Accelerator: HPU Training¶

This document offers instructions to Gaudi chip users who want to conserve memory and scale models using mixed-precision training.

Enable Mixed Precision¶

With Lightning, you can leverage mixed precision training on HPUs. By default, HPU training uses 32-bit precision. To enable mixed precision, set the precision flag.

from lightning_habana.pytorch.accelerator import HPUAccelerator

trainer = Trainer(devices=1, accelerator=HPUAccelerator(), precision="bf16-mixed")

Customize Mixed Precision Using Autocast¶

Lightning supports following methods to enable mixed precision training with HPU.

HPUPrecisionPlugin

HPUPrecisionPlugin, HPUPrecisionPlugin enables mixed precision training on Habana devices.

In addition to the default settings, you can choose to override these defaults and provide your own BF16 (LOWER_LIST) and FP32 (FP32_LIST) The LOWER_LIST and FP32_LIST environment variables must be set before any instances begin. HPUPrecisionPlugin supports bf16-mixed and 16-mixed for mixed precision training. It is advised to use bf16-mixed over 16-mixed where possible.

The following is an excerpt from an MNIST example implemented on a single HPU.

from lightning import Trainer
from lightning_habana.pytorch.accelerator import HPUAccelerator
from lightning_habana.pytorch.plugins.precision import HPUPrecisionPlugin

# Initialize a trainer with HPU accelerator for HPU strategy for single device,
# with HPU precision plugin for autocast
trainer = Trainer(
    accelerator=HPUAccelerator(),
    devices=1,
    plugins=[
        HPUPrecisionPlugin(
            precision="bf16-mixed",
        )
    ],
)

# Init our model
model = LitClassifier()
# Init the data
dm = MNISTDataModule(batch_size=batch_size)

# Train the model ⚡
trainer.fit(model, datamodule=dm)

Native PyTorch torch.autocast()

For more granular control over with mixed precision training, one can use torch.autocast from native PyTorch.

Instances of autocast serve as context managers or decorators that allow regions of your script to run in mixed precision.

import torch
from lightning import Trainer

class AutocastModelCM(nn.Module):
    # Autocast can be used as a context manager to the required code block.
    def forward(self, input):
        with torch.autocast(device_type="hpu", dtype=torch.bfloat16):
        ...
        return

class AutocastModelDecorator(nn.Module):
    # Autocast can be used as a decorator to the required code block.
    @torch.autocast(device_type="hpu", dtype=torch.bfloat16)
    def forward(self, input):
        ...
        return

# Initialize a trainer with HPU accelerator for HPU strategy for single device,
# with mixed precision using overridden HMP settings
trainer = Trainer(
    accelerator="hpu",
    devices=1,
)

# Init our model
model = AutocastModelCM()
# Init the data
dm = MNISTDataModule(batch_size=batch_size)

# Train the model ⚡
trainer.fit(model, datamodule=dm)

torch.autocast context manager allows fine-tuning of mixed precision training with enabled parameter. It can be used alongside HPUPrecisionPlugin, which globally enables mixed precision, while local torch.autocast contexts can disable it for particular model parts. Alternatively, users can forgo HPUPrecisionPlugin and use only torch.autocast to control precision for every Op. For nested contexts, the scope of a given context and its enabled parameter determine whether mixed precision is enabled or disabled in that region.

For more details, please refer to Native PyTorch Autocast. and Automatic Mixed Precision Package: torch.autocast.

fp8 Training¶

Lightning supports fp8 training using HPUPrecisionPlugin, HPUPrecisionPlugin.

fp8 training is only available on Gaudi2 and above. Output from fp8 supported modules is in torch.bfloat16.

For fp8 training, call plugin.convert_modules(). The function accepts following args for the fp8 training:

replace_layers : Set True to let the plugin replace torch.nn.Modules with transformer_engine equivalent modules. You can directly import and use modules from transformer_engine as well.
recipe : fp8 recipe used in training.

from lightning import Trainer
from lightning_habana.pytorch.accelerator import HPUAccelerator
from lightning_habana.pytorch.plugins.precision import HPUPrecisionPlugin
from habana_frameworks.torch.hpex.experimental.transformer_engine import recipe

model = BoringModel()

# init the precision plugin for fp8 training.
plugin = HPUPrecisionPlugin(precision="fp8")

# Replace torch.nn.Modules with transformer engine equivalent modules
plugin.convert_modules(model, replace_layers=True, recipe=recipe.DelayedScaling())

# Initialize a trainer with HPUPrecisionPlugin
trainer = Trainer(
    accelerator=HPUAccelerator(),
    plugins=plugin
)

# Train the model ⚡
trainer.fit(model)

Note

To use transformer_engine directly for training:

Import transformer_engine and replace your modules with transformer_engine modules in the model.
Wrap the forward pass of the training with fp8_autocast.

