feat(model): add Nemotron 3.5 Super VL support

[cuichenx]

What does this PR do ?

Adds Megatron Bridge support for Nemotron 3.5 Super VL, combining the Nemotron 3 Super language/MTP stack with the shared Nemotron Omni image and temporal-video stack, without an audio encoder.

The Bridge translates one serialized HF MTP block into two repeated Megatron prediction depths and preserves the HF block count on export. This PR is ready for code review; current-head model verification is still being refreshed and is not a merge-readiness claim.

Changelog

• Register NemotronH_Omni_Reasoning_V3 and normalize its hybrid/MTP configuration through the shared Nemotron-H bridge.
• Reuse Omni's RADIO tower, dynamic-resolution media support, and separate temporal video embedder without audio or EVS token pruning.
• Support nested MTP checkpoint weights, the vision projector final norm, and the deterministic RADIO summary-index buffer, including strict re-export of an exported HF checkpoint.
• Add H100/GB200 pretrain, SFT, PEFT, and GB200 long-context SFT recipes under recipes/, with independent language, vision, and projector freeze controls.
• Add focused model, conversion, recipe, and inference tests and a model verification card with generated model documentation.
• Fix shared visual inference helpers to retain post-resize image/video metadata and use the same native processor inputs for HF comparisons. These helpers serve both Omni and Super VL.
• Extend the public conversion, inference, and training launchers with bounded Slurm polling, shared allocation-level options, and an opt-in direct-Enroot runtime. Pyxis remains the default. These reusable changes make the documented verification workflows usable without private launcher forks and have focused backward-compatibility tests.
• Merge current main, including its pinned-HF-revision loader fix, and regenerate the model indexes to retain both branches' entries.

GitHub Actions CI

See the CI section in the Contributing doc for how to trigger the CI. A Nvidia developer will need to approve and trigger the CI for external contributors.

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) — No new required dependencies.
  • Reviewer: Does the PR have correct import guards for all optional libraries? — Existing optional boundaries retained.

Additional Information

• The final merged tree passed 348 focused launcher, visual-input, inference-helper, comparison, and generated-document tests; the model-card validator, generated-catalog consistency check, and all-files pre-commit also passed.
• Fresh independent review covered model/conversion/training, recipes/documentation, and shared launchers/helpers. Both findings were fixed and independently reviewed: strict re-export now preserves the RADIO summary buffer, and direct-Enroot cache defaults now honor forwarded HF_HOME and explicit cache overrides. Four export regression cases passed an isolated CPU check using the real strict safetensors writer; full conversion-test collection still requires the model runtime dependencies.
• A separate recipe-registry test fails identically on main and this branch: test_known_cross_package_collisions_have_explicit_benchmark_precedence reports the existing nemotron_3_super_pretrain_64gpu_gb200_bf16_config name collision. This unrelated base failure is not patched in the model-addition PR.
• The verification card deliberately retains historical evidence at its recorded revisions. Current-head numerical parity, inference, and affected training/export/resume verification remain pending; no old result is relabeled as a result of this merge. Full GPU model tests were not rerun by the local CPU checks above.

[cuichenx]

/ok to test b049c95

[claude]

Light review — LGTM

Reviewed the diff against the mandatory workflow (read diff, loaded adding-model-support + review-pr skills, then reviewed). Scope: critical bugs/logic, typos, test coverage, doc accuracy.

No blocking issues found. Notes on the one item I chased down:

• Nemotron35SuperVLBridge.megatron_to_hf_config replaces hf_config[llm_config] wholesale with {num_nextn_predict_layers: 1}. This is safe: export runs the dict through conform_config_to_reference, which recurses into the nested llm_config, keeps the exported num_nextn_predict_layers, and backfills every other llm_config key from the reference HF config. Covered by test_nemotron_35_super_vl_export_preserves_shared_mtp_serialization.
• MTP shared-block invariants (one serialized HF block maps to two repeated Megatron depths, pattern star-E), RADIO summary-buffer shard write (atomic tmp+replace, index/total_size update, idempotent re-export), list-based layers_block_type schema normalization, and vision recompute validation are all guarded with explicit errors and have direct unit coverage.
• No typos, doc drift, or missing-coverage gaps spotted in the changed code. is_nemotron_omni correctly matches NemotronH_Omni_Reasoning_V3.

Suggested test cases

No perf tests impacted (no scripts/performance/configs/ targets touched by this PR).

Unit cases exercised/derived from the diff (all present in the PR, listed for the record):

• test_nemotron_35_super_vl_architecture_is_registered
• test_nemotron_35_super_vl_provider_reuses_omni_with_list_based_mtp
• test_nemotron_35_super_vl_rejects_unexpected_serialized_mtp_depth
• test_nemotron_35_super_vl_export_preserves_shared_mtp_serialization
• test_nemotron_35_super_vl_export_closes_summary_idxs_buffer
• test_nemotron_35_super_vl_export_requires_weight_index
• test_nemotron_35_super_vl_strict_reexport_preserves_summary_buffer
• test_nemotron_35_super_vl_export_requires_direct_modeling_entrypoint
• test_nemotron_35_super_vl_mapping_uses_nested_mtp_and_vision_final_norm
• test_list_based_transformers_layer_schema_is_normalized
• test_list_based_transformers_layer_schema_rejects_unknown_types
• test_canonical_model_exposes_nested_language_decoder_for_cuda_graph_helper
• test_canonical_model_rejects_variable_sequences_with_cuda_graphs
• test_vision_recompute_requires_effective_granularity
• test_full_vision_recompute_requires_complete_policy
• test_full_vision_recompute_rejects_invalid_layer_count
• test_vision_recompute_rejects_forced_radio_eval_mode
• test_super_vl_rectangular_image_and_video_use_exact_visual_token_counts (GPU)