Upload 12 files

.gitattributes

@@ -33,3 +33,9 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+wheel/python3.12/torch2.8.0/SpargeAttn/spas_sage_attn-0.1.0-cp312-cp312-linux_x86_64.whl filter=lfs diff=lfs merge=lfs -text
+wheel/python3.12/torch2.9.0/flash-attention/flash_attn-2.7.4-cp312-cp312-linux_x86_64.whl filter=lfs diff=lfs merge=lfs -text
+wheel/python3.12/torch2.9.0/flash-attention/flash_attn-2.8.2-cp312-cp312-linux_x86_64.whl filter=lfs diff=lfs merge=lfs -text
+wheel/python3.12/torch2.9.0/nunchaku/nunchaku-1.3.0.dev20260304+cu13.0torch2.9-cp312-cp312-linux_x86_64.whl filter=lfs diff=lfs merge=lfs -text
+wheel/python3.12/torch2.9.0/SageAttention/sageattention-2.2.0-cp312-cp312-linux_x86_64.whl filter=lfs diff=lfs merge=lfs -text
+wheel/python3.12/torch2.9.0/SpargeAttn/spas_sage_attn-0.1.0-cp312-cp312-linux_x86_64.whl filter=lfs diff=lfs merge=lfs -text

wheel/python3.12/torch2.8.0/SpargeAttn/setup.py

@@ -0,0 +1,232 @@
+"""
+Copyright (c) 2025 by SpargeAttn team.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+"""
+
+import os
+from pathlib import Path
+import subprocess
+from packaging.version import parse, Version
+from typing import List, Set
+import warnings
+
+from setuptools import setup, find_packages
+import torch
+from torch.utils.cpp_extension import BuildExtension, CUDAExtension, CUDA_HOME
+
+HAS_SM90 = False
+SAGE2PP_ENABLED = True
+
+def run_instantiations(src_dir: str):
+    base_path = Path(src_dir)
+    py_files = [
+        path for path in base_path.rglob('*.py')
+        if path.is_file()
+    ]
+
+    for py_file in py_files:
+        print(f"Running: {py_file}")
+        os.system(f"python {py_file}")
+
+def get_instantiations(src_dir: str):
+    # get all .cu files under src_dir
+    base_path = Path(src_dir)
+    return [
+        os.path.join(src_dir, str(path.relative_to(base_path)))
+        for path in base_path.rglob('*')
+        if path.is_file() and path.suffix == ".cu"
+    ]
+
+# Supported NVIDIA GPU architectures.
+SUPPORTED_ARCHS = {"8.0", "8.6", "8.7", "8.9", "9.0"}
+
+# Compiler flags.
+CXX_FLAGS = ["-g", "-O3", "-fopenmp", "-lgomp", "-std=c++17", "-DENABLE_BF16"]
+NVCC_FLAGS = [
+    "-O3",
+    "-std=c++17",
+    "-U__CUDA_NO_HALF_OPERATORS__",
+    "-U__CUDA_NO_HALF_CONVERSIONS__",
+    "--use_fast_math",
+    "--threads=8",
+    "-Xptxas=-v",
+    "-diag-suppress=174", # suppress the specific warning
+    # "-Xcompiler", "-include,cassert", # fix error occurs when compiling for SM90+ with newer CUDA toolkits RTX 40 系列是 SM89可以安全移除该选项
+]
+
+ABI = 1 if torch._C._GLIBCXX_USE_CXX11_ABI else 0
+CXX_FLAGS += [f"-D_GLIBCXX_USE_CXX11_ABI={ABI}"]
+NVCC_FLAGS += [f"-D_GLIBCXX_USE_CXX11_ABI={ABI}"]
+
+if CUDA_HOME is None:
+    raise RuntimeError(
+        "Cannot find CUDA_HOME. CUDA must be available to build the package.")
+
+def get_nvcc_cuda_version(cuda_dir: str) -> Version:
+    """Get the CUDA version from nvcc.
+
+    Adapted from https://github.com/NVIDIA/apex/blob/8b7a1ff183741dd8f9b87e7bafd04cfde99cea28/setup.py
+    """
+    nvcc_output = subprocess.check_output([cuda_dir + "/bin/nvcc", "-V"],
+                                          universal_newlines=True)
+    output = nvcc_output.split()
+    release_idx = output.index("release") + 1
+    nvcc_cuda_version = parse(output[release_idx].split(",")[0])
+    return nvcc_cuda_version
+
+def get_torch_arch_list() -> Set[str]:
+    # TORCH_CUDA_ARCH_LIST can have one or more architectures,
+    # e.g. "8.0" or "7.5,8.0,8.6+PTX". Here, the "8.6+PTX" option asks the
+    # compiler to additionally include PTX code that can be runtime-compiled
+    # and executed on the 8.6 or newer architectures. While the PTX code will
+    # not give the best performance on the newer architectures, it provides
+    # forward compatibility.
+    env_arch_list = os.environ.get("TORCH_CUDA_ARCH_LIST", None)
+    if env_arch_list is None:
+        return set()
+
+    # List are separated by ; or space.
+    torch_arch_list = set(env_arch_list.replace(" ", ";").split(";"))
+    if not torch_arch_list:
+        return set()
+
+    # Filter out the invalid architectures and print a warning.
+    valid_archs = SUPPORTED_ARCHS.union({s + "+PTX" for s in SUPPORTED_ARCHS})
+    arch_list = torch_arch_list.intersection(valid_archs)
+    # If none of the specified architectures are valid, raise an error.
+    if not arch_list:
+        raise RuntimeError(
+            "None of the CUDA architectures in `TORCH_CUDA_ARCH_LIST` env "
+            f"variable ({env_arch_list}) is supported. "
+            f"Supported CUDA architectures are: {valid_archs}.")
