[ROCm][MiniMax-M3] Add AITER sparse paged attention (#47287)

Signed-off-by: Tan Pin Siang <tanpinsiang@gmail.com>
Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com>
Co-authored-by: vllmellm <vllm.ellm@embeddedllm.com>
Co-authored-by: Hongxia Yang <hongxia.yang@amd.com>
Co-authored-by: Jun Kang Chow <junkangchow@gmail.com>
Co-authored-by: tjtanaa <tunjian.tan@embeddedllm.com>
This commit is contained in:
Tan Pin Siang
2026-07-12 19:27:29 -07:00
committed by GitHub
co-authored by vllmellm Hongxia Yang Jun Kang Chow tjtanaa
parent e26264f3ef
commit ee5a89f4d7
15 changed files with 1554 additions and 128 deletions
@@ -820,6 +820,47 @@ def test_main_backend_layout_contract():
)
def test_aiter_sparse_pa_layout_contract(monkeypatch):
"""The shuffle-only AITER path retains separately contiguous K/V storage."""
import vllm.models.minimax_m3.common.sparse_attention as sparse_attn_mod
monkeypatch.setattr(sparse_attn_mod.rocm_aiter_ops, "is_enabled", lambda: True)
monkeypatch.setattr(
sparse_attn_mod.rocm_aiter_ops,
"is_shuffle_kv_cache_enabled",
lambda: True,
)
nb, bs, h, d = 7, BLOCK_SIZE, 1, HEAD_DIM
logical = MiniMaxM3SparseBackend.get_kv_cache_shape(nb, bs, h, d)
order = MiniMaxM3SparseBackend.get_kv_cache_stride_order()
assert logical == (nb, 2, bs, h, d)
assert order == (1, 0, 2, 3, 4)
physical_shape = tuple(logical[i] for i in order)
inv_order = [order.index(i) for i in range(len(order))]
raw = torch.empty(physical_shape, device="cuda", dtype=DTYPE)
logical_view = raw.permute(*inv_order)
key_cache, value_cache = logical_view.unbind(1)
assert key_cache.is_contiguous()
assert value_cache.is_contiguous()
def test_aiter_sparse_pa_rejects_multiple_kv_heads(monkeypatch):
"""Do not pair AITER's separated cache layout with the Triton fallback."""
import vllm.models.minimax_m3.common.sparse_attention as sparse_attn_mod
monkeypatch.setattr(sparse_attn_mod.rocm_aiter_ops, "is_enabled", lambda: True)
monkeypatch.setattr(
sparse_attn_mod.rocm_aiter_ops,
"is_shuffle_kv_cache_enabled",
lambda: True,
)
with pytest.raises(ValueError, match="num_kv_heads == 1"):
MiniMaxM3SparseBackend.get_kv_cache_shape(7, BLOCK_SIZE, 2, HEAD_DIM)
def test_main_backend_unknown_layout_raises(monkeypatch):
"""An unrecognized layout (injected past env-var validation) is rejected."""
import vllm.models.minimax_m3.common.sparse_attention as sparse_attn_mod