Users may still use HPUPrecisionPlugin to train in mixed precision for modules not supported by transformer_engine.

Note

To enable fp8 training with HPUDeepSpeedStrategy, use HPUDeepSpeedPrecisionPlugin, instead of HPUPrecisionPlugin, while keeping all other steps the same.

For more details on transformer_engine and recipes, refer to FP8 Training with Intel Gaudi Transformer Engine.

fp8 Inference¶

Lightning supports fp8 inference using HPUPrecisionPlugin, HPUPrecisionPlugin. fp8 inference is only available on Gaudi2 and above.

Intel Neural Compressor (INC) is required to run fp8 inference.

python -um pip install neural-compressor

Measurement and Quantization mechanisms

Inference in fp8 is a two step process.

Measurement mode: This step injects PyTorch measurement hooks to the model. Model is run on a portion of the dataset, and these measurement hooks measure the data statistics (e.g. max abs) and outputs them at the filepath specified by json.
Quantization mode: This is achieved by replacing modules with quantized modules implemented in INC that quantize and dequantize the tensors. It includes multiple steps, viz:
- Loading the measurements file.
- Calculating the scale of each tensor from its measurement.
- Injecting scale and cast ops to the model around ops that were selected to run in FP8.

Measurement

Get measurement data by running inference on a portion on data with HPUPrecisionPlugin.convert_modules(model, inference=True, quant=False).

from lightning import Trainer
from lightning_habana.pytorch.accelerator import HPUAccelerator
from lightning_habana.pytorch.plugins.precision import HPUPrecisionPlugin
from habana_frameworks.torch.hpex.experimental.transformer_engine import recipe

model = BoringModel()

# init the precision plugin for fp8 inference.
plugin = HPUPrecisionPlugin(precision="fp8")

# Replace module for fp8 inference measurements
plugin.convert_modules(model, inference=True, quant=False)

# Initialize a trainer with HPUPrecisionPlugin
trainer = Trainer(
    accelerator=HPUAccelerator(),
    plugins=plugin,
    limit_predict_batches=0.1,
)

# Run inference and dump measurements ⚡
trainer.predict(model)

Quantization

Run inference with HPUPrecisionPlugin.convert_modules(model, inference=True, quant=True).

# Replace module for fp8 inference measurements
plugin.convert_modules(model, inference=True, quant=True)

# Run inference ⚡
trainer.predict(model)

JSONs for quant and measure modes

INC uses configuration jsons for selecting between quant and measurement modes. This can be toggled via quant param in HPUPrecisionPlugin.convert_modules(). quant also accepts user defined config dictionary or a json path.

Refer to Supported JSON Config File Options for more information on supported json configs.

Note

To enable fp8 inference with HPUDeepSpeedStrategy, use HPUDeepSpeedPrecisionPlugin, instead of HPUPrecisionPlugin, while keeping all other steps the same.

For more details, refer to Inference Using FP8. For a list of data types supported with HPU, refer to PyTorch Support Matrix.

Enabling DeviceStatsMonitor with HPUs¶

DeviceStatsMonitor is a callback that automatically monitors and logs device stats during the training stage. This callback can be passed for training with HPUs. It returns a map of the following metrics with their values in bytes of type uint64:

Metric	Value
Limit	Amount of total memory on HPU.
InUse	Amount of allocated memory at any instance.
MaxInUse	Amount of total active memory allocated.
NumAllocs	Number of allocations.
NumFrees	Number of freed chunks.
ActiveAllocs	Number of active allocations.
MaxAllocSize	Maximum allocated size.
TotalSystemAllocs	Total number of system allocations.
TotalSystemFrees	Total number of system frees.
TotalActiveAllocs	Total number of active allocations.

The below shows how DeviceStatsMonitor can be enabled.

from lightning import Trainer
from lightning.callbacks import DeviceStatsMonitor
from lightning_habana.pytorch.accelerator import HPUAccelerator

device_stats = DeviceStatsMonitor()
trainer = Trainer(accelerator=HPUAccelerator(), callbacks=[device_stats])

For more details, please refer to Memory Stats APIs.

Runtime Environment Variables¶

Habana runtime environment flags are used to change the behavior as well as enable or disable some features.

For more information, refer to Runtime Flags.

Using LightningCLI¶

LightningCLI supports HPU. Following configurations from Lightning Habana are supported:

accelerator: “auto”, “hpu”.
strategies: “auto”, “hpu_single”, “hpu_parallel”.
plugins: class instances of HPUPrecisionPlugin and HPUCheckpointIO.

Limitations with HPU¶

LightningCLI cannot use class instances of accelerator and strategies. #19682. Applies to Lightning accelerator and strategies as well.
HPUProfiler does not work with LightningCLI since it is unable to patch torch.profiler.ProfilerActivity list.