mlxsmith 0.1.1__tar.gz → 0.1.3__tar.gz

{mlxsmith-0.1.1/src/mlxsmith.egg-info → mlxsmith-0.1.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mlxsmith
-Version: 0.1.1
+Version: 0.1.3
 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
@@ -36,18 +36,19 @@ 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: lora
+Requires-Dist: mlx-lm-lora>=1.0.0; extra == "lora"
 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: mlx-lm-lora>=1.0.0; 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"
@@ -59,7 +60,7 @@ Dynamic: license-file
 
 Apple Silicon MLX fine-tuning toolkit — SFT, DPO/ORPO, GRPO, distillation, and OpenAI-compatible serving.
 
-**Status:** alpha (v0.1.0). Full training pipeline validated on Qwen3-4B.
+**Status:** alpha (v0.1.2). Full training pipeline validated on Qwen3-4B.
 
 ## Install
 
@@ -76,6 +77,9 @@ pip install mlxsmith
 # Apple Silicon training + serving
 pip install "mlxsmith[mlx,llm,serve]"
 
+# mlx-lm-lora passthrough (advanced training methods)
+pip install "mlxsmith[lora]"
+
 # Everything
 pip install "mlxsmith[all]"
 ```
@@ -85,7 +89,7 @@ pip install "mlxsmith[all]"
 ```bash
 mlxsmith init myproj
 cd myproj
-mlxsmith doctor        # check Python, MLX, Metal, ZMLX
+mlxsmith doctor        # check Python, MLX, Metal
 ```
 
 ## Training
@@ -133,6 +137,22 @@ mlxsmith distill --teacher large-model --student small-model --mode opd
 mlxsmith pipeline
 ```
 
+### mlx-lm-lora parity (all methods)
+
+Use the passthrough to access mlx-lm-lora features (DPO variants, GRPO variants,
+PPO, synthetic datasets, judge training, etc.):
+
+```bash
+# Train with mlx-lm-lora directly
+mlxsmith lora train --model Qwen/Qwen3-4B-Instruct-2507 --data data/prefs --train-mode dpo -- --beta 0.1
+
+# Generate synthetic datasets
+mlxsmith lora synthetic prompts -- --model mlx-community/Qwen3-4B-Instruct-2507-4bit --num-samples 1000
+
+# Train judge model
+mlxsmith lora judge -- --model mlx-community/Qwen3-4B-Instruct-2507-4bit --data data/prefs
+```
+
 ## Serving
 
 OpenAI-compatible `/v1/chat/completions` endpoint.
@@ -204,6 +224,7 @@ Built-in verifiers for eval, RFT, and preference tuning:
 - **pytest** — sandboxed test execution
 - **docker** — containerized verification
 - **compose** — multi-verifier composition (AND/OR/weighted)
+- **llm_judge** — LLM-based self-verification / ThinkPRM-style verifier
 
 See `docs/VERIFIERS.md` for the verifier API.
 
@@ -232,6 +253,9 @@ mlxsmith config env               # show environment variable mapping
 
 Config sources (in priority order): CLI flags > environment variables (`MLXSMITH__SECTION__KEY`) > config file > defaults.
 
+Training optimizers are configurable via `train.optimizer` and `train.optimizer_kwargs`
+(for example `adamw`, `adam`, `qhadam`, `muon` when available in MLX).
+
 ## SDK (programmatic API)
 
 For building custom training loops:
@@ -269,14 +293,6 @@ 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`.
 
-### ZMLX acceleration
-
-Optional zero-copy MLX acceleration backend.
-
-```bash
-mlxsmith accel status
-```
-
 ## Docs
 
 - `docs/PROJECT_FORMAT.md` — project layout and artifacts

{mlxsmith-0.1.1 → mlxsmith-0.1.3}/README.md

@@ -2,7 +2,7 @@
 
 Apple Silicon MLX fine-tuning toolkit — SFT, DPO/ORPO, GRPO, distillation, and OpenAI-compatible serving.
 
-**Status:** alpha (v0.1.0). Full training pipeline validated on Qwen3-4B.
+**Status:** alpha (v0.1.2). Full training pipeline validated on Qwen3-4B.
 
