    def on_epoch_begin(
+        self,
+        args: TrainingArguments,  # pylint: disable=unused-argument
+        state: TrainerState,  # pylint: disable=unused-argument
+        control: TrainerControl,  # pylint: disable=unused-argument
+        **kwargs,  # pylint: disable=unused-argument
+    ):
+        """Handle epoch start."""
+        self.current_epoch += 1
+        self.tracker.start_epoch(self.current_epoch)
+
+    def on_epoch_end(
+        self,
+        args: TrainingArguments,  # pylint: disable=unused-argument
+        state: TrainerState,  # pylint: disable=unused-argument
+        control: TrainerControl,  # pylint: disable=unused-argument
+        **kwargs,  # pylint: disable=unused-argument
+    ):
+        """Handle epoch end."""
+        self.tracker.end_epoch(self.current_epoch)
+
+    def on_step_end(
+        self,
+        args: TrainingArguments,  # pylint: disable=unused-argument
+        state: TrainerState,
+        control: TrainerControl,  # pylint: disable=unused-argument
+        **kwargs,  # pylint: disable=unused-argument
+    ):
+        """Handle step end."""
+        step = state.global_step
+        self.tracker.update_step(step)
+
+        # Check if we should report metrics
+        should_report = (
+            step % self.report_interval_steps == 0
+            or step == 1  # Always report first step
+            or step - self.last_report_step >= self.report_interval_steps
+        )
+
+        if should_report:
+            current_time = time.time()
+            if self.last_report_time is not None:
+                time_since_last_report = current_time - self.last_report_time
+            else:
+                time_since_last_report = current_time - self.start_time
+            steps_since_last_report = step - self.last_report_step
+
+            # Only report if enough time has passed to avoid flooding
+            if (
+                step == 1
+                or time_since_last_report >= TIME_SINCE_LAST
+                or steps_since_last_report >= self.report_interval_steps
+            ):
+                # Calculate steps per second for this interval
+                if time_since_last_report > 0 and steps_since_last_report > 0:
+                    steps_per_second = steps_since_last_report / time_since_last_report
+                else:
+                    steps_per_second = 0
+
+                # Update memory metrics
+                self.tracker.update_memory_metrics()
+
+                loss = state.log_history[-1].get("loss", 0) if state.log_history else 0
+                learning_rate = (
+                    state.log_history[-1].get("learning_rate", 0)
+                    if state.log_history
+                    else 0
+                )
+
+                # Prepare metrics to report
+                metrics = {
+                    "step": step,
+                    "epoch": self.current_epoch,
+                    "progress": state.epoch,  # Fractional epoch progress
+                    "loss": loss,
+                    "learning_rate": learning_rate,
+                    "steps_per_second": steps_per_second,
+                    "elapsed_time": current_time - self.start_time,
+                    "time_since_last_report": time_since_last_report,
+                }
+
+                # Add memory metrics
+                memory_metrics = self.tracker.get_memory_metrics()
+                metrics.update({"memory": memory_metrics})
+
+                # Send telemetry
+                self.telemetry_manager.send_event(
+                    event_type="train-progress", properties=metrics
+                )
+
+                # Update last report time and step
+                self.last_report_time = current_time
+                self.last_report_step = step

src/axolotl/telemetry/errors.py

@@ -0,0 +1,160 @@
+"""Telemetry utilities for exception and traceback information."""
+
+import logging
+import os
+import re
+import traceback
+from functools import wraps
+from inspect import getmodule
+from typing import Any, Callable
+
+from axolotl.telemetry.manager import TelemetryManager
+
+LOG = logging.getLogger(__name__)
+
+ERROR_HANDLED = False
+
+
+def sanitize_stack_trace(stack_trace: str) -> str:
+    """
+    Remove personal information from stack trace messages while keeping Python package codepaths.
+
+    This function identifies Python packages by looking for common patterns in virtual environment
+    and site-packages directories, preserving the package path while removing user-specific paths.
+
+    Args:
+        stack_trace: The original stack trace string.
+
+    Returns:
+        A sanitized version of the stack trace with Python package paths preserved.
+    """
+    # Split the stack trace into lines to process each file path separately
+    lines = stack_trace.split("\n")
+    sanitized_lines = []
+
+    # Regular expression to find file paths in the stack trace
+    path_pattern = re.compile(r'(?:File ")(.*?)(?:")')
+
+    # Regular expression to identify paths in site-packages or dist-packages
+    # This matches path segments like "site-packages/package_name" or "dist-packages/package_name"
+    site_packages_pattern = re.compile(
+        r"(?:site-packages|dist-packages)[/\\]([\w\-\.]+)"
+    )
+
+    # Additional common virtual environment patterns
+    venv_lib_pattern = re.compile(
+        r"(?:lib|Lib)[/\\](?:python\d+(?:\.\d+)?[/\\])?(?:site-packages|dist-packages)[/\\]([\w\-\.]+)"
+    )
+
+    for line in lines:
+        # Check if this line contains a file path
+        path_match = path_pattern.search(line)
+
+        if path_match:
+            full_path = path_match.group(1)
+            sanitized_path = ""
+
+            # Try to match site-packages pattern
+            site_packages_match = site_packages_pattern.search(full_path)
+            venv_lib_match = venv_lib_pattern.search(full_path)
+
+            if site_packages_match:
+                # Find the index where the matched pattern starts
+                idx = full_path.find("site-packages")
+                if idx == -1:
+                    idx = full_path.find("dist-packages")
+
+                # Keep from 'site-packages' onward
+                if idx >= 0:
+                    sanitized_path = full_path[idx:]
+            elif venv_lib_match:
+                # For other virtual environment patterns, find the package directory
+                match_idx = venv_lib_match.start(1)
+                if match_idx > 0:
+                    # Keep from the package name onward
+                    package_name = venv_lib_match.group(1)
+                    idx = full_path.rfind(
+                        package_name, 0, match_idx + len(package_name)
+                    )
+                    if idx >= 0:
+                        sanitized_path = full_path[idx:]
+
+            # If we couldn't identify a package pattern but path contains 'axolotl'
+            elif "axolotl" in full_path:
+                idx = full_path.rfind("axolotl")
+                if idx >= 0:
+                    sanitized_path = full_path[idx:]
+
+            # Apply the sanitization to the line
+            if sanitized_path:
+                line = line.replace(full_path, sanitized_path)
+            else:
+                # If we couldn't identify a package pattern, just keep the filename
+                filename = os.path.basename(full_path)
+                if filename:
+                    line = line.replace(full_path, filename)
+                else:
+                    line = line.replace(full_path, "")
+
+        sanitized_lines.append(line)
+
+    return "\n".join(sanitized_lines)
+
+
+def send_errors(func: Callable) -> Callable:
+    """
+    Decorator to send exception info in a function. If an exception is raised, we send
+    telemetry containing the stack trace and error message.
+
+    If an error occurs in a decorated function that is called by another decorated
+    function, we'll only send telemetry corresponding to the lower-level function.
+
+    Args:
+        func: Function to decorate.
+
+    Returns:
+        Decorated function.
+    """
+
+    @wraps(func)
+    def wrapper(*args, **kwargs) -> Any:
+        telemetry_manager = TelemetryManager.get_instance()
+
+        if not telemetry_manager.enabled:
+            return func(*args, **kwargs)
+
+        try:
+            return func(*args, **kwargs)
+        except Exception as exception:
+            # Only track if we're not already handling an error. This prevents us from
+            # capturing an error more than once in nested decorated function calls.=
+            global ERROR_HANDLED  # pylint: disable=global-statement
+            if not ERROR_HANDLED:
+                ERROR_HANDLED = True
+
+                # Get function module path
+                module = getmodule(func)
+                module_path = (
+                    f"{module.__name__}.{func.__name__}" if module else func.__name__
+                )
+
+                # Get stack trace
+                stack_trace = "".join(
+                    traceback.format_exception(
+                        type(exception), exception, exception.__traceback__
+                    )
+                )
+                stack_trace = sanitize_stack_trace(stack_trace)
+
+                # Send error telemetry
+                telemetry_manager.send_event(
+                    event_type=f"{module_path}-errored",
+                    properties={
+                        "exception": str(exception),
+                        "stack_trace": stack_trace,
+                    },
+                )
+
+            raise
+
+    return wrapper

