Add UD-Q4_K_XL run instructions (build llama.cpp from PR #25165)

README.md

@@ -156,6 +156,53 @@ cd llama.cpp && cmake -B build && cmake --build build -j
   --spec-type draft-dflash --spec-draft-n-max 15 -fa on --jinja --port 8000
 ```
 
+### Running Unsloth's `UD-Q4_K_XL` with llama.cpp (PR #25165)
+
+The GGUFs in this repo are Unsloth [Dynamic 2.0](https://docs.unsloth.ai/basics/unsloth-dynamic-2.0-ggufs)
+quants (imatrix calibrated). Laguna support is not in a tagged llama.cpp release
+yet, so build llama.cpp from
+[ggml-org/llama.cpp#25165](https://github.com/ggml-org/llama.cpp/pull/25165):
+
+```shell
+git clone https://github.com/ggml-org/llama.cpp
+cd llama.cpp
+gh pr checkout 25165
+# build with CUDA (drop -DGGML_CUDA=ON for a CPU-only build)
+cmake -B build -DGGML_CUDA=ON
+cmake --build build -j --config Release --target llama-cli llama-server
+cd ..
+```
+
+Download the `UD-Q4_K_XL` shards (~40GB, split into 3 files):
+
+```shell
+huggingface-cli download unsloth/Laguna-S-2.1-GGUF \
+  --include "UD-Q4_K_XL/*" \
+  --local-dir Laguna-S-2.1-GGUF
+```
+
+Serve it with `llama-server` (pass the first shard; the remaining shards load
+automatically):
+
+```shell
+./llama.cpp/build/bin/llama-server \
+    --model Laguna-S-2.1-GGUF/UD-Q4_K_XL/Laguna-S-2.1-UD-Q4_K_XL-00001-of-00003.gguf \
+    --jinja -fa on -ngl 99 --ctx-size 16384 --port 8000
+```
+
+Or run a one-off generation with `llama-cli`:
+
+```shell
+./llama.cpp/build/bin/llama-cli \
+    --model Laguna-S-2.1-GGUF/UD-Q4_K_XL/Laguna-S-2.1-UD-Q4_K_XL-00001-of-00003.gguf \
+    --jinja -ngl 99 -p "Write a Flappy Bird game in Python."
+```
+
+> [!NOTE]
+> `-ngl 99` offloads all layers to GPU; lower it (or drop it) if you run out of
+> VRAM. `UD-Q4_K_XL` is roughly 40GB, so it fits on a single 48GB+ GPU or splits
+> across several GPUs.
+
 ## Controlling reasoning
 
 Laguna S 2.1 has native reasoning support and works best with *preserved thinking*: