[inureyes]

Summary

Ports Nemotron-Parse (nemotron_parse): C-RADIOv2-H ViT-H/16 tower plus compression neck (3329 encoder states for a 2048x1664 page) and a 10-layer pre-norm mBART decoder seeded with the task prompt. It reuses the Florence-2 attention sublayers, dual KV cache, and causal mask, and is wired into detection, loaders, mlxcel generate (early exit like Florence-2), mlxcel run refusal, the mlxcel arch catalog, a dedicated single-stream server worker (admission, warmup skip, built-in text-only template), and docs/supported-models.md.

Behavior worth reviewing

• Tokenizer loading now drops padding / truncation serialized in tokenizer.json, matching transformers. This checkpoint serializes padding: Fixed(9000), which turned every prompt into 9000 ids. This touches every tokenizer.json model; the tokenizer, loading, and chat-template test modules pass.
• The decoder always runs activations in f32. With bf16 activations the 8-bit export flips two near-ties (reference top-2 margins 0.04 to 0.07); f32 reproduces the reference.
• For this family the built-in template replaces the Hub checkpoint's chat_template.jinja, which prints message['content'] raw and would render the typed image list as the decoder seed.
• Server bounds on the task prompt: 1024 bytes, no control characters, at most 256 seed tokens; exactly one image; audio/video refused.

Deviations from the issue text

• Seeding: the reference tokenizes with add_special_tokens=False and generate prepends decoder_start when missing, so that is implemented instead of the encode-then-strip rule (same result for the default prompt).
• Repetition penalty covers the whole decoder sequence including the seed, as transformers does, not only generated ids.
• Line numbers in the issue have drifted (for example the Florence-2 dispatch is at generate.rs near 2968, not 2431); the code was followed.
• cargo test --workspace / clippy --workspace --all-targets were replaced by the narrow gate below (run rules for this machine).

Test plan

• Real checkpoints (tests/nemotron_parse_real_model.rs, ignored-gated): Hub f32 and 8-bit match a CPU fp32 run of the checkpoint's own transformers code for all 79 model-chosen greedy tokens, at penalty 1.0 and 1.1 (8-bit reference ran on its dequantized weights; the reference's 80th token is a forced EOS). 4-bit reads the page.
• mlxcel-server with the Hub checkpoint: /v1/chat/completions returns the reference text; empty prompt uses the default; two images, zero images, and control characters are refused; streaming works.
• Unit tests: models::nemotron_parse (22), server::nemotron_parse_worker (7), detection, chat template, tokenizer padding; plus tokenizer::tests, loading::tests, server::chat_template, models::registry, model_metadata, execution::memory_estimate.
• cargo fmt --all -- --check, cargo clippy --release -p mlxcel --lib --tests -- -D warnings, cli_help_consistency, dead_doc_pointers, llama_compat_manifest.

Pages larger than the target box are resized with image bilinear filtering, which is close to but not bit-identical with PIL; the parity page avoids resampling. The v1.x untied-head path is implemented but not run against a real v1.x checkpoint.

Closes #1369