mlxsmith 0.1.0__py3-none-any.whl → 0.1.2__py3-none-any.whl

mlxsmith/accel/__init__.py

@@ -1,10 +1,7 @@
 from __future__ import annotations
 from .none import NoneBackend
-from .zmlx_backend import ZMLXBackend
 
 def get_backend(name: str):
     if name == "none":
         return NoneBackend()
-    if name == "zmlx":
-        return ZMLXBackend()
     raise ValueError(f"Unknown accel backend: {name}")

mlxsmith/api/handlers.py

@@ -20,11 +20,10 @@ import uuid
 from pathlib import Path
 from typing import Any, AsyncGenerator, Callable, Dict, List, Optional
 
-from fastapi import APIRouter, Depends, FastAPI, HTTPException, Request, Security, status
+from fastapi import APIRouter, FastAPI, HTTPException, Request, Security, status
 from starlette.middleware.base import BaseHTTPMiddleware
 from fastapi.responses import StreamingResponse
 from fastapi.security import HTTPBearer, HTTPAuthorizationCredentials
-from pydantic import BaseModel
 
 from .schemas import (
     AdapterReloadRequest,
@@ -1102,12 +1101,10 @@ def create_router(
         models, use the list endpoint to check completion status.
         """
         cache_dir = _get_cache_dir()
-        local_path = cache_dir / "mlx" / request.model_id.replace("/", "__")
-        
+
         try:
             # Import here to avoid circular dependencies
             from ..models import hf_pull
-            from ..config import ProjectConfig
             
             # Get HF token if available
             hf_token = None

mlxsmith/cli.py

@@ -43,7 +43,7 @@ from .envs import (
 
 app = typer.Typer(
     add_completion=False,
-    help="mlxsmith — MLX fine-tuning + OpenAI-compatible serving (SFT stable; preference/RL experimental)",
+    help="mlxsmith — Apple Silicon MLX fine-tuning toolkit: SFT, DPO/ORPO, GRPO, distillation, and OpenAI-compatible serving.",
 )
 console = Console()
 
@@ -83,7 +83,6 @@ def doctor():
     table.add_row("cpu_count", str(info.cpu_count))
     table.add_row("metal", str(info.has_metal))
     table.add_row("mlx", f"{info.has_mlx} {info.mlx_version or ''}".strip())
-    table.add_row("zmlx", str(info.has_zmlx))
     console.print(table)
 
 
@@ -564,7 +563,7 @@ def config_validate(
     
     try:
         cfg = load_config(cfg_path, require=True)
-        console.print(f"[green]✓ Configuration is valid[/green]")
+        console.print("[green]✓ Configuration is valid[/green]")
         
         # Show summary
         table = Table(title="Configuration Summary")
@@ -593,9 +592,9 @@ def config_env(
     """Show available environment variables."""
     cfg = ProjectConfig()
     
-    console.print(f"\n[bold]Environment Variable Configuration[/bold]")
+    console.print("\n[bold]Environment Variable Configuration[/bold]")
     console.print(f"Prefix: [cyan]{prefix}[/cyan]")
-    console.print(f"Nested delimiter: [cyan]__[/cyan] (double underscore)\n")
+    console.print("Nested delimiter: [cyan]__[/cyan] (double underscore)\n")
     
     table = Table(title=f"Available {prefix}* Environment Variables")
     table.add_column("Environment Variable")
@@ -729,7 +728,7 @@ def rlm_history(limit: int = typer.Option(10, "--limit")):
 
 @accel_app.command("status")
 def accel_status():
-    backends = ["none", "zmlx"]
+    backends = ["none"]
     table = Table(title="mlxsmith accel status")
     table.add_column("backend")
     table.add_column("available")

mlxsmith/config.py

@@ -18,7 +18,6 @@ Config files support @path syntax:
 from __future__ import annotations
 
 import json
-import os
 from pathlib import Path
 from typing import Any, Dict, List, Optional, Tuple, Union
 
@@ -94,7 +93,7 @@ class ProjectSettings(BaseSettings):
 
 
 # Import CLI aliases from models
-from .config_models import CLI_ALIASES as _CLI_ALIASES
+from .config_models import CLI_ALIASES as _CLI_ALIASES  # noqa: E402
 
 
 def resolve_config_path(config: Union[str, Path], root: Optional[Path] = None) -> Path:

mlxsmith/config_models.py

@@ -6,7 +6,7 @@ from typing import Dict, List, Literal, Optional, Any
 
 from pydantic import BaseModel, Field, field_validator
 
-AccelBackendName = Literal["none", "zmlx"]
+AccelBackendName = Literal["none"]
 
 
 class ModelConfig(BaseModel):

mlxsmith/llm/interface.py

@@ -182,11 +182,9 @@ def compute_logprobs(
     else:
         prompt_len = len(prompt_ids)
     
-    # Get generation with logprobs
-    full_text = backend.decode(ids)
-    
-    # Use backend's sequence_logprob if available
-    seq_logprob = backend.sequence_logprob(ids, prompt_len=prompt_len)
+    # Decode and compute sequence-level logprob (used by callers via backend state)
+    backend.decode(ids)
+    backend.sequence_logprob(ids, prompt_len=prompt_len)
     
     # For per-token logprobs, we'd need to do a forward pass
     # This is a simplified version

mlxsmith/llm/mock_backend.py

@@ -2,7 +2,7 @@ from __future__ import annotations
 
 import random
 from dataclasses import dataclass
-from typing import Sequence, Any, List, Dict, Optional
+from typing import Sequence, Any, List, Dict
 
 from .backend import Generation
 

mlxsmith/orchestrator/daemon.py

@@ -9,27 +9,22 @@ Manages rollout requests, training batches, and weight updates.
 
