[AMD] Keep the FP4 varctx schedule test off AITER's 2048-row compile (#41793)
This commit is contained in:
@@ -497,14 +497,25 @@ def test_decode_cta_count_stays_within_available_chunks(
|
||||
assert cta_count <= max(1024, num_queries * 4)
|
||||
|
||||
|
||||
@pytest.mark.parametrize("num_queries", [1, 1536])
|
||||
def test_decode_schedule_matches_aiter_varctx_schedule(num_queries: int) -> None:
|
||||
@pytest.mark.parametrize("num_queries", [1, 200])
|
||||
def test_decode_schedule_matches_aiter_varctx_schedule(
|
||||
num_queries: int, monkeypatch: pytest.MonkeyPatch
|
||||
) -> None:
|
||||
"""The in-tree decode schedule writes AITER compute_varctx_schedule's cta_info rows and
|
||||
split factor, including zero-length rows and prefix sums carried across row blocks."""
|
||||
split factor, including zero-length rows and prefix sums carried across row blocks.
|
||||
|
||||
The row block is shrunk so 200 rows span four blocks. Crossing the production block
|
||||
(1024) needs over a thousand rows, and AITER's reference kernel holds a [256, rows]
|
||||
tile; at 2048 it spent ~9 minutes compiling on the ROCm 10 image."""
|
||||
from aiter.ops.flydsl.kernels.mqa_logits.pa_mqa_logits_fp4 import (
|
||||
compute_varctx_schedule,
|
||||
)
|
||||
|
||||
monkeypatch.setattr(
|
||||
"sglang.kernels.ops.attention.dsv4.fp4_indexer_schedule_hip._DECODE_ROW_BLOCK",
|
||||
64,
|
||||
)
|
||||
|
||||
torch.manual_seed(num_queries)
|
||||
max_seq_len = 65536
|
||||
c4_seq_lens = torch.randint(
|
||||
|
||||
Reference in New Issue
Block a user