attacklm 0.6.0__tar.gz → 0.6.2__tar.gz

{attacklm-0.6.0 → attacklm-0.6.2}/.github/workflows/release.yml

@@ -14,8 +14,26 @@ on:
       - "v*"
 
 jobs:
+  validate:
+    name: Validate version 🔍
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Check tag matches __version__.py
+        run: |
+          TAG_VERSION="${GITHUB_REF#refs/tags/v}"
+          FILE_VERSION=$(python3 -c "exec(open('src/attacklm/__version__.py').read()); print(__version__)")
+          if [ "$TAG_VERSION" != "$FILE_VERSION" ]; then
+            echo "ERROR: Tag v${TAG_VERSION} does not match __version__.py (${FILE_VERSION})"
+            echo "Bump src/attacklm/__version__.py to ${TAG_VERSION} before pushing the tag."
+            exit 1
+          fi
+          echo "Version validated: v${TAG_VERSION} == __version__.py (${FILE_VERSION})"
+
   build:
     name: Build distribution 📦
+    needs: validate
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v4

{attacklm-0.6.0 → attacklm-0.6.2}/CHANGELOG.md

@@ -4,17 +4,23 @@ All notable changes to AttackLM are documented in this file. Versions follow [Se
 
 ---
 
-## [0.6.0] — 2026-06-28 — Q-GaLore, Spectrum, PiSSA init
+## [0.6.1] — 2026-06-28 — Fix --use-galore API
 
-### Added
+### Fixed
 
-- **Q-GaLore (default)**: INT4 quantized gradient projection matrices with stochastic rounding. Cuts optimizer memory ~4x vs vanilla GaLore. Enables 7B full-parameter training on 16GB GPUs. Paper: arXiv:2407.08296. Use `--galore-fp16` for vanilla GaLore.
-- **Spectrum**: SNR-based layer freezing. Computes signal-to-noise ratio per layer and freezes the lowest-SNR layers. `--spectrum` keeps top 50%, `--spectrum 0.25` keeps top 25%. Reduces VRAM proportionally. Compatible with any training method. Paper: arXiv:2406.06623.
-- **PiSSA init**: SVD-based LoRA initialization. `--pissa-init` initializes LoRA weights from the SVD of pre-trained weights instead of random Kaiming init. Faster convergence, lower final loss. Paper: arXiv:2404.02948.
+- **`--use-galore` restored to vanilla GaLore** (FP16 projections). Q-GaLore is now `--use-qgalore` (separate flag). This preserves backward compatibility — existing `--use-galore` commands continue to work.
+- **Removed `--galore-fp16`** (redundant — `--use-galore` is FP16 by default).
+- **CI validation**: added version check job that fails the workflow if the git tag doesn't match `__version__.py`, preventing duplicate PyPI uploads.
 
-### Changed
+---
+
+## [0.6.0] — 2026-06-28 — Q-GaLore, Spectrum, PiSSA init
+
+### Added
 
-- **Default change**: `--use-galore` now defaults to Q-GaLore (INT4). Use `--galore-fp16` for vanilla GaLore.
+- **Q-GaLore** (`--use-qgalore`): INT4 quantized gradient projection matrices with stochastic rounding. Cuts optimizer memory ~4x vs vanilla GaLore. Enables 7B full-parameter training on 16GB GPUs. Requires `--use-galore`. Paper: arXiv:2407.08296.
+- **Spectrum** (`--spectrum`): SNR-based layer freezing. Computes signal-to-noise ratio per layer and freezes the lowest-SNR layers. `--spectrum` keeps top 50%, `--spectrum 0.25` keeps top 25%. Reduces VRAM proportionally. Compatible with any training method. Paper: arXiv:2406.06623.
+- **PiSSA init** (`--pissa-init`): SVD-based LoRA initialization. Initializes LoRA weights from the SVD of pre-trained weights instead of random Kaiming init. Faster convergence, lower final loss. Paper: arXiv:2404.02948.
 