 from __future__ import annotations
 
-import json
 import multiprocessing as mp
 import signal
-import sys
 import time
 import traceback
 from dataclasses import dataclass, field
 from pathlib import Path
-from typing import Any, Dict, List, Optional, Callable
+from typing import Any, Dict, List, Optional
 
 from rich.console import Console
 
 from ..config import ProjectConfig
-from ..rlm.corpus import append_corpus, load_corpus, sample_corpus
-from ..rlm.gating import load_state, save_state, should_accept, update_state
-from ..rlm.history import append_history
-from ..rlm.inference import Rollout, build_tasks
+from ..rlm.gating import load_state
 from ..rlm.weights import WeightPointerStore, WeightPointerIPC
 from ..runs import new_run, snapshot_config
-from ..util import ensure_dir, now_ts, write_jsonl
-from .queue import MessageQueue, MessageType, Message
+from ..util import ensure_dir, now_ts
+from .queue import MessageQueue, MessageType
 from .inference_worker import InferenceConfig, run_inference_worker
 from .trainer_worker import TrainerConfig, run_trainer_worker
 

mlxsmith/orchestrator/inference_worker.py

@@ -26,7 +26,7 @@ from fastapi.responses import StreamingResponse
 from ..config import ProjectConfig
 from ..llm.registry import get_llm_backend
 from ..models import resolve_model_spec
-from ..rlm.weights import WeightPointerStore, WeightPointerIPC
+from ..rlm.weights import WeightPointerStore
 from .queue import MessageQueue, MessageType, Message
 
 

mlxsmith/orchestrator/queue.py

@@ -11,11 +11,11 @@ import json
 import multiprocessing as mp
 import time
 import uuid
-from dataclasses import asdict, dataclass, field
+from dataclasses import dataclass, field
 from enum import Enum, auto
 from pathlib import Path
 from queue import Empty
-from typing import Any, Dict, List, Optional, Union
+from typing import Any, Dict, List, Optional
 
 
 class MessageType(Enum):

mlxsmith/orchestrator/trainer_worker.py

@@ -9,7 +9,6 @@ from __future__ import annotations
 
 import json
 import signal
-import sys
 import time
 import traceback
 from collections import defaultdict
@@ -17,7 +16,6 @@ from dataclasses import dataclass, field
 from pathlib import Path
 from typing import Any, Dict, List, Optional
 
-from ..config import ProjectConfig
 from ..llm.registry import get_llm_backend
 from ..rlm.inference import Rollout
 from ..rlm.weights import WeightPointerStore, WeightPointerIPC

mlxsmith/rlm/loop.py

@@ -15,12 +15,11 @@ from __future__ import annotations
 import json
 import multiprocessing as mp
 import signal
-import sys
 import time
 import traceback
-from dataclasses import dataclass, asdict
+from dataclasses import dataclass
 from pathlib import Path
-from typing import Any, Dict, List, Optional, Callable
+from typing import Dict, List, Optional
 
 from rich.console import Console
 
@@ -35,10 +34,9 @@ from ..verifiers.docker_verifier import verify as docker_verify
 from ..verifiers.pytest_verifier import verify as pytest_verify
 from .corpus import append_corpus, load_corpus, sample_corpus
 from .gating import load_state, save_state, should_accept, update_state
-from .generate import GeneratedTask, generate_tasks, filter_tasks
+from .generate import GeneratedTask
 from .history import append_history
 from .inference import Rollout, build_tasks
-from .mutate import mutate_tasks
 from .trainer import train_on_rollouts
 from .weights import (
     WeightPointer,
@@ -373,9 +371,9 @@ def run_rlm(
 # Multi-Process Orchestrated RLM
 # =============================================================================
 
-from ..orchestrator.queue import MessageQueue, MessageType, Message
-from ..orchestrator.inference_worker import InferenceConfig, run_inference_worker
-from ..orchestrator.trainer_worker import TrainerConfig, run_trainer_worker
+from ..orchestrator.queue import MessageQueue, MessageType, Message  # noqa: E402
+from ..orchestrator.inference_worker import InferenceConfig, run_inference_worker  # noqa: E402
+from ..orchestrator.trainer_worker import TrainerConfig, run_trainer_worker  # noqa: E402
 
 
 @dataclass

mlxsmith/rlm/weights.py

@@ -9,7 +9,7 @@ from __future__ import annotations
 import json
 import multiprocessing as mp
 import time
-from dataclasses import dataclass, asdict
+from dataclasses import dataclass
 from pathlib import Path
 from typing import Optional, Callable
 

mlxsmith/sdk/future.py

@@ -7,9 +7,8 @@ and progress tracking for async operations.
 from __future__ import annotations
 
 import threading
-import time
 from concurrent.futures import Future, ThreadPoolExecutor
-from typing import Any, Callable, Generic, Iterable, Optional, TypeVar, Union
+from typing import Any, Callable, Generic, Iterable, Optional, TypeVar
 
 from ..llm.backend import DecodingConfig
 
@@ -344,7 +343,7 @@ class APIFuture(Generic[T]):
             self._run_finally_callbacks()
             return True
     
-    def cancelled(self) -> bool:  # type: ignore
+    def cancelled(self) -> bool:  # type: ignore  # noqa: F811
         """Check if the future was cancelled."""
         with self._lock:
             return self._state == APIFutureState.CANCELLED

mlxsmith/sdk/sampling_client.py

@@ -27,8 +27,7 @@ Example:
 from __future__ import annotations
 
 from dataclasses import dataclass, field
-from typing import Any, Dict, List, Optional, Sequence, Union, Callable
-import concurrent.futures
+from typing import Any, Dict, List, Optional, Sequence
 
 from .future import APIFuture, SdkFuturePool
 

mlxsmith/sdk/training_client.py

@@ -22,7 +22,7 @@ Example:
 from __future__ import annotations
 
 from dataclasses import dataclass
-from typing import Any, Callable, Dict, List, Optional, Sequence, Tuple, Union
+from typing import Any, Callable, Dict, List, Optional
 
 from .future import APIFuture, SdkFuturePool
 
@@ -194,7 +194,7 @@ class TrainingClient:
             if batch.is_preference:
                 # Preference training (DPO, ORPO, etc.)
                 if batch.rejected_responses is None:
-                    raise ValueError(f"Preference batch requires rejected_responses")
+                    raise ValueError("Preference batch requires rejected_responses")
                 
                 for prompt, chosen, rejected in zip(
                     batch.prompts, 
@@ -345,7 +345,6 @@ class TrainingClient:
         def _run_save() -> CheckpointResult:
             try:
                 from pathlib import Path
-                import json
                 
                 save_path = Path(path)
                 save_path.parent.mkdir(parents=True, exist_ok=True)
@@ -673,4 +672,4 @@ class DistillationTrainingClient(TrainingClient):
 