src/axolotl/telemetry/manager.py

@@ -0,0 +1,399 @@
+"""Telemetry manager and associated utilities."""
+
+import atexit
+import importlib
+import logging
+import os
+import platform
+import time
+import uuid
+from pathlib import Path
+from typing import Any
+
+import posthog
+import psutil
+import torch
+import yaml
+
+LOG = logging.getLogger(__name__)
+
+POSTHOG_HOST = "https://app.posthog.com"
+POSTHOG_WRITE_KEY = "phc_1kUR0o04oJKKTTeSsIz2Mfm5mpiVsQEf2WOlzljMD7y"
+
+ENABLED_WARNING_SLEEP_SECONDS = 15
+ENABLED_WARNING = (
+    "\nTelemetry is enabled. This helps Axolotl's maintainers by providing insights into:\n"
+    "- Which models and configurations are most commonly used\n"
+    "- What hardware setups need to be supported\n"
+    "- Where users encounter errors\n\n"
+    "This data helps us prioritize features, optimize performance, and fix bugs.\n\n"
+    "To disable telemetry, set either:\n"
+    "- AXOLOTL_DO_NOT_TRACK=1 (Axolotl-specific)\n"
+    "- DO_NOT_TRACK=1 (Global standard; see https://consoledonottrack.com/)\n\n"
+    "To remove this warning and continue with telemetry enabled,"
+    "explicitly set AXOLOTL_DO_NOT_TRACK=0 (and leave DO_NOT_TRACK unset / set to 0)\n\n"
+    "No personally identifiable information is collected."
+    "For details, see: https://axolotl-ai-cloud.github.io/axolotl/docs/telemetry.html\n\n"
+    f"Sleeping for {ENABLED_WARNING_SLEEP_SECONDS}s..."
+)
+
+WHITELIST_PATH = str(Path(__file__).parent / "whitelist.yaml")
+
+# NOTE: Keep these up to date with any config schema changes
+FIELDS_WITH_ORGS = {
+    "base_model",
+    "tokenizer_config",
+    "base_model_config",
+    "pretraining_dataset",  # NOTE: this field may be a string or a dictionary
+}
+FIELDS_TO_REDACT = {"resume_from_checkpoint", "hub_model_id"}
+PREFIXES_TO_REDACT = {"wandb_", "comet_", "mlflow_", "gradio_"}
+PATH_INDICATORS = {"path", "dir"}
+
+# pylint: disable=duplicate-code
+RELEVANT_PACKAGES = {
+    "torch",
+    "transformers",
+    "trl",
+    "datasets",
+    "peft",
+    "bitsandbytes",
+    "accelerate",
+    "optimum",
+    "deepspeed",
+    "ray",
+    "axolotl",
+    "triton",
+    "mamba-ssm",
+    "flash-attn",
+    "xformers",
+    "autoawq",
+    "tokenizers",
+    "sentencepiece",
+    "torchao",
+    "lm_eval",
+}
+
+
+def is_main_process() -> bool:
+    """
+    Check whether we're running in the main process.
+
+    Note:
+        We're using this function instead of `torch.utils.distributed.is_main_process`
+        causes issues with DeepSpeed world_size since. This function avoids that issue
+        by checking env vars that are set by various launchers.
+
+    Returns:
+        Whether we're running in the main process.
+    """
+    # If PyTorch distributed is already initialized, use it
+    if torch.distributed.is_initialized():
+        return torch.distributed.get_rank() == 0
+
+    # Otherwise check environment variables for global rank
+    # NOTE: need to verify this in SLURM / OpenMPI environments
+    global_rank = int(
+        os.environ.get(
+            "RANK",
+            os.environ.get(
+                "GLOBAL_RANK",
+                os.environ.get(
+                    "SLURM_PROCID",
+                    os.environ.get(
+                        "OMPI_COMM_WORLD_RANK",
+                        "0",
+                    ),
+                ),
+            ),
+        )
+    )
+
+    return global_rank == 0
+
+
+class TelemetryManager:
+    """Manages telemetry collection and transmission"""
+
+    _instance = None
+    _initialized = False
+
+    def __new__(cls):
+        """
+        Telemetry manager constructor. Creates the singleton instance of this class if
+        it doesn't already exist.
+        """
+        if cls._instance is None:
+            cls._instance = super(TelemetryManager, cls).__new__(cls)
+            cls._instance._initialized = False
+
+        return cls._instance
+
+    def __init__(self):
+        """Telemetry manager initializer"""
+        if self._initialized:
+            return
+
+        self.enabled, self.explicit_enable = self._check_telemetry_enabled()
+
+        if self.enabled:
+            self.run_id = str(uuid.uuid4())
+            self.whitelist = self._load_whitelist()
+
+            try:
+                self.system_info = self._get_system_info()
+            except Exception as e:  # pylint: disable=broad-exception-caught
+                LOG.warning(f"Error during system info collection: {e}")
+                self.system_info = None
+
+            self._init_posthog()
+
+            # Register shutdown method to flush posthog telemetry
+            atexit.register(self.shutdown)
+
+        self._initialized = True
+
+    @classmethod
+    def get_instance(cls) -> "TelemetryManager":
+        if cls._instance is None:
+            cls._instance = TelemetryManager()
+
+        return cls._instance
+
+    def _check_telemetry_enabled(self) -> tuple[bool, bool]:
+        """
+        Check if telemetry is enabled based on environment variables. We also check
+        whether this is the main process (for the distributed setting and to avoid
+        sending duplicate PostHog events per GPU).
+
+        Note: This is enabled by default on an opt-out basis. Set either
+        `AXOLOTL_DO_NOT_TRACK=1` or `DO_NOT_TRACK=1` to disable telemetry. For more
+        details, see https://axolotl-ai-cloud.github.io/axolotl/docs/telemetry.html.
+
+        Returns:
+            Tuple containing:
+                - Boolean denoting whether telemetry is enabled or disabled.
+                - Boolean denoting whether telemetry is explicitly enabled or not.
+        """
+        # Parse relevant env vars and fill opt-out default values
+        axolotl_do_not_track = os.getenv("AXOLOTL_DO_NOT_TRACK")
+        do_not_track = os.getenv("DO_NOT_TRACK")
+
+        # If explicitly enabled, we'll disable the telemetry warning message
+        explicit_enabled = axolotl_do_not_track in ["0", "false"]
+
+        if axolotl_do_not_track is None:
+            axolotl_do_not_track = "0"
+
+        if do_not_track is None:
+            do_not_track = "0"
+
+        # Respect AXOLOTL_DO_NOT_TRACK, DO_NOT_TRACK if enabled
+        enabled = axolotl_do_not_track.lower() not in (
+            "1",
+            "true",
+        ) and do_not_track.lower() not in ("1", "true")
+
+        # Show warning (and sleep on all ranks) unless explicitly enabled
+        if enabled and not explicit_enabled:
+            if is_main_process():
+                LOG.warning(ENABLED_WARNING)
+            time.sleep(ENABLED_WARNING_SLEEP_SECONDS)
+
+        # Only rank 0 will send telemetry
+        if not is_main_process():
+            return False, False
+
+        return enabled, explicit_enabled
+
+    def _load_whitelist(self) -> dict:
+        """Load HuggingFace Hub organization whitelist"""
+        with open(WHITELIST_PATH, encoding="utf-8") as f:
+            return yaml.safe_load(f)
+
+    def _is_whitelisted(self, base_model: str) -> bool:
+        """Check if model org is in whitelist"""
+        if not base_model:
+            return False
+
+        base_model = base_model.lower()
+        return any(
+            org.lower() in base_model for org in self.whitelist.get("organizations", [])
+        )
+
+    def _init_posthog(self):
+        """Initialize PostHog client"""
+        posthog.host = POSTHOG_HOST
+        posthog.project_api_key = POSTHOG_WRITE_KEY
+
+    def _redact_paths(self, properties: dict[str, Any]) -> dict[str, Any]:
+        """
+        Redact properties to remove any paths, so as to avoid inadvertently collecting
+        private or personally identifiable information (PII). We also remove
+        information related to Wandb, MLflow, etc. configuration.
+
+        Args:
+            properties: Dictionary of properties to redact.
+
+        Returns:
+            Properties dictionary with redaction applied.
+        """
+        if not properties:
+            return {}
+
+        def redact_value(value: Any, key: str = "") -> Any:
+            """Recursively sanitize values, redacting those with path-like keys"""
+            if isinstance(key, str) and isinstance(value, str):
+                # Fields that should be redacted if org is not whitelisted
+                if key in FIELDS_WITH_ORGS:
+                    org = value.split("/")[0]
+                    if org not in self.whitelist["organizations"]:
+                        return "[REDACTED]"
+
+                # Other redaction special cases
+                if (
+                    key in FIELDS_TO_REDACT
+                    or any(prefix in key for prefix in PREFIXES_TO_REDACT)
+                    or any(indicator in key.lower() for indicator in PATH_INDICATORS)
+                ):
+                    return "[REDACTED]"
+
+            # Handle nested structures
+            if isinstance(value, dict):
+                return {k: redact_value(v, k) for k, v in value.items()}
+            if isinstance(value, list):
+                return [redact_value(item) for item in value]
+
+            return value
+
+        # Create new dict with redacted values
+        redacted = {k: redact_value(v, k) for k, v in properties.items()}
+
+        return redacted
+
+    def _get_system_info(self) -> dict[str, Any]:
+        """Collect system information for various hardware accelerators"""
+        gpu_info = []
+        accelerator_type = "none"
+
+        # NVIDIA GPUs
+        if torch.cuda.is_available():
+            accelerator_type = "cuda"
+            for i in range(torch.cuda.device_count()):
+                gpu_info.append(
+                    {
+                        "name": torch.cuda.get_device_name(i),
+                        "memory": torch.cuda.get_device_properties(i).total_memory,
+                    }
+                )
+
+        # AMD GPUs
+        elif hasattr(torch, "hip") and torch.hip.is_available():
+            accelerator_type = "hip"
+            for i in range(torch.hip.device_count()):
+                gpu_info.append(
+                    {
+                        "name": torch.hip.get_device_name(i),
+                        "memory": torch.hip.get_device_properties(i).total_memory
+                        if hasattr(torch.hip, "get_device_properties")
+                        else None,
+                    }
+                )
+
+        # Apple Silicon
+        elif hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
+            accelerator_type = "mps"
+            gpu_info.append(
+                {
+                    "name": "Apple Silicon",
+                    # NOTE: this is memory allocated to this process, not total memory
+                    "memory": torch.mps.driver_allocated_memory(),
+                }
+            )
+
+        # Intel GPUs
+        elif hasattr(torch, "xpu") and torch.xpu.is_available():
+            accelerator_type = "xpu"
+            for i in range(torch.xpu.device_count()):
+                memory = None
+                if hasattr(torch.xpu, "get_device_properties"):
+                    memory = torch.xpu.get_device_properties(i).total_memory
+
+                gpu_info.append(
+                    {
+                        "name": torch.xpu.get_device_name(i),
+                        "memory": memory,
+                    }
+                )
+
+        # NPUs
+        elif hasattr(torch, "npu") and torch.npu.is_available():
+            accelerator_type = "npu"
+            for i in range(torch.npu.device_count()):
+                memory = None
+                if hasattr(torch.npu, "get_device_properties"):
+                    memory = torch.npu.get_device_properties(i).total_memory
+
+                gpu_info.append(
+                    {
+                        "name": torch.npu.get_device_name(i),
+                        "memory": memory,
+                    }
+                )
+
+        # Get relevant package versions
+        installed_packages = {}
+        for package in RELEVANT_PACKAGES:
+            try:
+                version = importlib.metadata.version(package)
+                installed_packages[f"{package}_version"] = version
+            except importlib.metadata.PackageNotFoundError:
+                pass
+
+        return {
+            "os": platform.system(),
+            "python_version": platform.python_version(),
+            "cpu_count": psutil.cpu_count(),
+            "memory_total": psutil.virtual_memory().total,
+            "accelerator_type": accelerator_type,
+            "accelerator_count": len(gpu_info),
+            "accelerator_info": gpu_info,
+            **installed_packages,
+        }
+
+    def send_event(self, event_type: str, properties: dict[str, Any] | None = None):
+        """Send a telemetry event"""
+        if not self.enabled:
+            return
+
+        if properties is None:
+            properties = {}
+
+        # Sanitize properties to remove PII
+        properties = self._redact_paths(properties)
+
+        # Wrap PostHog errors in try / except to not raise errors during Axolotl usage
+        try:
+            # Send event via PostHog
+            posthog.capture(
+                distinct_id=self.run_id,
+                event=event_type,
+                properties=properties,
+                disable_geoip=True,
+            )
+        except Exception as e:  # pylint: disable=broad-exception-caught
+            LOG.warning(f"Failed to send telemetry event: {e}")
+
+        # Additionally, send system info telemetry when loading config.
+        # NOTE: Is this the best place for this?
+        if event_type == "config-loaded":
+            self.send_system_info()
+
+    def send_system_info(self):
+        """Helper method for sending system info"""
+        self.send_event(event_type="system-info", properties=self.system_info)
+
+    def shutdown(self):
+        """Ensure all queued events are processed before shutdown"""
+        if self.enabled:
+            posthog.flush()

