ci: Update to transformers v5.5

[athitten]

What does this PR do ?

Upgrade to transformers v5.5 and updates mistral_common. Both of them are required for gemma4.

Changelog

• tests/unit_tests/models/deepseek_v3/test_dsv3_layers.py — Updated mock configs to set rope_parameters = rope_scaling to match v5.5's DeepseekV3Config behavior.
• nemo_automodel/components/models/gemma4_moe/model.py — Added bidirectional compat between expert_intermediate_size (pre-v5.5) and moe_intermediate_size (v5.5 rename) in Gemma4ForConditionalGeneration.init and Gemma4MoETextModelBackend.init.
• tests/unit_tests/models/gemma4/test_gemma4_model.py — Updated test configs to use moe_intermediate_size instead of expert_intermediate_size.
• tests/unit_tests/models/gemma4/test_gemma4_state_dict_adapter.py — Updated mock state dict to use v5.5 key format (experts.gate_up_proj, router.per_expert_scale).
• nemo_automodel/components/models/nemotron_parse/model.py — Handle MBartDecoderLayer.forward returning a single tensor in v5.5 (was tuple in v5.3).
• nemo_automodel/components/models/nemotron_v3/model.py — Replace pre-created DynamicCache from GenerationMixin with NemotronHybridCache in prepare_inputs_for_generation.
• _reinit_non_persistent_buffers in nemo_automodel/components/checkpoint/checkpointing.py since in v5.5 Gemma3RotaryEmbedding's per-layer-type inv_freq buffers, embed_scale, and SigLIP position_ids buffers are non-persistent (not saved in checkpoint) and so they have to be initialized explicitly. Creates a generalized method instead of the previous once which was specific to rope. Also adds appropriate unit test.
• Adds new message pattern added in transformers v5.4 "attempted to run this operator with Meta tensors" to _build_model method in nemo_automodel/_transformers/auto_model.py to retry building without meta device. Transformers v5.5 added a torch.equal() call inside tie_weights() (line 2539 of modeling_utils.py) to compare tied parameter values. When _build_model wraps from_pretrained inside init_empty_weights() (meta device context), the model parameters are meta tensors, and torch.equal doesn't support them — producing a NotImplementedError. Retry logic is already there in _build_model method, the PR adds this error message so that retry is done even for this case. (this is done in NotImplementedError: aten::equal on meta tensors during multi-GPU model init with transformers >= 5.4.0 #1765 )
• Remove cache_position arg from Qwen3_5MoeGatedDeltaNet.forward() method in nemo_automodel/components/models/qwen3_5_moe/cp_linear_attn.py as its no longer supported in transformers v5.5
• Fix Gemma3 VLM checkpoint loading with transformers 5.5: the generic llava key mapping (aliased by transformers for gemma3) produced wrong FQNs (Fully Qualified Name) for Gemma3's model.language_model.* hierarchy, causing all base weights to be silently dropped by set_model_state_dict(strict=False). Replaced the fallback mechanism in get_combined_key_mapping with an explicit priority override in _VLM_KEY_MAPPINGS that uses the correct ^-anchored regex patterns for Gemma3's module structure.

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?

If you haven't finished some of the above items you can still open "Draft" PR.

Additional Information

• Related to # (issue)

[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.

[athitten]

/ok to test e4c7b30

[akoumpa]

/ok to test 30ca0e9

[athitten]

/ok to test 933ff7c

[athitten]

/ok to test 17c9a03

[akoumpa]

can we make this model specific?

ie run the function only on select models where it applies. I think it's well intended, catch as many cases as possible, but I also want to be very risk averse here, given that there's a wild west on hf hub :)

[athitten]

Addressed in 07f775a

[akoumpa]

Note-to-self: I think this is ok for the time being, but long term, it might be a good idea to move all the model-specific patching under nemo_automodel/components/models/, so that each model has its own patches. No action is needed.

[athitten]

/ok to test 07f775a

[athitten]

/ok to test 3ea3ff2

[akoumpa]

Thanks a lot @athitten !

[athitten]

/ok to test e1bf007

[akoumpa]

/ok to test 2cd5ee6