[ROCm][Perf] Drop two dead per-step ops from the AITER sparse MLA metadata build (#58354)
Signed-off-by: ZhengGong-amd <zgong@amd.com> Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
@@ -223,7 +223,7 @@ def _patch_build_deps(monkeypatch, events=None):
|
||||
on CPU."""
|
||||
|
||||
def fake_generate_sparse_seqlen_triton(
|
||||
query_lens, seq_lens, cu_query_lens, topk_token, num_tokens, max_query_len
|
||||
seq_lens, cu_query_lens, topk_token, num_tokens, max_query_len
|
||||
):
|
||||
return torch.zeros(num_tokens, dtype=torch.int32, device="cpu")
|
||||
|
||||
@@ -276,6 +276,25 @@ def test_build_populates_mixed_split_fields(monkeypatch):
|
||||
assert md.prefill is None
|
||||
|
||||
|
||||
def test_build_fills_paged_kv_indptr_tail(monkeypatch):
|
||||
"""Rows past num_tokens must read as empty: the tail repeats the total."""
|
||||
builder = _make_builder()
|
||||
_patch_build_deps(monkeypatch)
|
||||
monkeypatch.setattr(
|
||||
sparse_mod,
|
||||
"generate_sparse_seqlen_triton",
|
||||
lambda *args, **kwargs: torch.tensor([3, 5], dtype=torch.int32),
|
||||
)
|
||||
builder.paged_kv_indptr[1:].fill_(-1)
|
||||
|
||||
md = builder.build(
|
||||
common_prefix_len=0, common_attn_metadata=_make_common_metadata()
|
||||
)
|
||||
|
||||
assert md.paged_kv_indptr.tolist() == [0, 3, 8]
|
||||
assert builder.paged_kv_indptr.tolist() == [0, 3, 8, 8, 8, 8, 8, 8, 8]
|
||||
|
||||
|
||||
def test_sink_build_skips_persistent_metadata(monkeypatch):
|
||||
builder = _make_builder()
|
||||
builder._use_persistent_metadata = False
|
||||
@@ -385,10 +404,10 @@ def test_dense_prefill_builds_sparse_metadata_for_decode_subset(monkeypatch):
|
||||
sparse_call: dict[str, torch.Tensor] = {}
|
||||
|
||||
def fake_generate_sparse_seqlen(
|
||||
query_lens, sparse_seq_lens, sparse_query_start_loc, topk, num_tokens, max_q
|
||||
sparse_seq_lens, sparse_query_start_loc, topk, num_tokens, max_q
|
||||
):
|
||||
sparse_call.update(
|
||||
query_lens=query_lens.clone(),
|
||||
query_lens=torch.diff(sparse_query_start_loc),
|
||||
seq_lens=sparse_seq_lens.clone(),
|
||||
query_start_loc=sparse_query_start_loc.clone(),
|
||||
num_tokens=num_tokens,
|
||||
@@ -471,10 +490,10 @@ def test_long_prefill_shares_full_sparse_metadata_with_decode(monkeypatch):
|
||||
sparse_call: dict[str, torch.Tensor] = {}
|
||||
|
||||
def fake_generate_sparse_seqlen(
|
||||
query_lens, sparse_seq_lens, sparse_query_start_loc, topk, num_tokens, max_q
|
||||
sparse_seq_lens, sparse_query_start_loc, topk, num_tokens, max_q
|
||||
):
|
||||
sparse_call.update(
|
||||
query_lens=query_lens.clone(),
|
||||
query_lens=torch.diff(sparse_query_start_loc),
|
||||
seq_lens=sparse_seq_lens.clone(),
|
||||
query_start_loc=sparse_query_start_loc.clone(),
|
||||
num_tokens=num_tokens,
|
||||
@@ -699,7 +718,7 @@ def test_sparse_metadata_workload_length_dtype_matrix(
|
||||
monkeypatch.setattr(
|
||||
sparse_mod,
|
||||
"generate_sparse_seqlen_triton",
|
||||
lambda *args, **kwargs: torch.ones(args[4], dtype=torch.int32),
|
||||
lambda *args, **kwargs: torch.ones(args[3], dtype=torch.int32),
|
||||
)
|
||||
get_persistent_metadata = Mock()
|
||||
fake_aiter = _FakeAiter("aiter")
|
||||
|
||||
Reference in New Issue
Block a user