+    invalid_arch_list = torch_arch_list - valid_archs
+    if invalid_arch_list:
+        warnings.warn(
+            f"Unsupported CUDA architectures ({invalid_arch_list}) are "
+            "excluded from the `TORCH_CUDA_ARCH_LIST` env variable "
+            f"({env_arch_list}). Supported CUDA architectures are: "
+            f"{valid_archs}.")
+    return arch_list
+
+# First, check the TORCH_CUDA_ARCH_LIST environment variable.
+compute_capabilities = get_torch_arch_list()
+if not compute_capabilities:
+    # If TORCH_CUDA_ARCH_LIST is not defined or empty, target all available
+    # GPUs on the current machine.
+    device_count = torch.cuda.device_count()
+    for i in range(device_count):
+        major, minor = torch.cuda.get_device_capability(i)
+        if major < 8:
+            raise RuntimeError(
+                "GPUs with compute capability below 8.0 are not supported.")
+        compute_capabilities.add(f"{major}.{minor}")
+
+nvcc_cuda_version = get_nvcc_cuda_version(CUDA_HOME)
+if not compute_capabilities:
+    raise RuntimeError("No GPUs found. Please specify the target GPU architectures or build on a machine with GPUs.")
+
+# Validate the NVCC CUDA version.
+if nvcc_cuda_version < Version("12.0"):
+    raise RuntimeError("CUDA 12.0 or higher is required to build the package.")
+if nvcc_cuda_version < Version("12.4"):
+    if any(cc.startswith("8.9") for cc in compute_capabilities):
+        raise RuntimeError(
+            "CUDA 12.4 or higher is required for compute capability 8.9.")
+    if any(cc.startswith("9.0") for cc in compute_capabilities):
+        raise RuntimeError(
+            "CUDA 12.4 or higher is required for compute capability 9.0.")
+if nvcc_cuda_version < Version("12.8"):
+    warnings.warn("CUDA 12.8 or higher is required for Sage2++")
+    SAGE2PP_ENABLED = False
+
+# Add target compute capabilities to NVCC flags.
+for capability in compute_capabilities:
+    num = capability.replace(".", "")
+    if num == '90':
+        num = '90a'
+        HAS_SM90 = True
+        CXX_FLAGS += ["-DHAS_SM90"]
+    if num == '80' or num == '86' or num == '87':
+        SAGE2PP_ENABLED = False
+    
+    NVCC_FLAGS += ["-gencode", f"arch=compute_{num},code=sm_{num}"]
+    if capability.endswith("+PTX"):
+        NVCC_FLAGS += ["-gencode", f"arch=compute_{num},code=compute_{num}"]
+
+if SAGE2PP_ENABLED:
+    CXX_FLAGS += ["-DSAGE2PP_ENABLED"]
+
+ext_modules = []
+
+run_instantiations("csrc/qattn/instantiations_sm80")
+run_instantiations("csrc/qattn/instantiations_sm89")
+run_instantiations("csrc/qattn/instantiations_sm90")
+
+sources = [
+    "csrc/qattn/pybind.cpp",
+    "csrc/qattn/qk_int_sv_f16_cuda_sm80.cu",
+    "csrc/qattn/qk_int_sv_f8_cuda_sm89.cu",
+] + get_instantiations("csrc/qattn/instantiations_sm80") + get_instantiations("csrc/qattn/instantiations_sm89")
+
+if HAS_SM90:
+    sources += ["csrc/qattn/qk_int_sv_f8_cuda_sm90.cu", ]
+    sources += get_instantiations("csrc/qattn/instantiations_sm90")
+
+qattn_extension = CUDAExtension(
+    name="spas_sage_attn._qattn",
+    sources=sources,
+    extra_compile_args={
+        "cxx": CXX_FLAGS,
+        "nvcc": NVCC_FLAGS,
+    },
+    extra_link_args=['-lcuda'],
+)
+ext_modules.append(qattn_extension)
+
+fused_extension = CUDAExtension(
+    name="spas_sage_attn._fused",
+    sources=["csrc/fused/pybind.cpp", "csrc/fused/fused.cu"],
+    extra_compile_args={
+        "cxx": CXX_FLAGS,
+        "nvcc": NVCC_FLAGS,
+    },
+)
+ext_modules.append(fused_extension)
+
+setup(
+    name='spas_sage_attn', 
+    version='0.1.0',  
+    author='Jintao Zhang, Chendong Xiang, Haofeng Huang',  
+    author_email='jt-zhang6@gmail.com', 
+    packages=find_packages(),  
+    description='Accurate and efficient Sparse SageAttention.',  
+    long_description=open('README.md', encoding='utf-8').read(),  
+    long_description_content_type='text/markdown', 
+    url='https://github.com/thu-ml/SpargeAttn', 
+    license='BSD 3-Clause License', 
+    python_requires='>=3.9', 
+    classifiers=[  
+        'Development Status :: 3 - Alpha', 
+        'Intended Audience :: Developers',  
+        'Topic :: Software Development :: Libraries :: Python Modules',
+        'License :: OSI Approved :: BSD License',
+        'Programming Language :: Python :: 3', 
+        'Programming Language :: Python :: 3.9',
+        'Programming Language :: Python :: 3.10',
+        'Programming Language :: Python :: 3.11',
+        'Operating System :: OS Independent',
+    ],
+    ext_modules=ext_modules,
+    cmdclass={"build_ext": BuildExtension},
+)

