Fix type annotation and checkpoint conversion script

[okoge-kaz]

What is the issue

I have found that if you are using TransformerEngine v1.10 or later, CUDA error: initialization error will occur in the following location in saver_mcore.py.

Megatron-LM/tools/checkpoint/saver_mcore.py

Line 299 in 076972e

models[pp_rank][ep_rank][tp_rank] = model_provider(pre_process, post_process).to(md.params_dtype)

If the version of TransformerEngine is 1.10 or later, torch.cuda.Stream() is called in the following places. This causes an initialization error.

Megatron-LM/megatron/core/extensions/transformer_engine.py

Lines 1218 to 1221 in 076972e

	if is_te_min_version("1.10.0.dev0"):
	context, sync_func = _get_cpu_offload_context(
	enabled, num_layers, model_layers, activation_offloading, weight_offloading
	)

https://github.com/NVIDIA/TransformerEngine/blob/c9ea6be92948e1ec553037f1a04900617b9f7f6b/transformer_engine/pytorch/cpu_offload.py#L314-L315

How to solve the issue

Add mp.set_start_method(method='spawn') and torch.cuda.init()