Add a 'Rebuild llama.cpp' Cookbook action to force a fresh GPU build (#1787)

The serve bootstrap builds llama-server from source only when it is missing
from PATH, so a host that first compiled CPU-only (no nvcc present at build
time) reuses that CPU-only binary on every later serve and never gets a GPU
build, even after a CUDA/ROCm toolkit is installed. There was no UI lever to
force a rebuild.

Adds a 'Rebuild llama.cpp' button to the Cookbook Dependencies tab. It clears
the cached ~/bin/llama-server symlink and ~/llama.cpp/build directory (locally
or on the selected remote server) so the next serve recompiles and picks up
CUDA/HIP if a toolchain is now present. It installs and downloads nothing.

- routes/cookbook_helpers.py: _llama_cpp_rebuild_cmd() (single source of truth)
- routes/shell_routes.py: POST /api/cookbook/rebuild-engine (admin-only, reuses
  the existing SSH plumbing for remote hosts)
- static/js/cookbook.js: header button + handler honoring the deps server selector
- tests: cover the command shape and a clean run on a fresh HOME

Motivated by #831 (RTX 4070 user stuck on a CPU-only build with no way to
re-trigger the build).

Co-authored-by: ghreprimand <203024559+ghreprimand@users.noreply.github.com>

routes/cookbook_helpers.py

@@ -552,6 +552,27 @@ def _append_llama_cpp_linux_accel_build_lines(runner_lines: list[str]) -> None:
     runner_lines.append('      cmake -B build -DCMAKE_BUILD_TYPE=Release && cmake --build build -j"$NPROC" --target llama-server && ln -sf ~/llama.cpp/build/bin/llama-server ~/bin/llama-server')
     runner_lines.append('    fi')
 
+
+def _llama_cpp_rebuild_cmd() -> str:
+    """Shell command that clears the Cookbook-managed llama.cpp build.
+
+    Removes the cached ``llama-server`` symlink and the ``~/llama.cpp/build``
+    directory so the next llama.cpp serve recompiles from source, picking up a
+    CUDA or HIP toolchain if one is now available. The serve bootstrap only
+    builds when ``llama-server`` is missing from PATH, so without this an
+    existing CPU-only build is reused forever. It deliberately installs and
+    downloads nothing; the rebuild itself happens on the next serve.
+    """
+    return (
+        'mkdir -p "$HOME/bin" && '
+        'rm -f "$HOME/bin/llama-server" && '
+        'rm -rf "$HOME/llama.cpp/build" && '
+        'echo "[odysseus] Cleared the cached llama.cpp build. '
+        'Re-launch the serve task to rebuild llama-server from source '
+        '(CUDA or HIP will be used if a toolchain is now available)."'
+    )
+
+
 class ModelDownloadRequest(BaseModel):
     repo_id: str
     include: str | None = None  # glob pattern e.g. "*Q4_K_M*"

routes/shell_routes.py

@@ -1058,4 +1058,39 @@ def setup_shell_routes() -> APIRouter:
             return {"ok": True, "output": stdout.decode()[-200:]}
         return {"ok": False, "error": stderr.decode()[-300:]}
 
+    @router.post("/api/cookbook/rebuild-engine")
+    async def rebuild_engine(request: Request):
+        """Clear the cached llama.cpp build so the next serve recompiles.
+
+        Admin only — this removes the Cookbook-managed ``~/bin/llama-server``
+        symlink and ``~/llama.cpp/build`` directory, locally or on the selected
+        remote server. It installs and downloads nothing; the next llama.cpp
+        serve rebuilds from source and picks up CUDA/HIP if a toolchain is now
+        present. This is the missing "force a fresh GPU build" lever for hosts
+        stuck on a CPU-only llama-server.
+        """
+        _require_admin(request)
+        from routes.cookbook_helpers import _llama_cpp_rebuild_cmd
+        body = await request.json()
+        engine = str(body.get("engine") or "llamacpp").strip()
+        if engine != "llamacpp":
+            return {"ok": False, "error": f"Unsupported engine: {engine}"}
+        host = str(body.get("remote_host") or "").strip()
+        ssh_port = body.get("ssh_port")
+        cmd = _llama_cpp_rebuild_cmd()
+        try:
+            argv = (_ssh_base_argv(host, ssh_port) + [cmd]) if host else ["bash", "-lc", cmd]
+        except ValueError as e:
+            raise HTTPException(400, str(e))
+        try:
+            proc = await asyncio.create_subprocess_exec(
+                *argv, stdout=asyncio.subprocess.PIPE, stderr=asyncio.subprocess.PIPE
+            )
+            out, err = await asyncio.wait_for(proc.communicate(), timeout=30)
+        except asyncio.TimeoutError:
+            return {"ok": False, "error": "Rebuild-engine command timed out."}
+        if proc.returncode == 0:
+            return {"ok": True, "output": out.decode("utf-8", errors="replace")[-400:]}
+        return {"ok": False, "error": err.decode("utf-8", errors="replace")[-400:]}
+
     return router