 
 # Import at end to avoid circular dependency
-from .sampling_client import SamplingClient
+from .sampling_client import SamplingClient  # noqa: E402

mlxsmith/server.py

@@ -10,15 +10,10 @@ This module provides:
 
 from __future__ import annotations
 
-import json
-import time
-import uuid
 from pathlib import Path
-from typing import Any, Dict, List, Optional
 
 from fastapi import FastAPI
-from fastapi.responses import StreamingResponse, HTMLResponse
-from pydantic import BaseModel
+from fastapi.responses import HTMLResponse
 
 from .config import ProjectConfig
 from .models import resolve_model_spec

mlxsmith/train/lora.py

@@ -160,6 +160,8 @@ def apply_lora(model: Any, cfg: LoRAConfig) -> dict:
         keys = sorted(_keys_for_target_modules(model, cfg.target_modules))
 
     if tuner_utils is not None and hasattr(tuner_utils, "linear_to_lora_layers"):
+        # Freeze all base weights first so only LoRA params are trainable
+        model.freeze()
         # MLX-LM format
         config = {
             "rank": int(cfg.r),
@@ -256,6 +258,8 @@ def apply_adapter(model: Any, adapter_dir: str | Path) -> dict | None:
     adapter_cfg = load_adapter_config(adapter_dir)
     tuner_utils, _ = _try_mlx_lm_utils()
     if adapter_cfg is not None and tuner_utils is not None and hasattr(tuner_utils, "load_adapters"):
+        # Freeze base weights so only adapter params are trainable
+        model.freeze()
         tuner_utils.load_adapters(model, str(adapter_dir))
         return adapter_cfg
 

mlxsmith/util.py

@@ -46,7 +46,6 @@ class SystemInfo:
     has_metal: Optional[bool]
     has_mlx: bool
     mlx_version: Optional[str]
-    has_zmlx: bool
 
 def detect_system() -> SystemInfo:
     has_mlx = False
@@ -58,10 +57,6 @@ def detect_system() -> SystemInfo:
     except Exception:
         pass
 
-    import importlib.util
-
-    has_zmlx = importlib.util.find_spec("zmlx") is not None
-
     # Metal detection (best-effort): on macOS we assume Metal is present; for CI, this is not reliable.
     has_metal = None
     if sys.platform == "darwin":
@@ -83,7 +78,6 @@ def detect_system() -> SystemInfo:
         has_metal=has_metal,
         has_mlx=has_mlx,
         mlx_version=mlx_version,
-        has_zmlx=has_zmlx,
     )
 
 def require(cond: bool, msg: str):

