[ravel7524]

Overview

This PR fixes GGUF save/reload failures for models with MTP (nextn) layers. llama_model_save_to_file() wrote per-layer arrays for the main layers only, while the loader expects all layers, including MTP.

Per-layer arrays now use n_layer_all, except attention.indexer.types, which keeps the main-layer length and is always an array. All four readers (glm-dsa, glm5-next, hy-v4 and dots3note) accept arrays; dots3note requires one.

The saver emits this key for every architecture: its scalar-to-array encoding changed in 132/134 generated GGUFs, with identical tensor bytes; architectures that do not read the key ignore it.

Additional information

Reproduction: save Qwen3.5-0.8B Q8_0 with llama_model_save_to_file(), then reload it: qwen35.attention.recurrent_layers has wrong array length; expected 25, got 24. With the new regression test applied to baseline master f711c8f30, test-llama-archs -a qwen35 exits 255 with expected 3, got 2; it passes with the fix.

CPU testing:

• Full test-llama-archs: 139/139 passed, seeds 1 and 42, including MTP coverage for qwen35 and mixed-indexer glm-dsa.
• Public Qwen3.5-0.8B Q8_0 saves and reloads with MTP loading enabled and disabled.

The regression dates from the hparams refactor #24060. #30880 overlaps the saver and test refactor; I will rebase this fix if that lands first.

Requirements

• Yes, I have read and agree with the contributing guidelines.
• AI usage disclosure: YES, AI was used for a Workflow, implementation, tests, validation, and review.

[m4zar]

Tested this PR against a case not covered in the description: the Gemma 4 assistant (E4B, Q8_0, AtomicChat/gemma-4-E4B-it-assistant-GGUF). In this model, nextn_predict_layers == block_count, resulting in n_layer() == 0.

• On master: The saver drops head_count, head_count_kv, and feed_forward_length. The saved file fails to load.
• With this PR: All three keys are preserved with their original values, and the model reloads successfully with matching hparams.

gemma-4-E4B-it-assistant.Q8_0.gguf (block_count 4, nextn_predict_layers 4, n_layer() 0)

key                                      original   saved (master)   saved (#30891)
gemma4-assistant.attention.head_count    u32 4      missing          u32 4
gemma4-assistant.attention.head_count_kv u32 2      missing          u32 2
gemma4-assistant.feed_forward_length     u32 2048   missing          u32 2048

reload, master:  check_tensor_dims: tensor 'blk.0.attn_q.weight' has wrong shape; expected 256, 0, got 256, 1024
reload, #30891:  loads; n_head 4, n_head_kv 2, n_ff 2048, n_embd_head_k 512 (same as original)
tensors: 49/49 identical (name, type, shape, sha256) in all three files