tvm
tvm copied to clipboard
apache
Open deep learning compiler stack for cpu, gpu and specialized accelerators
Hello, this is my first time using TVM. An error appears when running realy.build. Can you suggest a solution? The error is as follows tvm.error.InternalError: Traceback (most recent call last):...
This PR brings up new reflection support in python. The new reflection now directly attaches property and methods to the class object themselves, making more efficient accessing than old mechanism....
Adds per layer sliding window functionality to the KV Cache. Correctness is mostly achieved, but there are some cases where single tokens are strange. The corresponding MLC-LLM PR is https://github.com/mlc-ai/mlc-llm/pull/3248...
Fixes #18002 When all dimensions of layer_norm input are reduction axes, the variable `non_reduce_indices` becomes empty, while the previous variable `target_shape` has `atleast1d` parameter enabled, causing shape inconsistency. https://github.com/apache/tvm/blob/2d964b4133aac2f92e4185b3f095df4eb3bf3a90/include/tvm/topi/nn/layer_norm.h#L111-L121 https://github.com/apache/tvm/blob/2d964b4133aac2f92e4185b3f095df4eb3bf3a90/include/tvm/topi/nn/layer_norm.h#L67-L68...
This PR adds support for bucketize op which is used in many vision models like Phi4, SmolVLM etc.,
### Expected behavior The onnx frontend should import the model correctly. ### Actual behavior For the following model, it can be executed by onnxruntime.  the results of onnxruntime are...
# Summary I was trying to resolve https://github.com/apache/tvm/issues/18004, where an ONNX model causes a segmentation fault in TVM but not in onnxruntime. **Why the seg fault occurs** This occurs because...
This commit adds support for CUDA device function calls by: 1. Modifying the calling convention handling in CUDA codegen to support both device kernel launches and device function calls 2....
In Arm China, we are migrating the old work to the newest code, recently the FFI refactor introduce lots of changes in very low level infrastructure, now only the `cython`...
When compiling an ONNX model using TVM's Relax, the generated .so files are identical when setting opt_level=0, opt_level=1, opt_level=3, and opt_level=4. The compilation code I used is as follows: `onnx_model...