 ## Install
 
@@ -19,6 +19,9 @@ pip install mlxsmith
 # Apple Silicon training + serving
 pip install "mlxsmith[mlx,llm,serve]"
 
+# mlx-lm-lora passthrough (advanced training methods)
+pip install "mlxsmith[lora]"
+
 # Everything
 pip install "mlxsmith[all]"
 ```
@@ -28,7 +31,7 @@ pip install "mlxsmith[all]"
 ```bash
 mlxsmith init myproj
 cd myproj
-mlxsmith doctor        # check Python, MLX, Metal, ZMLX
+mlxsmith doctor        # check Python, MLX, Metal
 ```
 
 ## Training
@@ -76,6 +79,22 @@ mlxsmith distill --teacher large-model --student small-model --mode opd
 mlxsmith pipeline
 ```
 
+### mlx-lm-lora parity (all methods)
+
+Use the passthrough to access mlx-lm-lora features (DPO variants, GRPO variants,
+PPO, synthetic datasets, judge training, etc.):
+
+```bash
+# Train with mlx-lm-lora directly
+mlxsmith lora train --model Qwen/Qwen3-4B-Instruct-2507 --data data/prefs --train-mode dpo -- --beta 0.1
+
+# Generate synthetic datasets
+mlxsmith lora synthetic prompts -- --model mlx-community/Qwen3-4B-Instruct-2507-4bit --num-samples 1000
+
+# Train judge model
+mlxsmith lora judge -- --model mlx-community/Qwen3-4B-Instruct-2507-4bit --data data/prefs
+```
+
 ## Serving
 
 OpenAI-compatible `/v1/chat/completions` endpoint.
@@ -147,6 +166,7 @@ Built-in verifiers for eval, RFT, and preference tuning:
 - **pytest** — sandboxed test execution
 - **docker** — containerized verification
 - **compose** — multi-verifier composition (AND/OR/weighted)
+- **llm_judge** — LLM-based self-verification / ThinkPRM-style verifier
 
 See `docs/VERIFIERS.md` for the verifier API.
 
@@ -175,6 +195,9 @@ mlxsmith config env               # show environment variable mapping
 
 Config sources (in priority order): CLI flags > environment variables (`MLXSMITH__SECTION__KEY`) > config file > defaults.
 
+Training optimizers are configurable via `train.optimizer` and `train.optimizer_kwargs`
+(for example `adamw`, `adam`, `qhadam`, `muon` when available in MLX).
+
 ## SDK (programmatic API)
 
 For building custom training loops:
@@ -212,14 +235,6 @@ 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`.
 
-### ZMLX acceleration
-
-Optional zero-copy MLX acceleration backend.
-
-```bash
-mlxsmith accel status
-```
-
 ## Docs
 
 - `docs/PROJECT_FORMAT.md` — project layout and artifacts

{mlxsmith-0.1.1 → mlxsmith-0.1.3}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "mlxsmith"
-version = "0.1.1"
+version = "0.1.3"
 description = "Apple Silicon MLX fine-tuning toolkit — SFT, DPO/ORPO, GRPO, distillation, and OpenAI-compatible serving."
 readme = {file = "README.md", content-type = "text/markdown"}
 requires-python = ">=3.10"
@@ -47,16 +47,17 @@ llm = [
   "transformers>=5.0.0",
   "datasets>=3.0.0",
 ]
+lora = ["mlx-lm-lora>=1.0.0"]
 serve = [
   "fastapi>=0.128.0",
   "uvicorn>=0.40.0",
   "httpx>=0.28.0",
 ]
-zmlx = ["zmlx"]
 dev = ["pytest>=9.0.0", "ruff>=0.14.0"]
 all = [
   "mlx>=0.30.4",
   "mlx-lm>=0.30.5",
+  "mlx-lm-lora>=1.0.0",
   "transformers>=5.0.0",
   "datasets>=3.0.0",
   "fastapi>=0.128.0",

{mlxsmith-0.1.1 → mlxsmith-0.1.3}/src/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-0.1.1 → mlxsmith-0.1.3}/src/mlxsmith/bench.py

@@ -44,7 +44,12 @@ def run_bench(
     mode = (mode or "inference").lower()
 
     if mode == "trainer":
-        opt, _params = llm.optimizer_and_params(lr=cfg.train.lr, weight_decay=cfg.train.weight_decay)
+        opt, _params = llm.optimizer_and_params(
+            lr=cfg.train.lr,
+            weight_decay=cfg.train.weight_decay,
+            optimizer=cfg.train.optimizer,
+            optimizer_kwargs=cfg.train.optimizer_kwargs,
+        )
         prompt_ids = llm.encode(prompt)
         ids = llm.encode(prompt + " " + "x" * max_tokens)
         for i in range(max(1, reps)):
@@ -59,7 +64,12 @@ def run_bench(
             elapsed = max(time.time() - t0, 1e-6)
             results.append({"rep": i, "steps": steps, "time_s": elapsed, "steps_per_s": steps / elapsed})
     elif mode == "end_to_end":
-        opt, _params = llm.optimizer_and_params(lr=cfg.train.lr, weight_decay=cfg.train.weight_decay)
+        opt, _params = llm.optimizer_and_params(
+            lr=cfg.train.lr,
+            weight_decay=cfg.train.weight_decay,
+            optimizer=cfg.train.optimizer,
+            optimizer_kwargs=cfg.train.optimizer_kwargs,
+        )
         for i in range(max(1, reps)):
             t0 = time.time()
             gen = llm.generate(prompt, max_new_tokens=max_tokens, temperature=0.0)

{mlxsmith-0.1.1 → mlxsmith-0.1.3}/src/mlxsmith/cli.py

@@ -24,6 +24,8 @@ from .train.sft import run_sft
 from .train.pref import run_pref
 from .train.rft import run_rft
 from .train.distill import run_distill
+from .train.online_dpo import run_online_dpo
+from .train.self_verify import run_self_verify
 from .eval import run_eval
 from .bench import run_bench
 from .rlm import run_rlm, run_rlm_orchestrated
@@ -40,6 +42,13 @@ from .envs import (
     resolve_env_path as resolve_env_path_plugin,
     load_manifest as load_env_manifest,
 )
+from .integrations.mlx_lm_lora import (
+    build_train_command as build_mlx_lm_lora_train_command,
+    build_synthetic_command as build_mlx_lm_lora_synth_command,
+    build_judge_command as build_mlx_lm_lora_judge_command,
+    build_reward_functions_command as build_mlx_lm_lora_reward_functions_command,
+    run_command as run_mlx_lm_lora_command,
+)
 
 app = typer.Typer(
     add_completion=False,
@@ -65,6 +74,9 @@ def init(path: str = typer.Argument(..., help="Project directory to create")):
     (p / "verifiers" / "regex.py").write_text(_sample_verifier_regex(), encoding="utf-8")
     (p / "verifiers" / "pytest.py").write_text(_sample_verifier_pytest(), encoding="utf-8")
     (p / "verifiers" / "jsonschema.py").write_text(_sample_verifier_jsonschema(), encoding="utf-8")
+    (p / "verifiers" / "llm_judge.py").write_text(_sample_verifier_llm_judge(), encoding="utf-8")
+    (p / "verifiers" / "rubrics").mkdir(parents=True, exist_ok=True)
+    (p / "verifiers" / "rubrics" / "coding.txt").write_text(_sample_judge_rubric(), encoding="utf-8")
     (p / "eval" / "suites" / "coding.yaml").write_text(_sample_eval_suite(), encoding="utf-8")
     console.print(f"[green]Initialized[/green] {p.resolve()}")
 
@@ -83,7 +95,6 @@ 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)
 