 ---
 

{attacklm-0.6.0 → attacklm-0.6.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: attacklm
-Version: 0.6.0
+Version: 0.6.2
 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

{attacklm-0.6.0 → attacklm-0.6.2}/scripts/train_template.py

@@ -389,17 +389,29 @@ Examples:
         action="store_true",
         default=False,
         help=(
-            "Use Q-GaLore (Quantized Gradient Low-Rank Projection) for "
-            "full-parameter fine-tuning. Q-GaLore quantizes gradient "
-            "projection matrices to INT4 with stochastic rounding, "
-            "enabling 7B full-parameter training on 16GB GPUs. "
-            "Mutually exclusive with --use-unsloth. "
+            "Use GaLore (Gradient Low-Rank Projection) for full-parameter "
+            "fine-tuning. GaLore projects gradients into a low-rank space "
+            "during optimization, enabling full-parameter learning on "
+            "consumer GPUs without LoRA adapters. Mutually exclusive with "
+            "--use-unsloth (GaLore trains ALL parameters, no LoRA needed). "
             "Requires 'uv pip install attacklm[galore]'. "
-            "Default: Q-GaLore (INT4 projections + 8-bit optimizer). "
-            "Use --galore-fp16 for vanilla GaLore (FP16 projections). "
+            "Default: 8-bit optimizer + per-layer hooks (fits 3B on 16GB). "
+            "Use --use-qgalore for INT4 projections (fits 7B on 16GB). "
             "Use --galore-32bit for full-precision (needs ~20GB+ for 3B)."
         ),
     )
+    parser.add_argument(
+        "--use-qgalore",
+        action="store_true",
+        default=False,
+        help=(
+            "Use Q-GaLore: INT4 quantized gradient projection matrices "
+            "with stochastic rounding. Cuts optimizer memory by ~4x vs "
+            "vanilla GaLore, enabling 7B full-parameter training on 16GB. "
+            "Mutually exclusive with --use-galore and --use-unsloth. "
+            "Paper: arXiv:2407.08296."
+        ),
+    )
     parser.add_argument(
         "--galore-32bit",
         action="store_true",
@@ -433,19 +445,6 @@ Examples:
             "scales with rank². Only meaningful with --use-galore."
         ),
     )
-    parser.add_argument(
-        "--galore-fp16",
-        action="store_true",
-        default=False,
-        help=(
-            "Use vanilla GaLore with FP16 projection matrices instead of "
-            "the default Q-GaLore (INT4 quantized projections). "
-            "Q-GaLore cuts optimizer memory by ~4x vs vanilla GaLore, "
-            "enabling 7B models on 16GB. Use --galore-fp16 if you need "
-            "maximum precision (e.g., for a rented A100 with plenty of VRAM). "
-            "Only meaningful with --use-galore."
-        ),
-    )
     parser.add_argument(
         "--spectrum",
         type=float,
@@ -483,6 +482,11 @@ Examples:
 # ---------------------------------------------------------------------------
 
 
+def _is_galore(args) -> bool:
+    """True if either --use-galore or --use-qgalore is active."""
+    return args.use_galore or args.use_qgalore
+
+
 def _resolve_model_path(model_id_or_path: str) -> str:
     """Resolve a user-supplied base-model argument to something `from_pretrained`
     can actually load.
@@ -1631,7 +1635,7 @@ def main() -> None:
     )
     _galore_display = (
         f"ON (full-parameter, rank={args.galore_rank}, update_proj_gap=200, scale=0.25)"
-        if args.use_galore
+        if _is_galore(args)
         else "OFF"
     )
 
@@ -1641,13 +1645,13 @@ def main() -> None:
     _plan.add_row("Model", args.base_model)
     if args.use_unsloth:
         _plan.add_row("Quantization", "4-bit (Unsloth internal)")
-    elif args.use_galore:
+    elif _is_galore(args):
         _plan.add_row("Quantization", "NONE (bf16 full-parameter for GaLore)")
     elif not args.moe_safe_target:
         _plan.add_row("Quantization", "4-bit NF4 (double quant)")
     else:
         _plan.add_row("Quantization", "NONE (bf16 full-precision for MoE)")
-    if args.use_galore:
+    if _is_galore(args):
         _plan.add_row("Training mode", "GaLore full-parameter (no LoRA)")
     else:
         _plan.add_row("LoRA rank", str(args.lora_r))
@@ -1700,7 +1704,7 @@ def main() -> None:
                 "~4-5 GB (batch=2, len=2048) / ~3-4 GB (batch=1, len=1024)",
             )
             _dry_tbl.add_row("Note", "13B model fits in 16GB with Unsloth QLoRA")
-        if args.use_galore:
+        if _is_galore(args):
             _dry_tbl.add_row(
                 "Est. VRAM (GaLore)",
                 "~10-12 GB (3B) / ~14-16 GB (7B), batch=1, len=2048",
@@ -1799,13 +1803,13 @@ def main() -> None:
     except Exception as e:
         console.print(f"  [dim](Skipped state.json write: {e})[/dim]")
 
