fix(model): apply dict config overrides with custom dispatch

[akoumpa]

NVBug: 6259955

What

get_hf_config / __init_model (nemo_automodel/_transformers/model_init.py) now treat a plain-dict
config kwarg — how the config loader materializes --model.config.* CLI overrides (and YAML
model.config: maps) — as config overrides rather than a finished config object.

• get_hf_config: a dict config is popped and its keys folded into the kwargs AutoConfig.from_pretrained
  applies as overrides, so a real config (e.g. GlmMoeDsaConfig with num_hidden_layers=N) is built. Its
  architectures survive, so custom-model dispatch is preserved.
• __init_model: the now-consumed dict config is dropped from kwargs so it isn't forwarded into the model
  constructor (which receives the config positionally).
  Adds two regression tests in tests/unit_tests/_transformers/test_model_init.py.

Why

Fixes NVBugs 6259955 Defect 2. Passing --model.config.num_hidden_layers 16 (the natural shrink knob for
GLM-5.1 / GlmMoeDsaForCausalLM) crashed with:

TypeError: GlmMoeDsaForCausalLM.__init__() got an unexpected keyword argument 'backend'

Root cause: the override was materialized as a dict and returned verbatim by get_hf_config.
get_architectures(dict) returns [], so _resolve_custom_model_cls_for_config returned None and
dispatch flipped from the custom GlmMoeDsa implementation to the stock-HF path — which forwards the
custom-only backend kwarg into the built-in transformers GlmMoeDsaForCausalLM.__init__ and rejects it.
(PR #1784 coerced dict→BackendConfig only on the custom path, so it didn't cover this stock-HF fallback.)
This made it impossible to shrink GLM-5.1 below its 256-H100 footprint via the supported CLI. The fix is
the report's preferred remedy: --model.config.* overrides no longer change which implementation is
dispatched, and they are correctly applied. Affects every registered custom model that accepts custom-only
constructor kwargs, not just GlmMoeDsa.

How tested

On cw-dfw (H100) in the rc6 container, latest main 2b96596a, transformers 5.8.1:

• Weightless repro driving NeMoAutoModelForCausalLM.from_pretrained on a tiny GlmMoeDsa config (no
  GLM-5.1 download): before → control custom path OK, override → TypeError: ... 'backend' (exact bug
  traceback); after → control unchanged, override keeps the custom path and applies the override (1 layer
  built vs the file's 2), exit 0.
• pytest tests/unit_tests/_transformers/test_model_init.py → 39 passed (37 existing + 2 new), 0 regressions.

Notes

• Defect 3 of the same bug (DeepEP internode dispatch timeout at ep>8) is already fixed on main by refactor(moe): remove enable_deepep, switch failing ep recipes to hybridep #2630
  (recipe switched deepep→hybridep); no change needed here.
• Defect 1 (ep8/4-node OOM) is expected under-provisioning, not a code bug; this fix unblocks shrinking
  GLM-5.1 for a small-footprint smoke.

[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.

[akoumpa]

/ok to test 6e9fdb5