refactor: remove one-click profiles system

docs/model-configuration.md

@@ -9,7 +9,6 @@ Models are configured in a YAML file (default: `config/models.yaml`). Each entry
 | Argument | Env Var | Default | Description |
 |---|---|---|---|
 | `--config` | — | `config/models.yaml` | Path to models config file |
-| `--model-stack` | `MSHIP_MODEL_STACK` | — | Auto-generate a config from a [profile](#profiles-mship_model_stack) (`chat`/`assistant`/`studio`/`everything`) sized to detected hardware |
 | `--gateway-name` | `MSHIP_GATEWAY_NAME` | `modelship api` | Name for the API gateway app |
 | `--use-existing-ray-cluster` | `MSHIP_USE_EXISTING_RAY_CLUSTER` | `false` | Connect to a Ray cluster you manage (must run on a cluster node) instead of starting one. Implies deploy-and-exit (no teardown) |
 | `--prune-ray-sessions` | `MSHIP_PRUNE_RAY_SESSIONS` | `true` | When starting its own Ray head, delete stale `session_*` dirs left under the Ray temp root (default `/tmp/ray`) by previous, no-longer-running heads — Ray never cleans these up, so they fill the disk across restarts. A live head's session is always kept. Set `false` to keep them (e.g. for debugging). No effect with `--use-existing-ray-cluster` |
@@ -64,47 +63,6 @@ python mship_deploy.py --config config/llm.yaml --gateway-name "llm-api"
 python mship_deploy.py --config config/tts.yaml --gateway-name "tts-api"
 ```
 
-## Profiles (`MSHIP_MODEL_STACK`)
-
-Don't want to pick models, write YAML, or hand-allocate GPU/CPU? Set a **profile** and modelship generates a config sized to your detected hardware:
-
-```bash
-MSHIP_MODEL_STACK=studio uv run mship_deploy.py
-# equivalently:
-uv run mship_deploy.py --model-stack studio
-```
-
-A profile is a set of capabilities to serve. Every model in the catalog is **ungated**, so the one-click path needs no `HF_TOKEN`.
-
-| Profile | Capabilities |
-|---|---|
-| `chat` | generate + embed |
-| `assistant` | generate + transcription + tts |
-| `studio` | generate + image + embed |
-| `everything` | generate + image + embed + transcription + tts |
-
-### Models per tier
-
-The `generate` and `image` models scale with a **tier** (small/medium/large) picked from your hardware — GPU VRAM if a GPU is present, otherwise RAM and core count. Bigger box → bigger tier. The other capabilities are the same on every box.
-
-| Capability | small | medium | large |
-|---|---|---|---|
-| generate (GPU, vLLM/AWQ) | Qwen2.5-7B | Qwen2.5-14B | Qwen2.5-32B |
-| generate (CPU, llama.cpp/GGUF) | Llama-3.2-3B | Qwen2.5-7B | Qwen2.5-14B |
-| image (GPU, diffusers) | SD-Turbo | SDXL-Turbo | playground-v2.5 |
-| image (CPU, stable-diffusion.cpp) | SD-Turbo | SDXL-Turbo | SDXL-base |
-| embed (CPU, llama.cpp) | nomic-embed-text-v1.5 | ← | ← |
-| tts (CPU, `kokoroonnx`) | Kokoro-82M | ← | ← |
-| transcription (CPU, `whispercpp`) | whisper `base` | whisper `small` | ← |
-
-Selection is **all-or-nothing**: the highest tier whose *complete* stack fits is chosen. A capability is never dropped to squeeze something in — if even the smallest tier won't fit, the deploy refuses with a clear message instead of serving a partial stack. Resource requests (`num_cpus`/`num_gpus`) are filled in automatically; on a shared GPU the LLM gets the larger slice.
-
-### File behavior and precedence
-
-- The generated `config/models_stack_<profile>.yaml` is **regenerated from scratch on every start** while the profile is set — switch profiles just by changing the value, with no stale file to delete.
-- Resolution order: an explicit `--config` always wins; otherwise `MSHIP_MODEL_STACK`/`--model-stack`; otherwise the default `config/models.yaml`.
-- The file is normal, editable YAML, but **edits are overwritten** on the next profile-driven start. To keep changes, copy it to `config/models.yaml` (or pass `--config <file>`) and unset `MSHIP_MODEL_STACK`.
