FlexAttention throws in multi-GPU env: ValueError: Pointer argument (at 3) cannot be accessed from Triton (cpu tensor?) #141010
Description
🐛 Describe the bug
Problem
When running compiled FlexAttention in a multi-GPU environment, if the device being used is not the first GPU (i.e., not cuda or cuda:0, but cuda:1, etc.), a ValueError will be raised: ValueError: Pointer argument (at 3) cannot be accessed from Triton (cpu tensor?).
I am using PyTorch Nightly built on 2024-11-15.
Output
Traceback (most recent call last):
File "/home/user/test.py", line 20, in <module>
attn_output = mod(q, k, v, mask)
^^^^^^^^^^^^^^^^^^
File "/home/user/.conda/envs/w568wnext/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1740, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/user/.conda/envs/w568wnext/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1751, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/user/.conda/envs/w568wnext/lib/python3.12/site-packages/torch/_dynamo/eval_frame.py", line 556, in _fn
return fn(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^
File "/home/user/.conda/envs/w568wnext/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1740, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/user/.conda/envs/w568wnext/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1751, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/user/test.py", line 5, in forward
def forward(self, q, k, v, block_mask):
File "/home/user/.conda/envs/w568wnext/lib/python3.12/site-packages/torch/_dynamo/eval_frame.py", line 721, in _fn
return fn(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^
File "/home/user/.conda/envs/w568wnext/lib/python3.12/site-packages/torch/_functorch/aot_autograd.py", line 1132, in forward
return compiled_fn(full_args)
^^^^^^^^^^^^^^^^^^^^^^
File "/home/user/.conda/envs/w568wnext/lib/python3.12/site-packages/torch/_functorch/_aot_autograd/runtime_wrappers.py", line 324, in runtime_wrapper
all_outs = call_func_at_runtime_with_args(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/user/.conda/envs/w568wnext/lib/python3.12/site-packages/torch/_functorch/_aot_autograd/utils.py", line 126, in call_func_at_runtime_with_args
out = normalize_as_list(f(args))
^^^^^^^
File "/home/user/.conda/envs/w568wnext/lib/python3.12/site-packages/torch/_functorch/_aot_autograd/runtime_wrappers.py", line 673, in inner_fn
outs = compiled_fn(args)
^^^^^^^^^^^^^^^^^
File "/home/user/.conda/envs/w568wnext/lib/python3.12/site-packages/torch/_functorch/_aot_autograd/runtime_wrappers.py", line 491, in wrapper
return compiled_fn(runtime_args)
^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/user/.conda/envs/w568wnext/lib/python3.12/site-packages/torch/_inductor/codecache.py", line 1686, in __call__
return self.current_callable(inputs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/user/.conda/envs/w568wnext/lib/python3.12/site-packages/torch/_inductor/utils.py", line 2065, in run
return model(new_inputs)
^^^^^^^^^^^^^^^^^
File "/tmp/torchinductor_w568w/g5/cg5bqruzf7jcvb5yvr3e3esy3ny77qf6l2phuzkvrwe54tn62zuc.py", line 522, in call
triton_tem_fused_0.run(arg0_1, arg1_1, arg2_1, buf0, arg3_1, arg4_1, arg5_1, arg6_1, buf1, grid=torch._inductor.kernel.flex_attention.flex_attention_grid(1, 1, 2048, 128, meta0), stream=stream1)
File "/home/user/.conda/envs/w568wnext/lib/python3.12/site-packages/torch/_inductor/runtime/triton_heuristics.py", line 1044, in run
return launcher(
^^^^^^^^^
File "<string>", line 13, in launcher
File "/home/user/.conda/envs/w568wnext/lib/python3.12/site-packages/triton/backends/nvidia/driver.py", line 365, in __call__
self.launch(*args, **kwargs)
ValueError: Pointer argument (at 3) cannot be accessed from Triton (cpu tensor?)
Expected Output
Print the shape of attn_output and exit normally.
