tvm
tvm copied to clipboard
apache
Open deep learning compiler stack for cpu, gpu and specialized accelerators
### Expected behavior When feed the same input into the same model, one in TVM, another in ONNXRuntime, we supposed that their results are the same. ### Actual behavior TVM's...
1. add a dlight rule LowBatchGEMV to schedule low-batch GEMM just like GEMV. 2. fix some issues when lowering low-batch GEMM
The provided usage example for `tvm.ir.make_node` has become outdated. Creating an IR node of type `IntImm` requires the field `span`. Compare: [`make_node` Unit Test](https://github.com/apache/tvm/blob/main/tests/python/ir/test_node_reflection.py#L73)
In GPT-2, attention calculation requires an additional feature `scale_attn_by_inverse_layer_idx`. It provides a scaling factor per attention layer when calculating the attention score, before applying the softmax function. This PR supports...
This commit enhances the AOT test harness to print the "actual" and "reference" values when there is a mismatch. This helps when debugging a failing test. Sample output: ``` Actual,...
This change will allow us to express scalable vectors through Ramp and Broadcast nodes, e.g. ``` vec = tvm.tir.expr.Ramp(0, 1, 4 * tvm.tir.vscale()) ``` We will use negative values for...
It is possible to measure GPU bandwidth through RPC on an Android device. cc: @junrushao, @masahi
Adds 2D Discontiguous alloc tensor hexagon builtin to support 2D allocations for hexagon at relax level. This is needed when the ops are implemented to take advantage of 2d indirections...
This commit adds the RNNState class to the Relax VM, similar to the PagedKVCache, for space state models like RWKV and mamba cc @tqchen @MasterJH5574