 
@@ -342,14 +353,19 @@ def pref(
     data: str = typer.Option("data/prefs", "--data"),
     model: str = typer.Option(..., "--model", help="Base adapter or model path (e.g., runs/sft_0001/adapter)"),
     accel: Optional[str] = typer.Option(None, "--accel", help="Override accel.backend"),
-    algo: Optional[str] = typer.Option(None, "--algo", help="Override pref.algo (dpo|orpo|grpo)"),
+    algo: Optional[str] = typer.Option(None, "--algo", help="Override pref.algo (legacy)"),
+    loss_type: Optional[str] = typer.Option(None, "--loss-type", help="dpo|cpo|orpo|ipo|hinge"),
 ):
     root = project_root_from_cwd()
+    overrides = {}
+    if loss_type is not None:
+        overrides["pref.loss_type"] = loss_type
     cfg = get_config(
         config_path=config,
         root=root,
         accel_backend=accel,
         algo=algo,
+        **overrides,
     )
     data_dir = root / data
     run = run_pref(root, cfg, data_dir, Path(model), cfg.accel.backend)
@@ -364,13 +380,27 @@ def rft(
     model: str = typer.Option(..., "--model"),
     accel: Optional[str] = typer.Option(None, "--accel", help="Override accel.backend"),
     rollouts: Optional[int] = typer.Option(None, "--rollouts", help="Override rft.rollouts"),
+    loss_type: Optional[str] = typer.Option(None, "--loss-type", help="grpo|dr_grpo|dapo"),
+    epsilon_low: Optional[float] = typer.Option(None, "--epsilon-low"),
+    epsilon_high: Optional[float] = typer.Option(None, "--epsilon-high"),
+    token_level_loss: Optional[bool] = typer.Option(None, "--token-level-loss/--sequence-level-loss"),
 ):
     root = project_root_from_cwd()
+    overrides = {}
+    if loss_type is not None:
+        overrides["rft.loss_type"] = loss_type
+    if epsilon_low is not None:
+        overrides["rft.epsilon_low"] = epsilon_low
+    if epsilon_high is not None:
+        overrides["rft.epsilon_high"] = epsilon_high
+    if token_level_loss is not None:
+        overrides["rft.token_level_loss"] = token_level_loss
     cfg = get_config(
         config_path=config,
         root=root,
         accel_backend=accel,
         rollouts=rollouts,
+        **overrides,
     )
     run = run_rft(root, cfg, root / env, root / verifier, Path(model), cfg.accel.backend)
     console.print(f"[bold]Run:[/bold] {run.run_dir}")
@@ -438,6 +468,142 @@ def distill(
     console.print(f"[bold]Run:[/bold] {run.run_dir}")
 
 
+@app.command("online-dpo")
+def online_dpo(
+    data: str = typer.Option(..., "--data", help="JSONL with prompts"),
+    model: str = typer.Option(..., "--model"),
+    judge_model: Optional[str] = typer.Option(None, "--judge-model"),
+    judge_backend: str = typer.Option("mlx-lm", "--judge-backend"),
+    rubric: Optional[str] = typer.Option(None, "--rubric"),
+    group_size: Optional[int] = typer.Option(None, "--group-size"),
+    max_new_tokens: Optional[int] = typer.Option(None, "--max-new-tokens"),
+    temperature: Optional[float] = typer.Option(None, "--temperature"),
+    config: str = typer.Option("mlxsmith.yaml", "-c", "--config", help="Config file path"),
+    accel: Optional[str] = typer.Option(None, "--accel", help="Override accel.backend"),
+):
+    root = project_root_from_cwd()
+    cfg = get_config(config_path=config, root=root, accel_backend=accel)
+    run = run_online_dpo(
+        root,
+        cfg,
+        Path(data),
+        model,
+        cfg.accel.backend,
+        judge_model=judge_model,
+        judge_backend=judge_backend,
+        rubric=rubric,
+        group_size=group_size,
+        max_new_tokens=max_new_tokens,
+        temperature=temperature,
+    )
+    console.print(f"[bold]Run:[/bold] {run.run_dir}")
+
+
+@app.command("self-verify")
+def self_verify(
+    data: str = typer.Option(..., "--data", help="JSONL with prompts"),
+    model: str = typer.Option(..., "--model"),
+    verifier_model: Optional[str] = typer.Option(None, "--verifier-model"),
+    verifier_backend: str = typer.Option("mlx-lm", "--verifier-backend"),
+    rubric: Optional[str] = typer.Option(None, "--rubric"),
+    max_new_tokens: Optional[int] = typer.Option(None, "--max-new-tokens"),
+    temperature: Optional[float] = typer.Option(None, "--temperature"),
+    config: str = typer.Option("mlxsmith.yaml", "-c", "--config", help="Config file path"),
+    accel: Optional[str] = typer.Option(None, "--accel", help="Override accel.backend"),
+):
+    root = project_root_from_cwd()
+    cfg = get_config(config_path=config, root=root, accel_backend=accel)
+    run = run_self_verify(
+        root,
+        cfg,
+        Path(data),
+        model,
+        cfg.accel.backend,
+        verifier_model=verifier_model,
+        verifier_backend=verifier_backend,
+        rubric=rubric,
+        max_new_tokens=max_new_tokens,
+        temperature=temperature,
+    )
+    console.print(f"[bold]Run:[/bold] {run.run_dir}")
+
+
+lora_app = typer.Typer(help="mlx-lm-lora passthrough commands")
+app.add_typer(lora_app, name="lora")
+
+
+@lora_app.command(
+    "train",
+    context_settings={"allow_extra_args": True, "ignore_unknown_options": True},
+)
+def lora_train(
+    ctx: typer.Context,
+    config: Optional[str] = typer.Option(None, "--config", help="mlx-lm-lora config path"),
+    model: Optional[str] = typer.Option(None, "--model", help="Model id or path"),
+    data: Optional[str] = typer.Option(None, "--data", help="Dataset path or HF dataset"),
+    train_mode: Optional[str] = typer.Option(None, "--train-mode", help="sft|dpo|orpo|grpo|ppo|..."),
+    train_type: Optional[str] = typer.Option(None, "--train-type", help="lora|dora|full"),
+    dry_run: bool = typer.Option(False, "--dry-run"),
+):
+    """Run mlx-lm-lora training with passthrough args.
+
+    Use `--` to pass through any additional mlx-lm-lora flags.
+    """
+    root = project_root_from_cwd()
+    cmd = build_mlx_lm_lora_train_command(
+        config=config,
+        model=model,
+        data=data,
+        train_mode=train_mode,
+        train_type=train_type,
+        extra_args=list(ctx.args),
+    )
+    run_mlx_lm_lora_command(cmd, dry_run=dry_run, cwd=root)
+
+
+@lora_app.command(
+    "synthetic",
+    context_settings={"allow_extra_args": True, "ignore_unknown_options": True},
+)
+def lora_synthetic(
+    ctx: typer.Context,
+    kind: str = typer.Argument(..., help="prompts|sft|dpo"),
+    dry_run: bool = typer.Option(False, "--dry-run"),
+):
+    """Run mlx-lm-lora synthetic dataset generation."""
+    root = project_root_from_cwd()
+    cmd = build_mlx_lm_lora_synth_command(kind, extra_args=list(ctx.args))
+    run_mlx_lm_lora_command(cmd, dry_run=dry_run, cwd=root)
+
+
+@lora_app.command(
+    "judge",
+    context_settings={"allow_extra_args": True, "ignore_unknown_options": True},
+)
+def lora_judge(
+    ctx: typer.Context,
+    dry_run: bool = typer.Option(False, "--dry-run"),
+):
+    """Run mlx-lm-lora judge model training."""
+    root = project_root_from_cwd()
+    cmd = build_mlx_lm_lora_judge_command(extra_args=list(ctx.args))
+    run_mlx_lm_lora_command(cmd, dry_run=dry_run, cwd=root)
+
+
+@lora_app.command(
+    "reward-functions",
+    context_settings={"allow_extra_args": True, "ignore_unknown_options": True},
+)
+def lora_reward_functions(
+    ctx: typer.Context,
+    dry_run: bool = typer.Option(False, "--dry-run"),
+):
+    """List mlx-lm-lora reward functions."""
+    root = project_root_from_cwd()
+    cmd = build_mlx_lm_lora_reward_functions_command(extra_args=list(ctx.args))
+    run_mlx_lm_lora_command(cmd, dry_run=dry_run, cwd=root)
+
+
 @app.command()
 def eval(
     suite: str = typer.Option("eval/suites/coding.yaml", "--suite"),
@@ -729,7 +895,7 @@ 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")
@@ -934,6 +1100,25 @@ def verify(prompt: str, completion: str, workdir: str, **kwargs):
 """
 