Minimal Reproducible Example
import torch
from torch.nn.attention.flex_attention import flex_attention, create_block_mask
class TestModule(torch.nn.Module):
def forward(self, q, k, v, block_mask):
return flex_attention(q, k, v, block_mask=block_mask)
q = torch.randn(1, 1, 2048, 128, device="cuda:1", dtype=torch.bfloat16)
k = torch.randn(1, 1, 2048, 128, device="cuda:1", dtype=torch.bfloat16)
v = torch.randn(1, 1, 2048, 128, device="cuda:1", dtype=torch.bfloat16)
mask = create_block_mask(
lambda b, h, q_idx, kv_idx: q_idx >= kv_idx,
B=None,
H=None,
Q_LEN=2048,
KV_LEN=2048,
device="cuda:1",
)
mod = torch.compile(TestModule())
attn_output = mod(q, k, v, mask)
print(attn_output.shape)Additional Information
The script generated by PyTorch Inductor /tmp/torchinductor_w568w/g5/cg5bqruzf7jcvb5yvr3e3esy3ny77qf6l2phuzkvrwe54tn62zuc.py is uploaded here: tmp.py.txt
Seems that triton is complaining that buf0 is inaccessible from cuda:1, which was created before setting default device to 1:
buf0 = empty_strided_cuda((1, 1, 2048), (2048, 2048, 1), torch.float32)
with torch.cuda._DeviceGuard(1):
torch.cuda.set_device(1)
buf1 = empty_strided_cuda((1, 1, 2048, 128), (262144, 262144, 128, 1), torch.bfloat16)
# Topologically Sorted Source Nodes: [flex_attention], Original ATen: []
stream1 = get_raw_stream(1)
triton_tem_fused_0.run(arg0_1, arg1_1, arg2_1, buf0, arg3_1, arg4_1, arg5_1, arg6_1, buf1, grid=torch._inductor.kernel.flex_attention.flex_attention_grid(1, 1, 2048, 128, meta0), stream=stream1)Related Issues
#134739 reported an issue where another parameter (i.e., the block mask) was not on the same device, which was caused by forgetting to set the correct device when calling create_block_mask.
#136078 reported the same error but failed to provide any useful debugging information or a minimal reproducible example.
Versions
PyTorch version: 2.6.0.dev20241115+cu124
Is debug build: False
CUDA used to build PyTorch: 12.4
ROCM used to build PyTorch: N/A
OS: Ubuntu 20.04.6 LTS (x86_64)
GCC version: (Ubuntu 9.4.0-1ubuntu1~20.04.2) 9.4.0
Clang version: Could not collect
CMake version: Could not collect
Libc version: glibc-2.31
Python version: 3.12.7 | packaged by Anaconda, Inc. | (main, Oct 4 2024, 13:27:36) [GCC 11.2.0] (64-bit runtime)
Python platform: Linux-5.15.0-107-generic-x86_64-with-glibc2.31
Is CUDA available: True
CUDA runtime version: 12.1.66
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration:
GPU 0: NVIDIA A800 80GB PCIe
GPU 1: NVIDIA A800 80GB PCIe
GPU 2: NVIDIA A800 80GB PCIe
GPU 3: NVIDIA A800 80GB PCIe
GPU 4: NVIDIA A800 80GB PCIe
GPU 5: NVIDIA A800 80GB PCIe
GPU 6: NVIDIA A800 80GB PCIe
GPU 7: NVIDIA A800 80GB PCIe
Nvidia driver version: 535.183.01
cuDNN version: Probably one of the following:
/usr/lib/x86_64-linux-gnu/libcudnn.so.9.1.0
/usr/lib/x86_64-linux-gnu/libcudnn_adv.so.9.1.0
/usr/lib/x86_64-linux-gnu/libcudnn_cnn.so.9.1.0
/usr/lib/x86_64-linux-gnu/libcudnn_engines_precompiled.so.9.1.0
/usr/lib/x86_64-linux-gnu/libcudnn_engines_runtime_compiled.so.9.1.0
/usr/lib/x86_64-linux-gnu/libcudnn_graph.so.9.1.0
/usr/lib/x86_64-linux-gnu/libcudnn_heuristic.so.9.1.0
/usr/lib/x86_64-linux-gnu/libcudnn_ops.so.9.1.0
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True
CPU:
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Byte Order: Little Endian
Address sizes: 46 bits physical, 57 bits virtual
