Compilation error

[smokey-afk]

system: nightly pytorch 2.9+cu128 RTX 5090

C:/Users/rst22/.venv/Lib/site-packages/torch/include\torch/csrc/dynamo/compiled_autograd.h(1134): error C2872: 'std': ambiguous symbol
C:/Program Files/Microsoft Visual Studio/2022/Community/VC/Tools/MSVC/14.44.35207/include\valarray(20): note: could be 'std'
C:/Users/rst22/.venv/Lib/site-packages/torch/include\torch/csrc/dynamo/compiled_autograd.h(1134): note: or 'std'
C:/Users/rst22/.venv/Lib/site-packages/torch/include\torch/csrc/dynamo/compiled_autograd.h(1134): note: the template instantiation context (the oldest one first) is
C:/Users/rst22/.venv/Lib/site-packages/torch/include\torch/csrc/dynamo/compiled_autograd.h(1181): note: see reference to class template instantiation 'torch::dynamo::autograd::IValuePacker<__int64>' being compiled
C:/Users/rst22/.venv/Lib/site-packages/torch/include\torch/csrc/dynamo/compiled_autograd.h(1108): note: while compiling class template member function 'c10::TypePtr torch::dynamo::autograd::IValuePacker<__int64>::packed_type(void)'

  C:/Users/rst22/.venv/Lib/site-packages/torch/include\torch/csrc/dynamo/compiled_autograd.h(1181): note: see the first reference to 'torch::dynamo::autograd::IValuePacker<__int64>::packed_type' in 'torch::dynamo::autograd::IValuePacker<unsigned __int64>::packed_type'

  ninja: build stopped: subcommand failed.
  Traceback (most recent call last):
    File "C:\Users\rst22\.venv\Lib\site-packages\torch\utils\cpp_extension.py", line 2595, in _run_ninja_build
      subprocess.run(
    File "C:\Program Files\Python312\Lib\subprocess.py", line 571, in run
      raise CalledProcessError(retcode, process.args,
  subprocess.CalledProcessError: Command '['ninja', '-v']' returned non-zero exit status 1.

DISCLAIMER: I know sod all about python or coding I was just troubleshooting with google gemini and it said:

error C2872: 'std': ambiguous symbol
This is a C++ compiler error. It means that the compiler found two different definitions for the symbol 'std'. The compiler can't decide which one to use, so it stops and reports an error.

The error message also tells us exactly where the conflict is:

One definition of std is coming from a standard C++ header file in your Visual Studio installation (valarray).

The other definition is coming from a PyTorch header file (compiled_autograd.h).

This is a compatibility issue between PyTorch and the C++ standard library headers in your Visual Studio installation. It's a bug in the SageAttention source code, not in your system configuration

[rugabunda]

Same error here. was working fine two weeks ago with 628a89f

[rugabunda]

Git checkout to the aforementioned did not fix this problem; thus possibly a compatibility error with nightly pytorch builds...

[rugabunda]

This combo with torch seems to fix the problem, and works with the nightly build of sage attention

pip install torch==2.9.0.dev20250716+cu128 torchvision==0.24.0.dev20250717+cu128 torchaudio==2.8.0.dev20250717+cu128 --index-url https://download.pytorch.org/whl/nightly/cu128 --no-cache-dir --force-reinstall

[gitgud5000]

is combo with torch seems to fix the problem, and works with the nightly build of sage attention

pip install torch==2.9.0.dev20250716+cu128 torchvision==0.24.0.dev20250717+cu128 torchaudio==2.8.0.dev20250717+cu128 --index-url https://download.pytorch.org/whl/nightly/cu128 --no-cache-dir --force-reinstall

Not the case for Debian and sm_120 😔

doesn't compile with stable torch==2.8 & cuda==128 either.
if the logs can be of any help, I can share them upon request.

edit:

this worked — compiling only for one arch (sm_120) worked.
i was trying to compile for sm_89, sm_90, and sm_120 all in one build — that was also an issue.

pip install torch==2.9.0.dev20250716+cu128 torchvision==0.24.0.dev20250716+cu128 torchaudio==2.8.0.dev20250716+cu128 --index-url https://download.pytorch.org/whl/nightly/cu128 --no-cache-dir --force-reinstall

[rugabunda]

this worked — compiling only for one arch (sm_120) worked. i was trying to compile for sm_89, sm_90, and sm_120 all in one build — that was also an issue.

pip install torch==2.9.0.dev20250716+cu128 torchvision==0.24.0.dev20250716+cu128 torchaudio==2.8.0.dev20250716+cu128 --index-url https://download.pytorch.org/whl/nightly/cu128 --no-cache-dir --force-reinstall

Tested, same date works for cuda129 also, problem seems to be cuda version agnostic, rather date specific

pip install --pre torch==2.9.0.dev20250716+cu129 torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu129

dev20250 latest fails
dev20250725+cu129 fails
dev20250720+cu129 fails
dev20250718+cu129 fails
dev20250717+cu129 fails
dev20250716+cu129 works with sageattention-2.2.0

[joojie]

Thanks you wonderful human - you pulled me outta the mouth of madness with that lifeline. crazy because it was working - we probably all have a "it was working fine and then.." story. I compiled SageAttention3 but without a logic node to start/stop/swap between attenuators its a fierce step down. Btw if anyone has a workflow that can do that sort of switching of attention models puhhhhhlease share it with me.
it takes a village. oh and let me know if this looks like a good result for a 5090 laptop (curiously missing sm_120. the code hones it down to etiher 80/89/or fused in the end right? right?...:

Finished generating code
copying build\lib.win-amd64-cpython-312\sageattention\_qattn_sm80.cp312-win_amd64.pyd -> sageattention
copying build\lib.win-amd64-cpython-312\sageattention\_qattn_sm89.cp312-win_amd64.pyd -> sageattention
copying build\lib.win-amd64-cpython-312\sageattention\_fused.cp312-win_amd64.pyd -> sageattention

[rugabunda]

FYI this no longer working, latest torchvision/audio breaks it:

pip install --pre torch==2.9.0.dev20250716+cu129 torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu129

For 128 Use:

pip install torch==2.9.0.dev20250716+cu128 torchvision==0.24.0.dev20250717+cu128 torchaudio==2.8.0.dev20250717+cu128 --index-url https://download.pytorch.org/whl/nightly/cu128

or 129

pip install torch==2.9.0.dev20250716+cu129 torchvision==0.24.0.dev20250717+cu129 torchaudio==2.8.0.dev20250717+cu129 --index-url https://download.pytorch.org/whl/nightly/cu129

[joojie]

thanks for the info - I really want to generate with cu130 but it breaks everything it touches rn..

btw torch 3.0.1+cu130 nightly are available for Linux only for the time being.

[rugabunda]

Update: See my better suggestion two posts below

Above no longer works; Old working builds were truncated; PyTorch purges old nightly wheels after ~3 months, so they're gone from the index.

The official thu-ml/SageAttention lacks Windows wheels and requires source builds, which fail here. However, the community-maintained fork woct0rdho/SageAttention provides stable, pre-built .whl files for Windows (AMD64, Python 3.12 via ABI3 compatibility) that work with PyTorch 2.9.0 stable. These use SageAttention 2.2.0.post4 (latest as of Oct 16, 2025), support CUDA >=12.8 (including cu128/cu129), and avoid the header bug entirely. It also needs triton-windows (another wheel from the same maintainer) for the Triton backend. https://github.com/woct0rdho/triton-windows/releases

pip install torch==2.9.0 torchvision==0.24.0 torchaudio==2.9.0 --index-url https://download.pytorch.org/whl/cu128

Go to https://github.com/woct0rdho/SageAttention/releases/tag/v2.2.0-windows.post4
Download the wheel tagged +torch2.9.0andhigher (e.g., SageAttention-2.2.0.post4+torch2.9.0andhigher-cp39-abi3-win_amd64.whl—ABI3 works for Python 3.12).

pip install path/to/SageAttention-2.2.0.post4+torch2.9.0andhigher-cp39-abi3-win_amd64.whl

[leomaxwell972]

On windows 11, Python 3.12.10 PyTorch 2.9.0+cu130, MSVC 14.44.35207 (with custom env and intel compile tools added).

I solved the ambiguous problem by:
fused.cu and qk_int_sv_f16_cuda_sm80.cu -
both added include vector and typeindex

switched flags in setup for C++ to:
CXX_FLAGS = ["/O2", "/openmp", "/std:c++20", "/DENABLE_BF16"]
and NVCC std=c++20

only other thing I can think of is I played around with the order of my includes and lib, trimmed excess only windows kits, MSVC, CUDA, Intel, Python, in that order for both include and lib i believe. If this matters.

the idea here was to solve the ambiguity by adapting to what torch was using, i noticed the only non torch includes it was using was the two mentioned above, so figured it was routing std through those. Did not work on its own, and flags except C++20 were already set. set C++20 for compatibility with namespaces and features, compiled no issues that I can see.

