mlxsmith 0.1.0__tar.gz → 0.1.2__tar.gz

mlxsmith-0.1.2/PKG-INFO

@@ -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-0.1.2/README.md

@@ -0,0 +1,228 @@
+# 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 → mlxsmith-0.1.2}/pyproject.toml

@@ -4,8 +4,8 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "mlxsmith"
-version = "0.1.0"
-description = "Apple Silicon MLX fine-tuning and OpenAI-compatible serving (SFT stable; preference/RL experimental)."
+version = "0.1.2"
+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"
 license = {text = "MIT"}
@@ -52,7 +52,6 @@ serve = [
   "uvicorn>=0.40.0",
   "httpx>=0.28.0",
 ]
-zmlx = ["zmlx"]
 dev = ["pytest>=9.0.0", "ruff>=0.14.0"]
 all = [
   "mlx>=0.30.4",

{mlxsmith-0.1.0 → mlxsmith-0.1.2}/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-0.1.0 → mlxsmith-0.1.2}/src/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-0.1.0 → mlxsmith-0.1.2}/src/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-0.1.0 → mlxsmith-0.1.2}/src/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-0.1.0 → mlxsmith-0.1.2}/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):

{mlxsmith-0.1.0 → mlxsmith-0.1.2}/src/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-0.1.0 → mlxsmith-0.1.2}/src/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