src/axolotl/telemetry/runtime_metrics.py

@@ -0,0 +1,209 @@
+"""Telemetry utilities for runtime and memory metrics."""
+
+import logging
+import time
+from dataclasses import dataclass, field
+from typing import Any
+
+import psutil
+import torch
+
+from axolotl.telemetry.manager import TelemetryManager
+
+LOG = logging.getLogger(__name__)
+
+
+@dataclass
+class RuntimeMetrics:
+    """Container for runtime metrics to be tracked throughout training."""
+
+    # Timing metrics
+    start_time: float
+    epoch_start_times: dict[int, float] = field(init=False)
+    epoch_end_times: dict[int, float] = field(init=False)
+
+    # Memory metrics
+    peak_cpu_memory: int = 0
+    peak_gpu_memory: dict[int, int] = field(init=False)
+
+    # Progress metrics
+    total_steps: int = 0
+    current_epoch: int = 0
+    current_step: int = 0
+
+    def __post_init__(self):
+        """Initialize empty metric mappings."""
+        self.epoch_start_times = {}
+        self.epoch_end_times = {}
+        self.peak_gpu_memory = {}
+
+    @property
+    def elapsed_time(self) -> float:
+        """Calculate total elapsed time in seconds."""
+        return time.time() - self.start_time
+
+    def epoch_time(self, epoch: int) -> float | None:
+        """Calculate time taken for a specific epoch in seconds."""
+        if epoch in self.epoch_start_times and epoch in self.epoch_end_times:
+            return self.epoch_end_times[epoch] - self.epoch_start_times[epoch]
+
+        return None
+
+    def average_epoch_time(self) -> float | None:
+        """Calculate average time per epoch in seconds."""
+        completed_epochs = [
+            epoch for epoch in self.epoch_start_times if epoch in self.epoch_end_times
+        ]
+        if not completed_epochs:
+            return None
+
+        total_time = 0.0
+        for epoch in completed_epochs:
+            epoch_time = self.epoch_time(epoch)
+            if epoch_time is not None:  # Check to avoid mypy warning
+                total_time += epoch_time
+
+        return total_time / len(completed_epochs)
+
+    def steps_per_second(self) -> float | None:
+        """Calculate average steps per second across all training."""
+        if self.total_steps == 0 or self.elapsed_time == 0:
+            return None
+
+        return self.total_steps / self.elapsed_time
+
+    def to_dict(self) -> dict[str, Any]:
+        """Convert metrics to a dictionary for telemetry reporting."""
+        metrics = {
+            "total_time_seconds": self.elapsed_time,
+            "total_steps": self.total_steps,
+            "steps_per_second": self.steps_per_second(),
+            "epochs_completed": len(
+                [
+                    epoch
+                    for epoch in self.epoch_start_times
+                    if epoch in self.epoch_end_times
+                ]
+            ),
+            "peak_cpu_memory_bytes": self.peak_cpu_memory,
+        }
+
+        # Add per-epoch timing if available
+        epoch_times: dict[str, float] = {}
+        for epoch in sorted(self.epoch_end_times.keys()):
+            time_taken = self.epoch_time(epoch)
+            if time_taken is not None:
+                epoch_times[f"epoch_{epoch}_seconds"] = time_taken
+
+        if epoch_times:
+            metrics["epoch_times"] = epoch_times  # type: ignore
+            metrics["average_epoch_time_seconds"] = self.average_epoch_time()
+
+        # Add GPU memory metrics if available
+        if self.peak_gpu_memory:
+            gpu_metrics: dict[str, int] = {}
+            for gpu_id, memory in self.peak_gpu_memory.items():
+                gpu_metrics[f"gpu_{gpu_id}_peak_memory_bytes"] = memory
+            metrics["gpu_memory"] = gpu_metrics  # type: ignore
+
+        return metrics
+
+
+class RuntimeMetricsTracker:
+    """Tracker for runtime metrics during training."""
+
+    update_interval = 100
+
+    def __init__(self):
+        """Initialize the runtime metrics tracker."""
+        self.metrics = RuntimeMetrics(start_time=time.time())
+        self.telemetry_manager = TelemetryManager.get_instance()
+
+    def start_epoch(self, epoch: int):
+        """Record the start of a new epoch."""
+        self.metrics.current_epoch = epoch
+        self.metrics.epoch_start_times[epoch] = time.time()
+        self.update_memory_metrics()
+
+    def end_epoch(self, epoch: int):
+        """Record the end of an epoch."""
+        self.metrics.epoch_end_times[epoch] = time.time()
+
+    def update_step(self, step: int):
+        """Update the current step count."""
+        self.metrics.current_step = step
+        self.metrics.total_steps += 1
+
+        # Periodically update memory metrics
+        if step % self.update_interval == 0:
+            self.update_memory_metrics()
+
+    def _get_allocated_memory(self) -> dict[int, int]:
+        """
+        Helper function for getting accelerator-agnostic allocated memory.
+
+        Returns:
+            A dictionary mapping device IDs to allocated memory in bytes
+        """
+        memory_used: dict[int, int] = {}
+
+        # NVIDIA GPUs
+        if torch.cuda.is_available():
+            for i in range(torch.cuda.device_count()):
+                memory_used[i] = torch.cuda.memory_allocated(i)
+
+        # AMD GPUs
+        elif hasattr(torch, "hip") and torch.hip.is_available():
+            for i in range(torch.hip.device_count()):
+                if hasattr(torch.hip, "memory_allocated"):
+                    memory_used[i] = torch.hip.memory_allocated(i)
+
+        # Apple Silicon
+        elif hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
+            # MPS doesn't have per-device memory stats since there's only one device
+            if hasattr(torch.mps, "current_allocated_memory"):
+                memory_used[0] = torch.mps.current_allocated_memory()
+
+        # Intel GPUs
+        elif hasattr(torch, "xpu") and torch.xpu.is_available():
+            for i in range(torch.xpu.device_count()):
+                if hasattr(torch.xpu, "memory_allocated"):
+                    memory_used[i] = torch.xpu.memory_allocated(i)
+
+        # NPUs
+        elif hasattr(torch, "npu") and torch.npu.is_available():
+            for i in range(torch.npu.device_count()):
+                if hasattr(torch.npu, "memory_allocated"):
+                    memory_used[i] = torch.npu.memory_allocated(i)
+
+        return memory_used
+
+    def update_memory_metrics(self):
+        """Update peak memory usage metrics."""
+        # CPU memory
+        cpu_memory = psutil.Process().memory_info().rss
+        self.metrics.peak_cpu_memory = max(self.metrics.peak_cpu_memory, cpu_memory)
+
+        # GPU memory (if available)
+        memory_used = self._get_allocated_memory()
+        for i, memory in memory_used.items():
+            self.metrics.peak_gpu_memory[i] = max(
+                self.metrics.peak_gpu_memory.get(i, 0), memory
+            )
+
+    def get_memory_metrics(self) -> dict[str, Any]:
+        """Get the current memory metrics as a dictionary."""
+        memory_metrics = {
+            "cpu_memory_bytes": psutil.Process().memory_info().rss,
+            "peak_cpu_memory_bytes": self.metrics.peak_cpu_memory,
+        }
+
+        # GPU memory (if available)
+        memory_used = self._get_allocated_memory()
+        for i, memory in memory_used.items():
+            memory_metrics[f"gpu_{i}_memory_bytes"] = memory
+            memory_metrics[
+                f"gpu_{i}_peak_memory_bytes"
+            ] = self.metrics.peak_gpu_memory.get(i, 0)
+
+        return memory_metrics

src/axolotl/telemetry/whitelist.yaml

@@ -0,0 +1,18 @@
+organizations:
+  - "axolotl-ai-co"
+  - "meta-llama"
+  - "huggingface"
+  - "nvidia"
+  - "facebook"
+  - "google"
+  - "microsoft"
+  - "deepseek-ai"
+  - "HuggingFaceTB"
+  - "mistralai"
+  - "Qwen"
+  - "briaai"
+  - "unsloth"
+  - "NousResearch"
+  - "allenai"
+  - "amd"
+  - "tiiuae"

src/axolotl/train.py

@@ -1,5 +1,6 @@
 """Prepare and train a model on a dataset. Can also infer from a model or merge lora"""
 
+import importlib
 import inspect
 import os
 import signal
@@ -13,7 +14,6 @@ import transformers.modelcard
 from accelerate.logging import get_logger
 from accelerate.utils import save_fsdp_model
 from peft import PeftModel
-from pkg_resources import get_distribution  # type: ignore
 from transformers import PreTrainedModel, PreTrainedTokenizer
 from transformers.integrations.deepspeed import is_deepspeed_zero3_enabled
 
@@ -22,6 +22,8 @@ from axolotl.contribs.lgpl.unsloth import (  # pylint: disable = no-name-in-modu
     fix_untrained_tokens,
 )
 from axolotl.logging_config import configure_logging
+from axolotl.telemetry.errors import send_errors
+from axolotl.telemetry.manager import TelemetryManager
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.freeze import freeze_layers_except
 from axolotl.utils.models import load_model, load_processor, load_tokenizer
@@ -39,7 +41,10 @@ sys.path.insert(0, src_dir)
 configure_logging()
 LOG = get_logger(__name__)
 
+TELEMETRY_MANAGER = TelemetryManager.get_instance()
 
+
+@send_errors
 def train(
     *, cfg: DictDefault, dataset_meta: TrainDatasetMeta
 ) -> Tuple[Union[PeftModel, PreTrainedModel], PreTrainedTokenizer]:
@@ -75,7 +80,7 @@ def train(
             )
     resume_from_checkpoint = cfg.resume_from_checkpoint
 
-    # Load the model and tokenizer
+    # Load model
     msg = "loading model"
     if cfg.adapter:
         msg += " and peft_config..."
@@ -84,6 +89,14 @@ def train(
     if model.generation_config is not None:
         model.generation_config.do_sample = True
 
+    TELEMETRY_MANAGER.send_event(
+        event_type="model-loaded", properties=model.config.to_dict()
+    )
+    if peft_config:
+        TELEMETRY_MANAGER.send_event(
+            event_type="peft-config-loaded", properties=peft_config.to_dict()
+        )
+
     model_ref = None
     if cfg.rl and cfg.rl != "orpo":
         if cfg.adapter and not cfg.rl_adapter_ref_model:
@@ -91,7 +104,7 @@ def train(
             LOG.debug("Passing model_ref: None to RL trainer")
             model_ref = None  # explicit setting to None
         else:
-            # load the model again for model_ref/baseline
+            # load the model again for model_ref / baseline
             model_ref, _ = load_model(cfg, tokenizer, reference_model=True)
 
     safe_serialization = cfg.save_safetensors is True
@@ -166,7 +179,7 @@ def train(
 
     if getattr(cfg, "axolotl_config_path"):
         raw_axolotl_cfg = Path(cfg.axolotl_config_path)
-        version = get_distribution("axolotl").version
+        version = importlib.metadata.version("axolotl")
         if raw_axolotl_cfg.is_file():
             transformers.modelcard.AUTOGENERATED_TRAINER_COMMENT += f"\n<details><summary>See axolotl config</summary>\n\naxolotl version: `{version}`\n```yaml\n{raw_axolotl_cfg.read_text(encoding='utf-8')}\n```\n\n</details><br>\n"
 