wheel/python3.12/torch2.8.0/SpargeAttn/spas_sage_attn-0.1.0-cp312-cp312-linux_x86_64.whl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:111662b1c06509c57bc3677ebcd877cc63a7055094046c811964f101299e4da2
+size 10037220

wheel/python3.12/torch2.9.0/SageAttention/sageattention-2.2.0-cp312-cp312-linux_x86_64.whl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c9096be647f7dc146a0c662115063b5e75b93dbfaf1c2a680f801acf92b95679
+size 15616303

wheel/python3.12/torch2.9.0/SpargeAttn/setup.py

@@ -0,0 +1,232 @@
+"""
+Copyright (c) 2025 by SpargeAttn team.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+"""
+
+import os
+from pathlib import Path
+import subprocess
+from packaging.version import parse, Version
+from typing import List, Set
+import warnings
+
+from setuptools import setup, find_packages
+import torch
+from torch.utils.cpp_extension import BuildExtension, CUDAExtension, CUDA_HOME
+
+HAS_SM90 = False
+SAGE2PP_ENABLED = True
+
+def run_instantiations(src_dir: str):
+    base_path = Path(src_dir)
+    py_files = [
+        path for path in base_path.rglob('*.py')
+        if path.is_file()
+    ]
+
+    for py_file in py_files:
+        print(f"Running: {py_file}")
+        os.system(f"python {py_file}")
+
+def get_instantiations(src_dir: str):
+    # get all .cu files under src_dir
+    base_path = Path(src_dir)
+    return [
+        os.path.join(src_dir, str(path.relative_to(base_path)))
+        for path in base_path.rglob('*')
+        if path.is_file() and path.suffix == ".cu"
+    ]
+
+# Supported NVIDIA GPU architectures.
+SUPPORTED_ARCHS = {"8.0", "8.6", "8.7", "8.9", "9.0"}
+
+# Compiler flags.
+CXX_FLAGS = ["-g", "-O3", "-fopenmp", "-lgomp", "-std=c++17", "-DENABLE_BF16"]
+NVCC_FLAGS = [
+    "-O3",
+    "-std=c++17",
+    "-U__CUDA_NO_HALF_OPERATORS__",
+    "-U__CUDA_NO_HALF_CONVERSIONS__",
+    "--use_fast_math",
+    "--threads=8",
+    "-Xptxas=-v",
+    "-diag-suppress=174", # suppress the specific warning
+    # "-Xcompiler", "-include,cassert", # fix error occurs when compiling for SM90+ with newer CUDA toolkits RTX 40 系列是 SM89可以安全移除该选项
+]
+
+ABI = 1 if torch._C._GLIBCXX_USE_CXX11_ABI else 0
+CXX_FLAGS += [f"-D_GLIBCXX_USE_CXX11_ABI={ABI}"]
+NVCC_FLAGS += [f"-D_GLIBCXX_USE_CXX11_ABI={ABI}"]
+
+if CUDA_HOME is None:
+    raise RuntimeError(
+        "Cannot find CUDA_HOME. CUDA must be available to build the package.")
+
+def get_nvcc_cuda_version(cuda_dir: str) -> Version:
+    """Get the CUDA version from nvcc.
+
+    Adapted from https://github.com/NVIDIA/apex/blob/8b7a1ff183741dd8f9b87e7bafd04cfde99cea28/setup.py
+    """
+    nvcc_output = subprocess.check_output([cuda_dir + "/bin/nvcc", "-V"],
+                                          universal_newlines=True)
+    output = nvcc_output.split()
+    release_idx = output.index("release") + 1
+    nvcc_cuda_version = parse(output[release_idx].split(",")[0])
+    return nvcc_cuda_version
+
+def get_torch_arch_list() -> Set[str]:
+    # TORCH_CUDA_ARCH_LIST can have one or more architectures,
+    # e.g. "8.0" or "7.5,8.0,8.6+PTX". Here, the "8.6+PTX" option asks the
+    # compiler to additionally include PTX code that can be runtime-compiled
+    # and executed on the 8.6 or newer architectures. While the PTX code will
