fix(model): apply Hugging Face YaRN defaults for omitted rope_scaling keys

[kamran-nvidia]

What does this PR do ?

Applies the Hugging Face YaRN defaults (beta_fast=32, beta_slow=1, truncate=True) in the shared HF → Megatron config mapping when a YaRN rope_scaling omits them, so GPT-family checkpoints with YaRN no longer crash at model construction.

The bug: the shared mapping only copies YaRN keys that are present. A config such as {"rope_type": "yarn", "factor": 4.0, "original_max_position_embeddings": 32768} is the standard form for Qwen2.5 / Qwen3 long context. With it, yarn_beta_fast / yarn_beta_slow stay None. Megatron Core's GPTModel passes these fields to YarnRotaryEmbedding explicitly, which overrides its defaults, so model construction fails:

TypeError: unsupported operand type(s) for *: 'NoneType' and 'int'   (yarn_rotary_pos_embedding.py:244)

Changelog

• MegatronModelBridge.YARN_HF_DEFAULTS: the defaults transformers' _compute_yarn_parameters uses for omitted keys.
• The GPT YaRN branch of the shared mapping fills those defaults with setdefault. Values the checkpoint sets explicitly always win.
• Unit tests: defaults are filled for omitted keys, explicit keys are preserved, and non-YaRN configs are unaffected.
• Qwen3Bridge: drops its local copy of the same three defaults, added by refactor(model): add Qwen3 builder-backed model config with YaRN support #6365, now that the shared mapping fills them. Qwen3's existing YaRN tests, which assert 32 / 1 / True for omitted keys and keep explicit values, now cover the shared path.

Scope and behaviour change

• Affected: generic GPT bridges whose YaRN config goes through the shared mapping (for example Qwen2/2.5, Qwen3, Qwen3-MoE).
• Not affected:
  • MLA models (separate MLA_ROPE_SCALING_MAPPING branch).
  • Mistral and Ministral3, which build their providers by hand.
  • GPT-OSS, which sets every YaRN field explicitly after the mapping.
• Behaviour change for configs that already worked: a YaRN config that sets the betas but omits truncate previously got yarn_correction_range_round_to_int=None (no rounding). It now gets True, matching Hugging Face's default. Agreement with HF logits improves (table below).

Verification

Provider path (AutoBridge.to_megatron_provider()), tiny random-weight checkpoints, single GPU, logits compared with Hugging Face on the same input:

HF rope_scaling	main	this PR
Qwen2, YaRN, no betas	TypeError at construction	builds; cosine 0.999993, argmax 100%
Qwen3, YaRN, no betas	TypeError at construction	builds; cosine 0.999989, argmax 100%
Qwen2, YaRN, betas set, no truncate	cosine 0.999980, argmax 99.2%	cosine 0.999993, argmax 100%

• tests/unit_tests/models/test_model_bridge.py: 57 passed.
• Regression run (GPT provider, Ministral3, GPT-OSS, StepFun, DeepSeek, Qwen, Llama, AutoBridge): 555 passed, 1 skipped.
• ruff check / ruff format --check: clean (ruff 0.16.7; pre-commit itself not run).
• After rebasing onto main (Megatron Core 8949def81) and removing the Qwen3 copy: 564 passed, 1 skipped across test_model_bridge.py, models/qwen, recipes/qwen, test_transformer_config.py, test_gpt_provider.py, and the DeepSeek, Ministral3, Kimi K2.5-VL, StepFun and GPT-OSS model tests.

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 add or update any necessary documentation?
• Does the PR affect components that are optional to install? (Ex: Numba, Pynini, Apex etc)
  • Reviewer: Does the PR have correct import guards for all optional libraries?

Additional Information

• refactor(model): add Qwen3 builder-backed model config with YaRN support #6365 (merged) added the same three defaults locally in Qwen3Bridge; this PR, merging second, removes that copy.

🤖 Generated with Claude Code

[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.

[kamran-nvidia]

/ok to test c9d6412

[kamran-nvidia]

/ok to test 04c9ff7