fix(state): make MLU backend part of the _prepare_backend elif chain (#4055)

[Anai-Guo]

What

In PartialState._prepare_backend, the MLU branch was guarded by a standalone if, immediately followed by a separate if is_sdaa_available(): ... elif is_musa_available(): ... chain:

if is_mlu_available():
    backend = "cncl"
    distributed_type = DistributedType.MULTI_MLU
if is_sdaa_available():        # <- new chain, not part of the MLU check
    ...
elif is_musa_available():
    ...
elif torch.cuda.is_available():
    ...

Because MLU sits outside the mutually-exclusive chain, on an MLU host the cncl / MULTI_MLU selection is not final: execution falls through into the second chain and a later branch (e.g. torch.cuda.is_available()) can silently overwrite backend/distributed_type. Every other accelerator (SDAA, MUSA, NPU, HPU, CUDA, XPU, NEURON) is already part of one if/elif chain — MLU is the lone outlier.

Fix

Change the second if is_sdaa_available() to elif, so all backends form a single mutually-exclusive chain with MLU as the head. This matches the structure already used by default_device.

• Non-MLU hosts: is_mlu_available() is False, so the elif is evaluated exactly as the old if was — no behavior change.
• MLU hosts: the chain now short-circuits correctly instead of being able to fall through to CUDA.

Closes #4055

🤖 Generated with Claude Code