mlwheels 0.1.0__tar.gz

mlwheels-0.1.0/PKG-INFO

@@ -0,0 +1,84 @@
+Metadata-Version: 2.4
+Name: mlwheels
+Version: 0.1.0
+Summary: Auto-detect and install pre-built wheels for Flash Attention & vLLM
+License: MIT
+Project-URL: Homepage, https://rs545837.github.io/Flash-Attn-wheels/
+Project-URL: Repository, https://github.com/rs545837/Flash-Attn-wheels
+Keywords: flash-attention,vllm,cuda,pytorch,wheels
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.8
+Classifier: Programming Language :: Python :: 3.9
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.8
+Description-Content-Type: text/markdown
+
+# Pre-Built Wheels
+
+Pre-built wheels for Flash Attention & vLLM. Skip the compilation.
+
+## Quick Install
+
+Auto-detect your environment and install the right wheel:
+
+```bash
+pip install mlwheels
+
+# Install Flash Attention
+mlwheels flash-attn
+
+# Install vLLM
+mlwheels vllm
+
+# Just detect environment (no install)
+mlwheels --detect
+```
+
+## Features
+
+- Search and filter by CUDA, Python, PyTorch, and Platform
+- One-click copy for `pip` and `uv` install commands
+- Direct download links
+- Flash Attention 2 & 3 support
+- vLLM wheels for multiple CUDA versions
+
+## Supported Configurations
+
+### Flash Attention 2
+- CUDA: 11.8, 12.1, 12.2, 12.3, 12.4, 12.6
+- PyTorch: 2.0 - 2.10
+- Python: 3.8 - 3.12
+- Platforms: Linux x86_64, Linux ARM64, Windows
+
+### Flash Attention 3
+- CUDA: 12.6, 12.8, 12.9, 13.0
+- PyTorch: 2.8 - 2.10
+- Python: 3.10 - 3.12
+- Platforms: Linux x86_64, Linux ARM64, Windows
+
+### vLLM
+- CUDA: 11.8, 12.1, 12.4, 12.6, 12.8, 12.9, 13.0, CPU
+- Python: 3.8+
+- Platforms: Linux x86_64, Linux ARM64
+
+## Sources
+
+**Flash Attention**
+- [flashattn.dev](https://flashattn.dev/)
+- [Flash Attention 3 Wheels](https://windreamer.github.io/flash-attention3-wheels/)
+- [mjun0812/flash-attention-prebuild-wheels](https://github.com/mjun0812/flash-attention-prebuild-wheels)
+- [Dao-AILab/flash-attention](https://github.com/Dao-AILab/flash-attention)
+
+**vLLM**
+- [vLLM GitHub Releases](https://github.com/vllm-project/vllm/releases)
+- [vLLM Documentation](https://docs.vllm.ai/)
+
+## License
+
+MIT

mlwheels-0.1.0/README.md

@@ -0,0 +1,63 @@
+# Pre-Built Wheels
+
+Pre-built wheels for Flash Attention & vLLM. Skip the compilation.
+
+## Quick Install
+
+Auto-detect your environment and install the right wheel:
+
+```bash
+pip install mlwheels
+
+# Install Flash Attention
+mlwheels flash-attn
+
+# Install vLLM
+mlwheels vllm
+
+# Just detect environment (no install)
+mlwheels --detect
+```
+
+## Features
+
+- Search and filter by CUDA, Python, PyTorch, and Platform
+- One-click copy for `pip` and `uv` install commands
+- Direct download links
+- Flash Attention 2 & 3 support
+- vLLM wheels for multiple CUDA versions
+
+## Supported Configurations
+
+### Flash Attention 2
+- CUDA: 11.8, 12.1, 12.2, 12.3, 12.4, 12.6
+- PyTorch: 2.0 - 2.10
+- Python: 3.8 - 3.12
+- Platforms: Linux x86_64, Linux ARM64, Windows
+
+### Flash Attention 3
+- CUDA: 12.6, 12.8, 12.9, 13.0
+- PyTorch: 2.8 - 2.10
+- Python: 3.10 - 3.12
+- Platforms: Linux x86_64, Linux ARM64, Windows
+
+### vLLM
+- CUDA: 11.8, 12.1, 12.4, 12.6, 12.8, 12.9, 13.0, CPU
+- Python: 3.8+
+- Platforms: Linux x86_64, Linux ARM64
+
+## Sources
+
+**Flash Attention**
+- [flashattn.dev](https://flashattn.dev/)
+- [Flash Attention 3 Wheels](https://windreamer.github.io/flash-attention3-wheels/)
+- [mjun0812/flash-attention-prebuild-wheels](https://github.com/mjun0812/flash-attention-prebuild-wheels)
+- [Dao-AILab/flash-attention](https://github.com/Dao-AILab/flash-attention)
+
+**vLLM**
+- [vLLM GitHub Releases](https://github.com/vllm-project/vllm/releases)
+- [vLLM Documentation](https://docs.vllm.ai/)
+
+## License
+
+MIT

