Accelerating Inference

Gaudi offers several possibilities to make inference faster.

Lazy Mode

Two execution modes are proposed:

• Lazy mode, where operations are accumulated in a graph whose execution is triggered in a lazy manner. This allows the graph compiler to optimize the device execution for these operations.
• Eager mode, where one operation at a time is executed.

In lazy mode, the graph compiler generates optimized binary code that implements the given model topology on Gaudi. It performs operator fusion, data layout management, parallelization, pipelining and memory management, as well as graph-level optimizations.

To execute inference in lazy mode, you must provide the following arguments:

args = GaudiTrainingArguments(
    # same arguments as in Transformers,
    use_habana=True,
    use_lazy_mode=True,
)

HPU Graphs

Gaudi provides a way to run fast inference with HPU Graphs. It consists in capturing a series of operations (i.e. graphs) in an HPU stream and then replaying them in an optimized way (more information here). Thus, you can apply this to the forward method of your model to run it efficiently at inference.

HPU Graphs are integrated into the GaudiTrainer and the GaudiStableDiffusionPipeline so that one can use them very easily:

• GaudiTrainer needs the training argument use_hpu_graphs_for_inference to be set to True as follows:

from optimum.habana import GaudiTrainer, GaudiTrainingArguments

# define the training arguments
training_args = GaudiTrainingArguments(
    use_habana=True,
    use_lazy_mode=True,
    use_hpu_graphs_for_inference=True,
    gaudi_config_name=gaudi_config_name,
    ...
)

# Initialize our Trainer
trainer = GaudiTrainer(
    model=model,
    args=training_args,
    train_dataset=train_dataset
    ... # other arguments
)

• GaudiStableDiffusionPipeline needs its argument use_hpu_graphs to be set to True such as:

from optimum.habana.diffusers import GaudiDDIMScheduler, GaudiStableDiffusionPipeline

model_name = "runwayml/stable-diffusion-v1-5"

scheduler = GaudiDDIMScheduler.from_pretrained(model_name, subfolder="scheduler")

pipeline = GaudiStableDiffusionPipeline.from_pretrained(
    model_name,
    scheduler=scheduler,
    use_habana=True,
    use_hpu_graphs=True,
    gaudi_config="Habana/stable-diffusion",
)

outputs = generator(
    ["An image of a squirrel in Picasso style"],
    num_images_per_prompt=16,
    batch_size=4,
)