attacklm 0.7.0__tar.gz → 0.7.1__tar.gz

{attacklm-0.7.0 → attacklm-0.7.1}/CHANGELOG.md

@@ -4,6 +4,30 @@ All notable changes to AttackLM are documented in this file. Versions follow [Se
 
 ---
 
+## [0.7.1] — 2026-06-30 — GUI, VRAM fixes, dataset caching
+
+### Added
+- **attacklm-gui**: Terminal-based TUI wrapper (Textual) for all CLI commands
+  - Form-based training config with 40+ params in 5 tabs (Basic, LoRA, GaLore, Advanced, Hardware)
+  - Live training monitor with loss sparkline, VRAM gauge, progress bar, log viewer
+  - One-click screens for extract, balance, infer, merge, build, pipeline, init
+  - 5 built-in presets (3B Q-GaLore, 3B LoRA, 7B Q-GaLore, 7B QLoRA, etc.)
+  - Pause/resume/quit controls during training
+  - Works in terminal-only environments (no X11/WSLg required)
+- **Dataset caching**: Tokenized datasets saved to disk after first run, reloaded in ~1s on subsequent runs (keyed by dataset + model + max_length + packing)
+- **config.json fix**: Training runs now preserve HF model architecture fields (model_type, architectures, hidden_size, etc.) for GGUF conversion compatibility
+
+### Fixed
+- **VRAM fragmentation**: Removed `device_map="auto"` for single-GPU training — was splitting model across GPU+CPU, creating meta tensors, and leaving VRAM fragmented. Model now loads directly to GPU.
+- **Spectrum crash on 7B**: Skip meta-device tensors in SNR computation (caused by CPU-offloaded layers from device_map="auto")
+- **VRAM parsing**: Added regex for compact VRAM format (`VRAM X/Y GB (A alloc + C cache)`) in training output parser
+
+### Changed
+- Moved attacklm-gui into the main AttackLM repo under `attacklm-gui/`
+- 17 GUI tests passing
+
+---
+
 ## [0.7.0] — 2026-06-29 — Defensive data extraction, test fixes, script cleanup
 
 ### Added

{attacklm-0.7.0 → attacklm-0.7.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: attacklm
-Version: 0.7.0
+Version: 0.7.1
 Summary: QLoRA-fine-tuned Qwen2.5-Coder for offensive/defensive security and AI red-teaming
 Project-URL: Homepage, https://github.com/Veedubin/AttackLM
 Project-URL: Repository, https://github.com/Veedubin/AttackLM
@@ -205,6 +205,27 @@ AttackLM — QLoRA fine-tuning pipeline for a MITRE ATT&CK-grounded security AI
 
 ---
 
+## GUI (New in v0.7.1)
+
+AttackLM now includes a terminal-based GUI for all commands. No more memorizing 40+ CLI flags.
+
+```bash
+pip install attacklm-gui
+attacklm-gui
+```
+
+Features:
+- **Training form** with 40+ params in 5 tabs (Basic, LoRA, GaLore, Advanced, Hardware)
+- **Live training monitor** with loss sparkline, VRAM gauge, progress bar, and log viewer
+- **One-click screens** for extract, balance, infer, merge, build, pipeline, and init
+- **5 built-in presets** (3B Q-GaLore, 3B LoRA, 7B Q-GaLore, 7B QLoRA, etc.)
+- **Pause/resume/quit** controls during training
+- Works in **terminal-only** environments — no X11, no browser, no GPU required. Works over SSH and WSL.
+
+The CLI still works exactly as before. The GUI is a thin wrapper that constructs and runs CLI commands.
+
+---
+
 ## Install
 The recommended way to install the full CUDA training stack:
 