@@ -174,8 +187,6 @@ def train(
     if cfg.group_by_length:
         LOG.info("hang tight... sorting dataset for group_by_length")
 
-    pretrain_hooks(cfg, trainer)
-
     if cfg.flash_optimum:
         with torch.backends.cuda.sdp_kernel(
             # TODO configure these from the YAML w/ sdp_kernel_kwargs: ...
@@ -186,9 +197,6 @@ def train(
             trainer.train(resume_from_checkpoint=resume_from_checkpoint)
     else:
         trainer.train(resume_from_checkpoint=resume_from_checkpoint)
-
-    post_train_hooks(cfg, trainer)
-
     LOG.info(f"Training Completed!!! Saving pre-trained model to {cfg.output_dir}")
 
     # post training
@@ -292,21 +300,3 @@ def train(
         trainer.push_to_hub()
 
     return model, tokenizer
-
-
-def pretrain_hooks(_cfg, _trainer):
-    """
-    Run hooks right before kicking off the training
-    :param cfg:
-    :param trainer:
-    :return:
-    """
-
-
-def post_train_hooks(_cfg, _trainer):
-    """
-    Run hooks right after training completes
-    :param cfg:
-    :param trainer:
-    :return:
-    """

src/axolotl/utils/config/models/input/v0_4_1/__init__.py

@@ -1683,7 +1683,7 @@ class AxolotlInputConfig(
 
 
 class AxolotlConfigWCapabilities(AxolotlInputConfig):
-    """wrapper to valdiate gpu capabilities with the configured options"""
+    """Wrapper to validate GPU capabilities with the config options"""
 
     capabilities: GPUCapabilities
     env_capabilities: EnvCapabilities

src/axolotl/utils/models.py

@@ -54,6 +54,7 @@ from axolotl.monkeypatch.multipack import (
     patch_for_multipack,
 )
 from axolotl.prompt_tokenizers import LLAMA_DEFAULT_EOS_TOKEN
+from axolotl.telemetry.errors import send_errors
 from axolotl.utils.bench import log_gpu_memory_usage
 from axolotl.utils.chat_templates import get_chat_template_from_config
 from axolotl.utils.dict import DictDefault
@@ -165,6 +166,7 @@ def load_model_config(cfg):
     return model_config
 
 
+@send_errors
 def load_tokenizer(cfg):
     model_config = load_model_config(cfg)
     tokenizer_kwargs = {}
@@ -318,6 +320,7 @@ def load_tokenizer(cfg):
     return tokenizer
 
 
+@send_errors
 def load_processor(cfg: DictDefault, tokenizer: PreTrainedTokenizerBase):
     processor_kwargs: Dict[str, Any] = {}  # do we actually need this?
 
@@ -1192,18 +1195,17 @@ class ModelLoader:
         return self.model, lora_config
 
 
+@send_errors
 def load_model(
     cfg: DictDefault,
     tokenizer: PreTrainedTokenizerBase,
     *,
-    processor: ProcessorMixin = None,  # pylint: disable=unused-argument
+    processor: ProcessorMixin = None,
     inference: bool = False,
     reference_model: bool = False,
-    **kwargs,  # pylint: disable=unused-argument
-) -> Tuple[PreTrainedModel, Optional[PeftConfig]]:
-    """
-    Load a model for a given configuration and tokenizer.
-    """
+    **kwargs,
+) -> Tuple[PreTrainedModel, PeftConfig | None]:
+    """Load a model for a given configuration and tokenizer"""
     loader = ModelLoader(
         cfg,
         tokenizer,
@@ -1215,6 +1217,7 @@ def load_model(
     return loader.load_model()
 
 
+@send_errors
 def load_adapter(model, cfg, adapter, inference=False):
     # type: (PreTrainedModel, DictDefault, Optional[str], bool) -> Tuple[PreTrainedModel, Optional[PeftConfig]]
 

tests/conftest.py

@@ -1,6 +1,5 @@
-"""
-shared pytest fixtures
-"""
+"""Shared pytest fixtures"""
+
 import functools
 import importlib
 import shutil
@@ -171,3 +170,9 @@ def cleanup_monkeypatches():
             module_globals = module_name_tuple[1]
             for module_global in module_globals:
                 globals().pop(module_global, None)
+
+
+@pytest.fixture(autouse=True)
+def disable_telemetry(monkeypatch):
+    monkeypatch.setenv("AXOLOTL_DO_NOT_TRACK", "1")
+    yield

tests/telemetry/conftest.py

@@ -0,0 +1,9 @@
+"""Shared pytest fixtures for telemetry tests."""
+
+import pytest
+
+
+@pytest.fixture(autouse=True)
+def disable_telemetry(monkeypatch):
+    monkeypatch.delenv("AXOLOTL_DO_NOT_TRACK", raising=False)
+    yield

