New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[release/2.5][ROCm] Fix largeIndexBlockSize #1659

Open

pragupta wants to merge 1 commit into ROCm:release/2.5 from pragupta:pg-msft-fix-2.5

pragupta commented Oct 30, 2024

On ROCm, hipification converts std::min to ::min, but ::min is not returning the right result. This impacts index_add_ operation on a large tensor, we end up picking the large values instead of max supported block size (128). This leads to GPU accessing memory out of bounds.

While we wait for ::min to be fixed, we can use < operator to compare instead of relying on ::min.

Example Code w/ failure:

D=6144
hidden_states = torch.zeros([16384, 6144],           device="cuda:0", dtype=torch.bfloat16)
index         = torch.randint(0, 16384, (1, 32, 16384), device="cuda:0", dtype=torch.int64)
output        = torch.empty([1, 32, 16384, 6144],    device="cuda:0", dtype=torch.bfloat16)
hidden_states.index_add_(0, index.view(-1), output.view(-1, D))

Traceback (most recent call last):
RuntimeError: HIP error: invalid configuration argument

(cherry picked from commit c0266db)


          [release/2.5][ROCm] Fix largeIndexBlockSize

db33c0f

On ROCm, hipification converts std::min to ::min, but ::min is not
returning the right result. In the meantime, use < operator to comapre.

(cherry picked from commit c0266db)

pragupta requested review from pruthvistony and jithunnair-amd

October 30, 2024 16:30

Collaborator

pruthvistony commented Oct 30, 2024

Not yet decided on cherry-picks into 2.5, so want to wait on this PR merge.

rocm-mici commented Oct 30, 2024

Jenkins build for db33c0f8917630a279e142c898a5011bdef163a1 commit finished as FAILURE
Links: Blue Ocean view / Build artifacts

rocm-mici commented Nov 5, 2024

Jenkins build for db33c0f8917630a279e142c898a5011bdef163a1 commit finished as FAILURE
Links: Blue Ocean view / Build artifacts

rocm-mici commented Nov 7, 2024

Jenkins build for db33c0f8917630a279e142c898a5011bdef163a1 commit finished as FAILURE
Links: Blue Ocean view / Build artifacts

rocm-mici commented Nov 9, 2024

Jenkins build for db33c0f8917630a279e142c898a5011bdef163a1 commit finished as FAILURE
Links: Blue Ocean view / Build artifacts

Detected error during Pytorch building:

[7969/8668] Building HIPCC object caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/hip/torch_hip_generated_ForeachBinaryOpScalar.hip.o
[7970/8668] Building HIPCC object caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/hip/torch_hip_generated_fused_adamw_amsgrad_impl.hip.o
[7971/8668] Building HIPCC object caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/hip/torch_hip_generated_scaled_modified_bessel_k1.hip.o
[7972/8668] Building HIPCC object caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/hip/torch_hip_generated_ForeachBinaryOpScalarList.hip.o
[7973/8668] Building HIPCC object caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/torch_hip_generated_attention.hip.o
FAILED: caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/torch_hip_generated_attention.hip.o /var/lib/jenkins/pytorch/build/caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/torch_hip_generated_attention.hip.o 
cd /var/lib/jenkins/pytorch/build/caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip && /opt/conda/envs/py_3.10/bin/cmake -E make_directory /var/lib/jenkins/pytorch/build/caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/. && /opt/conda/envs/py_3.10/bin/cmake -D verbose:BOOL=OFF -D build_configuration:STRING=RELEASE -D generated_file:STRING=/var/lib/jenkins/pytorch/build/caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/./torch_hip_generated_attention.hip.o -P /var/lib/jenkins/pytorch/build/caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/torch_hip_generated_attention.hip.o.cmake
In file included from /var/lib/jenkins/pytorch/aten/src/ATen/native/transformers/hip/attention.hip:84:
/var/lib/jenkins/pytorch/aten/src/ATen/native/transformers/hip/aotriton_adapter.h:120:10: error: no matching constructor for initialization of 'aotriton::TensorView<0>'
  120 |   return aotriton::TensorView<0>(reinterpret_cast<intptr_t>(q.data_ptr()),
      |          ^                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

rocm-mici commented Nov 13, 2024

Jenkins build for db33c0f8917630a279e142c898a5011bdef163a1 commit finished as FAILURE
Links: Blue Ocean view / Build artifacts

