unslothai/unsloth · error · SttUnavailableError

llama-server did not become ready for '{model_id}'.

Error message

llama-server did not become ready for '{model_id}'.

What it means

SttUnavailableError raised when _wait_for_server() returns False without cancellation: llama-server was spawned for model_id but did not answer its readiness probe in time. The failed process is reaped first (it was never assigned to self._process, so unload() could not reach it), then the error names the model.

Source

Thrown at studio/backend/core/inference/stt_mtmd_sidecar.py:927

                # Die with Studio, so a crash never orphans a server on the GPU.
                **child_popen_kwargs(),
            )
            # Published before the wait, so training can preempt a startup that
            # is already allocating; _process is not set for another 180s.
            with self._lock:
                self._starting_process = process
            adopt_pid(process.pid)  # terminate_all backstop for graceful exits
            if not self._wait_for_server(process, port, cancel_event):
                # Reap it here: _process was never assigned, so unload() cannot
                # reach a child that ignores SIGTERM and keeps port and VRAM.
                _reap(process)
                if cancel_event.is_set():
                    # 409 through the route, like the other sidecars: expected
                    # preemption, not a broken or missing runtime (501).
                    raise SttLoadCancelledError(
                        "Dictation model loading was cancelled so training could start."
                    )
                raise SttUnavailableError(f"llama-server did not become ready for '{model_id}'.")
            with self._lock:
                self._process = process
                self._port = port
                self._model_id = model_id
                self._gpu_disabled = training
                self._generation += 1
                self._schedule_idle_unload_locked()
        finally:
            with self._lock:
                self._loading = False
                self._load_cancel_event = None
                self._load_owner_cancel_event = None
                self._starting_process = None

    @staticmethod
    def _wait_for_server(
        process: subprocess.Popen,
        port: int,

View on GitHub (pinned to 203007d190)

Solutions

1. Retry once — cold-cache loads of large models can exceed the readiness window.
2. Check llama-server logs/stderr for crash causes (bad GGUF, CUDA OOM, missing mmproj).
3. Free GPU memory (close other model backends) or pick a smaller STT model.
4. Re-download the model to rule out a corrupted snapshot.

```python
import shutil, psutil
if psutil.virtual_memory().percent > 95 or gpu_memory_free() < model_min_vram(model_id):
    free_resources_before_load(model_id)
```

```python
try:
    sidecar.load(model_id)
except SttUnavailableError as exc:
    if "did not become ready" in str(exc):
        log_llama_server_stderr()
        retry_once_after(freeing_vram=True)
    raise
```

When it happens

Trigger: llama-server startup exceeding the readiness timeout: huge model loading from disk, GPU driver stall/OOM during CUDA init, corrupt GGUF file, or the server crashing on an incompatible mmproj/model pairing.

Common situations: First load of a large model on a slow disk; VRAM exhausted by other processes; mismatched model/mmproj revisions after a partial download; antivirus or sandbox slowing process spawn on Windows.

Related errors

• The local transcription runtime is being updated. Try dictat
• STT model '{model}' is not a curated llama.cpp dictation mod
• llama.cpp is not installed, so these dictation models cannot
• Transcription cancelled.
• Dictation model loading was cancelled so training could star

AI-assisted analysis of unslothai/unsloth@203007d190 (2026-08-15). Data as JSON: /api/errors/77b53ce91940b529. Report an issue: GitHub.