chore(model): deprecate additional legacy model support

[yaoyu-33]

Summary

• deprecate DeepSeek V2/V2 Lite, Llama Nemotron, Nemotron Nano v2, and Nemotron Nano v2 VL with removal scheduled for 0.7.0
• emit runtime warnings through AutoBridge and direct Nemotron Nano v2 recipe entry points while keeping implementations available during the deprecation window
• retire model-specific CI launchers and tests, and move shared NemotronH coverage to the active Nemotron-3 Nano configuration
• update the supported-model tables and model guides with matching deprecation notices

Testing

• uv run pre-commit run --all-files
• focused unit tests: 149 passed
• affected functional and collator test collection: 84 collected

Known baseline

The broader recipe metadata file currently has an unrelated main-branch collision-list failure for nemotron_3_nano_pretrain_8gpu_gb200_bf16_config. The metadata selector changed by this PR passes, and this PR does not modify the collision lists or assertion.

[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.

[yaoyu-33]

/ok to test f479f9f

[claude]

Light review — deprecation of DeepSeek V2/V2 Lite, Llama Nemotron, Nemotron Nano v2 (9B/12B), Nemotron Nano v2 VL 12B (removal in 0.7.0).

The core detection logic in src/megatron/bridge/models/_deprecation.py is correct:

• New architecture entries (DeciLMForCausalLM, DeepseekV2ForCausalLM, NemotronH_Nano_VL_V2) route through the exact-match table first.
• _NEMOTRON_NANO_V2_SHAPES = {(4480, 56), (5120, 62)} is disjoint from _NEMOTRON_H_V1_SHAPES, and the shape-based fallback is guarded behind the NemotronHForCausalLM architecture check, so there is no collision with active Nemotron-H v1 detection. The revised active fixture shape (2688, 52) correctly matches neither set.
• Identifier substring checks (nemotron-nano + -v2/_v2, and llama + nemotron) are sound and covered by the added positive/negative unit cases.

No critical bugs, typos, or outdated docs found. Docs were updated consistently across README.md, docs/models/**, and docs/fern/**.

One consistency observation (non-blocking): DeepSeek V2 and Nemotron Nano v2 VL are newly deprecated here and both ship recipe builders, but only the nemotronh nemotron_nano_v2.py recipe got direct warn_deprecated_model(...) calls. recipes/deepseek/deepseek_v2.py and recipes/nemotron_vl/nemotron_nano_v2_vl.py builders emit no warning, so a user invoking those recipe functions directly (not via HF-config load through warn_if_deprecated_model) will not be warned. Consider adding the same recipe-level warning for parity with the nemotronh treatment.

Suggested test cases

• tests/unit_tests/models/test_model_deprecation.py::test_deprecated_model_detection
• tests/unit_tests/models/test_model_deprecation.py::test_active_model_detection
• tests/unit_tests/models/test_model_deprecation.py::test_nemotron_nano_v2_recipe_warns
• tests/unit_tests/models/test_model_deprecation.py::test_nemotron_h_v1_recipe_warns
• tests/unit_tests/models/test_model_deprecation.py::test_auto_bridge_warns_for_deprecated_model
• tests/unit_tests/models/test_autobridge_registration_matrix.py::test_public_autobridge_import_registers_every_supported_model
• tests/unit_tests/models/nemotronh/test_nemotron_h_bridge.py::TestNemotronHBridge::test_provider_bridge_moe_config
• tests/unit_tests/models/nemotronh/test_nemotron_h_bridge.py::TestAutoBridgeIntegration::test_supports_nemotronh_architectures
• tests/unit_tests/recipes/test_deepseek_recipes.py::test_each_deepseek_recipe_builds_config
• tests/unit_tests/scripts/training/test_recipe_metadata.py::test_every_flat_recipe_has_registered_family_and_metadata

No perf tests impacted.

[yaoyu-33]

/ok to test 3261fcb

[yaoyu-33]

/ok to test 32b2d12