patches: add llama.cpp architecture patch + build instructions

patches/README.md

@@ -0,0 +1,71 @@
+# Architecture patch — build a llama.cpp that can load this model
+
+This model uses a `Q4_0_ROCMFP4_*` quantization type **and** a model architecture that is not in
+upstream llama.cpp. You need both. The patch in this directory adds the architecture on top of
+[`charlie12345/ROCmFPX`](https://github.com/charlie12345/ROCmFPX), which supplies the ROCmFP4
+quant types.
+
+Verified with `git apply --check` against the exact base commit named in the patch filename.
+
+## Build
+
+```bash
+git clone https://github.com/charlie12345/ROCmFPX.git
+cd ROCmFPX
+git checkout <BASE_COMMIT>          # the commit in the patch filename
+git apply /path/to/<patch-file>.patch
+cmake -B build -S . -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx1151 -DCMAKE_BUILD_TYPE=Release
+cmake --build build -j$(nproc)
+```
+
+`build/bin/llama-server` and `build/bin/llama-quantize` land in `build/bin/`. Set
+`LD_LIBRARY_PATH` to that directory when running them from elsewhere.
+
+⚠️ **If you add new files under `src/models/`, re-run `cmake -B build -S .`** — the `models/*.cpp`
+GLOB is evaluated at configure time, so `cmake --build` alone will not link them.
+
+## Which patch
+
+| patch | base commit | adds |
+|---|---|---|
+| `rocmfpx-2809dc5-add-bailingmoe3qlora-mellum-zaya.patch` | `2809dc5` | `mellum`, `zaya`, and the **Q-LoRA path** for `bailingmoe3` |
+| `rocmfpx-3edc3d3-add-muse-glimmer.patch` | `3edc3d3` | `muse-glimmer` |
+
+The `2809dc5` patch adds three architectures in one diff because they share seam files
+(`llama-arch`, `llama-model`, `llama-graph`, `gguf-py`); splitting them cleanly is error-prone,
+so it is shipped whole. Applying it gives you all three.
+
+## Notes on specific architectures
+
+**`bailingmoe3` (Ling-3.0)** — upstream ROCmFPX already has a `bailingmoe3.cpp`, but it was
+written for **Ling-3.0-flash**, which has `q_lora_rank: null`. **Ling-3.0-tiny sets
+`q_lora_rank: 256`** and compresses queries via `q_a_proj → q_a_layernorm → q_b_proj`. Without
+this patch, *every* GGUF of tiny fails — including BF16 and Q4_K_M — usually at
+`missing tensor 'blk.0.ssm_f.weight'`, before the Q-LoRA gap is even reached. The patch adds the
+Q-LoRA branch at both tensor-creation sites and both forward graphs, plus an `ssm_f`→`ssm_f_a`
+name fallback. Reference: upstream llama.cpp PR
+[#26608](https://github.com/ggml-org/llama.cpp/pull/26608).
+
+**`mellum` (Mellum2-12B-A2.5B)** — reference: upstream llama.cpp PR
+[#23966](https://github.com/ggml-org/llama.cpp/pull/23966). `layer_types` alternate
+sliding_attention ×3 → full_attention (`n_swa = 1024`); the patch honours that per-layer.
+
+**`zaya` (ZAYA1-8B)** — reference: upstream llama.cpp draft PR
+[#23112](https://github.com/ggml-org/llama.cpp/pull/23112). Uses Compressed Convolutional
+Attention. The patch **excludes `cca_conv_grp` from quantization** in `llama-quant.cpp` — those
+tensors are highly sensitive and must stay high precision. Also adds an `ssm-conv.cu` `case 2`
+kernel (`d_conv = 2`) and BF16-safe `conv_1d`.
+
+**`muse-glimmer`** — this architecture has no upstream PR we are aware of; the implementation in
+this patch is our own. ⚠️ **`-fa off` is mandatory** — flash attention breaks the vision path on
+gfx1151.
+
+## ⚠️ Do not use `strings` to check whether your build supports an architecture
+
+We hit this twice. A binary can contain the literal architecture name and still fail with
+`unknown model architecture` — the name lives in a tensor-name table while the loader is separate
+code. **Attempt the load.** That is the only reliable check.
+
+## Licence
+
+These patches are derivative of llama.cpp / ROCmFPX and carry the upstream **MIT** licence.