+    # not give the best performance on the newer architectures, it provides
+    # forward compatibility.
+    env_arch_list = os.environ.get("TORCH_CUDA_ARCH_LIST", None)
+    if env_arch_list is None:
+        return set()
+
+    # List are separated by ; or space.
+    torch_arch_list = set(env_arch_list.replace(" ", ";").split(";"))
+    if not torch_arch_list:
+        return set()
+
+    # Filter out the invalid architectures and print a warning.
+    valid_archs = SUPPORTED_ARCHS.union({s + "+PTX" for s in SUPPORTED_ARCHS})
+    arch_list = torch_arch_list.intersection(valid_archs)
+    # If none of the specified architectures are valid, raise an error.
+    if not arch_list:
+        raise RuntimeError(
+            "None of the CUDA architectures in `TORCH_CUDA_ARCH_LIST` env "
+            f"variable ({env_arch_list}) is supported. "
+            f"Supported CUDA architectures are: {valid_archs}.")
+    invalid_arch_list = torch_arch_list - valid_archs
+    if invalid_arch_list:
+        warnings.warn(
+            f"Unsupported CUDA architectures ({invalid_arch_list}) are "
+            "excluded from the `TORCH_CUDA_ARCH_LIST` env variable "
+            f"({env_arch_list}). Supported CUDA architectures are: "
+            f"{valid_archs}.")
+    return arch_list
+
+# First, check the TORCH_CUDA_ARCH_LIST environment variable.
+compute_capabilities = get_torch_arch_list()
+if not compute_capabilities:
+    # If TORCH_CUDA_ARCH_LIST is not defined or empty, target all available
+    # GPUs on the current machine.
+    device_count = torch.cuda.device_count()
+    for i in range(device_count):
+        major, minor = torch.cuda.get_device_capability(i)
+        if major < 8:
+            raise RuntimeError(
+                "GPUs with compute capability below 8.0 are not supported.")
+        compute_capabilities.add(f"{major}.{minor}")
+
+nvcc_cuda_version = get_nvcc_cuda_version(CUDA_HOME)
+if not compute_capabilities:
+    raise RuntimeError("No GPUs found. Please specify the target GPU architectures or build on a machine with GPUs.")
+
+# Validate the NVCC CUDA version.
+if nvcc_cuda_version < Version("12.0"):
+    raise RuntimeError("CUDA 12.0 or higher is required to build the package.")
+if nvcc_cuda_version < Version("12.4"):
+    if any(cc.startswith("8.9") for cc in compute_capabilities):
+        raise RuntimeError(
+            "CUDA 12.4 or higher is required for compute capability 8.9.")
+    if any(cc.startswith("9.0") for cc in compute_capabilities):
+        raise RuntimeError(
+            "CUDA 12.4 or higher is required for compute capability 9.0.")
+if nvcc_cuda_version < Version("12.8"):
+    warnings.warn("CUDA 12.8 or higher is required for Sage2++")
+    SAGE2PP_ENABLED = False
+
+# Add target compute capabilities to NVCC flags.
+for capability in compute_capabilities:
+    num = capability.replace(".", "")
+    if num == '90':
+        num = '90a'
+        HAS_SM90 = True
+        CXX_FLAGS += ["-DHAS_SM90"]
+    if num == '80' or num == '86' or num == '87':
+        SAGE2PP_ENABLED = False
+    
+    NVCC_FLAGS += ["-gencode", f"arch=compute_{num},code=sm_{num}"]
+    if capability.endswith("+PTX"):
