Skip to content

context_mixin

optimus_dl.recipe.train.mixins.execution.context_mixin

Training context mixin for AMP and gradient scaler setup.

TrainingContextMixin

Mixin for setting up the training context (precision, scaling, devices).

Responsible for initializing PyTorch's AMP (Automatic Mixed Precision) and GradScaler based on the optimization configuration. This ensures consistent precision settings across the training loop.

Parameters:

Name Type Description Default
optimization_config OptimizationConfig

Configuration containing AMP settings.

 required
Source code in optimus_dl/recipe/train/mixins/execution/context_mixin.py 
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
class TrainingContextMixin:
    """Mixin for setting up the training context (precision, scaling, devices).

    Responsible for initializing PyTorch's AMP (Automatic Mixed Precision) and
    GradScaler based on the optimization configuration. This ensures consistent
    precision settings across the training loop.

    Args:
        optimization_config: Configuration containing AMP settings.
    """

    def __init__(self, optimization_config: OptimizationConfig):
        self.optimization_config = optimization_config

    def setup_training_context(self, device: torch.device) -> dict[str, Any]:
        """Initialize AMP context and Gradient Scaler.

        Args:
            device: The target compute device.

        Returns:
            A dictionary containing:

            - "scaler": The torch.cuda.amp.GradScaler instance.
            - "amp_ctx": The torch.autocast context manager.
            - "amp_cfg": The raw AMP configuration object.
            - "device": The device being used.
        """
        amp_cfg = self.optimization_config.amp
        scaler = torch.GradScaler(
            device=device.type,
            enabled=amp_cfg.enabled and amp_cfg.enable_scaler,
            init_scale=amp_cfg.init_scale,
            growth_factor=amp_cfg.growth_factor,
            backoff_factor=amp_cfg.backoff_factor,
            growth_interval=amp_cfg.growth_interval,
        )
        logger.info(f"Using grad scaler: {scaler.is_enabled()}")
        # Safe dtype conversion without eval()
        dtype = str_to_dtype(amp_cfg.dtype)
        amp_ctx = torch.autocast(device.type, dtype=dtype, enabled=amp_cfg.enabled)

        return {
            "scaler": scaler,
            "amp_ctx": amp_ctx,
            "amp_cfg": amp_cfg,
            "device": device,
        }

setup_training_context(device)

Initialize AMP context and Gradient Scaler.

Parameters:

Name Type Description Default
device device

The target compute device.

 required

Returns:

Type Description
dict[str, Any]

A dictionary containing:
dict[str, Any]
  • "scaler": The torch.cuda.amp.GradScaler instance.
dict[str, Any]
  • "amp_ctx": The torch.autocast context manager.
dict[str, Any]
  • "amp_cfg": The raw AMP configuration object.
dict[str, Any]
  • "device": The device being used.
Source code in optimus_dl/recipe/train/mixins/execution/context_mixin.py 
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
def setup_training_context(self, device: torch.device) -> dict[str, Any]:
    """Initialize AMP context and Gradient Scaler.

    Args:
        device: The target compute device.

    Returns:
        A dictionary containing:

        - "scaler": The torch.cuda.amp.GradScaler instance.
        - "amp_ctx": The torch.autocast context manager.
        - "amp_cfg": The raw AMP configuration object.
        - "device": The device being used.
    """
    amp_cfg = self.optimization_config.amp
    scaler = torch.GradScaler(
        device=device.type,
        enabled=amp_cfg.enabled and amp_cfg.enable_scaler,
        init_scale=amp_cfg.init_scale,
        growth_factor=amp_cfg.growth_factor,
        backoff_factor=amp_cfg.backoff_factor,
        growth_interval=amp_cfg.growth_interval,
    )
    logger.info(f"Using grad scaler: {scaler.is_enabled()}")
    # Safe dtype conversion without eval()
    dtype = str_to_dtype(amp_cfg.dtype)
    amp_ctx = torch.autocast(device.type, dtype=dtype, enabled=amp_cfg.enabled)

    return {
        "scaler": scaler,
        "amp_ctx": amp_ctx,
        "amp_cfg": amp_cfg,
        "device": device,
    }