relax
relax copied to clipboard
tlc-pack
Temp repo for prototyping relax(relay next), the effort will be upstreamed. We use the wiki pages on this repo to host design docs.
Discussion Thread: https://github.com/tlc-pack/relax/issues/138
The following PrimFunc produces different results after tuning on hexagon. ```python @T.prim_func def conv2d(lv1: T.Buffer[(1, 230, 230, 3), "float16"], param_0: T.Buffer[(7, 7, 3, 64), "float16"], conv2d_nhwc: T.Buffer[(1, 112, 112, 64),...
From early on, Relax has adopted ANF* as a standard representation for passes and much of the compiler infrastructure is built around assuming that passes will be normalized into ANF....
As have been noticed in several places, we find it is helpful for clarity and development to bring dedicated Expr AST to relax, including If, Tuple, Constant
# Motivation & Goals Tensor data layout describes how the data is laid out in memory. It determines the memory access pattern and it can significantly impact performance and memory...
# 1. Motivation ## Problem Currently, once we lower op to `primfunc` implementation, it is hard to exploit op-level info (e.g., op name, op kind, op attribute..) although `primfunc` is...
This is a draft proposal to show high-level ideas about supporting constant POD (plain-old-data) values (e.g., int, float, bool) in Relax IR. Relax has first-class support of interacting with TVM...
I has one tf2 model which has node like below: node { name: “unknown” op: “Placeholder” attr { key: “_output_shapes” value { list { shape { } } } }...
Presently, any wildcard dimension in a shape annotation (namely, written as `_`) is parsed as a fresh shape variable. For example, let us consider the following function: ```python @tvm.script.ir_module class...
As we start to work on specific hardware, many operators would expect a specific kind of layout for both data and weight. Logically the layout start with simple ones. This...