CPU(s): 112
On-line CPU(s) list: 0-111
Thread(s) per core: 2
Core(s) per socket: 28
Socket(s): 2
NUMA node(s): 4
Vendor ID: GenuineIntel
CPU family: 6
Model: 106
Model name: Intel(R) Xeon(R) Gold 6330 CPU @ 2.00GHz
Stepping: 6
CPU MHz: 2600.000
CPU max MHz: 3100.0000
CPU min MHz: 800.0000
BogoMIPS: 4000.00
L1d cache: 2.6 MiB
L1i cache: 1.8 MiB
L2 cache: 70 MiB
L3 cache: 84 MiB
NUMA node0 CPU(s): 0-13,56-69
NUMA node1 CPU(s): 14-27,70-83
NUMA node2 CPU(s): 28-41,84-97
NUMA node3 CPU(s): 42-55,98-111
Vulnerability Gather data sampling: Mitigation; Microcode
Vulnerability Itlb multihit: Not affected
Vulnerability L1tf: Not affected
Vulnerability Mds: Not affected
Vulnerability Meltdown: Not affected
Vulnerability Mmio stale data: Mitigation; Clear CPU buffers; SMT vulnerable
Vulnerability Retbleed: Not affected
Vulnerability Spec rstack overflow: Not affected
Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp
Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2: Mitigation; Enhanced IBRS; IBPB conditional; RSB filling; PBRSB-eIBRS SW sequence; BHI Syscall hardening, KVM SW loop
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf pni pclmulqdq dtes64 monitor ds_cpl smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 invpcid_single intel_ppin ssbd mba ibrs ibpb stibp ibrs_enhanced fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb intel_pt avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local split_lock_detect wbnoinvd dtherm ida arat pln pts avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg tme avx512_vpopcntdq la57 rdpid fsrm md_clear pconfig flush_l1d arch_capabilities
Versions of relevant libraries:
[pip3] numpy==2.0.2
[pip3] nvidia-cublas-cu12==12.4.5.8
[pip3] nvidia-cuda-cupti-cu12==12.4.127
[pip3] nvidia-cuda-nvrtc-cu12==12.4.127
[pip3] nvidia-cuda-runtime-cu12==12.4.127
[pip3] nvidia-cudnn-cu12==9.1.0.70
[pip3] nvidia-cufft-cu12==11.2.1.3
[pip3] nvidia-curand-cu12==10.3.5.147
[pip3] nvidia-cusolver-cu12==11.6.1.9
[pip3] nvidia-cusparse-cu12==12.3.1.170
[pip3] nvidia-cusparselt-cu12==0.6.2
[pip3] nvidia-nccl-cu12==2.21.5
[pip3] nvidia-nvjitlink-cu12==12.4.127
[pip3] nvidia-nvtx-cu12==12.4.127
[pip3] nvtx==0.2.10
[pip3] pynvjitlink-cu12==0.4.0
[pip3] pytorch-triton==3.1.0+cf34004b8a
[pip3] torch==2.6.0.dev20241115+cu124
[pip3] torchvision==0.20.0.dev20241115+cu124
[conda] numpy 2.0.2 pypi_0 pypi
[conda] nvidia-cublas-cu12 12.4.5.8 pypi_0 pypi
[conda] nvidia-cuda-cupti-cu12 12.4.127 pypi_0 pypi
[conda] nvidia-cuda-nvrtc-cu12 12.4.127 pypi_0 pypi
[conda] nvidia-cuda-runtime-cu12 12.4.127 pypi_0 pypi
[conda] nvidia-cudnn-cu12 9.1.0.70 pypi_0 pypi
[conda] nvidia-cufft-cu12 11.2.1.3 pypi_0 pypi
[conda] nvidia-curand-cu12 10.3.5.147 pypi_0 pypi
[conda] nvidia-cusolver-cu12 11.6.1.9 pypi_0 pypi
[conda] nvidia-cusparse-cu12 12.3.1.170 pypi_0 pypi
[conda] nvidia-cusparselt-cu12 0.6.2 pypi_0 pypi
[conda] nvidia-nccl-cu12 2.21.5 pypi_0 pypi
[conda] nvidia-nvjitlink-cu12 12.4.127 pypi_0 pypi
[conda] nvidia-nvtx-cu12 12.4.127 pypi_0 pypi
[conda] nvtx 0.2.10 pypi_0 pypi
[conda] pynvjitlink-cu12 0.4.0 pypi_0 pypi
[conda] pytorch-triton 3.1.0+cf34004b8a pypi_0 pypi
[conda] torch 2.6.0.dev20241115+cu124 pypi_0 pypi
[conda] torchvision 0.20.0.dev20241115+cu124 pypi_0 pypi
cc @chauhang @penguinwu @zou3519 @ydwu4 @bdhirsh @yf225 @Chillee @drisspg @yanboliang @BoyuanFeng