SparseTIR
SparseTIR copied to clipboard
uwsampl
SparseTIR: Sparse Tensor Compiler for Deep Learning
# Milestone - [ ] Blitz course to SparseTIR (due Nov 17) - [ ] System Overview (for developers) (due Nov 20) - [ ] Tutorial: SpMM operator and optimization...
Hi, I've noticed that the 'tvm.sparse' module is missing from the '```__init.py__```' file, where it should presumably be included. ``` python # tvm.sparse from . import sparse ``` It works...
### Expected behavior params removed in remove_unused_args pass shouldn't be checked in make_packed_api pass ### Actual behavior In primfunc with dynamic behavior, **Variable** args used in describing an axis metadata...
# The Problem Currently, SparseTIR does not support lowering code to co-iteration structure, whenever we want to add/multiply two sparse tensors/vectors, we need to create another axis to indicate the...
# Motivation After format decomposition, many of the generated sparse iterations (or `blocks` after lowering) only perform partial updates: only some blocks need to initialize the output buffer with zeros....
The forward function of the RGMS kernel is (relation related information are ignored for simplicity): $$ Y = AXW $$ we already have its implementation written in SparseTIR using composable...
Milestone - [ ] Program transformations - [ ] Format Description - [ ] Format Conversion Routines
# Pitch Currently, we use `PreservePreprocess` and `RemovePreprocess` pass to filter and split a given IRModule by tag `preprocessing`, however, format decomposition does not always produce preprocess blocks, some of...
# Pitch Currently, we only support two modes `lower_bound`/`upper_bound` for binary search. In some cases, we need to identify whether the given coordinate matches with any of the values in...
# Problem Our current format annotations do not efficiently handle branches in the axes dependency tree, for example, the [irregular batched-GEMM operator] has the following dependency trees: ```python """ B...