mlwheels-0.1.0/mlwheels/__init__.py

@@ -0,0 +1,7 @@
+"""ML Wheels - Auto-detect and install pre-built wheels for Flash Attention & vLLM."""
+
+__version__ = "0.1.0"
+
+from .detector import detect_environment, get_wheel_url, install_wheel
+
+__all__ = ["detect_environment", "get_wheel_url", "install_wheel"]

mlwheels-0.1.0/mlwheels/cli.py

@@ -0,0 +1,81 @@
+"""CLI for mlwheels."""
+
+import argparse
+import sys
+from .detector import detect_environment, get_wheel_url, install_wheel, get_platform
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Auto-detect and install pre-built wheels for Flash Attention & vLLM"
+    )
+    parser.add_argument(
+        "library",
+        nargs="?",
+        choices=["flash-attn", "vllm"],
+        help="Library to install (flash-attn or vllm)"
+    )
+    parser.add_argument(
+        "--detect", "-d",
+        action="store_true",
+        help="Only detect environment, don't install"
+    )
+    parser.add_argument(
+        "--dry-run", "-n",
+        action="store_true",
+        help="Show what would be installed without installing"
+    )
+    parser.add_argument(
+        "--url", "-u",
+        action="store_true",
+        help="Only print the wheel URL"
+    )
+
+    args = parser.parse_args()
+
+    env = detect_environment()
+    platform = get_platform()
+
+    if args.detect or not args.library:
+        print("Detected environment:")
+        print(f"  Python:   {env['python']}")
+        print(f"  PyTorch:  {env['torch'] or 'not installed'}")
+        print(f"  CUDA:     {env['cuda'] or 'not detected'}")
+        print(f"  Platform: {platform or 'unknown'}")
+
+        if not args.library:
+            print("\nRecommended wheels:")
+            for lib in ["flash-attn", "vllm"]:
+                wheel = get_wheel_url(lib, env)
+                if wheel:
+                    print(f"\n  {lib} {wheel['version']}:")
+                    print(f"    pip install {wheel['url']}")
+                else:
+                    print(f"\n  {lib}: no matching wheel found")
+
+            print("\nTo install, run:")
+            print("  mlwheels flash-attn")
+            print("  mlwheels vllm")
+            return 0
+
+    if args.library:
+        wheel = get_wheel_url(args.library, env)
+
+        if args.url:
+            if wheel:
+                print(wheel["url"])
+                return 0
+            else:
+                print(f"No matching wheel found", file=sys.stderr)
+                return 1
+
+        if args.dry_run:
+            return 0 if install_wheel(args.library, dry_run=True) else 1
+
+        return 0 if install_wheel(args.library) else 1
+
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