-    # --- Mutual exclusivity: GaLore vs Unsloth ---
-    if args.use_galore and args.use_unsloth:
+    # --- Mutual exclusivity: GaLore/Q-GaLore vs Unsloth ---
+    if (args.use_galore or args.use_qgalore) and args.use_unsloth:
         console.print(
-            "[red]ERROR:[/red] --use-galore and --use-unsloth are mutually exclusive.\n"
-            "  GaLore is full-parameter training (no LoRA adapters).\n"
+            "[red]ERROR:[/red] --use-galore/--use-qgalore and --use-unsloth are mutually exclusive.\n"
+            "  GaLore/Q-GaLore is full-parameter training (no LoRA adapters).\n"
             "  Unsloth is optimized QLoRA (LoRA adapters on quantized base).\n"
-            "  Choose one: --use-galore OR --use-unsloth, not both."
+            "  Choose one: --use-galore/--use-qgalore OR --use-unsloth, not both."
         )
         sys.exit(1)
 
@@ -1830,7 +1834,7 @@ def main() -> None:
 
     # --- GaLore: import galore-torch for full-parameter training ---
     _galore_available = False
-    if args.use_galore:
+    if _is_galore(args):
         try:
             # Fix: bitsandbytes calls torch.utils._pytree.register_constant()
             # on Enum subclasses, which is deprecated in PyTorch 2.12+.
@@ -1864,7 +1868,7 @@ def main() -> None:
 
     # --multi-gpu + --use-galore: auto-enable 32-bit (per-layer hooks
     # are incompatible with DDP's gradient all-reduce).
-    if args.multi_gpu and args.use_galore and not args.galore_32bit:
+    if args.multi_gpu and _is_galore(args) and not args.galore_32bit:
         console.print(
             "  [yellow]GaLore:[/yellow] --multi-gpu detected — auto-enabling "
             "--galore-32bit (per-layer weight updates are incompatible with DDP)"
@@ -1920,7 +1924,9 @@ def main() -> None:
     # load_in_4bit parameter. We skip BitsAndBytes entirely.
     # --use-galore: GaLore is full-parameter training — no quantization
     # needed. Model is loaded in bf16/fp16 for full-parameter updates.
-    skip_quantization = args.moe_safe_target or args.use_unsloth or args.use_galore
+    skip_quantization = (
+        args.moe_safe_target or args.use_unsloth or args.use_galore or args.use_qgalore
+    )
 
     if skip_quantization:
         bnb_config = None  # No quantization for MoE-safe or Unsloth mode
@@ -2084,7 +2090,7 @@ def main() -> None:
                 # trust_remote_code=False — the model is architecturally
                 # identical to standard Qwen2.5-Coder, just abliterated weights.
                 # The standard Qwen2ForCausalLM class respects torch_dtype.
-                if args.use_galore:
+                if _is_galore(args):
                     # Standard models (Qwen, Llama, etc.) load in bf16 with
                     # dtype= directly. No config patching needed.
                     # Note: huihui-ai abliterated models have custom code
@@ -2277,8 +2283,8 @@ def main() -> None:
         sys.exit(1)
 
     # --- Apply LoRA (or skip for GaLore full-parameter training) ---
-    if args.use_galore and _galore_available:
-        # GaLore: full-parameter training — no LoRA adapter needed.
+    if _is_galore(args) and _galore_available:
+        # GaLore/Q-GaLore: full-parameter training — no LoRA adapter needed.
         # GaLore projects gradients into a low-rank space during optimization,
         # so ALL model parameters are trained directly. No PEFT wrapping.
         print("GaLore: full-parameter training (no LoRA adapter)")
@@ -3304,17 +3310,17 @@ def main() -> None:
         hpo_callbacks.append(HPOMetricsCSVCallback(args.hpo_metrics_csv, hpo_label))
         print(f"\n  HPO metrics CSV: {args.hpo_metrics_csv}  (label={hpo_label})")
 