Detected error during Pytorch building:

/opt/rocm-6.2.3/lib/llvm/bin/../../../include/hip/hip_runtime_api.h:580:41: note: expanded from macro 'DEPRECATED'
  580 | #define DEPRECATED(msg) __attribute__ ((deprecated(msg)))
      |                                         ^
1 warning generated when compiling for gfx908.
[8005/8668] Building HIPCC object caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/flash_attn/torch_hip_generated_flash_api.hip.o
FAILED: caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/flash_attn/torch_hip_generated_flash_api.hip.o /var/lib/jenkins/pytorch/build/caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/flash_attn/torch_hip_generated_flash_api.hip.o 
cd /var/lib/jenkins/pytorch/build/caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/flash_attn && /opt/conda/envs/py_3.10/bin/cmake -E make_directory /var/lib/jenkins/pytorch/build/caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/flash_attn/. && /opt/conda/envs/py_3.10/bin/cmake -D verbose:BOOL=OFF -D build_configuration:STRING=RELEASE -D generated_file:STRING=/var/lib/jenkins/pytorch/build/caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/flash_attn/./torch_hip_generated_flash_api.hip.o -P /var/lib/jenkins/pytorch/build/caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/flash_attn/torch_hip_generated_flash_api.hip.o.cmake
In file included from /var/lib/jenkins/pytorch/aten/src/ATen/native/transformers/hip/flash_attn/flash_api.hip:57:
/var/lib/jenkins/pytorch/aten/src/ATen/native/transformers/hip/aotriton_adapter.h:120:10: error: no matching constructor for initialization of 'aotriton::TensorView<0>'
  120 |   return aotriton::TensorView<0>(reinterpret_cast<intptr_t>(q.data_ptr()),
      |          ^                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

rocm-mici commented Nov 13, 2024

Jenkins build for db33c0f8917630a279e142c898a5011bdef163a1 commit finished as FAILURE
Links: Blue Ocean view / Build artifacts

Detected error during Pytorch building:

[7968/8668] Building HIPCC object caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/hip/torch_hip_generated_ForeachBinaryOpScalarList.hip.o
[7969/8668] Building HIPCC object caffe2/CMakeFiles/torch_hip.dir/__/torch/csrc/distributed/c10d/torch_hip_generated_CUDASymmetricMemoryOps.cu.o
[7970/8668] Building HIPCC object caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/hip/torch_hip_generated_UnfoldBackwardKernel.hip.o
[7971/8668] Building HIPCC object caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/hip/torch_hip_generated_scaled_modified_bessel_k1.hip.o
[7972/8668] Building HIPCC object caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/flash_attn/torch_hip_generated_flash_api.hip.o
FAILED: caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/flash_attn/torch_hip_generated_flash_api.hip.o /var/lib/jenkins/pytorch/build/caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/flash_attn/torch_hip_generated_flash_api.hip.o 
cd /var/lib/jenkins/pytorch/build/caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/flash_attn && /opt/conda/envs/py_3.10/bin/cmake -E make_directory /var/lib/jenkins/pytorch/build/caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/flash_attn/. && /opt/conda/envs/py_3.10/bin/cmake -D verbose:BOOL=OFF -D build_configuration:STRING=RELEASE -D generated_file:STRING=/var/lib/jenkins/pytorch/build/caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/flash_attn/./torch_hip_generated_flash_api.hip.o -P /var/lib/jenkins/pytorch/build/caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/flash_attn/torch_hip_generated_flash_api.hip.o.cmake
In file included from /var/lib/jenkins/pytorch/aten/src/ATen/native/transformers/hip/flash_attn/flash_api.hip:57:
/var/lib/jenkins/pytorch/aten/src/ATen/native/transformers/hip/aotriton_adapter.h:120:10: error: no matching constructor for initialization of 'aotriton::TensorView<0>'
  120 |   return aotriton::TensorView<0>(reinterpret_cast<intptr_t>(q.data_ptr()),
      |          ^                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

rocm-mici commented Nov 15, 2024

Jenkins build for db33c0f8917630a279e142c898a5011bdef163a1 commit finished as FAILURE
Links: Blue Ocean view / Build artifacts