@@ -348,7 +369,8 @@ This layout allows the pipeline to deterministically extract data from upstream
 ## CLI Reference
 | Command | Description |
 |---------|-------------|
-| `attacklm-init` | One-shot init: clone $\rightarrow$ extract $\rightarrow$ attribute $\rightarrow$ buckets |
+| `attacklm-gui` | Terminal GUI for all commands (new in v0.7.1) |
+| `attacklm-init` | One-shot init: clone → extract → attribute → buckets |
 | `attacklm-train` | Train a single QLoRA adapter |
 | `attacklm-train-all` | Orchestrated training for all buckets / HPO |
 | `attacklm-balance` | Build a balanced subset of buckets |
@@ -356,7 +378,7 @@ This layout allows the pipeline to deterministically extract data from upstream
 | `attacklm-infer` | Smoke-test inference |
 | `attacklm-merge` | Merge LoRA adapter into base model |
 | `attacklm-gguf` | Convert merged model to GGUF (llama.cpp) |
-| `attacklm-build` | One-shot merge $\rightarrow$ GGUF $\rightarrow$ install |
+| `attacklm-build` | One-shot merge → GGUF → install |
 | `attacklm-demo` | Multi-agent orchestrator demo |
 | `attacklm-extract` | Extract data from cloned repositories |
 | `attacklm-buckets` | Organize data into bucket structure |

{attacklm-0.7.0 → attacklm-0.7.1}/README.md

@@ -7,6 +7,27 @@ AttackLM — QLoRA fine-tuning pipeline for a MITRE ATT&CK-grounded security AI
 
 ---
 
+## GUI (New in v0.7.1)
+
+AttackLM now includes a terminal-based GUI for all commands. No more memorizing 40+ CLI flags.
+
+```bash
+pip install attacklm-gui
+attacklm-gui
+```
+
+Features:
+- **Training form** with 40+ params in 5 tabs (Basic, LoRA, GaLore, Advanced, Hardware)
+- **Live training monitor** with loss sparkline, VRAM gauge, progress bar, and log viewer
+- **One-click screens** for extract, balance, infer, merge, build, pipeline, and init
+- **5 built-in presets** (3B Q-GaLore, 3B LoRA, 7B Q-GaLore, 7B QLoRA, etc.)
+- **Pause/resume/quit** controls during training
+- Works in **terminal-only** environments — no X11, no browser, no GPU required. Works over SSH and WSL.
+
+The CLI still works exactly as before. The GUI is a thin wrapper that constructs and runs CLI commands.
+
+---
+
 ## Install
 The recommended way to install the full CUDA training stack:
 
@@ -150,7 +171,8 @@ This layout allows the pipeline to deterministically extract data from upstream
 ## CLI Reference
 | Command | Description |
 |---------|-------------|
-| `attacklm-init` | One-shot init: clone $\rightarrow$ extract $\rightarrow$ attribute $\rightarrow$ buckets |
+| `attacklm-gui` | Terminal GUI for all commands (new in v0.7.1) |
+| `attacklm-init` | One-shot init: clone → extract → attribute → buckets |
 | `attacklm-train` | Train a single QLoRA adapter |
 | `attacklm-train-all` | Orchestrated training for all buckets / HPO |
 | `attacklm-balance` | Build a balanced subset of buckets |
@@ -158,7 +180,7 @@ This layout allows the pipeline to deterministically extract data from upstream
 | `attacklm-infer` | Smoke-test inference |
 | `attacklm-merge` | Merge LoRA adapter into base model |
 | `attacklm-gguf` | Convert merged model to GGUF (llama.cpp) |
-| `attacklm-build` | One-shot merge $\rightarrow$ GGUF $\rightarrow$ install |
+| `attacklm-build` | One-shot merge → GGUF → install |
 | `attacklm-demo` | Multi-agent orchestrator demo |
 | `attacklm-extract` | Extract data from cloned repositories |
 | `attacklm-buckets` | Organize data into bucket structure |