Edit: Test looks good

# sagetest.py
import torch
from sageattention import sageattn   # <-- this is the API

device = "cuda"
dtype  = torch.float16

# (B, H, N, D) with default tensor_layout="HND"
q = torch.randn(1, 8, 512, 64, device=device, dtype=dtype)
k = torch.randn(1, 8, 512, 64, device=device, dtype=dtype)
v = torch.randn(1, 8, 512, 64, device=device, dtype=dtype)

# Eager
out_eager = sageattn(q, k, v, tensor_layout="HND", is_causal=False)

# torch.compile path (optional)
compiled = torch.compile(lambda Q,K,V: sageattn(Q,K,V, tensor_layout="HND", is_causal=False), dynamic=False)
out_comp = compiled(q,k,v)

print("sageattn OK:", out_eager.shape, torch.allclose(out_eager, out_comp, atol=1e-2, rtol=1e-2))


C:\AI>python sagetest.py
__triton_launcher.c
   Creating library C:\x\tmp4lh33m7n\__triton_launcher.cp312-win_amd64.lib and object C:\x\tmp4lh33m7n\__triton_launcher.cp312-win_amd64.exp
__triton_launcher.c
   Creating library C:\x\tmp739v44ci\__triton_launcher.cp312-win_amd64.lib and object C:\x\tmp739v44ci\__triton_launcher.cp312-win_amd64.exp
__triton_launcher.c
   Creating library C:\x\tmp_p7nw7tv\__triton_launcher.cp312-win_amd64.lib and object C:\x\tmp_p7nw7tv\__triton_launcher.cp312-win_amd64.exp
__triton_launcher.c
   Creating library C:\x\tmpf38wnk5a\__triton_launcher.cp312-win_amd64.lib and object C:\x\tmpf38wnk5a\__triton_launcher.cp312-win_amd64.exp
sageattn OK: torch.Size([1, 8, 512, 64]) True

[rugabunda]

On windows 11, Python 3.12.10 PyTorch 2.9.0+cu130, MSVC 14.44.35207 (with custom env and intel compile tools added).

I solved the ambiguous problem by: fused.cu and qk_int_sv_f16_cuda_sm80.cu - both added include vector and typeindex

switched flags in setup for C++ to: CXX_FLAGS = ["/O2", "/openmp", "/std:c++20", "/DENABLE_BF16"] and NVCC std=c++20

@leomaxwell972

That works! I made a script that automates this process that works with the latest bleeding cuda 13.0; pytorch 2.9; Python 3.12.11. Probably everything that was broken earlier.

Activate your environment

git clone https://github.com/thu-ml/SageAttention.git

Drop the code below into a file in its root and run:

python build_sage.py

# Sageattention build patcher
import os
import subprocess
import sys

# Apply patches
print("Applying patches...")
files = ['csrc/fused/fused.cu', 'csrc/qattn/qk_int_sv_f16_cuda_sm80.cu']

for f in files:
    if os.path.exists(f):
        with open(f, 'r') as file:
            content = file.read()
        if '#include <vector>' not in content:
            content = '#include <vector>\n#include <typeindex>\n' + content
            with open(f, 'w') as file:
                file.write(content)
            print(f"✓ Patched {f}")

# Patch setup.py
with open('setup.py', 'r') as f:
    setup_content = f.read()

setup_content = setup_content.replace(
    'CXX_FLAGS = ["/O2", "/openmp"]',
    'CXX_FLAGS = ["/O2", "/openmp", "/std:c++20", "/DENABLE_BF16"]'
)
setup_content = setup_content.replace('"-std=c++17"', '"-std=c++20"')

with open('setup.py', 'w') as f:
    f.write(setup_content)

print("✓ Patches applied successfully")

# Set environment variables
os.environ['EXT_PARALLEL'] = '4'
os.environ['NVCC_APPEND_FLAGS'] = '--threads 8'
os.environ['MAX_JOBS'] = '32'

# Uninstall old version first
subprocess.run([sys.executable, '-m', 'pip', 'uninstall', 'sageattention', '-y'])

# Build and install with pip
print("\nBuilding with pip...")
subprocess.run([sys.executable, '-m', 'pip', 'install', '.', '--no-build-isolation'], check=True)
print("\n✓ Build complete!")

[rugabunda]

thanks for the info - I really want to generate with cu130 but it breaks everything it touches rn..

btw torch 3.0.1+cu130 nightly are available for Linux only for the time being.

@joojie try the method above.