mlxsmith-0.1.2.dist-info/METADATA

@@ -0,0 +1,283 @@
+Metadata-Version: 2.4
+Name: mlxsmith
+Version: 0.1.2
+Summary: Apple Silicon MLX fine-tuning toolkit — SFT, DPO/ORPO, GRPO, distillation, and OpenAI-compatible serving.
+Author-email: Shannon Labs <hmbown@gmail.com>
+License: MIT
+Project-URL: Homepage, https://github.com/Hmbown/MLXSmith
+Project-URL: Repository, https://github.com/Hmbown/MLXSmith
+Project-URL: Issues, https://github.com/Hmbown/MLXSmith/issues
+Keywords: mlx,apple-silicon,llm,fine-tuning,lora,openai-compatible
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Developers
+Classifier: Intended Audience :: Science/Research
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3 :: Only
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Operating System :: MacOS :: MacOS X
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: typer>=0.9.0
+Requires-Dist: rich>=13.7.0
+Requires-Dist: pyyaml>=6.0.1
+Requires-Dist: pydantic>=2.5.0
+Requires-Dist: pydantic-settings>=2.2.1
+Requires-Dist: tomli>=2.0.1; python_version < "3.11"
+Requires-Dist: huggingface_hub>=1.3.4
+Requires-Dist: jsonschema>=4.21.0
+Provides-Extra: mlx
+Requires-Dist: mlx>=0.30.4; extra == "mlx"
+Provides-Extra: llm
+Requires-Dist: mlx-lm>=0.30.5; extra == "llm"
+Requires-Dist: transformers>=5.0.0; extra == "llm"
+Requires-Dist: datasets>=3.0.0; extra == "llm"
+Provides-Extra: serve
+Requires-Dist: fastapi>=0.128.0; extra == "serve"
+Requires-Dist: uvicorn>=0.40.0; extra == "serve"
+Requires-Dist: httpx>=0.28.0; extra == "serve"
+Provides-Extra: dev
+Requires-Dist: pytest>=9.0.0; extra == "dev"
+Requires-Dist: ruff>=0.14.0; extra == "dev"
+Provides-Extra: all
+Requires-Dist: mlx>=0.30.4; extra == "all"
+Requires-Dist: mlx-lm>=0.30.5; extra == "all"
+Requires-Dist: transformers>=5.0.0; extra == "all"
+Requires-Dist: datasets>=3.0.0; extra == "all"
+Requires-Dist: fastapi>=0.128.0; extra == "all"
+Requires-Dist: uvicorn>=0.40.0; extra == "all"
+Requires-Dist: httpx>=0.28.0; extra == "all"
+Dynamic: license-file
+
+# mlxsmith
+
+Apple Silicon MLX fine-tuning toolkit — SFT, DPO/ORPO, GRPO, distillation, and OpenAI-compatible serving.
+
+**Status:** alpha (v0.1.2). Full training pipeline validated on Qwen3-4B.
+
+## Install
+
+MLX training and serving require macOS on Apple Silicon.
+Other platforms can use data tools and mock backends.
+
+```bash
+python -m venv .venv && source .venv/bin/activate
+pip install -U pip
+
+# Core CLI (data tools, config, project scaffolding)
+pip install mlxsmith
+
+# Apple Silicon training + serving
+pip install "mlxsmith[mlx,llm,serve]"
+
+# Everything
+pip install "mlxsmith[all]"
+```
+
+## Quickstart
+
+```bash
+mlxsmith init myproj
+cd myproj
+mlxsmith doctor        # check Python, MLX, Metal
+```
+
+## Training
+
+### SFT (LoRA/QLoRA)
+
+```bash
+mlxsmith sft --model cache/mlx/Qwen__Qwen3-4B-Instruct-2507 --data data/sft
+```
+
+Produces run artifacts under `runs/sft_NNNN/` (adapter weights, `metrics.jsonl`, config snapshot).
+
+### Preference tuning (DPO/ORPO)
+
+```bash
+mlxsmith pref --model cache/mlx/Qwen__Qwen3-4B-Instruct-2507 \
+  --data data/prefs --algo dpo
+```
+
+Supports DPO and ORPO algorithms with configurable beta and KL coefficients. Expects `{prompt, chosen, rejected}` data format.
+
+### Reinforced fine-tuning (GRPO)
+
+```bash
+mlxsmith rft --model cache/mlx/Qwen__Qwen3-4B-Instruct-2507 \
+  --env envs/coding.yaml --verifier verifiers/pytest.py
+```
+
+GRPO-style RL training with token-level environment integration and verifier-based rewards. Rollout acceptance/rejection gating with reward tracking.
+
+### Knowledge distillation
+
+```bash
+# Offline distillation (teacher generates, student learns)
+mlxsmith distill --teacher large-model --student small-model --mode offline
+
+# Online preference distillation (OPD)
+mlxsmith distill --teacher large-model --student small-model --mode opd
+```
+
+### Full pipeline
+
+```bash
+# Run SFT → Pref → RFT in sequence
+mlxsmith pipeline
+```
+
+## Serving
+
+OpenAI-compatible `/v1/chat/completions` endpoint.
+
+```bash
+mlxsmith serve --model runs/sft_0001/adapter --port 8080
+```
+
+```bash
+curl http://localhost:8080/v1/chat/completions \
+  -H 'Content-Type: application/json' \
+  -d '{"messages":[{"role":"user","content":"Hello"}],"max_tokens":64}'
+```
+
+Supports streaming (`"stream": true`), logprobs, stop sequences, and an optional UI dashboard (`serve.ui: true` in config).
+
+## Data tools
+
+```bash
+mlxsmith data presets                                     # list built-in datasets
+mlxsmith data pull alpaca                                 # pull a preset
+mlxsmith data import raw.json --out data/sft/train.jsonl  # import ShareGPT → JSONL
+mlxsmith data split data/sft/train.jsonl --fractions 0.9 0.05 0.05
+mlxsmith data stats data/sft/train.jsonl                  # token counts, field analysis
+mlxsmith data validate data/sft/train.jsonl               # structure check
+```
+
+Built-in presets: `alpaca`, `hh-rlhf`, `ultrachat-200k`, `ultrafeedback-binarized-prefs`, `ultrafeedback-binarized-sft`.
+
+## Model management
+
+```bash
+# Pull + convert HF model to MLX
+mlxsmith pull Qwen/Qwen3-4B-Instruct-2507
+
+# With quantization
+mlxsmith pull Qwen/Qwen3-4B-Instruct-2507 --quantize --q-bits 4
+
+# Merge adapters
+mlxsmith adapters merge runs/sft_0001/adapter runs/pref_0001/adapter --weights 0.7 0.3
+```
+
+## HF auth
+
+```bash
+mlxsmith auth login --token "$HF_TOKEN"
+mlxsmith auth status
+mlxsmith auth logout
+```
+
+## Eval and bench
+
+```bash
+# Evaluation suite (pass@k with verifier checks)
+mlxsmith eval --suite eval/suites/coding.yaml
+
+# Benchmark inference or training throughput
+mlxsmith bench --mode inference
+mlxsmith bench --mode trainer
+mlxsmith bench --mode end_to_end
+```
+
+## Verifiers
+
+Built-in verifiers for eval, RFT, and preference tuning:
+
+- **regex** — pattern matching on completions
+- **jsonschema** — JSON structure validation
+- **pytest** — sandboxed test execution
+- **docker** — containerized verification
+- **compose** — multi-verifier composition (AND/OR/weighted)
+
+See `docs/VERIFIERS.md` for the verifier API.
+
+## Environment plugin system
+
+```bash
+mlxsmith env list                  # list available environments
+mlxsmith env info envs/coding.yaml # show manifest (tasks, verifier, version)
+mlxsmith env init my_env           # scaffold a new environment
+mlxsmith env install ./my_env      # install from directory
+mlxsmith env package ./my_env      # create distributable tarball
+mlxsmith env run envs/coding.yaml  # execute RFT with this environment
+```
+
+Environments define tasks, verifiers, and reward functions for RFT training. See `docs/ENVIRONMENTS.md`.
+
+## Config system
+
+```bash
+mlxsmith config show              # display merged config (YAML/JSON/TOML)
+mlxsmith config show --sources    # show where each value comes from
+mlxsmith config init              # create default mlxsmith.yaml
+mlxsmith config validate          # check config structure
+mlxsmith config env               # show environment variable mapping
+```
+
+Config sources (in priority order): CLI flags > environment variables (`MLXSMITH__SECTION__KEY`) > config file > defaults.
+
+## SDK (programmatic API)
+
+For building custom training loops:
+
+```python
+from mlxsmith.sdk import load_model, SamplingClient, TrainingClient, TrainingBatch
+
+loaded = load_model("path/to/model", config)
+
+# Sampling with logprobs
+sampler = SamplingClient(loaded.backend)
+result = sampler.sample("prompt", logprobs_k=5)
+
+# Training operations
+trainer = TrainingClient(loaded.backend)
+trainer.create_optimizer(lr=1e-4, weight_decay=0.01)
+fb = trainer.forward_backward(batch)
+trainer.optim_step(fb.result().grads)
+```
+
+Loss functions: DPO, ORPO, GRPO, CISPO, DRO, PPO, importance sampling, cross-entropy.
+
+## Research
+
+### RLM self-play loop
+
+RLM (Recursive Language Model) is a research feature — the infrastructure runs but has not produced measured gains yet.
+
+```bash
+mlxsmith rlm                       # single-process RLM
+mlxsmith pipeline --orchestrated   # multi-process orchestrated RLM
+mlxsmith rlm status                # check iteration state
+mlxsmith rlm history               # view history
+```
+
+Includes task generation, mutation for data diversity, corpus management, EMA-based gating, and weight pointer IPC for multi-process coordination. See `docs/orchestrator.md`.
+
+## Docs
+
+- `docs/PROJECT_FORMAT.md` — project layout and artifacts
+- `docs/VERIFIERS.md` — verifier API and sandbox behavior
+- `docs/COMPATIBILITY.md` — tested versions and model families
+- `docs/ENVIRONMENTS.md` — environment plugin system
+- `docs/orchestrator.md` — multi-process RLM orchestrator
+- `docs/rlm-ctl.md` — RLM training guide
+- `docs/ROADMAP.md` — product direction and milestones
+- `docs/README.md` — full docs index
+
+## License
+
+MIT