-    # --- Q-GaLore / GaLore: full-parameter training with gradient projection ---
-    # Q-GaLore (default): INT4 quantized projection matrices + stochastic
-    # rounding + layer-adaptive SVD skipping. Cuts optimizer memory ~4x vs
-    # vanilla GaLore. Enables 7B on 16GB.
-    # Vanilla GaLore (--galore-fp16): FP16 projection matrices.
-    # Paper: arXiv:2407.08296
-    if args.use_galore and _galore_available:
+    # --- GaLore / Q-GaLore: full-parameter training with gradient projection ---
+    # GaLore (--use-galore): FP16 projection matrices, 8-bit optimizer.
+    #   Fits 3B on 16GB, 7B on 24GB.
+    # Q-GaLore (--use-qgalore): INT4 quantized projection matrices with
+    #   stochastic rounding. Cuts optimizer memory ~4x vs vanilla GaLore.
+    #   Enables 7B on 16GB. Paper: arXiv:2407.08296.
+    if _is_galore(args) and _galore_available:
         from transformers import Trainer
 
         _use_32bit = args.galore_32bit
-        _use_qgalore = not args.galore_fp16  # Q-GaLore is the default
+        _use_qgalore = args.use_qgalore
         _galore_rank = args.galore_rank
 
         if _use_qgalore:
@@ -3669,7 +3675,7 @@ def main() -> None:
     # use train_result.metrics["epoch"] for that because HF rounds it
     # to an int and we want the fractional value from log_history.
 
-    if args.use_galore and _galore_available:
+    if _is_galore(args) and _galore_available:
         print(f"\nSaving GaLore full-parameter model to: {args.output}")
     else:
         print(f"\nSaving best LoRA adapter to: {args.output}")
@@ -3682,7 +3688,7 @@ def main() -> None:
     best_ckpt = Path(args.output) / "checkpoint-best"
     if best_ckpt.exists() and (best_ckpt / "model.safetensors").exists():
         print(f"  Loading best checkpoint from: {best_ckpt}")
-        if args.use_galore and _galore_available:
+        if (args.use_galore or args.use_qgalore) and _galore_available:
             # GaLore: full model, reload from checkpoint
             model = AutoModelForCausalLM.from_pretrained(
                 str(best_ckpt),
@@ -3703,7 +3709,7 @@ def main() -> None:
         "base_model": args.base_model,
         "dataset": args.dataset,
         "training_mode": "galore_full_parameter"
-        if (args.use_galore and _galore_available)
+        if (_is_galore(args) and _galore_available)
         else "qlora",
         "lora_r": args.lora_r,
         "lora_alpha": args.lora_alpha,
@@ -3728,7 +3734,7 @@ def main() -> None:
         if not skip_quantization
         else (
             "none (bf16 full-parameter for GaLore)"
-            if (args.use_galore and _galore_available)
+            if (_is_galore(args) and _galore_available)
             else "none (bf16 full-precision for MoE)"
         ),
         "epochs": args.epochs,
@@ -3738,13 +3744,13 @@ def main() -> None:
         "packing": args.packing,
         "compute_dtype": compute_type,
         "optimizer": "galore_adamw"
-        if (args.use_galore and _galore_available)
+        if (_is_galore(args) and _galore_available)
         else args.optim,
         "final_loss": final_loss,
         "training_time_seconds": round(elapsed, 2),
         "num_examples": len(dataset),
     }
-    if args.use_galore and _galore_available:
+    if _is_galore(args) and _galore_available:
         config["galore_rank"] = 128
         config["galore_update_proj_gap"] = 200
         config["galore_scale"] = 0.25
@@ -3894,7 +3900,7 @@ def main() -> None:
     )
     summary_tbl.add_row("Config written to", config_path)
 
-    if args.use_galore and _galore_available:
+    if _is_galore(args) and _galore_available:
         code = f"""from transformers import AutoModelForCausalLM, AutoTokenizer
 
 model = AutoModelForCausalLM.from_pretrained("{args.output}", device_map="auto")

attacklm-0.6.2/src/attacklm/__version__.py

@@ -0,0 +1 @@
+__version__ = "0.6.2"

attacklm-0.6.0/src/attacklm/__version__.py

@@ -1 +0,0 @@
-__version__ = "0.6.0"