+        NVCC_FLAGS += ["-gencode", f"arch=compute_{num},code=compute_{num}"]
+
+if SAGE2PP_ENABLED:
+    CXX_FLAGS += ["-DSAGE2PP_ENABLED"]
+
+ext_modules = []
+
+run_instantiations("csrc/qattn/instantiations_sm80")
+run_instantiations("csrc/qattn/instantiations_sm89")
+run_instantiations("csrc/qattn/instantiations_sm90")
+
+sources = [
+    "csrc/qattn/pybind.cpp",
+    "csrc/qattn/qk_int_sv_f16_cuda_sm80.cu",
+    "csrc/qattn/qk_int_sv_f8_cuda_sm89.cu",
+] + get_instantiations("csrc/qattn/instantiations_sm80") + get_instantiations("csrc/qattn/instantiations_sm89")
+
+if HAS_SM90:
+    sources += ["csrc/qattn/qk_int_sv_f8_cuda_sm90.cu", ]
+    sources += get_instantiations("csrc/qattn/instantiations_sm90")
+
+qattn_extension = CUDAExtension(
+    name="spas_sage_attn._qattn",
+    sources=sources,
+    extra_compile_args={
+        "cxx": CXX_FLAGS,
+        "nvcc": NVCC_FLAGS,
+    },
+    extra_link_args=['-lcuda'],
+)
+ext_modules.append(qattn_extension)
+
+fused_extension = CUDAExtension(
+    name="spas_sage_attn._fused",
+    sources=["csrc/fused/pybind.cpp", "csrc/fused/fused.cu"],
+    extra_compile_args={
+        "cxx": CXX_FLAGS,
+        "nvcc": NVCC_FLAGS,
+    },
+)
+ext_modules.append(fused_extension)
+
+setup(
+    name='spas_sage_attn', 
+    version='0.1.0',  
+    author='Jintao Zhang, Chendong Xiang, Haofeng Huang',  
+    author_email='jt-zhang6@gmail.com', 
+    packages=find_packages(),  
+    description='Accurate and efficient Sparse SageAttention.',  
+    long_description=open('README.md', encoding='utf-8').read(),  
+    long_description_content_type='text/markdown', 
+    url='https://github.com/thu-ml/SpargeAttn', 
+    license='BSD 3-Clause License', 
+    python_requires='>=3.9', 
+    classifiers=[  
+        'Development Status :: 3 - Alpha', 
+        'Intended Audience :: Developers',  
+        'Topic :: Software Development :: Libraries :: Python Modules',
+        'License :: OSI Approved :: BSD License',
+        'Programming Language :: Python :: 3', 
+        'Programming Language :: Python :: 3.9',
+        'Programming Language :: Python :: 3.10',
+        'Programming Language :: Python :: 3.11',
+        'Operating System :: OS Independent',
+    ],
+    ext_modules=ext_modules,
+    cmdclass={"build_ext": BuildExtension},
+)

wheel/python3.12/torch2.9.0/SpargeAttn/spas_sage_attn-0.1.0-cp312-cp312-linux_x86_64.whl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0cb2be8204cbe67a9b61f249d6e26986ddaeeb43512cce90aea5be7ad91db97f
+size 10345136

wheel/python3.12/torch2.9.0/flash-attention/flash_attn-2.7.4-cp312-cp312-linux_x86_64.whl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:10d88c0e49eb139a74bf0cc6ed71ca09d7bf82f6ff27a2cfb80b516a8afd7e9b
+size 87973409

wheel/python3.12/torch2.9.0/flash-attention/flash_attn-2.8.2-cp312-cp312-linux_x86_64.whl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:85cc068cbd1d0b2762833a61d261e7938c5c531744b058ffd6a1abb0c370c46d
+size 56739530

wheel/python3.12/torch2.9.0/nunchaku/nunchaku-1.3.0.dev20260304+cu13.0torch2.9-cp312-cp312-linux_x86_64.whl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d27d235797f821b2674164833da09ce2f8ccd1dc1f373e5270716f8670b43bd4
+size 47571075