Detected error during Pytorch building:

[7963/8668] Building HIPCC object caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/hip/torch_hip_generated_ForeachBinaryOpScalarList.hip.o
[7964/8668] Building HIPCC object caffe2/CMakeFiles/torch_hip.dir/__/torch/csrc/distributed/c10d/torch_hip_generated_CUDASymmetricMemoryOps.cu.o
[7965/8668] Building HIPCC object caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/hip/torch_hip_generated_modified_bessel_k0.hip.o
[7966/8668] Building HIPCC object caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/hip/torch_hip_generated_ForeachBinaryOpScalar.hip.o
[7967/8668] Building HIPCC object caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/torch_hip_generated_attention.hip.o
FAILED: caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/torch_hip_generated_attention.hip.o /var/lib/jenkins/pytorch/build/caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/torch_hip_generated_attention.hip.o 
cd /var/lib/jenkins/pytorch/build/caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip && /opt/conda/envs/py_3.10/bin/cmake -E make_directory /var/lib/jenkins/pytorch/build/caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/. && /opt/conda/envs/py_3.10/bin/cmake -D verbose:BOOL=OFF -D build_configuration:STRING=RELEASE -D generated_file:STRING=/var/lib/jenkins/pytorch/build/caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/./torch_hip_generated_attention.hip.o -P /var/lib/jenkins/pytorch/build/caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/torch_hip_generated_attention.hip.o.cmake
In file included from /var/lib/jenkins/pytorch/aten/src/ATen/native/transformers/hip/attention.hip:84:
/var/lib/jenkins/pytorch/aten/src/ATen/native/transformers/hip/aotriton_adapter.h:120:10: error: no matching constructor for initialization of 'aotriton::TensorView<0>'
  120 |   return aotriton::TensorView<0>(reinterpret_cast<intptr_t>(q.data_ptr()),
      |          ^                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Collaborator

okakarpa commented Nov 21, 2024

Jenkins build for db33c0f8917630a279e142c898a5011bdef163a1 commit finished as FAILURE
Links: Blue Ocean view / Build artifacts

Detected error during Pytorch building:

      |                                         ^
1 warning generated when compiling for gfx908.
[8004/8668] Building HIPCC object caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/hip/torch_hip_generated_spherical_bessel_j0.hip.o
[8005/8668] Building HIPCC object caffe2/CMakeFiles/torch_hip.dir/__/torch/csrc/distributed/c10d/torch_hip_generated_CUDASymmetricMemoryOps.cu.o
[8006/8668] Building HIPCC object caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/torch_hip_generated_attention_backward.hip.o
FAILED: caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/torch_hip_generated_attention_backward.hip.o /var/lib/jenkins/pytorch/build/caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/torch_hip_generated_attention_backward.hip.o 
cd /var/lib/jenkins/pytorch/build/caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip && /opt/conda/envs/py_3.10/bin/cmake -E make_directory /var/lib/jenkins/pytorch/build/caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/. && /opt/conda/envs/py_3.10/bin/cmake -D verbose:BOOL=OFF -D build_configuration:STRING=RELEASE -D generated_file:STRING=/var/lib/jenkins/pytorch/build/caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/./torch_hip_generated_attention_backward.hip.o -P /var/lib/jenkins/pytorch/build/caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/torch_hip_generated_attention_backward.hip.o.cmake
In file included from /var/lib/jenkins/pytorch/aten/src/ATen/native/transformers/hip/attention_backward.hip:49:
/var/lib/jenkins/pytorch/aten/src/ATen/native/transformers/hip/aotriton_adapter.h:120:10: error: no matching constructor for initialization of 'aotriton::TensorView<0>'
  120 |   return aotriton::TensorView<0>(reinterpret_cast<intptr_t>(q.data_ptr()),
      |          ^                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Collaborator

okakarpa commented Nov 22, 2024

Jenkins build for db33c0f8917630a279e142c898a5011bdef163a1 commit finished as FAILURE
Links: Blue Ocean view / Build artifacts