attacklm-0.7.1/attacklm-gui/README.md

@@ -0,0 +1,148 @@
+# AttackLM GUI
+
+Terminal-based GUI wrapper for the [AttackLM](https://github.com/Veedubin/AttackLM) CLI training tools. Built with [Textual](https://textual.textualize.io/).
+
+## Why?
+
+AttackLM has 21 CLI commands and `attacklm-train` alone has 40+ parameters. Remembering all the flags is painful. The GUI provides:
+
+- **Form-based parameter input** with tabs, presets, and validation
+- **Live training monitoring** with loss charts, VRAM gauges, and log streaming
+- **One-click commands** for extract, balance, infer, merge, build, and more
+- **Works everywhere** — terminal-only, no X11 required (WSL, SSH, headless)
+
+The CLI still works exactly as before. The GUI is a thin wrapper that constructs and runs CLI commands.
+
+## Installation
+
+```bash
+pip install attacklm-gui
+```
+
+Or from source:
+
+```bash
+git clone https://github.com/Veedubin/AttackLM.git
+cd AttackLM-Models/attacklm-gui
+pip install -e .
+```
+
+Requires `attacklm` (the CLI package) and `textual>=2.0`.
+
+## Usage
+
+```bash
+attacklm-gui
+```
+
+### Main Menu
+
+```
+┌──────────────────────────────────────┐
+│         AttackLM GUI v0.1.0          │
+│                                      │
+│  🏋️  Train Model                     │
+│  📊  Extract Data                    │
+│  ⚖️  Balance Dataset                 │
+│  🧠  Run Inference                   │
+│  🔗  Merge Adapter                   │
+│  📦  Build & Install                 │
+│  🔧  Pipeline                        │
+│  🚀  Init Dataset                    │
+└──────────────────────────────────────┘
+```
+
+### Training Form
+
+The training form organizes 40+ parameters into tabs:
+
+| Tab | Parameters |
+|-----|-----------|
+| **Basic** | dataset, output, base model, epochs, batch size, max length |
+| **LoRA** | rank, alpha, dropout, DoRA, RSLoRA, target modules, PiSSA |
+| **GaLore** | Q-GaLore, rank, 32-bit, Spectrum layer freezing |
+| **Advanced** | eval split, early stopping, save steps, gradient accum, optimizer, packing, live LR |
+| **Hardware** | precision (BF16/FP16/FP32), multi-GPU, MoE safe, auto-tune |
+
+**Presets** save and load complete configurations. Built-in presets:
+- `3B Q-GaLore Spectrum` — full-parameter training on 16GB GPU
+- `3B Q-GaLore Rank 128` — higher quality variant
+- `3B LoRA Default` — standard QLoRA
+- `7B Q-GaLore` — for 24GB GPUs
+- `7B QLoRA Default` — standard QLoRA for 7B
+
+### Live Training Monitor
+
+```
+┌─────────────────────────────────────────────────────┐
+│  Training: attacklm-3b-qgalore-spectrum             │
+│  Epoch 12/30  |  Step 1896  |  Elapsed: 1h 23m     │
+├──────────────────────┬──────────────────────────────┤
+│  Loss: 1.1143 ▂▃▅▆▇  │  VRAM: 5.9/15.6 GB (62%)   │
+│  Eval Loss: 1.176    │  alloc: 5.9  cache: 4.8     │
+│  Trend: ↓ -0.0116    │  ████████░░░░               │
+├──────────────────────┴──────────────────────────────┤
+│  Log Output (scrolling)                             │
+│  Epoch 12/20  58% | loss 1.1143 | 2,844 tok/s ...  │
+├─────────────────────────────────────────────────────┤
+│  [P]ause  [S]top at checkpoint  [Q]uit              │
+└─────────────────────────────────────────────────────┘
+```
+
+Controls:
+- **Pause/Resume** — SIGSTOP/SIGCONT the training process
+- **Stop at Checkpoint** — graceful stop after current eval
+- **Quit** — kill training immediately
+
+### Other Commands
+
+| Screen | What it does |
+|--------|-------------|
+| **Extract** | Run data extractors on upstream sources |
+| **Balance** | Balance a dataset with cap size |
+| **Inference** | Run a trained model with a prompt |
+| **Merge** | Merge LoRA adapter into base model |
+| **Build** | Merge → GGUF → install to LM Studio |
+| **Pipeline** | Run a YAML training pipeline |
+| **Init** | One-click dataset initialization |
+
+## Keyboard Shortcuts
+
+| Key | Action |
+|-----|--------|
+| `q` | Quit |
+| `m` | Return to main menu |
+| `Tab` | Next field |
+| `Enter` | Activate button |
+| `Esc` | Go back |
+
+## Architecture
+
+```
+attacklm-gui/
+├── src/attacklm_gui/
+│   ├── app.py              # Textual App entry point
+│   ├── cli.py              # CLI entry point
+│   ├── runner.py           # Subprocess manager + output parser
+│   ├── presets.py          # Save/load training presets
+│   ├── screens/
+│   │   ├── main_menu.py    # Command launcher
+│   │   ├── train_form.py   # Training parameter form
+│   │   ├── train_live.py   # Live training monitor
+│   │   └── command_forms.py # Other command screens
+│   └── widgets/
+│       └── __init__.py
+└── tests/
+    └── test_gui.py
+```
+
+The GUI never touches AttackLM's internal logic. It constructs CLI command strings and runs them via `asyncio.create_subprocess_exec`. All training, extraction, and inference logic stays in `scripts/*.py`.
+
+## Requirements
+
+- Python 3.10+
+- Linux (any flavor) or WSL
+- Terminal with 256-color support
+- AttackLM installed (`pip install attacklm[all]`)
+
+No X11, no GPU, no browser required. Works over SSH.