{mlxsmith-0.1.0.dist-info → mlxsmith-0.1.2.dist-info}/RECORD

@@ -2,55 +2,54 @@ mlxsmith/__init__.py,sha256=CJZKl9Hp16DYlQR7yqstir-cL4n7GCw90d-meXliCHk,48
 mlxsmith/adapters.py,sha256=wkQ2q_ugaxCviNARSmxehwBcc2_NKVJ7mOofT-y30TY,1318
 mlxsmith/auth.py,sha256=_j_gx5ccZfpHs0_Xmpcgh_ELhX3ZBJLg2YYpjA-aPRI,2195
 mlxsmith/bench.py,sha256=VBgY9uOGm3xhmL3UrNNOnUoa3P0yaVQQ7wxykIEmDEw,3778
-mlxsmith/cli.py,sha256=YHjrPti1OCe0m2gpEwygiJ6_x-xeycKHRDYCksdcOuM,35750
-mlxsmith/config.py,sha256=iFDhO7KcAtq1KySGD7TfNMYqbQcec0Tc0-VJdr5Gwo4,16385
-mlxsmith/config_models.py,sha256=pMsLGyC9J9T9jqs5ipYFUuEkfir7iiDA4EZdcb65k5g,8407
+mlxsmith/cli.py,sha256=MAS21bSz4DTkdMAao99KdtI0FtZFUEVQatSsS7VHPQ0,35713
+mlxsmith/config.py,sha256=K1VbN-3WoWf4vzZ6BNeTgEz2DFH8s9YminqwyNBVLj0,16389
+mlxsmith/config_models.py,sha256=x827pTs-SfPyqXI3mPR6h-cz7KQ69zQICPAaeGlW-A4,8399
 mlxsmith/data.py,sha256=3ZlNS8bnD7LlWACEmULbf8RGQzCuf0QulFpI1PWvNuI,16160
 mlxsmith/eval.py,sha256=nSARSEKKwZM8Ot5rUDDpGikaClGNxvg0ifgGkTA6mM0,3792
 mlxsmith/infer.py,sha256=ekpHhTird0dnTJzFOc-O98rjwkEKgAr9AFicKlaB3MA,4610
 mlxsmith/models.py,sha256=BRaPTxzqy-5KEKdccveMgjpbRP0ZmbRnA_su8rz2P4k,8033
 mlxsmith/runs.py,sha256=2voYBryGGMlAKskHJ7TDiIPQL2_fFxSQ8RgtfGZ7ccg,1409
-mlxsmith/server.py,sha256=CBxmV9WwpyivsHytVALgT4DOGh1vqjwm4W_iHBH79rg,10821
-mlxsmith/util.py,sha256=8fagFtMP2YS1qlWkCt8bPWAz2jAgfvAf1ipNGJZAgIk,4544
-mlxsmith/accel/__init__.py,sha256=Mv2mj-2bLqUILhMUCjMLu7JORcu0-cGBOri32j7O5Xo,291
+mlxsmith/server.py,sha256=Fk-i9xK_Teq0Z0m-W1GRJVtcG0nYvd4bQ85lnAUuT1w,10690
+mlxsmith/util.py,sha256=Cke2FxIVNmvpW9ElPxL5bc0JO_YFVHWtBwpR3QRtfWQ,4410
+mlxsmith/accel/__init__.py,sha256=fBY65q66jql1q5YMT9aIo96dgvVLHqS_LoJbVNA2xHY,201
 mlxsmith/accel/base.py,sha256=o0kmxV68KbxOmucC3eDjKbFA8dfTT2ETqN0XD_l3mM0,435
 mlxsmith/accel/none.py,sha256=WhxECIBv-pE63Vh1Iv86ObgT9JHOi4hA4BUyJc__sKU,362
-mlxsmith/accel/zmlx_backend.py,sha256=JfzQ44v9hrCJgcqU018ZD7qLNlubIe09CwYRpKyfwR8,1529
 mlxsmith/api/__init__.py,sha256=IrpIXDkUJm4BZqahYOK_0BkxvomlscEvCoLCm4GDxo8,998
-mlxsmith/api/handlers.py,sha256=5YKQfHLG2mWjCugOkG1L3veiPMz_Qg_s4h12xx6eibs,47068
+mlxsmith/api/handlers.py,sha256=94Spq4glFp7mRwmKrFqt7erd1Af_PxVP_vpxCo2UFdQ,46896
 mlxsmith/api/schemas.py,sha256=Q18kF9FKtvT1vdnXy6feSNTtCV2FiRWDzfdsPzc0te8,19316
 mlxsmith/envs/__init__.py,sha256=t7QiEHtfyP1dUCj-4TJUN0hd9lRqBKYd5Ek7dgEwus4,671
 mlxsmith/envs/system.py,sha256=2bChkOxm2S7d0WCrweHGhoI6-xOYDxlC0YbHH6Ibjq4,12782
 mlxsmith/envs/token_env.py,sha256=rhv2o3eI1GyTtfAXG72z-31amNGaLv0KW56mEsWkXlY,6709
 mlxsmith/llm/__init__.py,sha256=jWEkXGdvwZ8tUYHVqWW3SYHXG-LSWaGbdwOR0mF_4Zw,396
 mlxsmith/llm/backend.py,sha256=_xGfdJ30_6Nq8zIqMjTemCZWV-vUi9qV0djfwxny5SM,4335
-mlxsmith/llm/interface.py,sha256=zmqJuyiEQqjEpDjZwooS1hp1GfTPCTiWYJ5nPw8xSqg,6633
+mlxsmith/llm/interface.py,sha256=udQl_R7ecmM4Nh20P50Nmnv2h853ByrgevjQIRDxX4g,6601
 mlxsmith/llm/mlx_lm_backend.py,sha256=OitqY_6LqnfqbN7GZz39Ma3cItjjNqHXF3SV3AZsHxk,18284
-mlxsmith/llm/mock_backend.py,sha256=DXXnMlXZGCtf0datgjFc5z_X8JMEKaYDRPXKU1erGdQ,7449
+mlxsmith/llm/mock_backend.py,sha256=ZTkPRnRmCXSkhOJ1TLbSgTnAcLTKa9sq6-zzbTEX6Qc,7439
 mlxsmith/llm/registry.py,sha256=ZmYE-WclyMo6z0HwUufqt3tKT4E84xZ6I-PFu1Z5nws,309
 mlxsmith/orchestrator/__init__.py,sha256=oc4qIkISZMGMvYeOqU8lDDmFL3uxDYJHsv_rra9DH-E,792
-mlxsmith/orchestrator/daemon.py,sha256=MRQf84sCjeShBFcOcF7yfP6HhHl7IHrnmw0nV50mJrI,16360
-mlxsmith/orchestrator/inference_worker.py,sha256=4AI_j7qnrnwXObBxSa7YHlZHnGfqou-W1fpqHYt8jpQ,17981
-mlxsmith/orchestrator/queue.py,sha256=8c1n-fiW3ITcLbwOeFlH_AEZdJYvZumy8EczJ0lh4NA,11301
-mlxsmith/orchestrator/trainer_worker.py,sha256=uvJQesXjfdsiNRsO2FVHhUk0WmMB_PQNSPff5U9Swp4,16061
+mlxsmith/orchestrator/daemon.py,sha256=VJFF8s_85h4C-YM14wRUlzDHrhhVPfQOztmNLKyRk30,16107
+mlxsmith/orchestrator/inference_worker.py,sha256=PfmsanrBnx9HZNqG00jTQQTKqDa2bl-wUtYAWtxfzvs,17963
