Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

【CINN】Add DSL test case for rope of llama #61130

Merged
merged 2 commits into from
Jan 25, 2024
Merged

【CINN】Add DSL test case for rope of llama #61130

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
002d3e1
add slice
Courtesy-Xs Jan 24, 2024
d7b1afa
rename test case and use int64 to indexing
Courtesy-Xs Jan 24, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
118 changes: 117 additions & 1 deletion test/cinn/ir/test_llir_schedule_compute_inline.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,7 +14,7 @@

from test.cinn.utils.testing import assert_llir_equal

from cinn import ir, to_cinn_llir
from cinn import common, ir, to_cinn_llir
from cinn.runtime.data_array import DataArray
from cinn.schedule import IRSchedule as sch

Expand Down Expand Up @@ -165,8 +165,124 @@ def elementwise_add_inline_gt(
assert_llir_equal(elementwise_add_inline, elementwise_add_inline_gt)


def test_gather_slice_concat_mul_pattern_compute_inline():
class origin:
@to_cinn_llir
def fn_rotary(
A: DataArray((1, -1, 1, 512)),
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Add another test case that replaces the first 1 to 64?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We should replace it with a symbol indeed, but dsl don't support it, so temporarily we use a constant shape 1, maybe this can be a TODO

var_1: DataArray((1, -1, 1, 256)),
var_2: DataArray((1, -1, 1, 256)),
var_3: DataArray((1, -1, 1, 512)),
C: DataArray((1, -1, 1, 512)),
index: DataArray((2, 1), dtype=common.Int(64)),
Out: DataArray((1, -1, 1, 512)),
seq_len: ir.Var(),
):
for i in range(0, 1):
for j in range(0, seq_len):
for k in range(0, 1):
for a in range(0, 256):
with ir.ScheduleBlockContext(
"var_1"
) as var_1_block:
i0, i1, i2, i3 = ir.AxisMap(
"SSSS", [i, j, k, a]
)
var_1[i0, i1, i2, i3] = A[
ir.Expr(index[i0, ir.Expr(0)]), i1, i2, i3
]
for i in range(0, 1):
for j in range(0, seq_len):
for k in range(0, 1):
for a in range(0, 256):
with ir.ScheduleBlockContext(
"var_2"
) as var_2_block:
i0, i1, i2, i3 = ir.AxisMap(
"SSSS", [i, j, k, a]
)
var_2[i0, i1, i2, i3] = A[
ir.Expr(index[i0 + ir.Expr(1), ir.Expr(0)]),
i1,
i2,
ir.Expr(256) + i3,
]
for i in range(0, 1):
for j in range(0, seq_len):
for k in range(0, 1):
for a in range(0, 512):
with ir.ScheduleBlockContext(
"var_3"
) as var_3_block:
i0, i1, i2, i3 = ir.AxisMap(
"SSSS", [i, j, k, a]
)
var_3[i0, i1, i2, i3] = ir.Select.make(
(i3 < 256),
var_1[i0, i1, i2, i3],
var_2[i0, i1, i2, (i3 - ir.Expr(256))],
)
for i in range(0, 1):
for j in range(0, seq_len):
for k in range(0, 1):
for a in range(0, 512):
with ir.ScheduleBlockContext("out") as out_block:
i0, i1, i2, i3 = ir.AxisMap(
"SSSS", [i, j, k, a]
)
Out[i0, i1, i2, i3] = (
C[i0, i1, i2, i3] * var_3[i0, i1, i2, i3]
)
block_var_1 = sch.get_block("var_1")
sch.compute_inline(block_var_1)
block_var_2 = sch.get_block("var_2")
sch.compute_inline(block_var_2)
block_var_3 = sch.get_block("var_3")
sch.compute_inline(block_var_3)

class expected:
@to_cinn_llir
def fn_rotary(
A: DataArray((1, -1, 1, 512)),
C: DataArray((1, -1, 1, 512)),
index: DataArray((2, 1), dtype=common.Int(64)),
Out: DataArray((1, -1, 1, 512)),
seq_len: ir.Var(),
):
for i in range(0, 1):
for j in range(0, seq_len):
for k in range(0, 1):
for a in range(0, 512):
with ir.ScheduleBlockContext("out") as out_block:
i0, i1, i2, i3 = ir.AxisMap(
"SSSS", [i, j, k, a]
)
Out[i0, i1, i2, i3] = C[
i0, i1, i2, i3
] * ir.Select.make(
(i3 < 256),
A[
ir.Expr(index[i0, ir.Expr(0)]),
i1,
i2,
i3,
],
A[
ir.Expr(
index[(i0 + ir.Expr(1)), ir.Expr(0)]
),
i1,
i2,
ir.Expr(256) + (i3 - ir.Expr(256)),
],
)

assert_llir_equal(origin.fn_rotary, expected.fn_rotary)


if __name__ == "__main__":
test_compute_inline_elementwise()
test_reverse_compute_inline_elementwise()
test_compute_inline_elementwise_dynamic()
test_reverse_compute_inline_elementwise_dynamic()
test_gather_slice_concat_mul_pattern_compute_inline()