Closed aire-meta-bot closed 2 years ago
TVM CUDA backend seems broken. Here is an example:
python3 -m pytest tests/python/core/test_multi_function_flow.py
Results:
E TVMError: Cannot find a valid dispatch for op raf.op.reshape:
E [TVM] Failed to JIT: raf_op_tvm_reshape_1: RuntimeError:
E #ifdef _WIN32
E using uint = unsigned int;
E using uchar = unsigned char;
E using ushort = unsigned short;
E using int64_t = long long;
E using uint64_t = unsigned long long;
E #else
E #define uint unsigned int
E #define uchar unsigned char
E #define ushort unsigned short
E #define int64_t long long
E #define uint64_t unsigned long long
E #endif
E extern "C" __global__ void __launch_bounds__(256) fused_raf_op_tvm_reshape_kernel0(float* __restrict__ T_reshape, float* __restrict__ ) {
E T_reshape[((int)threadIdx.x)] = [((int)threadIdx.x)];
E }
E
E
E Compilation error:
E /tmp/tmpa1io3cw0/my_kernel.cu(16): error: expected an identifier
E
E /tmp/tmpa1io3cw0/my_kernel.cu(16): error: expected a "{" introducing a lambda body
E
E 2 errors detected in the compilation of "/tmp/tmpa1io3cw0/my_kernel.cu".
Fixed at https://github.com/apache/tvm/pull/12912. Closing this one and continue working at #128
This PR is auto-generated by Github Action Bot.
Please check the CI results before merge it. If one or more tests failed, please commit necessary changes to this branch (update-submodule-tvm-2022-09-19-13-18-53).
If there are more than one PRs like this, please keep the latest one and manually close others.
cc @awslabs/raf-committer