Skip to content

Grouped GEMM execution not possible with HW #163

Open
@cassanof

Description

@cassanof

When running the grouped gemm implementation and expert parallelism, i am faced with the following error:

[rank5]:   File "/env/lib/python3.11/site-packages/megablocks-0.8.0.dev0-py3.11-linux-x86_64.egg/megablocks/layers/glu.py", line 255, in forward
[rank5]:     x1 = gg.ops.gmm(x, w1, batch_sizes, trans_b=True)
[rank5]:          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank5]:   File "/env/lib/python3.11/site-packages/grouped_gemm-0.1.6-py3.11-linux-x86_64.egg/grouped_gemm/ops.py", line 33, in gmm
[rank5]:     return GroupedGemm.apply(a, b, batch_sizes, trans_b)
[rank5]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank5]:   File "/env/lib/python3.11/site-packages/torch/autograd/function.py", line 575, in apply
[rank5]:     return super().apply(*args, **kwargs)  # type: ignore[misc]
[rank5]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank5]:   File "/env/lib/python3.11/site-packages/grouped_gemm-0.1.6-py3.11-linux-x86_64.egg/grouped_gemm/ops.py", line 11, in forward
[rank5]:     return backend.gmm(a, b, batch_sizes, trans_a=False, trans_b=trans_b)
[rank5]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank5]:   File "/env/lib/python3.11/site-packages/grouped_gemm-0.1.6-py3.11-linux-x86_64.egg/grouped_gemm/backend.py", line 27, in gmm
[rank5]:     backend.gmm(a, b, c, batch_sizes, trans_a, trans_b)
[rank5]: RuntimeError: Grouped GEMM execution not possible with HW

this only happens when you combine the two. using either alone works fine. setup here is 8xh100.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions