fix: pass trust_remote_code=True to remaining AutoConfig.from_pretrained sites

[bzantium]

What does this PR do ?

Three call sites in the codebase still invoked AutoConfig.from_pretrained without trust_remote_code=True. This crashes for any model that ships custom modeling code via auto_map in config.json, since transformers then tries to interactively prompt the user, which raises EOFError inside a Ray worker (stdin is closed) and finally surfaces the underlying ValueError:

ValueError: The repository <path> contains custom code which must be executed
to correctly load the model. ... Please pass the argument
`trust_remote_code=True` to allow custom code to be run.

The rest of the codebase already passes trust_remote_code=True at every other from_pretrained site (flops_tracker.py, native_checkpoint.py, sglang_worker.py, vllm_worker.py, dtensor_policy_worker.py, automodel/setup.py, huggingface/common.py, megatron/setup.py, megatron/community_import.py, algorithms/utils.py), so this PR just restores consistency at the three remaining sites.

Patched sites:

File	Function	Why it crashes
nemo_rl/algorithms/distillation.py	check_vocab_equality	Probes student and teacher configs at setup. The crash is fatal because it happens before any model weights are loaded, so distillation is fully blocked for any custom-arch student or teacher.
nemo_rl/models/generation/vllm/quantization/fp8.py	init_fp8	Probes the model config to decide FP8 quant plumbing for vLLM.
nemo_rl/models/megatron/draft/utils.py	build_draft_model	Probes the draft model config for Megatron speculative decoding.

The distillation.py site is the one I actually hit in practice (on-policy distillation with a custom-arch HF student that ships modeling_*.py and configuration_*.py via auto_map). The other two are the same precondition violation in lower-traffic paths and were caught while looking at the surrounding AutoConfig.from_pretrained usages.

Issues

None filed; this is a small consistency fix discovered while debugging a custom-arch distillation run.

Usage

No new API or configuration. After this patch a distillation run with a custom-arch student or teacher, e.g.

policy:
  model_name: /path/to/custom-arch-student   # config.json has auto_map.AutoModelForCausalLM
distillation:
  teacher:
    model_name: /path/to/custom-arch-teacher

proceeds past check_vocab_equality instead of dying at setup.

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you run the unit tests and functional tests locally? Visit our Testing Guide for how to run tests
• Did you add or update any necessary documentation? Visit our Document Development Guide for how to write, build and test the docs.

Additional Information

• No new tests. This is a 3 line consistency fix mirroring existing call sites; happy to add a unit test that mocks AutoConfig.from_pretrained and asserts the kwarg is forwarded if the reviewers prefer.
• Verified locally by applying the same patch in a downstream fork and re-running the failing distillation recipe; check_vocab_equality passes and the driver proceeds normally.

[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.