[Relay] add ShapeFunc for tanh #6898
Conversation
|
lgtm. pls fix the test failure. cc @icemelon9 @kevinthesun |
|
I have a question. What do you mean "input's shape's second dim is unknown"? TVM is able to compile the kernel even though the dim is a sym var. |
This in_dim will be used by following AutoTVM template code which requires integer. |
|
@wweic @lixiaoquan thanks for the comment. @icemelon9
Hi, this is the case when I import static LSTM model from tensorflow with batch dim set to Crash Stack: It fails because the data's second dim is inferred to |
python/tvm/topi/cuda/dense.py
Outdated
| A, _ = C.op.input_tensors | ||
| _, in_dim = get_const_tuple(A.shape) | ||
| A, weights = C.op.input_tensors | ||
| _, in_dim = get_const_tuple(weights.shape) |
There was a problem hiding this comment.
Probably we can extent the logic a bit here to check both in_dim from data and weight. If both of them are symbolic, we can skip cfg define for in_dim since autotvm doesn't support dynamic shape kernel in general.
There was a problem hiding this comment.
Okay, I will update later.
There was a problem hiding this comment.
@kevinthesun updated
|
@monklof any updates? Thanks in advance! |
updated, check the lastest commit |
monklof
force-pushed
the
lstm-cell-dynbatch
branch
from
a492115
to
ee1a749
Compare
|
@kevinthesun could you merge this pr? thanks. |
|
Thanks @monklof @junrushao1994 @kevinthesun @wweic |
* add ShapeFunc for tanh * _schedule_dense_small_batch turn autotvm off when dense's inner dim is unknown * fix CI pylint
* add ShapeFunc for tanh * _schedule_dense_small_batch turn autotvm off when dense's inner dim is unknown * fix CI pylint
Enable run static lstm on relay vm by dynamic batch:
_schedule_dense_small_batchinstead of input in case input's shape's second dim is unknown.@wweic