mlwheels-0.1.0/mlwheels/detector.py

@@ -0,0 +1,207 @@
+"""Environment detection and wheel matching."""
+
+import subprocess
+import sys
+import re
+
+
+def get_python_version():
+    """Get Python version as string (e.g., '3.10')."""
+    return f"{sys.version_info.major}.{sys.version_info.minor}"
+
+
+def get_torch_version():
+    """Get PyTorch version if installed."""
+    try:
+        import torch
+        # Extract major.minor (e.g., "2.5" from "2.5.1+cu124")
+        version = torch.__version__
+        match = re.match(r"(\d+\.\d+)", version)
+        return match.group(1) if match else None
+    except ImportError:
+        return None
+
+
+def get_cuda_version():
+    """Get CUDA version from PyTorch or nvidia-smi."""
+    # Try PyTorch first
+    try:
+        import torch
+        if torch.cuda.is_available():
+            # Get CUDA version from PyTorch build
+            cuda_version = torch.version.cuda
+            if cuda_version:
+                # Extract major.minor (e.g., "12.4" from "12.4")
+                match = re.match(r"(\d+\.\d+)", cuda_version)
+                return match.group(1) if match else None
+    except ImportError:
+        pass
+
+    # Fallback to nvidia-smi
+    try:
+        result = subprocess.run(
+            ["nvidia-smi", "--query-gpu=driver_version", "--format=csv,noheader"],
+            capture_output=True, text=True, timeout=5
+        )
+        if result.returncode == 0:
+            # nvidia-smi doesn't directly give CUDA version, try nvcc
+            nvcc_result = subprocess.run(
+                ["nvcc", "--version"],
+                capture_output=True, text=True, timeout=5
+            )
+            if nvcc_result.returncode == 0:
+                match = re.search(r"release (\d+\.\d+)", nvcc_result.stdout)
+                if match:
+                    return match.group(1)
+    except (FileNotFoundError, subprocess.TimeoutExpired):
+        pass
+
+    return None
+
+
+def detect_environment():
+    """Detect current environment."""
+    return {
+        "python": get_python_version(),
+        "torch": get_torch_version(),
+        "cuda": get_cuda_version(),
+    }
+
+
+def get_platform():
+    """Get platform identifier."""
+    import platform
+    system = platform.system().lower()
+    machine = platform.machine().lower()
+
+    if system == "linux":
+        if machine in ("x86_64", "amd64"):
+            return "linux_x86_64"
+        elif machine in ("aarch64", "arm64"):
+            return "linux_aarch64"
+    elif system == "windows":
+        return "win_amd64"
+    elif system == "darwin":
+        return "macos"
+
+    return None
+
+
+# Base URLs for wheels
+WHEEL_SOURCES = {
+    "flash-attn": {
+        "base": "https://github.com/Dao-AILab/flash-attention/releases/download",
+        "alt": "https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download",
+    },
+    "vllm": {
+        "base": "https://github.com/vllm-project/vllm/releases/download",
+    }
+}
+
+# Known wheel configurations (subset - full list at flashattn.dev)
+FLASH_ATTN_WHEELS = {
+    # Format: (cuda, torch, python, platform): version
+    ("12.4", "2.5", "3.10", "linux_x86_64"): "2.7.4",
+    ("12.4", "2.5", "3.11", "linux_x86_64"): "2.7.4",
+    ("12.4", "2.5", "3.12", "linux_x86_64"): "2.7.4",
+    ("12.6", "2.6", "3.10", "linux_x86_64"): "2.7.4",
+    ("12.6", "2.6", "3.11", "linux_x86_64"): "2.7.4",
+    ("12.6", "2.6", "3.12", "linux_x86_64"): "2.7.4",
+    ("12.1", "2.4", "3.10", "linux_x86_64"): "2.6.3",
+    ("12.1", "2.4", "3.11", "linux_x86_64"): "2.6.3",
+    ("11.8", "2.3", "3.10", "linux_x86_64"): "2.6.3",
+    ("11.8", "2.3", "3.11", "linux_x86_64"): "2.6.3",
+}
+
+VLLM_WHEELS = {
+    # Format: (cuda, python, platform): (version, url)
+    ("12.6", "3.10", "linux_x86_64"): ("0.15.0", "https://github.com/vllm-project/vllm/releases/download/v0.15.0/vllm-0.15.0-cp38-abi3-manylinux1_x86_64.whl"),
+    ("12.6", "3.11", "linux_x86_64"): ("0.15.0", "https://github.com/vllm-project/vllm/releases/download/v0.15.0/vllm-0.15.0-cp38-abi3-manylinux1_x86_64.whl"),
+    ("12.6", "3.12", "linux_x86_64"): ("0.15.0", "https://github.com/vllm-project/vllm/releases/download/v0.15.0/vllm-0.15.0-cp38-abi3-manylinux1_x86_64.whl"),
+    ("12.4", "3.10", "linux_x86_64"): ("0.15.0", "https://github.com/vllm-project/vllm/releases/download/v0.15.0/vllm-0.15.0+cu124-cp38-abi3-manylinux1_x86_64.whl"),
+    ("12.4", "3.11", "linux_x86_64"): ("0.15.0", "https://github.com/vllm-project/vllm/releases/download/v0.15.0/vllm-0.15.0+cu124-cp38-abi3-manylinux1_x86_64.whl"),
+    ("12.1", "3.10", "linux_x86_64"): ("0.15.0", "https://github.com/vllm-project/vllm/releases/download/v0.15.0/vllm-0.15.0+cu121-cp38-abi3-manylinux1_x86_64.whl"),
+    ("12.1", "3.11", "linux_x86_64"): ("0.15.0", "https://github.com/vllm-project/vllm/releases/download/v0.15.0/vllm-0.15.0+cu121-cp38-abi3-manylinux1_x86_64.whl"),
+    ("11.8", "3.10", "linux_x86_64"): ("0.15.0", "https://github.com/vllm-project/vllm/releases/download/v0.15.0/vllm-0.15.0+cu118-cp38-abi3-manylinux1_x86_64.whl"),
+}
+
+
+def find_closest_cuda(target_cuda, available_versions):
+    """Find the closest CUDA version that's <= target."""
+    if not target_cuda:
+        return None
+
+    target = float(target_cuda)
+    available = sorted([float(v) for v in available_versions], reverse=True)
+
+    for v in available:
+        if v <= target:
+            return str(v) if v == int(v) else f"{v:.1f}"
+
+    return None
+
+
+def get_wheel_url(library="flash-attn", env=None):
+    """Get the best matching wheel URL for the current environment."""
+    if env is None:
+        env = detect_environment()
+
+    platform = get_platform()
+    python = env.get("python")
+    torch = env.get("torch")
+    cuda = env.get("cuda")
+
+    if library == "vllm":
+        # Find matching vLLM wheel
+        available_cuda = list(set(k[0] for k in VLLM_WHEELS.keys()))
+        matched_cuda = find_closest_cuda(cuda, available_cuda) or "12.6"
+
+        for py in [python, "3.11", "3.10"]:
+            key = (matched_cuda, py, platform)
+            if key in VLLM_WHEELS:
+                version, url = VLLM_WHEELS[key]
+                return {"url": url, "version": version, "cuda": matched_cuda, "python": py}
+
+        return None
+
+    else:  # flash-attn
+        available_cuda = list(set(k[0] for k in FLASH_ATTN_WHEELS.keys()))
+        matched_cuda = find_closest_cuda(cuda, available_cuda)
+
+        if not matched_cuda or not torch:
+            return None
+
+        for py in [python, "3.11", "3.10"]:
+            key = (matched_cuda, torch, py, platform)
+            if key in FLASH_ATTN_WHEELS:
+                version = FLASH_ATTN_WHEELS[key]
+                # Construct URL
+                url = f"https://github.com/Dao-AILab/flash-attention/releases/download/v{version}/flash_attn-{version}+cu{matched_cuda.replace('.', '')}torch{torch}-cp{py.replace('.', '')}-cp{py.replace('.', '')}-{platform}.whl"
+                return {"url": url, "version": version, "cuda": matched_cuda, "torch": torch, "python": py}
+
+        return None
+
+
+def install_wheel(library="flash-attn", dry_run=False):
+    """Install the matching wheel."""
+    env = detect_environment()
+    wheel = get_wheel_url(library, env)
+
+    if not wheel:
+        print(f"No matching {library} wheel found for your environment:")
+        print(f"  Python: {env.get('python')}")
+        print(f"  PyTorch: {env.get('torch') or 'not installed'}")
+        print(f"  CUDA: {env.get('cuda') or 'not detected'}")
+        print(f"\nVisit https://rs545837.github.io/Flash-Attn-wheels/ to find a compatible wheel.")
+        return False
+
+    cmd = f"pip install {wheel['url']}"
+
+    if dry_run:
+        print(f"Would install {library} {wheel['version']}:")
+        print(f"  {cmd}")
+        return True
+
+    print(f"Installing {library} {wheel['version']}...")
+    result = subprocess.run(cmd.split(), capture_output=False)
+    return result.returncode == 0

