[FR] Support config torch.compile()/TorchDynamoPlugin with dynamic=None

[IrineSistiana]

Hi accelerate devs.

The dynamic parameter in torch.compile() has 3 values. True/False/None. According to the torch documentation, the default value is "None".

Doc: https://pytorch.org/docs/2.5/generated/torch.compile.html#torch.compile

dynamic (bool or None) – Use dynamic shape tracing. When this is True, we will up-front attempt to generate a kernel that is as dynamic as possible to avoid recompilations when sizes change. This may not always work as some operations/optimizations will force specialization; use TORCH_LOGS=dynamic to debug overspecialization. When this is False, we will NEVER generate dynamic kernels, we will always specialize. By default (None), we automatically detect if dynamism has occurred and compile a more dynamic kernel upon recompile.

If my understanding is correct. Currently accelerate can only config it as True or False.

accelerate/src/accelerate/utils/dataclasses.py

Lines 964 to 965 in cb8b7c6

	if self.dynamic is None:
	self.dynamic = str_to_bool(os.environ.get(prefix + "USE_DYNAMIC", "False")) == 1

accelerate/src/accelerate/accelerator.py

Line 1613 in cb8b7c6

model = torch.compile(model, **self.state.dynamo_plugin.to_kwargs())

In our development environment, using dynamic = None has slightly better results. It would be nice if accelerate can support it out of the box, and maybe have the same default value/behavior as torch.compile().