 
+def _sample_verifier_llm_judge() -> str:
+    return """from mlxsmith.verifiers.llm_judge import verify as _verify
+
+def verify(prompt: str, completion: str, workdir: str, **kwargs):
+    # Pass model=... or set MLXSMITH_JUDGE_MODEL for the judge model id.
+    return _verify(prompt, completion, workdir, **kwargs)
+"""
+
+
+def _sample_judge_rubric() -> str:
+    return """Score from 0.0 to 1.0.
+- 1.0: Correct, complete, and safe.
+- 0.7: Mostly correct with small issues.
+- 0.4: Partial correctness or unclear reasoning.
+- 0.0: Incorrect or unsafe.
+Return JSON only.
+"""
+
+
 def _sample_eval_suite() -> str:
     return """name: coding-eval-sample
 notes: |

{mlxsmith-0.1.1 → mlxsmith-0.1.3}/src/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):
@@ -47,6 +47,8 @@ class TrainConfig(BaseModel):
     grad_accum: int = 8
     lr: float = 2e-4
     weight_decay: float = 0.0
+    optimizer: str = "adamw"
+    optimizer_kwargs: Dict[str, Any] = Field(default_factory=dict)
     iters: int = 1000
     save_every: int = 100
     eval_every: int = 100
@@ -61,6 +63,11 @@ class TrainConfig(BaseModel):
             raise ValueError("value must be non-negative")
         return v
 
+    @field_validator("optimizer")
+    @classmethod
+    def normalize_optimizer(cls, v: str) -> str:
+        return v.strip().lower()
+
 
 class LoraConfig(BaseModel):
     """LoRA/DoRA adapter configuration."""
@@ -89,11 +96,13 @@ class LoraConfig(BaseModel):
 
 
 class PrefConfig(BaseModel):
-    """Preference tuning configuration (DPO, ORPO, GRPO)."""
+    """Preference tuning configuration (DPO variants)."""
     
     algo: Literal["dpo", "orpo", "grpo"] = "dpo"
+    loss_type: Literal["dpo", "cpo", "orpo", "ipo", "hinge"] = "dpo"
     beta: float = 0.1
     kl_coeff: float = 0.0
+    delta: float = 0.0
     reference_model: Optional[str] = None
 
 
@@ -101,12 +110,16 @@ class RftConfig(BaseModel):
     """Reinforcement fine-tuning configuration."""
     
     algo: Literal["grpo"] = "grpo"
+    loss_type: Literal["grpo", "dr_grpo", "dapo"] = "grpo"
     rollouts: int = 8
     kl_coeff: float = 0.02
     max_steps_per_task: int = 1
     temperature: float = 0.8
     max_new_tokens: int = 256
     normalize_advantage: bool = True
+    epsilon_low: float = 0.2
+    epsilon_high: float = 0.2
+    token_level_loss: bool = False
     reference_model: Optional[str] = None
 
 
@@ -164,6 +177,7 @@ CLI_ALIASES: dict[str, tuple[str, ...]] = {
     "lr": ("train", "lr"),
     "batch_size": ("train", "batch_size"),
     "iters": ("train", "iters"),
+    "optimizer": ("train", "optimizer"),
     "model_id": ("model", "id"),
     "accel_backend": ("accel", "backend"),
     "host": ("serve", "host"),

mlxsmith-0.1.3/src/mlxsmith/integrations/__init__.py

@@ -0,0 +1,19 @@
+"""External integrations for mlxsmith."""
+
+from .mlx_lm_lora import (
+    build_train_command as build_mlx_lm_lora_train_command,
+    build_synthetic_command as build_mlx_lm_lora_synth_command,
+    build_judge_command as build_mlx_lm_lora_judge_command,
+    build_reward_functions_command as build_mlx_lm_lora_reward_functions_command,
+    run_command as run_mlx_lm_lora_command,
+    ensure_available as ensure_mlx_lm_lora_available,
+)
+
+__all__ = [
+    "build_mlx_lm_lora_train_command",
+    "build_mlx_lm_lora_synth_command",
+    "build_mlx_lm_lora_judge_command",
+    "build_mlx_lm_lora_reward_functions_command",
+    "run_mlx_lm_lora_command",
+    "ensure_mlx_lm_lora_available",
+]