attacklm-0.7.1/attacklm-gui/pyproject.toml

@@ -0,0 +1,19 @@
+[project]
+name = "attacklm-gui"
+version = "0.1.0"
+description = "Terminal GUI wrapper for AttackLM CLI training tools"
+requires-python = ">=3.10"
+dependencies = [
+    "textual>=2.0",
+    "attacklm",
+]
+
+[project.scripts]
+attacklm-gui = "attacklm_gui.cli:main"
+
+[build-system]
+requires = ["setuptools>=75"]
+build-backend = "setuptools.build_meta"
+
+[tool.setuptools.packages.find]
+where = ["src"]

attacklm-0.7.1/attacklm-gui/src/attacklm_gui/__init__.py

@@ -0,0 +1,3 @@
+"""AttackLM GUI — Terminal-based wrapper for the AttackLM CLI."""
+
+__version__ = "0.1.0"

attacklm-0.7.1/attacklm-gui/src/attacklm_gui/app.py

@@ -0,0 +1,58 @@
+"""Main Textual application for AttackLM GUI."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+from textual.app import App, ComposeResult
+from textual.binding import Binding
+from textual.widgets import Header, Footer
+
+from attacklm_gui.presets import ensure_builtin_presets
+from attacklm_gui.screens.main_menu import MainMenuScreen
+
+
+class AttackLMApp(App):
+    """Terminal GUI wrapper for AttackLM CLI tools."""
+
+    TITLE = "AttackLM GUI"
+    SUB_TITLE = "v0.1.0"
+    CSS_PATH = None  # We'll use inline CSS for now
+
+    BINDINGS = [
+        Binding("q", "quit", "Quit", show=True),
+        Binding("m", "main_menu", "Main Menu", show=True),
+        Binding("?", "help", "Help", show=False),
+    ]
+
+    def __init__(self) -> None:
+        super().__init__()
+        self.attacklm_dir = self._find_attacklm_dir()
+
+    def _find_attacklm_dir(self) -> Path:
+        """Find the AttackLM project directory."""
+        # Try common locations
+        candidates = [
+            Path.cwd(),
+            Path.home() / "Projects" / "reverse_engineering" / "AttackLM",
+            Path.home() / "AttackLM",
+        ]
+        for c in candidates:
+            if (c / "pyproject.toml").exists() and (
+                c / "scripts" / "train_template.py"
+            ).exists():
+                return c
+        return Path.cwd()
+
+    def compose(self) -> ComposeResult:
+        yield Header()
+        yield Footer()
+
+    def on_mount(self) -> None:
+        ensure_builtin_presets()
+        self.push_screen(MainMenuScreen())
+
+    def action_main_menu(self) -> None:
+        """Return to main menu."""
+        while len(self.screen_stack) > 1:
+            self.pop_screen()

attacklm-0.7.1/attacklm-gui/src/attacklm_gui/cli.py

@@ -0,0 +1,17 @@
+"""CLI entry point for attacklm-gui."""
+
+import sys
+from pathlib import Path
+
+
+def main() -> int:
+    """Launch the AttackLM TUI."""
+    from attacklm_gui.app import AttackLMApp
+
+    app = AttackLMApp()
+    app.run()
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