Detected error during Pytorch building:

Warning: Unused direct dependencies:
	/var/lib/jenkins/pytorch/build/lib/libshm.so
	/lib/x86_64-linux-gnu/libm.so.6
[8005/8668] Building HIPCC object caffe2/CMakeFiles/torch_hip.dir/__/torch/csrc/distributed/c10d/torch_hip_generated_CUDASymmetricMemoryOps.cu.o
[8006/8668] Building HIPCC object caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/flash_attn/torch_hip_generated_flash_api.hip.o
FAILED: caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/flash_attn/torch_hip_generated_flash_api.hip.o /var/lib/jenkins/pytorch/build/caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/flash_attn/torch_hip_generated_flash_api.hip.o 
cd /var/lib/jenkins/pytorch/build/caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/flash_attn && /opt/conda/envs/py_3.10/bin/cmake -E make_directory /var/lib/jenkins/pytorch/build/caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/flash_attn/. && /opt/conda/envs/py_3.10/bin/cmake -D verbose:BOOL=OFF -D build_configuration:STRING=RELEASE -D generated_file:STRING=/var/lib/jenkins/pytorch/build/caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/flash_attn/./torch_hip_generated_flash_api.hip.o -P /var/lib/jenkins/pytorch/build/caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/flash_attn/torch_hip_generated_flash_api.hip.o.cmake
In file included from /var/lib/jenkins/pytorch/aten/src/ATen/native/transformers/hip/flash_attn/flash_api.hip:57:
/var/lib/jenkins/pytorch/aten/src/ATen/native/transformers/hip/aotriton_adapter.h:120:10: error: no matching constructor for initialization of 'aotriton::TensorView<0>'
  120 |   return aotriton::TensorView<0>(reinterpret_cast<intptr_t>(q.data_ptr()),
      |          ^                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Collaborator

okakarpa commented Nov 23, 2024

Jenkins build for db33c0f8917630a279e142c898a5011bdef163a1 commit finished as FAILURE
Links: Blue Ocean view / Build artifacts

Detected error during Pytorch building:

/opt/rocm-6.2.3/lib/llvm/bin/../../../include/hip/hip_runtime_api.h:580:41: note: expanded from macro 'DEPRECATED'
  580 | #define DEPRECATED(msg) __attribute__ ((deprecated(msg)))
      |                                         ^
1 warning generated when compiling for gfx908.
[8007/8668] Building HIPCC object caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/flash_attn/torch_hip_generated_flash_api.hip.o
FAILED: caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/flash_attn/torch_hip_generated_flash_api.hip.o /var/lib/jenkins/pytorch/build/caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/flash_attn/torch_hip_generated_flash_api.hip.o 
cd /var/lib/jenkins/pytorch/build/caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/flash_attn && /opt/conda/envs/py_3.10/bin/cmake -E make_directory /var/lib/jenkins/pytorch/build/caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/flash_attn/. && /opt/conda/envs/py_3.10/bin/cmake -D verbose:BOOL=OFF -D build_configuration:STRING=RELEASE -D generated_file:STRING=/var/lib/jenkins/pytorch/build/caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/flash_attn/./torch_hip_generated_flash_api.hip.o -P /var/lib/jenkins/pytorch/build/caffe2/CMakeFiles/torch_hip.dir/__/aten/src/ATen/native/transformers/hip/flash_attn/torch_hip_generated_flash_api.hip.o.cmake
In file included from /var/lib/jenkins/pytorch/aten/src/ATen/native/transformers/hip/flash_attn/flash_api.hip:57:
/var/lib/jenkins/pytorch/aten/src/ATen/native/transformers/hip/aotriton_adapter.h:120:10: error: no matching constructor for initialization of 'aotriton::TensorView<0>'
  120 |   return aotriton::TensorView<0>(reinterpret_cast<intptr_t>(q.data_ptr()),
      |          ^                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet