Enhance LLM configuration and installation scripts for improved flexibility

- Added new environment variables in `env.example` and `gpu-rent.vars.example` for fine-tuning LLM settings, including CUDA build options and model version pinning.
- Updated `llm.md` documentation to include detailed descriptions of new configuration options and usage cases for LLM setups.
- Enhanced the `provision.py` script to forward new environment variables during remote installations, improving the installation process for LLM components.
- Modified the `install_llamacpp.sh` script to support conditional CUDA builds and asset URL overrides, ensuring better compatibility with various environments.
- Improved logging in the installation scripts to provide clearer feedback during the setup process.

src/gpu_rent/provision.py

@@ -29,6 +29,32 @@ from gpu_rent.sync_files import pull_tree, push_tree
 Log = Callable[[str], None]
 DATA = "/mnt/swarm_data"
 
+# Forwarded to remote install_*.sh (from .env / gpu-rent.vars → os.environ).
+_OLLAMA_INSTALL_ENV = ("OLLAMA_VERSION", "OLLAMA_SHA256")
+_LLAMACPP_INSTALL_ENV = (
+    "LLAMACPP_TAG",
+    "LLAMACPP_ASSET_URL",
+    "LLAMACPP_SHA256",
+    "LLAMACPP_BUILD_CUDA",
+    "LLAMACPP_FORCE_REINSTALL",
+    "LLAMACPP_NGL",
+    "LLAMACPP_CTX",
+    "LLAMACPP_HOST",
+    "LLAMACPP_PORT",
+    "LLAMACPP_EXTRA_ARGS",
+)
+
+
+def _remote_llm_env(cfg: Config, *keys: str) -> dict[str, str]:
+    import os
+
+    env: dict[str, str] = {"SWARM_USER": cfg.ssh_user}
+    for key in keys:
+        value = (os.environ.get(key) or "").strip()
+        if value:
+            env[key] = value
+    return env
+
 
 def _pkg_text(name: str) -> str:
     return files("gpu_rent.remote").joinpath(name).read_text(encoding="utf-8")
@@ -442,7 +468,7 @@ def provision_llm(cfg: Config, host: str, log: Log) -> None:
             _pkg_text("install_ollama.sh"),
             remote_path="/tmp/gpu-rent-install_ollama.sh",
             timeout=900,
-            env={"SWARM_USER": cfg.ssh_user},
+            env=_remote_llm_env(cfg, *_OLLAMA_INSTALL_ENV),
             log=log,
         )
         entries = parse_ollama_models(cfg.ollama_models_manifest)
@@ -516,13 +542,16 @@ def provision_llm(cfg: Config, host: str, log: Log) -> None:
         else:
             log("llamacpp-models.yaml пуст — GGUF skip (положи вручную)")
         log("LLM: ставим/запускаем llama.cpp server")
+        import os
+
+        build_cuda = (os.environ.get("LLAMACPP_BUILD_CUDA") or "").strip() == "1"
         run_script_sudo(
             cfg,
             host,
             _pkg_text("install_llamacpp.sh"),
             remote_path="/tmp/gpu-rent-install_llamacpp.sh",
-            timeout=1200,
-            env={"SWARM_USER": cfg.ssh_user},
+            timeout=3600 if build_cuda else 1200,
+            env=_remote_llm_env(cfg, *_LLAMACPP_INSTALL_ENV),
             log=log,
         )
     st = load_state()