+mlxsmith/orchestrator/queue.py,sha256=E8VymvJi2zEpuTwsG7JB-vROJGGS5evPPhIpkmdwtq4,11286
+mlxsmith/orchestrator/trainer_worker.py,sha256=IM7vOhkYlCKoQSC1tNobkeIVnkmsz4DtAWfu24bjJXY,16015
 mlxsmith/rlm/__init__.py,sha256=Q09oRONXWTFXuWwMJOpGWg0I-UDkuib0OA1O_cNFp2U,236
 mlxsmith/rlm/corpus.py,sha256=-p12H650_ybe2kXC219M4wXYpD08QHUpY92ErVjSfX8,2112
 mlxsmith/rlm/gating.py,sha256=L18niYKEezphASdsgzW6pz3PN7ylA-L5Wu4_GLLVfHw,2455
 mlxsmith/rlm/generate.py,sha256=q1v_TP8sqVj05omhoF60Ns1iX6yClgc7lP6njz4lK18,7601
 mlxsmith/rlm/history.py,sha256=Vm4JtWqsZnqB-fuo3zWfweeogmmLTL3VHaYZ45vrkz8,299
 mlxsmith/rlm/inference.py,sha256=ntCEKxD1KrkIXgZNQbD4jhS5rJPtwcVYc8qLc5E5cnc,5297
-mlxsmith/rlm/loop.py,sha256=WLRP1PI5PX7OjSEZsurrCwSTVVPNVIx4-7td1ihSMY0,49955
+mlxsmith/rlm/loop.py,sha256=nkAR5KYErcq893kQYigFkl3NG469ZUu1CkN_sS1ObAM,49900
 mlxsmith/rlm/mutate.py,sha256=_NUNMpVCRaEDgtzI8J2NOTcj5NnycZnP_UoUpFacjTs,2553
 mlxsmith/rlm/trainer.py,sha256=RRXPlJy4SySpLZGge0ORMYs7HeiWgfGQNeMBOBfG4Ys,3014
-mlxsmith/rlm/weights.py,sha256=NO7wjl2T0eXTVFoYrzPT_IUmaLvD2z-zSSyKpcX93kY,8463
+mlxsmith/rlm/weights.py,sha256=tgl4Uc80QF9YpCCr3ewBmL7uru9As2fDA1Z1SgZn-e4,8455
 mlxsmith/sdk/__init__.py,sha256=42WpTgC309sYKp8SArULBWz0trVN51THcjvPdVh-thc,10777
-mlxsmith/sdk/future.py,sha256=mleqPgJ997hSuZuQegvS2GoOxqo_gd4pfh37gv70APc,16873
+mlxsmith/sdk/future.py,sha256=WmYB-fDstaEuv-FUNX_S7IJSENbVEsYYEEVzH02ImLk,16868
 mlxsmith/sdk/losses.py,sha256=lJi3R7Red_QO3IatbhKi_GBI7kM0yu-kS14xN2kX_04,7532
-mlxsmith/sdk/sampling_client.py,sha256=AkjJHEZ8OLU8SZoALk7ds3NHMquiqk9GGZHqni5vB2g,24942
-mlxsmith/sdk/training_client.py,sha256=4GXcswTJypkymPO3E5DREUesGzlumbmUrTnt0T80wEI,24380
+mlxsmith/sdk/sampling_client.py,sha256=o7jfgYpVWXrrIHo4-SrGAJx4FAlYdo198da27Jp0Yj4,24899
+mlxsmith/sdk/training_client.py,sha256=71bSgS65ofXL9X4qGSyDhVL5asDxENAf95LV66pwC3g,24341
 mlxsmith/train/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 mlxsmith/train/distill.py,sha256=9Xbzn6zt8uqYOwg_pJB2rZJj8y-nESP1OB8DLxnJ0jM,10378
-mlxsmith/train/lora.py,sha256=E8ymi1wUVsGp5-0DldvkSWDHNh_5Fhr6AelJOR5YoPw,8859
+mlxsmith/train/lora.py,sha256=k3aNqBjMyE6rPGS2CAJRSDsTJiUa1ztjrA3k9N87IjY,9046
 mlxsmith/train/pref.py,sha256=-z7mj-MQu6dPb8y0U6sRFbqKU0nNQ3YRpC3CcmS3l5k,6987
 mlxsmith/train/rft.py,sha256=bf6z-h6VQKvMDZ0XN_ayZV44YsIvuwSzP1YRYrwSQ2M,18956
 mlxsmith/train/sft.py,sha256=w3QmLLoscNQzz-xbtmrCw46PbYIApvgeQi0XjxCop90,5590
@@ -61,9 +60,9 @@ mlxsmith/verifiers/jsonschema.py,sha256=hG_8c07Hwv-tpN2g0oxELwmLRxS8QGzRFwabmo4y
 mlxsmith/verifiers/pytest_verifier.py,sha256=ARNajzxUPNwtzSow6I2d0mLopZyvY29_d3F1sYVwEUY,2514
 mlxsmith/verifiers/regex.py,sha256=N7z3koE8Iy-a4DBs4404iQCNX2WGxequm5g4umric2Y,524
 mlxsmith/verifiers/types.py,sha256=FytBxB1OnNX1EcqZXSSs3WvL0GRv7byW4mfBMf6xP68,240
-mlxsmith-0.1.0.dist-info/licenses/LICENSE,sha256=ESYyLizI0WWtxMeS7rGVcX3ivMezm-HOd5WdeOh-9oU,1056
-mlxsmith-0.1.0.dist-info/METADATA,sha256=yRnNRKdCFPZQRgAN7qtxCDBrUJxplOqK3mTV_Uoh2jA,4812
-mlxsmith-0.1.0.dist-info/WHEEL,sha256=wUyA8OaulRlbfwMtmQsvNngGrxQHAvkKcvRmdizlJi0,92
-mlxsmith-0.1.0.dist-info/entry_points.txt,sha256=ys7GcKEjhzhkkTMBmmaNavTgsjqOuFnCKIG2w8Wcn6w,46
-mlxsmith-0.1.0.dist-info/top_level.txt,sha256=hKBwc8bn7uoI-_5Yhcq1T3IuChFhUFdzItIkZK1up6A,9
-mlxsmith-0.1.0.dist-info/RECORD,,
+mlxsmith-0.1.2.dist-info/licenses/LICENSE,sha256=ESYyLizI0WWtxMeS7rGVcX3ivMezm-HOd5WdeOh-9oU,1056
+mlxsmith-0.1.2.dist-info/METADATA,sha256=mtwo3IlnvH9ffCKWSHbxRMTLQFyzrvDi29BCyzbteqk,8963
+mlxsmith-0.1.2.dist-info/WHEEL,sha256=wUyA8OaulRlbfwMtmQsvNngGrxQHAvkKcvRmdizlJi0,92
+mlxsmith-0.1.2.dist-info/entry_points.txt,sha256=ys7GcKEjhzhkkTMBmmaNavTgsjqOuFnCKIG2w8Wcn6w,46
+mlxsmith-0.1.2.dist-info/top_level.txt,sha256=hKBwc8bn7uoI-_5Yhcq1T3IuChFhUFdzItIkZK1up6A,9
+mlxsmith-0.1.2.dist-info/RECORD,,