mlwheels-0.1.0/mlwheels.egg-info/PKG-INFO

@@ -0,0 +1,84 @@
+Metadata-Version: 2.4
+Name: mlwheels
+Version: 0.1.0
+Summary: Auto-detect and install pre-built wheels for Flash Attention & vLLM
+License: MIT
+Project-URL: Homepage, https://rs545837.github.io/Flash-Attn-wheels/
+Project-URL: Repository, https://github.com/rs545837/Flash-Attn-wheels
+Keywords: flash-attention,vllm,cuda,pytorch,wheels
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.8
+Classifier: Programming Language :: Python :: 3.9
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.8
+Description-Content-Type: text/markdown
+
+# Pre-Built Wheels
+
+Pre-built wheels for Flash Attention & vLLM. Skip the compilation.
+
+## Quick Install
+
+Auto-detect your environment and install the right wheel:
+
+```bash
+pip install mlwheels
+
+# Install Flash Attention
+mlwheels flash-attn
+
+# Install vLLM
+mlwheels vllm
+
+# Just detect environment (no install)
+mlwheels --detect
+```
+
+## Features
+
+- Search and filter by CUDA, Python, PyTorch, and Platform
+- One-click copy for `pip` and `uv` install commands
+- Direct download links
+- Flash Attention 2 & 3 support
+- vLLM wheels for multiple CUDA versions
+
+## Supported Configurations
+
+### Flash Attention 2
+- CUDA: 11.8, 12.1, 12.2, 12.3, 12.4, 12.6
+- PyTorch: 2.0 - 2.10
+- Python: 3.8 - 3.12
+- Platforms: Linux x86_64, Linux ARM64, Windows
+
+### Flash Attention 3
+- CUDA: 12.6, 12.8, 12.9, 13.0
+- PyTorch: 2.8 - 2.10
+- Python: 3.10 - 3.12
+- Platforms: Linux x86_64, Linux ARM64, Windows
+
+### vLLM
+- CUDA: 11.8, 12.1, 12.4, 12.6, 12.8, 12.9, 13.0, CPU
+- Python: 3.8+
+- Platforms: Linux x86_64, Linux ARM64
+
+## Sources
+
+**Flash Attention**
+- [flashattn.dev](https://flashattn.dev/)
+- [Flash Attention 3 Wheels](https://windreamer.github.io/flash-attention3-wheels/)
+- [mjun0812/flash-attention-prebuild-wheels](https://github.com/mjun0812/flash-attention-prebuild-wheels)
+- [Dao-AILab/flash-attention](https://github.com/Dao-AILab/flash-attention)
+
+**vLLM**
+- [vLLM GitHub Releases](https://github.com/vllm-project/vllm/releases)
+- [vLLM Documentation](https://docs.vllm.ai/)
+
+## License
+
+MIT

mlwheels-0.1.0/mlwheels.egg-info/SOURCES.txt

@@ -0,0 +1,10 @@
+README.md
+pyproject.toml
+mlwheels/__init__.py
+mlwheels/cli.py
+mlwheels/detector.py
+mlwheels.egg-info/PKG-INFO
+mlwheels.egg-info/SOURCES.txt
+mlwheels.egg-info/dependency_links.txt
+mlwheels.egg-info/entry_points.txt
+mlwheels.egg-info/top_level.txt

mlwheels-0.1.0/mlwheels.egg-info/dependency_links.txt

@@ -0,0 +1 @@
+

mlwheels-0.1.0/mlwheels.egg-info/entry_points.txt

@@ -0,0 +1,2 @@
+[console_scripts]
+mlwheels = mlwheels.cli:main

mlwheels-0.1.0/mlwheels.egg-info/top_level.txt

@@ -0,0 +1 @@
+mlwheels

mlwheels-0.1.0/pyproject.toml

@@ -0,0 +1,34 @@
+[build-system]
+requires = ["setuptools>=61.0"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "mlwheels"
+version = "0.1.0"
+description = "Auto-detect and install pre-built wheels for Flash Attention & vLLM"
+readme = "README.md"
+license = {text = "MIT"}
+requires-python = ">=3.8"
+classifiers = [
+    "Development Status :: 4 - Beta",
+    "Intended Audience :: Developers",
+    "License :: OSI Approved :: MIT License",
+    "Programming Language :: Python :: 3",
+    "Programming Language :: Python :: 3.8",
+    "Programming Language :: Python :: 3.9",
+    "Programming Language :: Python :: 3.10",
+    "Programming Language :: Python :: 3.11",
+    "Programming Language :: Python :: 3.12",
+    "Topic :: Scientific/Engineering :: Artificial Intelligence",
+]
+keywords = ["flash-attention", "vllm", "cuda", "pytorch", "wheels"]
+
+[project.scripts]
+mlwheels = "mlwheels.cli:main"
+
+[project.urls]
+Homepage = "https://rs545837.github.io/Flash-Attn-wheels/"
+Repository = "https://github.com/rs545837/Flash-Attn-wheels"
+
+[tool.setuptools.packages.find]
+include = ["mlwheels*"]

mlwheels-0.1.0/setup.cfg

@@ -0,0 +1,4 @@
+[egg_info]
+tag_build = 
+tag_date = 0
+