Adding e2e tests for i1 mask attentions #19312
Conversation
lialan
force-pushed
the
lialan/attention_i1_e2e
branch
6 times, most recently
from
35f7356 to
9985827
Compare
lialan
requested review from
ScottTodd,
rsuderman,
Max191 and
raikonenfnu
and removed request for
ScottTodd and
rsuderman
lialan
force-pushed
the
lialan/attention_i1_e2e
branch
from
9985827 to
6983fa5
Compare
|
@raikonenfnu @ScottTodd comments are addressed. Can you review again? |
| %mask = util.unfoldable_constant dense<[[[true, true, true, true], | ||
| [true, true, true, true], | ||
| [true, true, true, true], | ||
| [true, true, true, true]]]> : tensor<1x4x4xi1> |
There was a problem hiding this comment.
For unit testing, would it be useful to set some of these mask values to false?
Maybe add a new test case attention1x4x4_i1_mask_some_ones?
There was a problem hiding this comment.
There is already a test which has some mask values set to false in the same file, so I am not testing it for general cases. This all ones test has a copy in attention_i1_mask.mlir, the purpose is to test that the results are exactly the same under both memory layout configurations.
There was a problem hiding this comment.
If we intend to set all values to true, why not just write something like dense<true>?
lialan
force-pushed
the
lialan/attention_i1_e2e
branch
from
372984a to
f2587f5
Compare
To test actual i1 handling with attention op. Signed-off-by: Alan Li <me@alanli.org>
lialan
force-pushed
the
lialan/attention_i1_e2e
branch
from
f2587f5 to
b26f268
Compare
There was a problem hiding this comment.
It is better to name it to attention_i1_packed_mask.mlir, which clearly make a distinction between this and the test in the other file.
--iree-experimental-packed-i1-storageturned on, which allows real packed i1 datatype in memory.