tests/telemetry/test_callbacks.py

@@ -0,0 +1,372 @@
+"""Tests for telemetry callback module."""
+# pylint: disable=redefined-outer-name
+
+import time
+from unittest.mock import MagicMock, patch
+
+import pytest
+from transformers import TrainerControl, TrainerState, TrainingArguments
+
+from axolotl.telemetry.callbacks import TIME_SINCE_LAST, TelemetryCallback
+
+
+def calc_expected_metrics(step, last_step, current_time, last_time, start_time=900.0):
+    """Calculate expected metrics values for tests"""
+    time_diff = current_time - last_time
+    step_diff = step - last_step
+    return {
+        "steps_per_second": step_diff / time_diff
+        if time_diff > 0 and step_diff > 0
+        else 0,
+        "time_since_last_report": time_diff,
+        "elapsed_time": current_time - start_time,
+    }
+
+
+@pytest.fixture
+def mock_time():
+    """Mock time.time() to have predictable values in tests"""
+    with patch("axolotl.telemetry.callbacks.time") as mock_time:
+        mock_time.time.return_value = 1000.0
+        yield mock_time
+
+
+@pytest.fixture
+def mock_telemetry_manager():
+    """Create a mock TelemetryManager"""
+    with patch("axolotl.telemetry.callbacks.TelemetryManager") as mock_manager_class:
+        mock_manager = MagicMock()
+        mock_manager_class.get_instance.return_value = mock_manager
+        yield mock_manager
+
+
+@pytest.fixture
+def mock_runtime_metrics_tracker():
+    """Create a mock RuntimeMetricsTracker"""
+    with patch(
+        "axolotl.telemetry.callbacks.RuntimeMetricsTracker"
+    ) as mock_tracker_class:
+        mock_tracker = MagicMock()
+        # Set up metrics property on the tracker
+        mock_metrics = MagicMock()
+        mock_metrics.to_dict.return_value = {
+            "total_steps": 100,
+            "peak_cpu_memory_bytes": 1024,
+        }
+        mock_tracker.metrics = mock_metrics
+
+        # Make the constructor return our mock
+        mock_tracker_class.return_value = mock_tracker
+        yield mock_tracker
+
+
+@pytest.fixture
+def training_args():
+    """Create a minimal TrainingArguments instance"""
+    return TrainingArguments(output_dir="./output")
+
+
+@pytest.fixture
+def trainer_state():
+    """Create a mock TrainerState"""
+    state = MagicMock(spec=TrainerState)
+    state.global_step = 10
+    state.epoch = 0.5  # halfway through first epoch
+    state.log_history = [{"loss": 2.5, "learning_rate": 5e-5}]
+    return state
+
+
+@pytest.fixture
+def trainer_control():
+    """Create a mock TrainerControl"""
+    return MagicMock(spec=TrainerControl)
+
+
+# pylint: disable=unused-argument
+@pytest.fixture
+def callback(mock_telemetry_manager, mock_runtime_metrics_tracker):
+    """Create a TelemetryCallback instance with mocked dependencies"""
+    return TelemetryCallback()
+
+
+class TestTelemetryCallback:
+    """Tests for the TelemetryCallback class."""
+
+    def test_initialization(self, callback, mock_runtime_metrics_tracker):
+        """Test callback initialization."""
+        assert callback.current_epoch == -1
+        assert callback.tracker == mock_runtime_metrics_tracker
+        assert callback.last_report_step == 0
+        assert hasattr(callback, "start_time")
+        assert hasattr(callback, "last_report_time")
+        assert callback.report_interval_steps == 100
+
+    def test_on_train_begin(
+        self,
+        callback,
+        mock_telemetry_manager,
+        training_args,
+        trainer_state,
+        trainer_control,
+    ):
+        """Test on_train_begin sends expected event."""
+        callback.on_train_begin(training_args, trainer_state, trainer_control)
+
+        mock_telemetry_manager.send_event.assert_called_once_with(
+            event_type="train-started"
+        )
+
+    def test_on_train_end(
+        self,
+        callback,
+        mock_telemetry_manager,
+        training_args,
+        trainer_state,
+        trainer_control,
+    ):
+        """Test on_train_end sends expected event with metrics."""
+        callback.on_train_end(training_args, trainer_state, trainer_control)
+
+        mock_telemetry_manager.send_event.assert_called_once()
+        call_args = mock_telemetry_manager.send_event.call_args[1]
+
+        assert call_args["event_type"] == "train-ended"
+        assert "loss" in call_args["properties"]
+        assert call_args["properties"]["loss"] == 2.5
+        assert "learning_rate" in call_args["properties"]
+        assert call_args["properties"]["learning_rate"] == 5e-5
+
+        # Check that metrics from RuntimeMetricsTracker are included
+        assert "total_steps" in call_args["properties"]
+        assert call_args["properties"]["total_steps"] == 100
+        assert "peak_cpu_memory_bytes" in call_args["properties"]
+        assert call_args["properties"]["peak_cpu_memory_bytes"] == 1024
+
+    def test_on_epoch_begin(
+        self,
+        callback,
+        mock_runtime_metrics_tracker,
+        training_args,
+        trainer_state,
+        trainer_control,
+    ):
+        """Test on_epoch_begin updates epoch counter and calls tracker."""
+        initial_epoch = callback.current_epoch
+
+        callback.on_epoch_begin(training_args, trainer_state, trainer_control)
+
+        assert callback.current_epoch == initial_epoch + 1
+        mock_runtime_metrics_tracker.start_epoch.assert_called_once_with(
+            initial_epoch + 1
+        )
+
+    def test_on_epoch_end(
+        self,
+        callback,
+        mock_runtime_metrics_tracker,
+        training_args,
+        trainer_state,
+        trainer_control,
+    ):
+        """Test on_epoch_end calls tracker."""
+        # Set current epoch
+        callback.current_epoch = 2
+
+        callback.on_epoch_end(training_args, trainer_state, trainer_control)
+
+        mock_runtime_metrics_tracker.end_epoch.assert_called_once_with(2)
+
+    def test_on_step_end_no_report(
+        self,
+        callback,
+        mock_telemetry_manager,
+        mock_runtime_metrics_tracker,
+        training_args,
+        trainer_state,
+        trainer_control,
+    ):
+        """Test on_step_end updates tracker but doesn't report if criteria not met."""
+        # Set up state to avoid reporting
+        trainer_state.global_step = 42  # Not divisible by report_interval_steps
+        callback.last_report_step = 41  # Just 1 step since last report
+        callback.last_report_time = time.time()  # Just now
+
+        callback.on_step_end(training_args, trainer_state, trainer_control)
+
+        # Should update tracker
+        mock_runtime_metrics_tracker.update_step.assert_called_once_with(42)
+
+        # Should not send telemetry
+        mock_telemetry_manager.send_event.assert_not_called()
+
+        # Should not update last report time/step
+        assert callback.last_report_step == 41
+
+    def test_on_step_end_report_interval_steps(
+        self,
+        callback,
+        mock_telemetry_manager,
+        mock_runtime_metrics_tracker,
+        mock_time,
+        training_args,
+        trainer_state,
+        trainer_control,
+    ):
+        """Test on_step_end reports when step interval is reached."""
+        # Set up state with clear values
+        current_step = 100  # Exactly matches report_interval_steps
+        last_step = 0
+        start_time = 900.0
+        current_time = 1000.0
+        time_diff = current_time - start_time  # 100 seconds
+
+        # Configure state and callback
+        trainer_state.global_step = current_step
+        callback.report_interval_steps = 100
+        callback.last_report_step = last_step
+        callback.start_time = start_time
+        callback.last_report_time = start_time
+
+        # Mock time.time() to return consistent values
+        mock_time.time.return_value = current_time
+
+        callback.on_step_end(training_args, trainer_state, trainer_control)
+
+        # Should update tracker
+        mock_runtime_metrics_tracker.update_step.assert_called_once_with(current_step)
+        mock_runtime_metrics_tracker.update_memory_metrics.assert_called_once()
+
+        # Should send telemetry
+        mock_telemetry_manager.send_event.assert_called_once()
+        call_args = mock_telemetry_manager.send_event.call_args[1]
+        assert call_args["event_type"] == "train-progress"
+
+        # Properties should include expected values
+        props = call_args["properties"]
+        assert props["step"] == current_step
+        assert props["elapsed_time"] == time_diff  # 1000 - 900 = 100
+        assert props["time_since_last_report"] == time_diff  # 1000 - 900 = 100
+        assert props["steps_per_second"] == 1.0  # 100 steps / 100 seconds
+
+        # Should update last report time/step
+        assert callback.last_report_step == current_step
+        assert callback.last_report_time == current_time
+
+    def test_on_step_end_report_time_elapsed(
+        self,
+        callback,
+        mock_telemetry_manager,
+        mock_runtime_metrics_tracker,  # pylint: disable=unused-argument
+        mock_time,
+        training_args,
+        trainer_state,
+        trainer_control,
+    ):
+        """Test on_step_end reports when enough time has elapsed."""
+        # Set up state with clear values
+        current_step = 120
+        last_step = 10
+        start_time = 900.0
+        current_time = 1000.0
+        time_diff = TIME_SINCE_LAST + 1  # Just over the threshold
+
+        # Configure state and callback
+        trainer_state.global_step = current_step
+        callback.report_interval_steps = 100
+        callback.last_report_step = last_step
+        callback.start_time = start_time
+        callback.last_report_time = current_time - time_diff
+
+        # Mock time.time() to return consistent values
+        mock_time.time.return_value = current_time
+
+        callback.on_step_end(training_args, trainer_state, trainer_control)
+
+        # Should send telemetry
+        mock_telemetry_manager.send_event.assert_called_once()
+
+        # Properties should include expected values
+        props = mock_telemetry_manager.send_event.call_args[1]["properties"]
+        expected_metrics = calc_expected_metrics(
+            current_step, last_step, current_time, current_time - time_diff, start_time
+        )
+        assert props["steps_per_second"] == expected_metrics["steps_per_second"]
+        assert (
+            props["time_since_last_report"]
+            == expected_metrics["time_since_last_report"]
+        )
+
+    def test_on_step_end_first_step(
+        self,
+        callback,
+        mock_telemetry_manager,
+        mock_runtime_metrics_tracker,  # pylint: disable=unused-argument
+        mock_time,
+        training_args,
+        trainer_state,
+        trainer_control,
+    ):
+        """Test on_step_end always reports on first step."""
+        # Set up state with clear values
+        current_step = 1  # First step
+        last_step = 0
+        start_time = 900.0
+        current_time = 1000.0
+        last_report_time = 999.0  # Just 1 second ago
+
+        # Configure state and callback
+        trainer_state.global_step = current_step
+        callback.report_interval_steps = 100
+        callback.last_report_step = last_step
+        callback.start_time = start_time
+        callback.last_report_time = last_report_time
+
+        # Mock time.time() to return consistent values
+        mock_time.time.return_value = current_time
+
+        callback.on_step_end(training_args, trainer_state, trainer_control)
+
+        # Should send telemetry even though not much time has passed
+        mock_telemetry_manager.send_event.assert_called_once()
+
+        # Properties should include expected values for first step
+        props = mock_telemetry_manager.send_event.call_args[1]["properties"]
+        assert props["step"] == current_step
+        expected_metrics = calc_expected_metrics(
+            current_step, last_step, current_time, last_report_time, start_time
+        )
+        assert props["steps_per_second"] == expected_metrics["steps_per_second"]
+
+    def test_log_history_empty(
+        self,
+        callback,
+        mock_telemetry_manager,
+        mock_runtime_metrics_tracker,  # pylint: disable=unused-argument
+        mock_time,
+        training_args,
+        trainer_state,
+        trainer_control,
+    ):
+        """Test handling of empty log history."""
+        # Set up state with clear values
+        current_step = 1
+        start_time = 900.0
+        current_time = 1000.0
+
+        # Configure state and callback
+        trainer_state.global_step = current_step
+        trainer_state.log_history = []
+        callback.start_time = start_time
+
+        # Mock time.time() to return consistent values
+        mock_time.time.return_value = current_time
+
+        callback.on_step_end(training_args, trainer_state, trainer_control)
+
+        # Should still send telemetry
+        mock_telemetry_manager.send_event.assert_called_once()
+
+        # Properties should have default values for missing log data
+        props = mock_telemetry_manager.send_event.call_args[1]["properties"]
+        assert props["loss"] == 0
+        assert props["learning_rate"] == 0

