attacklm 0.5.4__tar.gz → 0.5.6__tar.gz

{attacklm-0.5.4 → attacklm-0.5.6}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: attacklm
-Version: 0.5.4
+Version: 0.5.6
 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
@@ -25,10 +25,7 @@ Requires-Python: >=3.10
 Provides-Extra: all
 Requires-Dist: accelerate==1.13.0; extra == 'all'
 Requires-Dist: bitsandbytes==0.49.2; extra == 'all'
-Requires-Dist: causal-conv1d>=1.4; extra == 'all'
 Requires-Dist: datasets==5.0.0; extra == 'all'
-Requires-Dist: flash-attn>=2.7; extra == 'all'
-Requires-Dist: flash-linear-attention>=0.2; extra == 'all'
 Requires-Dist: gguf; extra == 'all'
 Requires-Dist: gitpython; extra == 'all'
 Requires-Dist: huggingface-hub==1.18.0; extra == 'all'
@@ -51,10 +48,7 @@ Requires-Dist: trl==1.5.1; extra == 'all'
 Provides-Extra: all-cuda
 Requires-Dist: accelerate==1.13.0; extra == 'all-cuda'
 Requires-Dist: bitsandbytes==0.49.2; extra == 'all-cuda'
-Requires-Dist: causal-conv1d>=1.4; extra == 'all-cuda'
 Requires-Dist: datasets==5.0.0; extra == 'all-cuda'
-Requires-Dist: flash-attn>=2.7; extra == 'all-cuda'
-Requires-Dist: flash-linear-attention>=0.2; extra == 'all-cuda'
 Requires-Dist: gguf; extra == 'all-cuda'
 Requires-Dist: gitpython; extra == 'all-cuda'
 Requires-Dist: huggingface-hub==1.18.0; extra == 'all-cuda'
@@ -95,9 +89,6 @@ Requires-Dist: bitsandbytes==0.49.2; extra == 'cuda'
 Requires-Dist: torch==2.12.0; extra == 'cuda'
 Requires-Dist: torchvision==0.27.0; extra == 'cuda'
 Provides-Extra: cuda-extras
-Requires-Dist: causal-conv1d>=1.4; extra == 'cuda-extras'
-Requires-Dist: flash-attn>=2.7; extra == 'cuda-extras'
-Requires-Dist: flash-linear-attention>=0.2; extra == 'cuda-extras'
 Provides-Extra: dev
 Requires-Dist: ipython; extra == 'dev'
 Requires-Dist: mypy; extra == 'dev'
@@ -110,6 +101,10 @@ Provides-Extra: extract
 Requires-Dist: gitpython; extra == 'extract'
 Requires-Dist: pyyaml; extra == 'extract'
 Requires-Dist: requests; extra == 'extract'
+Provides-Extra: flash-attn
+Requires-Dist: causal-conv1d>=1.4; extra == 'flash-attn'
+Requires-Dist: flash-attn>=2.7; extra == 'flash-attn'
+Requires-Dist: flash-linear-attention>=0.2; extra == 'flash-attn'
 Provides-Extra: infer
 Requires-Dist: bitsandbytes==0.49.2; extra == 'infer'
 Requires-Dist: huggingface-hub==1.18.0; extra == 'infer'
@@ -151,10 +146,7 @@ Provides-Extra: rocm-extras
 Provides-Extra: train
 Requires-Dist: accelerate==1.13.0; extra == 'train'
 Requires-Dist: bitsandbytes==0.49.2; extra == 'train'
-Requires-Dist: causal-conv1d>=1.4; extra == 'train'
 Requires-Dist: datasets==5.0.0; extra == 'train'
-Requires-Dist: flash-attn>=2.7; extra == 'train'
-Requires-Dist: flash-linear-attention>=0.2; extra == 'train'
 Requires-Dist: huggingface-hub==1.18.0; extra == 'train'
 Requires-Dist: kernels<0.13,>=0.12; extra == 'train'
 Requires-Dist: peft==0.19.1; extra == 'train'
@@ -177,10 +169,7 @@ Requires-Dist: trl==1.5.1; extra == 'train-base'
 Provides-Extra: train-cuda
 Requires-Dist: accelerate==1.13.0; extra == 'train-cuda'
 Requires-Dist: bitsandbytes==0.49.2; extra == 'train-cuda'
-Requires-Dist: causal-conv1d>=1.4; extra == 'train-cuda'
 Requires-Dist: datasets==5.0.0; extra == 'train-cuda'
-Requires-Dist: flash-attn>=2.7; extra == 'train-cuda'
-Requires-Dist: flash-linear-attention>=0.2; extra == 'train-cuda'
 Requires-Dist: huggingface-hub==1.18.0; extra == 'train-cuda'
 Requires-Dist: kernels<0.13,>=0.12; extra == 'train-cuda'
 Requires-Dist: peft==0.19.1; extra == 'train-cuda'
@@ -281,12 +270,14 @@ The full per-source map:
 # 1. Install uv (Python package manager, ~10MB)
 curl -LsSf https://astral.sh/uv/install.sh | sh
 
-# 2. Install PyTorch + Triton first (required by flash-attn and Unsloth)
-uv pip install torch==2.12.0 torchvision==0.27.0 triton
-
-# 3. Install AttackLM from PyPI with the full CUDA training stack
+# 2. Install AttackLM with the full CUDA training stack
 uv pip install "attacklm[all]"
 
+# 3. (Optional) Install flash-attn for 2-3x faster training on Qwen3-Next models
+#    --no-build-isolation is required because flash-attn needs torch at build time
+#    but doesn't declare it in its build-system.requires
+uv pip install "attacklm[flash-attn]" --no-build-isolation
+
 # 4. Clone the repo for the training data (the PyPI package is code-only)
 git clone https://github.com/Veedubin/AttackLM.git
 cd AttackLM
@@ -344,13 +335,23 @@ build anything by hand.
 
