[benchmarks] Small fixes for benchmarking script.

[ysiraichi]

This PR implements small fixes introduced by past PRs:

• Removes the incorrect self argument from the find_near_file function that was left-over in the last PR that moved it [benchmarks] Run some models with smaller batch sizes. #6542
• Replace the variable model_name by TorchBenchModel attribute of same name
• Select user specified batch size by default

@miladm

[zpcore]

Thanks, I think you forgot to import exists.

[zpcore]

There are also other issues I see in recent check-ins:

xla/benchmarks/torchbench_model.py

Line 309 in 0a2e5b5

model_name in config_data()["dont_change_batch_size"])

, need self.model_name.
For the batch_size, I can't find DEFAULT_EVAL_BSIZE and DEFAULT_TRAIN_BSIZE to be use anywhere.

[ysiraichi]

@zpcore @vanbasten23 @cota I have updated this PR fixing another small issues with the benchmarking script, as pointed out by @zpcore. I have also tested it by running BERT_pytorch. Could you take a look at it, again, when you have some time?

[zpcore]

Thanks for the fixing!

[zpcore]

Hi @ysiraichi , do you know where the default batch size from torchbench was used in the code?

[ysiraichi]

Do you mean where do we set the default batch size? As far as I understand, torchbench does it automatically when we instantiate a Model. Specifically, here:

xla/benchmarks/torchbench_model.py

Lines 329 to 333 in 4cde625

	return self.benchmark_cls()(
	test=self.benchmark_experiment.test,
	device=device,
	batch_size=batch_size,
	)

Which ends up in this function.

[zpcore]

Got it, thanks! Now the batch_size setting looks clear to me.