tests/telemetry/test_errors.py

@@ -0,0 +1,340 @@
+"""Tests for telemetry error utilities"""
+# pylint: disable=redefined-outer-name
+
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+from axolotl.telemetry.errors import sanitize_stack_trace, send_errors
+
+
+@pytest.fixture(autouse=True)
+def reset_error_flag(monkeypatch):
+    """Reset ERROR_HANDLED flag using monkeypatch"""
+    import axolotl.telemetry.errors
+
+    monkeypatch.setattr(axolotl.telemetry.errors, "ERROR_HANDLED", False)
+    yield
+    monkeypatch.setattr(axolotl.telemetry.errors, "ERROR_HANDLED", False)
+
+
+@pytest.fixture
+def example_stack_trace():
+    """Provide a sample stack trace with mixed paths"""
+    return """Traceback (most recent call last):
+  File "/home/user/.local/lib/python3.9/site-packages/axolotl/cli/train.py", line 83, in main
+    trainer = get_trainer(cfg)
+  File "/home/user/.local/lib/python3.9/site-packages/axolotl/train.py", line 214, in get_trainer
+    model = get_model(cfg, tokenizer)
+  File "/home/user/.local/lib/python3.9/site-packages/axolotl/utils/models.py", line 120, in get_model
+    raise ValueError("Model path not found")
+ValueError: Model path not found
+"""
+
+
+@pytest.fixture
+def windows_stack_trace():
+    """Provide a sample stack trace with Windows paths"""
+    return """Traceback (most recent call last):
+  File "C:\\Users\\name\\AppData\\Local\\Programs\\Python\\Python39\\lib\\site-packages\\axolotl\\cli\\train.py", line 83, in main
+    trainer = get_trainer(cfg)
+  File "C:\\Users\\name\\AppData\\Local\\Programs\\Python\\Python39\\lib\\site-packages\\axolotl\\train.py", line 214, in get_trainer
+    model = get_model(cfg, tokenizer)
+  File "C:\\Users\\name\\AppData\\Local\\Programs\\Python\\Python39\\lib\\site-packages\\transformers\\models\\auto\\modeling_auto.py", line 482, in from_pretrained
+    raise ValueError(f"Unrecognized configuration class {config.__class__}")
+ValueError: Unrecognized configuration class <class 'transformers.models.llama.configuration_llama.LlamaConfig'>
+"""
+
+
+@pytest.fixture
+def mixed_stack_trace():
+    """Provide a sample stack trace with both axolotl and non-axolotl paths"""
+    return """Traceback (most recent call last):
+  File "/home/user/.local/lib/python3.9/site-packages/axolotl/cli/train.py", line 83, in main
+    trainer = get_trainer(cfg)
+  File "/home/user/.local/lib/python3.9/site-packages/transformers/trainer.py", line 520, in train
+    self._inner_training_loop()
+  File "/home/user/.local/lib/python3.9/site-packages/axolotl/utils/trainer.py", line 75, in _inner_training_loop
+    super()._inner_training_loop()
+  File "/home/user/.local/lib/python3.9/site-packages/torch/utils/data/dataloader.py", line 631, in __next__
+    data = self._next_data()
+RuntimeError: CUDA out of memory
+"""
+
+
+@pytest.fixture
+def venv_stack_trace():
+    """Provide a sample stack trace with virtual environment paths"""
+    return """Traceback (most recent call last):
+  File "/home/user/venv/lib/python3.9/site-packages/transformers/trainer.py", line 1729, in train
+    self._inner_training_loop()
+  File "/home/user/venv/lib/python3.9/site-packages/transformers/trainer.py", line 2013, in _inner_training_loop
+    self.accelerator.backward(loss)
+  File "/home/user/venv/lib/python3.9/site-packages/accelerate/accelerator.py", line 1851, in backward
+    self.scaler.scale(loss).backward(**kwargs)
+  File "/home/user/venv/lib/python3.9/site-packages/torch/_tensor.py", line 487, in backward
+    torch.autograd.backward(self, gradient, retain_graph, create_graph, inputs=inputs)
+RuntimeError: CUDA out of memory
+"""
+
+
+@pytest.fixture
+def dist_packages_stack_trace():
+    """Provide a sample stack trace with dist-packages paths"""
+    return """Traceback (most recent call last):
+  File "/usr/local/lib/python3.8/dist-packages/torch/utils/data/dataloader.py", line 631, in __next__
+    data = self._next_data()
+  File "/usr/local/lib/python3.8/dist-packages/torch/utils/data/dataloader.py", line 675, in _next_data
+    data = self._dataset_fetcher.fetch(index)
+  File "/usr/local/lib/python3.8/dist-packages/torch/utils/data/_utils/fetch.py", line 51, in fetch
+    data = [self.dataset[idx] for idx in possibly_batched_index]
+  File "/usr/local/lib/python3.8/dist-packages/datasets/arrow_dataset.py", line 2808, in __getitem__
+    raise IndexError(f"Index {key} out of range for dataset of length {len(self)}.")
+IndexError: Index 10000 out of range for dataset of length 9832.
+"""
+
+
+@pytest.fixture
+def project_stack_trace():
+    """Provide a sample stack trace from a project directory (not a virtual env)"""
+    return """Traceback (most recent call last):
+  File "/home/user/projects/myproject/run.py", line 25, in <module>
+    main()
+  File "/home/user/projects/myproject/src/cli.py", line 45, in main
+    app.run()
+  File "/home/user/projects/myproject/src/app.py", line 102, in run
+    raise ValueError("Configuration missing")
+ValueError: Configuration missing
+"""
+
+
+def test_sanitize_stack_trace(example_stack_trace):
+    """Test that sanitize_stack_trace properly preserves axolotl paths"""
+    sanitized = sanitize_stack_trace(example_stack_trace)
+
+    # Check that personal paths are removed
+    assert "/home/user" not in sanitized
+    assert ".local/lib/python3.9" not in sanitized
+
+    # Check that site-packages is preserved
+    assert "site-packages/axolotl/cli/train.py" in sanitized
+    assert "site-packages/axolotl/train.py" in sanitized
+    assert "site-packages/axolotl/utils/models.py" in sanitized
+
+    # Check that error message is preserved
+    assert "ValueError: Model path not found" in sanitized
+
+
+def test_sanitize_windows_paths(windows_stack_trace):
+    """Test that sanitize_stack_trace handles Windows paths"""
+    sanitized = sanitize_stack_trace(windows_stack_trace)
+
+    # Check that personal paths are removed
+    assert "C:\\Users\\name" not in sanitized
+    assert "AppData\\Local\\Programs\\Python" not in sanitized
+
+    # Check that both axolotl and transformers packages are preserved
+    assert (
+        "site-packages\\axolotl\\cli\\train.py" in sanitized
+        or "site-packages/axolotl/cli/train.py" in sanitized
+    )
+    assert (
+        "site-packages\\axolotl\\train.py" in sanitized
+        or "site-packages/axolotl/train.py" in sanitized
+    )
+    assert (
+        "site-packages\\transformers\\models\\auto\\modeling_auto.py" in sanitized
+        or "site-packages/transformers/models/auto/modeling_auto.py" in sanitized
+    )
+
+    # Check that error message is preserved
+    assert "ValueError: Unrecognized configuration class" in sanitized
+
+
+def test_sanitize_mixed_paths(mixed_stack_trace):
+    """Test that sanitize_stack_trace preserves all package paths"""
+    sanitized = sanitize_stack_trace(mixed_stack_trace)
+
+    # Check that all package paths are preserved
+    assert "site-packages/axolotl/cli/train.py" in sanitized
+    assert "site-packages/transformers/trainer.py" in sanitized
+    assert "site-packages/axolotl/utils/trainer.py" in sanitized
+    assert "site-packages/torch/utils/data/dataloader.py" in sanitized
+
+    # Check that error message is preserved
+    assert "RuntimeError: CUDA out of memory" in sanitized
+
+
+def test_sanitize_venv_paths(venv_stack_trace):
+    """Test that sanitize_stack_trace preserves virtual environment package paths"""
+    sanitized = sanitize_stack_trace(venv_stack_trace)
+
+    # Check that personal paths are removed
+    assert "/home/user/venv" not in sanitized
+
+    # Check that all package paths are preserved
+    assert "site-packages/transformers/trainer.py" in sanitized
+    assert "site-packages/accelerate/accelerator.py" in sanitized
+    assert "site-packages/torch/_tensor.py" in sanitized
+
+    # Check that error message is preserved
+    assert "RuntimeError: CUDA out of memory" in sanitized
+
+
+def test_sanitize_dist_packages(dist_packages_stack_trace):
+    """Test that sanitize_stack_trace preserves dist-packages paths"""
+    sanitized = sanitize_stack_trace(dist_packages_stack_trace)
+
+    # Check that system paths are removed
+    assert "/usr/local/lib/python3.8" not in sanitized
+
+    # Check that all package paths are preserved
+    assert "dist-packages/torch/utils/data/dataloader.py" in sanitized
+    assert "dist-packages/torch/utils/data/_utils/fetch.py" in sanitized
+    assert "dist-packages/datasets/arrow_dataset.py" in sanitized
+
+    # Check that error message is preserved
+    assert (
+        "IndexError: Index 10000 out of range for dataset of length 9832." in sanitized
+    )
+
+
+def test_sanitize_project_paths(project_stack_trace):
+    """Test handling of project paths (non-virtual env)"""
+    sanitized = sanitize_stack_trace(project_stack_trace)
+
+    # Check that personal paths are removed
+    assert "/home/user/projects" not in sanitized
+
+    # For non-package paths, we should at least preserve the filename
+    assert "run.py" in sanitized
+    assert "cli.py" in sanitized
+    assert "app.py" in sanitized
+
+    # Check that error message is preserved
+    assert "ValueError: Configuration missing" in sanitized
+
+
+@pytest.fixture
+def mock_telemetry_manager():
+    """Create a mock TelemetryManager"""
+    with patch("axolotl.telemetry.errors.TelemetryManager") as mock_manager_class:
+        mock_manager = MagicMock()
+        mock_manager.enabled = True
+        mock_manager_class.get_instance.return_value = mock_manager
+        yield mock_manager
+
+
+def test_send_errors_successful_execution(mock_telemetry_manager):
+    """Test that send_errors doesn't send telemetry for successful function execution"""
+
+    @send_errors
+    def test_func():
+        return "success"
+
+    result = test_func()
+    assert result == "success"
+    mock_telemetry_manager.send_event.assert_not_called()
+
+
+def test_send_errors_with_exception(mock_telemetry_manager):
+    """Test that send_errors sends telemetry when an exception occurs"""
+    test_error = ValueError("Test error")
+
+    @send_errors
+    def test_func():
+        raise test_error
+
+    with pytest.raises(ValueError) as excinfo:
+        test_func()
+
+    assert excinfo.value == test_error
+    mock_telemetry_manager.send_event.assert_called_once()
+
+    # Check that the error info was passed correctly
+    call_args = mock_telemetry_manager.send_event.call_args[1]
+    assert "test_func-errored" in call_args["event_type"]
+    assert "Test error" in call_args["properties"]["exception"]
+    assert "stack_trace" in call_args["properties"]
+
+
+def test_send_errors_nested_calls(mock_telemetry_manager):
+    """Test that send_errors only sends telemetry once for nested decorated functions"""
+
+    @send_errors
+    def inner_func():
+        raise ValueError("Inner error")
+
+    @send_errors
+    def outer_func():
+        return inner_func()
+
+    with pytest.raises(ValueError):
+        outer_func()
+
+    # Telemetry should be sent only once for the inner function
+    assert mock_telemetry_manager.send_event.call_count == 1
+    call_args = mock_telemetry_manager.send_event.call_args[1]
+    assert "inner_func-error" in call_args["event_type"]
+
+
+def test_send_errors_telemetry_disable():
+    """Test that send_errors doesn't attempt to send telemetry when disabled"""
+
+    with patch("axolotl.telemetry.errors.TelemetryManager") as mock_manager_class:
+        mock_manager = MagicMock()
+        mock_manager.enabled = False
+        mock_manager_class.get_instance.return_value = mock_manager
+
+        @send_errors
+        def test_func():
+            raise ValueError("Test error")
+
+        with pytest.raises(ValueError):
+            test_func()
+
+        mock_manager.send_event.assert_not_called()
+
+
+def test_error_handled_reset():
+    """Test that ERROR_HANDLED flag is properly reset"""
+    with patch("axolotl.telemetry.errors.TelemetryManager") as mock_manager_class:
+        # Create and configure the mock manager
+        mock_manager = MagicMock()
+        mock_manager.enabled = True
+        mock_manager_class.get_instance.return_value = mock_manager
+
+        from axolotl.telemetry.errors import ERROR_HANDLED
+
+        @send_errors
+        def test_func():
+            raise ValueError("Test error")
+
+        assert not ERROR_HANDLED
+
+        with pytest.raises(ValueError):
+            test_func()
+
+        from axolotl.telemetry.errors import ERROR_HANDLED
+
+        assert ERROR_HANDLED
+
+
+def test_module_path_resolution(mock_telemetry_manager):
+    """Test that the module path is correctly resolved for the event type"""
+    import inspect
+
+    current_module = inspect.getmodule(test_module_path_resolution).__name__
+
+    @send_errors
+    def test_func():
+        raise ValueError("Test error")
+
+    with pytest.raises(ValueError):
+        test_func()
+
+    assert mock_telemetry_manager.send_event.called
+    event_type = mock_telemetry_manager.send_event.call_args[1]["event_type"]
+
+    expected_event_type = f"{current_module}.test_func-errored"
+    assert expected_event_type == event_type

tests/telemetry/test_manager.py

@@ -0,0 +1,245 @@
+"""Tests for TelemetryManager class and utilities"""
+# pylint: disable=redefined-outer-name,protected-access
+
+import os
+from unittest.mock import patch
+
+import pytest
+import yaml
+
+from axolotl.telemetry.manager import TelemetryManager
+
+
+@pytest.fixture
+def mock_whitelist(tmp_path):
+    """Create a temporary whitelist file for testing"""
+    whitelist_content = {
+        "organizations": ["meta-llama", "mistralai"],
+    }
+    whitelist_file = tmp_path / "whitelist.yaml"
+    with open(whitelist_file, "w", encoding="utf-8") as f:
+        yaml.dump(whitelist_content, f)
+
+    return str(whitelist_file)
+
+
+@pytest.fixture
+def telemetry_manager_class():
+    """Reset the TelemetryManager singleton between tests"""
+    original_instance = TelemetryManager._instance
+    original_initialized = TelemetryManager._initialized
+    TelemetryManager._instance = None
+    TelemetryManager._initialized = False
+    yield TelemetryManager
+    TelemetryManager._instance = original_instance
+    TelemetryManager._initialized = original_initialized
+
+
+@pytest.fixture
+def manager(telemetry_manager_class, mock_whitelist):
+    """Create a TelemetryManager instance with mocked dependencies"""
+    with patch("posthog.capture"), patch("posthog.flush"), patch("time.sleep"), patch(
+        "axolotl.telemetry.manager.WHITELIST_PATH", mock_whitelist
+    ), patch.dict(os.environ, {"RANK": "0"}):
+        manager = telemetry_manager_class()
+        # Manually enable for most tests
+        manager.enabled = True
+        return manager
+
+
+def test_singleton_instance(telemetry_manager_class):
+    """Test that TelemetryManager is a singleton"""
+    with patch("posthog.capture"), patch("time.sleep"), patch.dict(
+        os.environ, {"RANK": "0"}
+    ):
+        first = telemetry_manager_class()
+        second = telemetry_manager_class()
+        assert first is second
+        assert telemetry_manager_class.get_instance() is first
+
+
+def test_telemetry_disabled_with_axolotl_do_not_track(telemetry_manager_class):
+    """Test that telemetry is disabled when AXOLOTL_DO_NOT_TRACK=1"""
+    with patch.dict(os.environ, {"AXOLOTL_DO_NOT_TRACK": "1", "RANK": "0"}):
+        manager = telemetry_manager_class()
+        assert not manager.enabled
+
+
+def test_telemetry_disabled_with_do_not_track(telemetry_manager_class):
+    """Test that telemetry is disabled when DO_NOT_TRACK=1"""
+    with patch.dict(os.environ, {"DO_NOT_TRACK": "1", "RANK": "0"}):
+        manager = telemetry_manager_class()
+        assert not manager.enabled
+
+
+def test_telemetry_disabled_for_non_main_process(telemetry_manager_class):
+    """Test that telemetry is disabled for non-main processes"""
+    with patch.dict(os.environ, {"AXOLOTL_DO_NOT_TRACK": "0", "RANK": "1"}):
+        manager = telemetry_manager_class()
+        assert not manager.enabled
+
+
+def test_telemetry_enabled_by_default(telemetry_manager_class):
+    """Test that telemetry is enabled by default"""
+    with patch.dict(os.environ, {"RANK": "0"}, clear=True), patch("time.sleep"), patch(
+        "logging.Logger.warning"
+    ):
+        manager = telemetry_manager_class()
+        assert manager.enabled
+        assert not manager.explicit_enable
+
+
+def test_explicit_enable_disables_warning(telemetry_manager_class):
+    """Test that explicit enabling prevents warning"""
+    with patch.dict(os.environ, {"AXOLOTL_DO_NOT_TRACK": "0", "RANK": "0"}), patch(
+        "logging.Logger.warning"
+    ) as mock_warning, patch("time.sleep"):
+        manager = telemetry_manager_class()
+        assert manager.enabled
+        assert manager.explicit_enable
+        for call in mock_warning.call_args_list:
+            assert "Telemetry is enabled" not in str(call)
+
+
+def test_warning_displayed_for_implicit_enable(telemetry_manager_class):
+    """Test that warning is displayed when telemetry is implicitly enabled"""
+    with patch.dict(os.environ, {"RANK": "0"}, clear=True), patch(
+        "logging.Logger.warning"
+    ) as mock_warning, patch("time.sleep"):
+        manager = telemetry_manager_class()
+        assert manager.enabled
+        assert not manager.explicit_enable
+        warning_displayed = False
+        for call in mock_warning.call_args_list:
+            if "Telemetry is enabled" in str(call):
+                warning_displayed = True
+                break
+        assert warning_displayed
+
+
+def test_is_whitelisted(manager, mock_whitelist):
+    """Test org whitelist functionality"""
+    with patch("axolotl.telemetry.manager.WHITELIST_PATH", mock_whitelist):
+        # Should match organizations from the mock whitelist
+        assert manager._is_whitelisted("meta-llama/llama-7b")
+        assert manager._is_whitelisted("mistralai/mistral-7b-instruct")
+        # Should not match
+        assert not manager._is_whitelisted("unknown/model")
+        # Should handle case insensitively
+        assert manager._is_whitelisted("META-LLAMA/Llama-7B")
+        # Should handle empty input
+        assert not manager._is_whitelisted("")
+        assert not manager._is_whitelisted(None)
+
+
+def test_system_info_collection(manager):
+    """Test system information collection"""
+    system_info = manager._get_system_info()
+
+    # Check essential keys
+    assert "os" in system_info
+    assert "python_version" in system_info
+    assert "torch_version" in system_info
+    assert "transformers_version" in system_info
+    assert "axolotl_version" in system_info
+    assert "cpu_count" in system_info
+    assert "memory_total" in system_info
+    assert "accelerator_count" in system_info
+
+
+def test_send_event(manager):
+    """Test basic event sending"""
+    with patch("posthog.capture") as mock_capture:
+        # Test with clean properties (no PII)
+        manager.send_event("test_event", {"key": "value"})
+        assert mock_capture.called
+        assert mock_capture.call_args[1]["event"] == "test_event"
+        assert mock_capture.call_args[1]["properties"] == {"key": "value"}
+        assert mock_capture.call_args[1]["distinct_id"] == manager.run_id
+
+        # Test with default properties (None)
+        mock_capture.reset_mock()
+        manager.send_event("simple_event")
+        assert mock_capture.called
+        assert mock_capture.call_args[1]["properties"] == {}
+
+
+def test_send_system_info(manager):
+    """Test sending system info"""
+    with patch("posthog.capture") as mock_capture:
+        manager.send_system_info()
+        assert mock_capture.called
+        assert mock_capture.call_args[1]["event"] == "system-info"
+        assert mock_capture.call_args[1]["properties"] == manager.system_info
+
+
+def test_redacted_properties(manager):
+    """Test path redaction in send_event method"""
+    with patch("posthog.capture") as mock_capture:
+        # Test with properties containing various paths and non-paths
+        test_properties = {
+            "filepath": "/home/user/sensitive/data.txt",
+            "windows_path": "C:\\Users\\name\\Documents\\project\\file.py",
+            "output_dir": "/var/lib/data",
+            "path_to_model": "models/llama/7b",
+            "message": "Training started",  # Should not be redacted
+            "metrics": {"loss": 0.5, "accuracy": 0.95},  # Should not be redacted
+            "base_model": "models/local_model",
+            "nested": {
+                "model_path": "/models/my_model",
+                "root_dir": "/home/user/projects",
+                "stats": {"steps": 1000, "epochs": 3},  # Should not be redacted
+            },
+        }
+
+        manager.send_event("test_event", test_properties)
+
+        # Verify the call was made
+        assert mock_capture.called
+
+        # Get the sanitized properties that were sent
+        sanitized = mock_capture.call_args[1]["properties"]
+
+        # Check that path-like and base_model keys were redacted
+        assert sanitized["filepath"] == "[REDACTED]"
+        assert sanitized["windows_path"] == "[REDACTED]"
+        assert sanitized["path_to_model"] == "[REDACTED]"
+        assert sanitized["base_model"] == "[REDACTED]"
+
+        # Check that non-path values were preserved
+        assert sanitized["message"] == "Training started"
+        assert sanitized["metrics"] == {"loss": 0.5, "accuracy": 0.95}
+
+        # Check nested structure handling
+        assert sanitized["nested"]["model_path"] == "[REDACTED]"
+        assert sanitized["nested"]["root_dir"] == "[REDACTED]"
+        assert sanitized["nested"]["stats"] == {"steps": 1000, "epochs": 3}
+
+
+def test_disable_telemetry(manager):
+    """Test that disabled telemetry doesn't send events"""
+    with patch("posthog.capture") as mock_capture:
+        manager.enabled = False
+        manager.send_event("test_event")
+        assert not mock_capture.called
+
+
+def test_exception_handling_during_send(manager):
+    """Test that exceptions in PostHog are handled gracefully"""
+    with patch("posthog.capture", side_effect=Exception("Test error")), patch(
+        "logging.Logger.warning"
+    ) as mock_warning:
+        manager.send_event("test_event")
+        warning_logged = False
+        for call in mock_warning.call_args_list:
+            if "Failed to send telemetry event" in str(call):
+                warning_logged = True
+                break
+        assert warning_logged
+
+
+def test_shutdown(manager):
+    """Test shutdown behavior"""
+    with patch("posthog.flush") as mock_flush:
+        manager.shutdown()
+        assert mock_flush.called

tests/telemetry/test_runtime_metrics.py

@@ -0,0 +1,356 @@
+"""Tests for runtime metrics telemetry module"""
+# pylint: disable=redefined-outer-name
+
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+from axolotl.telemetry.runtime_metrics import RuntimeMetrics, RuntimeMetricsTracker
+
+
+@pytest.fixture
+def mock_time():
+    """Mock time.time() to have predictable values in tests"""
+    with patch("time.time") as mock_time:
+        # Start with time 1000.0 and increment by 10 seconds on each call
+        times = [1000.0 + i * 10 for i in range(10)]
+        mock_time.side_effect = times
+        yield mock_time
+
+
+@pytest.fixture
+def mock_telemetry_manager():
+    """Create a mock TelemetryManager"""
+    with patch(
+        "axolotl.telemetry.runtime_metrics.TelemetryManager"
+    ) as mock_manager_class:
+        mock_manager = MagicMock()
+        mock_manager.enabled = True
+        mock_manager_class.get_instance.return_value = mock_manager
+        yield mock_manager
+
+
+@pytest.fixture
+def mock_psutil():
+    """Mock psutil for memory information"""
+    with patch("axolotl.telemetry.runtime_metrics.psutil") as mock_psutil:
+        mock_process = MagicMock()
+        mock_memory_info = MagicMock()
+        # Set initial memory to 1GB
+        mock_memory_info.rss = 1024 * 1024 * 1024
+        mock_process.memory_info.return_value = mock_memory_info
+        mock_psutil.Process.return_value = mock_process
+        yield mock_psutil
+
+
+@pytest.fixture
+def mock_torch():
+    """Mock torch.cuda functions"""
+    with patch("axolotl.telemetry.runtime_metrics.torch") as mock_torch:
+        mock_torch.cuda.is_available.return_value = True
+        mock_torch.cuda.device_count.return_value = 2
+
+        # Mock memory allocated per device (1GB for device 0, 2GB for device 1)
+        mock_torch.cuda.memory_allocated.side_effect = (
+            lambda device: (device + 1) * 1024 * 1024 * 1024
+        )
+
+        yield mock_torch
+
+
+class TestRuntimeMetrics:
+    """Tests for RuntimeMetrics class."""
+
+    def test_initialization(self):
+        """Test RuntimeMetrics initialization."""
+        metrics = RuntimeMetrics(start_time=1000.0)
+
+        assert metrics.start_time == 1000.0
+        assert metrics.epoch_start_times == {}
+        assert metrics.epoch_end_times == {}
+        assert metrics.peak_gpu_memory == {}
+        assert metrics.total_steps == 0
+        assert metrics.current_epoch == 0
+        assert metrics.current_step == 0
+        assert metrics.peak_cpu_memory == 0
+
+    def test_elapsed_time(self, mock_time):
+        """Test elapsed_time property."""
+        metrics = RuntimeMetrics(start_time=1000.0)
+
+        # Mock time.time() to return 1050.0
+        mock_time.side_effect = [1050.0]
+
+        assert metrics.elapsed_time == 50.0
+
+    def test_epoch_time(self):
+        """Test epoch_time method."""
+        metrics = RuntimeMetrics(start_time=1000.0)
+
+        # No epoch data
+        assert metrics.epoch_time(0) is None
+
+        # Add epoch start but no end
+        metrics.epoch_start_times[0] = 1000.0
+        assert metrics.epoch_time(0) is None
+
+        # Add epoch end
+        metrics.epoch_end_times[0] = 1060.0
+        assert metrics.epoch_time(0) == 60.0
+
+    def test_average_epoch_time(self):
+        """Test average_epoch_time method."""
+        metrics = RuntimeMetrics(start_time=1000.0)
+
+        # No completed epochs
+        assert metrics.average_epoch_time() is None
+
+        # Add one completed epoch
+        metrics.epoch_start_times[0] = 1000.0
+        metrics.epoch_end_times[0] = 1060.0
+        assert metrics.average_epoch_time() == 60.0
+
+        # Add second completed epoch
+        metrics.epoch_start_times[1] = 1060.0
+        metrics.epoch_end_times[1] = 1140.0  # 80 seconds
+        assert metrics.average_epoch_time() == 70.0  # Average of 60 and 80
+
+        # Add incomplete epoch (should not affect average)
+        metrics.epoch_start_times[2] = 1140.0
+        assert metrics.average_epoch_time() == 70.0
+
+    def test_steps_per_second(self, mock_time):
+        """Test steps_per_second method."""
+        metrics = RuntimeMetrics(start_time=1000.0)
+
+        # No steps - first call to time.time()
+        mock_time.side_effect = None
+        mock_time.return_value = 1050.0
+        assert metrics.steps_per_second() is None
+
+        # Add steps - second call to time.time()
+        metrics.total_steps = 100
+        mock_time.return_value = 1050.0  # Keep same time for consistent result
+        assert metrics.steps_per_second() == 2.0  # 100 steps / 50 seconds
+
+    def test_to_dict_basic(self, mock_time):
+        """Test to_dict method with basic metrics."""
+        metrics = RuntimeMetrics(start_time=1000.0)
+        metrics.total_steps = 100
+        metrics.peak_cpu_memory = 2 * 1024 * 1024 * 1024  # 2GB
+
+        # Mock elapsed_time
+        mock_time.side_effect = None
+        mock_time.return_value = 1050.0
+
+        result = metrics.to_dict()
+
+        assert result["total_time_seconds"] == 50.0
+        assert result["total_steps"] == 100
+        assert result["steps_per_second"] == 2.0
+        assert result["epochs_completed"] == 0
+        assert result["peak_cpu_memory_bytes"] == 2 * 1024 * 1024 * 1024
+        assert "epoch_times" not in result
+        assert "gpu_memory" not in result
+
+    def test_to_dict_with_epochs(self, mock_time):
+        """Test to_dict method with epoch data."""
+        metrics = RuntimeMetrics(start_time=1000.0)
+        metrics.total_steps = 100
+
+        # Add epoch data
+        metrics.epoch_start_times[0] = 1000.0
+        metrics.epoch_end_times[0] = 1060.0
+        metrics.epoch_start_times[1] = 1060.0
+        metrics.epoch_end_times[1] = 1140.0
+
+        # Mock elapsed_time
+        mock_time.side_effect = None
+        mock_time.return_value = 1150.0
+
+        result = metrics.to_dict()
+
+        assert "epoch_times" in result
+        assert result["epoch_times"]["epoch_0_seconds"] == 60.0
+        assert result["epoch_times"]["epoch_1_seconds"] == 80.0
+        assert result["average_epoch_time_seconds"] == 70.0
+
+    def test_to_dict_with_gpu_memory(self, mock_time):
+        """Test to_dict method with GPU memory data."""
+        metrics = RuntimeMetrics(start_time=1000.0)
+        metrics.peak_gpu_memory = {
+            0: 1 * 1024 * 1024 * 1024,  # 1GB
+            1: 2 * 1024 * 1024 * 1024,  # 2GB
+        }
+
+        # Mock elapsed_time
+        mock_time.side_effect = [1050.0]
+
+        result = metrics.to_dict()
+
+        assert "gpu_memory" in result
+        assert result["gpu_memory"]["gpu_0_peak_memory_bytes"] == 1 * 1024 * 1024 * 1024
+        assert result["gpu_memory"]["gpu_1_peak_memory_bytes"] == 2 * 1024 * 1024 * 1024
+
+
+class TestRuntimeMetricsTracker:
+    """Tests for RuntimeMetricsTracker class."""
+
+    # pylint: disable=unused-argument
+    def test_initialization(self, mock_time, mock_telemetry_manager):
+        """Test RuntimeMetricsTracker initialization."""
+        tracker = RuntimeMetricsTracker()
+
+        assert isinstance(tracker.metrics, RuntimeMetrics)
+        assert tracker.metrics.start_time == 1000.0  # First value from mock_time
+
+    # pylint: disable=unused-argument
+    def test_start_epoch(
+        self, mock_time, mock_psutil, mock_torch, mock_telemetry_manager
+    ):
+        """Test start_epoch method."""
+        tracker = RuntimeMetricsTracker()
+
+        # Reset mock_time to control next value
+        mock_time.side_effect = [1010.0]
+
+        tracker.start_epoch(0)
+
+        assert tracker.metrics.current_epoch == 0
+        assert tracker.metrics.epoch_start_times[0] == 1010.0
+
+        # Verify memory metrics were updated
+        assert tracker.metrics.peak_cpu_memory == 1 * 1024 * 1024 * 1024
+        assert 0 in tracker.metrics.peak_gpu_memory
+        assert 1 in tracker.metrics.peak_gpu_memory
+
+    # pylint: disable=unused-argument
+    def test_end_epoch(self, mock_time, mock_telemetry_manager):
+        """Test end_epoch method."""
+        tracker = RuntimeMetricsTracker()
+
+        # Start epoch 0
+        mock_time.side_effect = [1010.0]
+        tracker.start_epoch(0)
+
+        # End epoch 0
+        mock_time.side_effect = [1060.0]
+        tracker.end_epoch(0)
+
+        assert 0 in tracker.metrics.epoch_end_times
+        assert tracker.metrics.epoch_end_times[0] == 1060.0
+
+    # pylint: disable=unused-argument
+    def test_update_step(
+        self, mock_time, mock_psutil, mock_torch, mock_telemetry_manager
+    ):
+        """Test update_step method."""
+        tracker = RuntimeMetricsTracker()
+
+        # Update step to a non-multiple of 100
+        tracker.update_step(42)
+
+        assert tracker.metrics.current_step == 42
+        assert tracker.metrics.total_steps == 1
+
+        # Memory metrics should not be updated for non-multiple of 100
+        assert tracker.metrics.peak_cpu_memory == 0
+
+        # Update step to a multiple of 100
+        tracker.update_step(100)
+
+        assert tracker.metrics.current_step == 100
+        assert tracker.metrics.total_steps == 2
+
+        # Memory metrics should be updated for multiple of 100
+        assert tracker.metrics.peak_cpu_memory == 1 * 1024 * 1024 * 1024
+
+    # pylint: disable=unused-argument
+    def test_update_memory_metrics(
+        self, mock_psutil, mock_torch, mock_telemetry_manager
+    ):
+        """Test update_memory_metrics method."""
+        tracker = RuntimeMetricsTracker()
+
+        # Initial memory state
+        assert tracker.metrics.peak_cpu_memory == 0
+        assert tracker.metrics.peak_gpu_memory == {}
+
+        # Update memory metrics
+        tracker.update_memory_metrics()
+
+        # Verify CPU memory
+        assert tracker.metrics.peak_cpu_memory == 1 * 1024 * 1024 * 1024
+
+        # Verify GPU memory
+        assert tracker.metrics.peak_gpu_memory[0] == 1 * 1024 * 1024 * 1024
+        assert tracker.metrics.peak_gpu_memory[1] == 2 * 1024 * 1024 * 1024
+
+        # Change mocked memory values to be lower
+        mock_process = mock_psutil.Process.return_value
+        mock_memory_info = mock_process.memory_info.return_value
+        mock_memory_info.rss = 0.5 * 1024 * 1024 * 1024  # 0.5GB
+
+        mock_torch.cuda.memory_allocated.side_effect = (
+            lambda device: (device + 0.5) * 1024 * 1024 * 1024
+        )
+
+        # Update memory metrics again
+        tracker.update_memory_metrics()
+
+        # Peak values should not decrease
+        assert tracker.metrics.peak_cpu_memory == 1 * 1024 * 1024 * 1024
+        assert tracker.metrics.peak_gpu_memory[0] == 1 * 1024 * 1024 * 1024
+        assert tracker.metrics.peak_gpu_memory[1] == 2 * 1024 * 1024 * 1024
+
+        # Change mocked memory values to be higher
+        mock_memory_info.rss = 2 * 1024 * 1024 * 1024  # 2GB
+
+        mock_torch.cuda.memory_allocated.side_effect = (
+            lambda device: (device + 2) * 1024 * 1024 * 1024
+        )
+
+        # Update memory metrics again
+        tracker.update_memory_metrics()
+
+        # Peak values should increase
+        assert tracker.metrics.peak_cpu_memory == 2 * 1024 * 1024 * 1024
+        assert tracker.metrics.peak_gpu_memory[0] == 2 * 1024 * 1024 * 1024
+        assert tracker.metrics.peak_gpu_memory[1] == 3 * 1024 * 1024 * 1024
+
+    # pylint: disable=unused-argument
+    def test_get_memory_metrics(self, mock_psutil, mock_torch, mock_telemetry_manager):
+        """Test get_memory_metrics method."""
+        tracker = RuntimeMetricsTracker()
+
+        # Set peak memory values
+        tracker.metrics.peak_cpu_memory = 2 * 1024 * 1024 * 1024
+        tracker.metrics.peak_gpu_memory = {
+            0: 3 * 1024 * 1024 * 1024,
+            1: 4 * 1024 * 1024 * 1024,
+        }
+
+        # Get memory metrics
+        memory_metrics = tracker.get_memory_metrics()
+
+        # Verify CPU memory
+        assert (
+            memory_metrics["cpu_memory_bytes"] == 1 * 1024 * 1024 * 1024
+        )  # Current value from mock
+        assert (
+            memory_metrics["peak_cpu_memory_bytes"] == 2 * 1024 * 1024 * 1024
+        )  # Peak value we set
+
+        # Verify GPU memory
+        assert (
+            memory_metrics["gpu_0_memory_bytes"] == 1 * 1024 * 1024 * 1024
+        )  # Current value from mock
+        assert (
+            memory_metrics["gpu_0_peak_memory_bytes"] == 3 * 1024 * 1024 * 1024
+        )  # Peak value we set
+        assert (
+            memory_metrics["gpu_1_memory_bytes"] == 2 * 1024 * 1024 * 1024
+        )  # Current value from mock
+        assert (
+            memory_metrics["gpu_1_peak_memory_bytes"] == 4 * 1024 * 1024 * 1024
+        )  # Peak value we set