Temporarily pin SALM Automodel dispatcher to torch

[DongjiGao]

What changed

• Pin examples/speechlm2/conf/salm_automodel.yaml to automodel_backend.dispatcher: torch.
• Document DeepEP as an explicit opt-in for GPUs with NVLINK/NVSHMEM.
• Add a focused config regression test.
• Add TODO (Dongji) to remove the override once Automodel performs topology-aware dispatcher selection.

Why

This is a temporary fix to unblock the release.

Automodel's BackendConfig currently selects DeepEP when the package is installed and CUDA is available. It does not validate whether the GPU topology supports DeepEP. On PCIe-only systems, SALM MoE expert-parallel training can therefore enter DeepEP kernels and fail with an unexplained CUDA unspecified launch failure / EPException.

The long-term fix belongs in Automodel: select DeepEP only when the required NVLINK/NVSHMEM topology is available, otherwise fall back to Torch with a warning.

Impact

The example now uses the portable Torch dispatcher by default. Systems with compatible NVLINK/NVSHMEM topology can opt into DeepEP with:

model.automodel_backend.dispatcher=deepep

Validation

Using nvcr.io/nvidian/nemo-speech:26.06.rc4 on a two-GPU PCIe-only host:

• Default BackendConfig resolved to dispatcher=deepep.
• Automodel DeepEP dispatch reproduced intranode.cu:126 'unspecified launch failure' and deep_ep.cpp:147 'unspecified launch failure', followed by EPException / SIGABRT.
• The portable Automodel Torch GroupedExperts forward/backward path passed on both ranks.
• The patched YAML resolved to dispatcher=torch.
• Focused config test passed: pytest -q --confcutdir=tests/collections/speechlm2 tests/collections/speechlm2/test_example_configs.py.
• git diff --check, Black, isort, and Python compilation passed.

The original 8x H200 QA case should still be rerun in the release environment.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[KunalDhawan]

/ok to test fe8310c

[DongjiGao]

/ok to test 9ac7aa6

[github-actions]

[🤖]: Hi @DongjiGao 👋,

We wanted to let you know that a CICD pipeline for this PR just finished successfully.

So it might be time to merge this PR or get some approvals.