mlxsmith/accel/zmlx_backend.py

@@ -1,42 +0,0 @@
-from __future__ import annotations
-
-from typing import Any, Dict
-from .base import AccelStats
-
-class ZMLXBackend:
-    name = "zmlx"
-
-    def __init__(self):
-        self._available = False
-        self._notes = {}
-        try:
-            import zmlx  # type: ignore
-            self._available = True
-            self._notes["zmlx_version"] = getattr(zmlx, "__version__", None)
-        except Exception as e:
-            self._available = False
-            self._notes["error"] = f"{type(e).__name__}: {e}"
-
-    def patch(self) -> None:
-        if not self._available:
-            # soft fail; caller should report status
-            return
-        # ZMLX can patch ops/modules. We keep this intentionally minimal and safe.
-        try:
-            import zmlx  # type: ignore
-            # If ZMLX provides a global patch hook, call it; otherwise, no-op.
-            patch_fn = getattr(zmlx, "patch", None)
-            if callable(patch_fn):
-                patch_fn()
-                self._notes["patched"] = True
-            else:
-                self._notes["patched"] = False
-                self._notes["hint"] = "No zmlx.patch() found; implement patch hook or integrate per-module."
-        except Exception as e:
-            self._notes["patched_error"] = f"{type(e).__name__}: {e}"
-
-    def warmup(self, model: Any, example_batch: Any) -> Dict[str, Any]:
-        return {"warmup": "not_implemented", "notes": self._notes}
-
-    def stats(self) -> AccelStats:
-        return AccelStats(backend="zmlx", notes=self._notes)