 **Install from PyPI (recommended):**
 ```bash
-# PyTorch + Triton must be installed first (flash-attn and Unsloth need them)
-uv pip install torch==2.12.0 torchvision==0.27.0 triton
+# Full CUDA training stack (torch, bitsandbytes, trl, peft, transformers, etc.)
+uv pip install "attacklm[all]"
+
+# Optional: flash-attn for 2-3x faster training (Qwen3-Next, packing mode)
+# --no-build-isolation is required because flash-attn needs torch at build time
+# but doesn't declare it in its build-system.requires
+uv pip install "attacklm[flash-attn]" --no-build-isolation
+
+# Optional: Unsloth for 2-5x faster training + 70% less VRAM
+# Must be separate because it conflicts with trl==1.5.1 (needs trl<=0.24.0)
+uv pip install "attacklm[unsloth]"
+
+# ROCm stack (AMD GPUs)
+uv pip install "attacklm[all-rocm]"
 
-# Then install AttackLM
-uv pip install "attacklm[all]"          # CUDA training stack
-uv pip install "attacklm[all-rocm]"     # ROCm training stack
-uv pip install "attacklm[infer]"        # CPU / Apple Silicon (inference only)
+# Inference only (CPU / Apple Silicon, no GPU needed)
+uv pip install "attacklm[infer]"
 ```
 
 Then clone the repo for the training data: `git clone https://github.com/Veedubin/AttackLM.git`

{attacklm-0.5.4 → attacklm-0.5.6}/README.md

@@ -72,12 +72,14 @@ The full per-source map:
 # 1. Install uv (Python package manager, ~10MB)
 curl -LsSf https://astral.sh/uv/install.sh | sh
 
-# 2. Install PyTorch + Triton first (required by flash-attn and Unsloth)
-uv pip install torch==2.12.0 torchvision==0.27.0 triton
-
-# 3. Install AttackLM from PyPI with the full CUDA training stack
+# 2. Install AttackLM with the full CUDA training stack
 uv pip install "attacklm[all]"
 
+# 3. (Optional) Install flash-attn for 2-3x faster training on Qwen3-Next models
+#    --no-build-isolation is required because flash-attn needs torch at build time
+#    but doesn't declare it in its build-system.requires
+uv pip install "attacklm[flash-attn]" --no-build-isolation
+
 # 4. Clone the repo for the training data (the PyPI package is code-only)
 git clone https://github.com/Veedubin/AttackLM.git
 cd AttackLM
@@ -135,13 +137,23 @@ build anything by hand.
 
 **Install from PyPI (recommended):**
 ```bash
-# PyTorch + Triton must be installed first (flash-attn and Unsloth need them)
-uv pip install torch==2.12.0 torchvision==0.27.0 triton
+# Full CUDA training stack (torch, bitsandbytes, trl, peft, transformers, etc.)
+uv pip install "attacklm[all]"
+
+# Optional: flash-attn for 2-3x faster training (Qwen3-Next, packing mode)
+# --no-build-isolation is required because flash-attn needs torch at build time
+# but doesn't declare it in its build-system.requires
+uv pip install "attacklm[flash-attn]" --no-build-isolation
+
+# Optional: Unsloth for 2-5x faster training + 70% less VRAM
+# Must be separate because it conflicts with trl==1.5.1 (needs trl<=0.24.0)
+uv pip install "attacklm[unsloth]"
+
+# ROCm stack (AMD GPUs)
+uv pip install "attacklm[all-rocm]"
 
-# Then install AttackLM
-uv pip install "attacklm[all]"          # CUDA training stack
-uv pip install "attacklm[all-rocm]"     # ROCm training stack
-uv pip install "attacklm[infer]"        # CPU / Apple Silicon (inference only)
+# Inference only (CPU / Apple Silicon, no GPU needed)
+uv pip install "attacklm[infer]"
 ```
 
 Then clone the repo for the training data: `git clone https://github.com/Veedubin/AttackLM.git`

{attacklm-0.5.4 → attacklm-0.5.6}/pyproject.toml

@@ -150,7 +150,14 @@ rocm = [
 # `causal_conv1d` and `flash_linear_attention` are CUDA C++ extensions.
 # ROCm users can SKIP these — the Qwen3NextForCausalLM modeling has pure-PyTorch
 # fallbacks (slower but works).
-cuda-extras = [
+#
+# IMPORTANT: flash-attn and causal-conv1d need torch at build time but don't
+# declare it in their build-system.requires. They are NOT included in [all]
+# or [train-cuda] — install them separately AFTER torch is in the environment:
+#   uv pip install attacklm[all]
+#   uv pip install attacklm[flash-attn]
+cuda-extras = []
+flash-attn = [
     "flash-attn>=2.7",                # requires CUDA dev tools to build
     "causal-conv1d>=1.4",             # CUDA-only
     "flash-linear-attention>=0.2",    # CUDA-only

attacklm-0.5.6/src/attacklm/__version__.py

@@ -0,0 +1 @@
+__version__ = "0.5.6"

attacklm-0.5.4/src/attacklm/__version__.py

@@ -1 +0,0 @@
-__version__ = "0.5.4"