jinhongyii opened a new pull request, #14132:
URL: https://github.com/apache/tvm/pull/14132
As we start to build more structure into the program, there is an increasing
need to incorporate globally static object that are referred by the IR itself.
So we introduce "GlobalInfo" here to represent these objects. Below is an
example of a possible representation of a distributed program with global info.
Pay attention to "mesh", which represents the global infos in the module. It
contains a list of two device meshes , and they are accessed using syntax like
"mesh[0]", which is a sugar for IRModule.global_infos["mesh"][0]
```
@I.ir_module
class TestModule:
I.module_attrs({"device_num": 10})
I.module_global_infos(
{
"mesh": [
R.device_mesh((2, 2), R.range(0, 4)), # mesh[0]
R.device_mesh((1,), R.range(4, 5)), # mesh[1]
]
}
)
@T.prim_func
def tir_func(
x: T.Buffer((T.int64(128), T.int64(128)), "float32"),
y: T.Buffer((T.int64(128), T.int64(128)), "float32"),
):
T.func_attr({"tir.noalias": True})
for i, j in T.grid(T.int64(128), T.int64(128)):
with T.block():
vi, vj = T.axis.remap("SS", [i, j])
y[vi, vj] = x[vi, vj] + 1.0
@R.function
def foo(
x: R.DTensor((128, 128), "float32", device_mesh="mesh[0]",
placement="S[0], R"),
) -> R.DTensor((128, 128), "float32", device_mesh="mesh[0]",
placement="S[0], R"):
gv0 = R.dist.call_tir(
tir_func,
x,
R.DTensor(
shape=(128, 128), dtype="float32",
device_mesh="mesh[0]", placement="S[0], R"
),
)
return gv0
```
To summarize the PR, it
1. add a base class GlobalInfo
2. add a global_infos: Map[str, List[GlobalInfo]] field in IRModule
3. support parsing/printing for I.module_attrs and I.module_global_infos.\
cc: @tqchen, @junrushao
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