-
 ## Fields
 
 | Field | Type | Description |
@@ -509,7 +467,6 @@ deployment.
 | Variable | Description | Default |
 |---|---|---|
 | `HF_TOKEN` | HuggingFace access token | — |
-| `MSHIP_MODEL_STACK` | [Profile](#profiles-mship_model_stack) to auto-generate a hardware-sized config from (`chat`/`assistant`/`studio`/`everything`) | — |
 | `MSHIP_CACHE_DIR` | Model cache directory (HuggingFace + plugins) | `/.cache` |
 | `MSHIP_STATE_STORE` | State-store connection URI for the effective config + deploy coordinator (see [State store](#state-store-mship_state_store)) | `memory://` |
 | `MSHIP_STATE_DIR` | Default directory for a `file://` state store with no path | `<cache-dir>/state` |

modelship/deploy/config.py

@@ -17,55 +17,20 @@ def resolve_config_path(arg_path: str | None, config_dir: Path | None = None) ->
 
     Precedence:
     1. An explicit ``--config`` path always wins (most specific signal); it must exist.
-    2. ``MSHIP_MODEL_STACK=<profile>`` (or ``--model-stack``) → regenerate
-       ``models_stack_<profile>.yaml`` from scratch on every start, sized to the
-       detected hardware, and deploy that. Regenerating fresh each time lets the
-       user switch profiles by just changing the value — no stale file to delete by
-       hand. Refuses with a clean exit (no partial deploy) if the profile can't fit.
-    3. Otherwise the default ``config/models.yaml`` must exist.
+    2. Otherwise the default ``config/models.yaml`` must exist.
     """
     config_dir = config_dir or Path(__file__).resolve().parent.parent.parent / "config"
-    stack = os.environ.get("MSHIP_MODEL_STACK")
 
     if arg_path:
         if not os.path.exists(arg_path):
             raise FileNotFoundError(f"--config {arg_path} not found.")
         return arg_path
 
-    if stack:
-        from modelship.deploy.profiles.catalog import PROFILES
-        from modelship.deploy.profiles.generator import generate_models_yaml
-        from modelship.deploy.profiles.selector import ProfileDoesNotFitError
-
-        # Validate against the known profiles BEFORE building a path or touching the
-        # filesystem — `stack` is operator-supplied (env var / CLI), and feeding it
-        # into the filename unchecked would allow path traversal on the unlink below.
-        if stack not in PROFILES:
-            raise SystemExit(f"MSHIP_MODEL_STACK={stack!r}: unknown profile; choose one of {sorted(PROFILES)}.")
-
-        path = config_dir / f"models_stack_{stack}.yaml"
-        logger.info("MSHIP_MODEL_STACK=%s: generating %s for the detected hardware...", stack, path)
-        try:
-            # Remove any prior generation first so a refusal never leaves a stale
-            # file behind that a later run could mistake for hand-authored config.
-            path.unlink(missing_ok=True)
-            generate_models_yaml(stack, str(path))
-        except (ProfileDoesNotFitError, ValueError) as e:
-            raise SystemExit(f"MSHIP_MODEL_STACK={stack}: {e}") from e
-        except OSError as e:
-            # Read-only / permission-denied config dir, etc. — fail cleanly instead
-            # of dumping a traceback.
-            raise SystemExit(f"MSHIP_MODEL_STACK={stack}: cannot write {path}: {e}") from e
-        return str(path)
-
     default = config_dir / "models.yaml"
     if default.exists():
         return str(default)
 
-    raise FileNotFoundError(
-        f"{default} not found. Set MSHIP_MODEL_STACK=<profile> (or pass --model-stack) to "
-        f"auto-generate one, or copy an example config from config/examples/ to config/models.yaml."
-    )
+    raise FileNotFoundError(f"{default} not found. Copy an example config from config/examples/ to config/models.yaml.")
 
 
 def load_yaml_config(arg_path: str | None) -> ModelshipConfig:
@@ -78,9 +43,7 @@ def load_raw_models(arg_path: str | None) -> list[dict]:
 
     The effective-config store keeps raw dicts (not validated configs, which don't
     round-trip through num_gpus/tp normalization), so the deploy path merges at the
-    raw-dict level and validates only the merged result. Shares resolve_config_path
-    with load_yaml_config so MSHIP_MODEL_STACK generation runs at most once per
-    deploy (callers should use one or the other, not both)."""
+    raw-dict level and validates only the merged result."""
     with open(resolve_config_path(arg_path)) as f:
         data = yaml.safe_load(f) or {}
     if not isinstance(data, dict):