fix(mistral3): preserve medium VLM checkpoint layout

[akoumpa]

Summary

• preserve identity full-VLM checkpoint key layout for the Mistral-Medium-3.5 128B FP8 config
• keep the nested language_model.model.* mapping for the smaller Ministral/Devstral-style layouts
• drive the layout choice from config.text_config.model_type plus config.text_config.num_hidden_layers, not from the repo id

Failure

The regression showed up while loading the base model for mistral3p5_128b_medpix_lora:

RuntimeError: Missing key in checkpoint state_dict: language_model.model.embed_tokens.weight.

Mistral-Medium-3.5-128B stores full-VLM body keys in the HF/runtime layout (model.language_model.*, model.vision_tower.*, model.multi_modal_projector.*), so remapping those destination keys to language_model.model.* made DCP request keys that are absent from that checkpoint.

Validation

• Local: pytest -q tests/unit_tests/models/mistral3_vlm/test_state_dict_adapter.py tests/unit_tests/models/mistral3_vlm/test_model.py -> 53 passed
• nemo-ci root pipeline: https://gitlab-master.nvidia.com/dl/JoC/nemo-ci/-/pipelines/55711938
• mistral3p5_128b_medpix_lora: https://gitlab-master.nvidia.com/dl/JoC/nemo-ci/-/jobs/347146651
• ministral3_3b_squad: https://gitlab-master.nvidia.com/dl/JoC/nemo-ci/-/jobs/347146977

[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]

Additional nemo-ci trigger for the original SFT MedPix job requested from https://gitlab-master.nvidia.com/dl/JoC/nemo-ci/-/jobs/346789236.

• branch/commit: akoumpa/fix/mistral3-vlm-peft-load / ffeff898dc0a116e52c5b03d7fe238187b8c24c0
• root pipeline: https://gitlab-master.nvidia.com/dl/JoC/nemo-ci/-/pipelines/55717003
• https://gitlab-master.nvidia.com/dl/JoC/nemo-ci/-/pipelines/55713786
• target regex: ^mistral3p5_128b_medpix: \[

Current state: root pipeline is running; automodel_test is still waiting on the root build stage, so the target leaf job has not materialized yet.

[akoumpa]

Superseding the earlier retrigger for mistral3p5_128b_medpix:

• 55717003 did not create a downstream test because it was triggered with AUTOMODEL_TEST_FOLDER=vlm_finetune; the top-level Automodel trigger expects the broad folder selector there.
• Retried with AUTOMODEL_TEST_FOLDER=all, ARCH=x86, and the same branch SHA ffeff898dc0a116e52c5b03d7fe238187b8c24c0.
• Root pipeline: https://gitlab-master.nvidia.com/dl/JoC/nemo-ci/-/pipelines/55730155
• Leaf job: https://gitlab-master.nvidia.com/dl/JoC/nemo-ci/-/jobs/347275996

The leaf job is currently running.

[akoumpa]

/ok to test ffeff89

[athitten]

LGTM, thank you @akoumpa !

[akoumpa]

Final follow-up for the extra mistral3p5_128b_medpix retest: passed.

• Root pipeline: https://gitlab-master.nvidia.com/dl/JoC/nemo-ci/-/pipelines/55730155
• Leaf job: https://gitlab-master.nvidia.com/dl/JoC/nemo-ci/-/jobs/347275996