llama : bump LLAMA_MAX_EXPERTS to 1024 for Kimi-K3 (896 experts)

[GrEarl]

LLAMA_MAX_EXPERTS has been 512 since Qwen3 Next. Kimi-K3 has 896 routed experts, so it trips the arch-generic assert in llama_model_base::load_hparams before any arch hook runs:

src/llama-hparams.h:10   #define LLAMA_MAX_EXPERTS 512 // Qwen3 Next
src/llama-model.cpp:1115 GGML_ASSERT(hparams.n_expert <= LLAMA_MAX_EXPERTS);

moonshotai/Kimi-K3 reports num_experts = 896 in text_config, and convert_hf_to_gguf.py on #26185 emits gguf: expert count = 896. This cannot be worked around on the file side: understating expert_count contradicts ne[2] = 896 on the ffn_*_exps tensors.

LLAMA_MAX_EXPERTS sizes exactly one object in the tree:

src/llama-graph.cpp:2128  ggml_tensor * cur_experts[LLAMA_MAX_EXPERTS] = { nullptr };

Only [0, n_expert_used) of it is ever written or read, so raising the bound to 1024 costs 4 KB of stack in build_moe_ffn and changes no behaviour for existing models.

This unblocks #26185 (model: add Kimi-K3 text model), which is otherwise complete on the conversion side — I verified its repack_mxfp4_blocks is bit-exact against an independent implementation (0 mismatches over 11,010,048 values on a real K3 expert tensor) and that the converter reaches prepare_tensors() with a fully mapped tensor set for all 93 layers. Details in #26185 (comment)

Not verified: I have not built or run inference with a Kimi-K3 GGUF, so this change is argued from the assert and the single array it bounds, not from a successful load.

AI usage disclosure: yes. The line references were read from master with an AI agent; the reasoning and the one-line change are as described above.

[ggml-gh-bot]

Hi @GrEarl, thanks for your contribution!

Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:

• PR Template not respected: Please respect the template when creating a new pull request. Make sure to fill out all required sections.

• AI-generated content: While code is allowed to be generated by AI, please write the PR description and commit messages on your own without the help of AI.

---

Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below.