Port Nemotron-H-4B-Base-8K (nemotron_h)

[DarthCeltic]

Summary

• Adds nvidia/Nemotron-H-4B-Base-8K to the most_models_ported track (identity nemotron_h), introducing the Nemotron-H hybrid Mamba2/Attention/MLP execution family to the board.
• Real converter bug found and fixed: NemotronHModel.__init__ in convert_hf_to_gguf.py decides MoE-vs-dense by checking key presence of num_experts_per_tok in ModelBase.load_hparams()'s merged config. For this checkpoint that reports a full, plausible-looking MoE config (num_experts_per_tok=2, moe_intermediate_size=7688, n_routed_experts=8) — but direct inspection of both safetensors shards (311 tensors total) finds zero tensors matching expert/moe/router. The config fields are stale, almost certainly inherited from Nemotron-H's larger 8B MoE parent during pruning/distillation. Fixed via a standalone monkeypatch wrapper (ported_models/nemotron_h_4b_base/convert_wrapper.py) that hardcodes is_moe=False, treating tensor evidence as authoritative over the misleading config — the vendored submodule is never touched, matching the established pythia410m precedent.
• Self-converted to Q8_0 GGUF (311 tensors, 4.8 GB, sha256=23cf524eaf7cc0c84ffe3dc168a5bacaad10bde2bb9fbaacba431a96b50c9d31), hosted on Hugging Face (darthceltic85/nemotron-h-4b-base-gguf) since it exceeds GitHub's 2 GB release-asset limit.
• Verified live: model loads cleanly on the CPU backend (llama.cpp-et, fused Gated Delta Net enabled, 1489-node compute graph, 1 split) and real llama-perplexity against WikiText-2 raw gives PPL = 9.4458 +/- 1.75431 — a good, in-range result that also confirms the quantization RuntimeWarnings seen during conversion were a false alarm.

Test plan

• benchmark_config in the claim JSON exactly matches .github/ci/benchmark_config.json's new nemotron_h_4b_base entry
• Source revision is a valid 40-char hex commit SHA
• RECIPE.md exists and documents the real bug found
• All new/modified JSON files are valid JSON
• Model loads and produces a real, in-range perplexity result (not fabricated)

[AFOliveira]

The converter fix and tensor inspection are useful, but Nemotron-H still needs a real full-offload ET-SoC1 execution and a committed deterministic oracle/expected output with an explicit comparison threshold. Host loading and CPU PPL alone cannot validate the ET port. Please pin the hosted GGUF immutably and include the converter-wrapper output hash. Maintainers will then review/adopt the identity/contract and runner policy.