mlxsmith-0.1.0.dist-info/METADATA

@@ -1,163 +0,0 @@
-Metadata-Version: 2.4
-Name: mlxsmith
-Version: 0.1.0
-Summary: Apple Silicon MLX fine-tuning and OpenAI-compatible serving (SFT stable; preference/RL experimental).
-Author-email: Shannon Labs <hmbown@gmail.com>
-License: MIT
-Project-URL: Homepage, https://github.com/Hmbown/MLXSmith
-Project-URL: Repository, https://github.com/Hmbown/MLXSmith
-Project-URL: Issues, https://github.com/Hmbown/MLXSmith/issues
-Keywords: mlx,apple-silicon,llm,fine-tuning,lora,openai-compatible
-Classifier: Development Status :: 3 - Alpha
-Classifier: Intended Audience :: Developers
-Classifier: Intended Audience :: Science/Research
-Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
-Classifier: License :: OSI Approved :: MIT License
-Classifier: Programming Language :: Python :: 3
-Classifier: Programming Language :: Python :: 3 :: Only
-Classifier: Programming Language :: Python :: 3.10
-Classifier: Programming Language :: Python :: 3.11
-Classifier: Programming Language :: Python :: 3.12
-Classifier: Operating System :: MacOS :: MacOS X
-Requires-Python: >=3.10
-Description-Content-Type: text/markdown
-License-File: LICENSE
-Requires-Dist: typer>=0.9.0
-Requires-Dist: rich>=13.7.0
-Requires-Dist: pyyaml>=6.0.1
-Requires-Dist: pydantic>=2.5.0
-Requires-Dist: pydantic-settings>=2.2.1
-Requires-Dist: tomli>=2.0.1; python_version < "3.11"
-Requires-Dist: huggingface_hub>=1.3.4
-Requires-Dist: jsonschema>=4.21.0
-Provides-Extra: mlx
-Requires-Dist: mlx>=0.30.4; extra == "mlx"
-Provides-Extra: llm
-Requires-Dist: mlx-lm>=0.30.5; extra == "llm"
-Requires-Dist: transformers>=5.0.0; extra == "llm"
-Requires-Dist: datasets>=3.0.0; extra == "llm"
-Provides-Extra: serve
-Requires-Dist: fastapi>=0.128.0; extra == "serve"
-Requires-Dist: uvicorn>=0.40.0; extra == "serve"
-Requires-Dist: httpx>=0.28.0; extra == "serve"
-Provides-Extra: zmlx
-Requires-Dist: zmlx; extra == "zmlx"
-Provides-Extra: dev
-Requires-Dist: pytest>=9.0.0; extra == "dev"
-Requires-Dist: ruff>=0.14.0; extra == "dev"
-Provides-Extra: all
-Requires-Dist: mlx>=0.30.4; extra == "all"
-Requires-Dist: mlx-lm>=0.30.5; extra == "all"
-Requires-Dist: transformers>=5.0.0; extra == "all"
-Requires-Dist: datasets>=3.0.0; extra == "all"
-Requires-Dist: fastapi>=0.128.0; extra == "all"
-Requires-Dist: uvicorn>=0.40.0; extra == "all"
-Requires-Dist: httpx>=0.28.0; extra == "all"
-Dynamic: license-file
-
-# mlxsmith
-
-Apple Silicon MLX fine-tuning and OpenAI-compatible serving.
-SFT + serving are stable. Preference/RL/RLM features are experimental.
-
-Status: alpha (2026-02-02).
-
-## Stable features
-- Project init, config, data tools, HF auth, model pull/convert.
-- SFT (LoRA/QLoRA) training with run tracking and adapters.
-- Inference and OpenAI-compatible /v1/chat/completions serving.
-- Basic eval/bench and verifier plumbing (regex/jsonschema/pytest).
-
-## Experimental features
-- Preference tuning (DPO/ORPO).
-- GRPO-style RFT.
-- RLM self-play loop (research).
-- Distill/OPD and orchestrated RLM.
-
-## Install
-
-MLX is only available on Apple Silicon. Other platforms can still use data tools
-and mock backends, but MLX training and serving require macOS on Apple Silicon.
-
-```bash
-python -m venv .venv && source .venv/bin/activate
-pip install -U pip
-
-# Core CLI
-pip install mlxsmith
-
-# Apple Silicon training + serving
-pip install "mlxsmith[mlx,llm,serve]"
-```
-
-## Quickstart
-
-```bash
-mlxsmith init myproj
-cd myproj
-mlxsmith doctor
-```
-
-## HF auth (optional)
-
-```bash
-mlxsmith auth login --token "$HF_TOKEN"
-mlxsmith auth status
-mlxsmith auth logout
-```
-
-## Pull + convert a model (HF -> MLX)
-
-```bash
-mlxsmith pull Qwen/Qwen3-4B-Instruct-2507
-# outputs to cache/mlx/Qwen__Qwen3-4B-Instruct-2507
-```
-
-Optional quantization:
-
-```bash
-mlxsmith pull Qwen/Qwen3-4B-Instruct-2507 --quantize --q-bits 4
-```
-
-## SFT (LoRA/QLoRA)
-
-```bash
-mlxsmith sft --model cache/mlx/Qwen__Qwen3-4B-Instruct-2507 --data data/sft
-```
-
-## Serve (OpenAI-compatible)
-
-```bash
-mlxsmith serve --model runs/sft_0001/adapter --port 8080
-```
-
-Sample request:
-
-```bash
-curl http://localhost:8080/v1/chat/completions \
-  -H 'Content-Type: application/json' \
-  -d '{"messages":[{"role":"user","content":"Hello"}],"max_tokens":64}'
-```
-
-To enable the optional UI/monitor dashboard, set `serve.ui: true` in `mlxsmith.yaml`.
-
-## Experimental commands
-
-- `mlxsmith pref` (DPO/ORPO)
-- `mlxsmith rft` (GRPO-style)
-- `mlxsmith rlm` / `mlxsmith pipeline` (self-play loop)
-- `mlxsmith distill` (offline/OPD)
-- `mlxsmith eval` / `mlxsmith bench`
-
-## Docs
-
-- `docs/PROJECT_FORMAT.md` for project layout and artifacts.
-- `docs/VERIFIERS.md` for verifier API and sandbox behavior.
-- `docs/COMPATIBILITY.md` for tested versions and model families.
-- `docs/ENVIRONMENTS.md` for the environment plugin system.
-- `docs/ROADMAP.md` for product direction and milestones.
-- `docs/README.md` for the full docs index.
-
-## License
-
-MIT