attacklm-0.7.1/attacklm-gui/src/attacklm_gui/presets.py

@@ -0,0 +1,170 @@
+"""Configuration preset management for AttackLM training parameters."""
+
+from __future__ import annotations
+
+import json
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any
+
+
+PRESETS_DIR = Path.home() / ".config" / "attacklm" / "presets"
+
+
+@dataclass
+class Preset:
+    """A saved training configuration."""
+
+    name: str
+    params: dict[str, Any] = field(default_factory=dict)
+    description: str = ""
+
+    @property
+    def filename(self) -> str:
+        return f"{self.name.lower().replace(' ', '_')}.json"
+
+    @property
+    def path(self) -> Path:
+        return PRESETS_DIR / self.filename
+
+    def save(self) -> None:
+        """Save preset to disk."""
+        PRESETS_DIR.mkdir(parents=True, exist_ok=True)
+        data = {
+            "name": self.name,
+            "description": self.description,
+            "params": self.params,
+        }
+        with open(self.path, "w") as f:
+            json.dump(data, f, indent=2)
+
+    @classmethod
+    def load(cls, name: str) -> Preset | None:
+        """Load a preset by name."""
+        filename = f"{name.lower().replace(' ', '_')}.json"
+        path = PRESETS_DIR / filename
+        if not path.exists():
+            return None
+        with open(path) as f:
+            data = json.load(f)
+        return cls(
+            name=data["name"],
+            params=data.get("params", {}),
+            description=data.get("description", ""),
+        )
+
+    @classmethod
+    def list_all(cls) -> list[str]:
+        """List all saved preset names."""
+        PRESETS_DIR.mkdir(parents=True, exist_ok=True)
+        presets = []
+        for p in sorted(PRESETS_DIR.glob("*.json")):
+            try:
+                with open(p) as f:
+                    data = json.load(f)
+                presets.append(data.get("name", p.stem))
+            except (json.JSONDecodeError, OSError):
+                pass
+        return presets
+
+    @classmethod
+    def delete(cls, name: str) -> bool:
+        """Delete a preset by name. Returns True if deleted."""
+        filename = f"{name.lower().replace(' ', '_')}.json"
+        path = PRESETS_DIR / filename
+        if path.exists():
+            path.unlink()
+            return True
+        return False
+
+
+# --- Built-in presets ---
+
+BUILTIN_PRESETS: list[Preset] = [
+    Preset(
+        name="3B Q-GaLore Spectrum",
+        description="Q-GaLore full-parameter training for 3B models on 16GB GPU",
+        params={
+            "use_qgalore": True,
+            "spectrum": 0.5,
+            "galore_rank": 64,
+            "epochs": 20,
+            "batch_size": 1,
+            "max_length": 12000,
+            "packing": True,
+            "early_stop_steps": 5,
+            "live_lr": True,
+            "optim": "paged_adamw_8bit",
+        },
+    ),
+    Preset(
+        name="3B Q-GaLore Rank 128",
+        description="Higher-rank Q-GaLore for better quality on 3B models",
+        params={
+            "use_qgalore": True,
+            "spectrum": 0.5,
+            "galore_rank": 128,
+            "epochs": 30,
+            "batch_size": 1,
+            "max_length": 12000,
+            "packing": True,
+            "early_stop_steps": 5,
+            "live_lr": True,
+            "optim": "paged_adamw_8bit",
+        },
+    ),
+    Preset(
+        name="3B LoRA Default",
+        description="Standard QLoRA training for 3B models",
+        params={
+            "lora_r": 16,
+            "lora_alpha": 32,
+            "lora_dropout": 0.05,
+            "epochs": 10,
+            "batch_size": 2,
+            "max_length": 2048,
+            "packing": False,
+            "early_stop_steps": 5,
+            "optim": "paged_adamw_8bit",
+        },
+    ),
+    Preset(
+        name="7B Q-GaLore",
+        description="Q-GaLore for 7B models on 24GB GPU",
+        params={
+            "use_qgalore": True,
+            "spectrum": 0.5,
+            "galore_rank": 128,
+            "epochs": 20,
+            "batch_size": 1,
+            "max_length": 8192,
+            "packing": True,
+            "early_stop_steps": 5,
+            "live_lr": True,
+            "optim": "paged_adamw_8bit",
+        },
+    ),
+    Preset(
+        name="7B QLoRA Default",
+        description="Standard QLoRA for 7B models",
+        params={
+            "lora_r": 16,
+            "lora_alpha": 32,
+            "lora_dropout": 0.05,
+            "epochs": 10,
+            "batch_size": 2,
+            "max_length": 2048,
+            "packing": False,
+            "early_stop_steps": 5,
+            "optim": "paged_adamw_8bit",
+        },
+    ),
+]
+
+
+def ensure_builtin_presets() -> None:
+    """Create built-in presets if they don't exist."""
+    existing = set(Preset.list_all())
+    for preset in BUILTIN_PRESETS:
+        if preset.name not in existing:
+            preset.save()