[ROCm][Spec Decode] Add Aiter MLA decode support non-causal draft block (#55966)

Signed-off-by: Poovaiah Palangappa <poovaiah.palangappa@amd.com>
Co-authored-by: Poovaiah Palangappa <poovaiah.palangappa@amd.com>
Co-authored-by: John Qin <zhengsheng1997@gmail.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Hongxia Yang <62075498+hongxiayang@users.noreply.github.com>
This commit is contained in:
ppalanga
2026-09-15 12:18:09 -05:00
committed by GitHub
co-authored by Poovaiah Palangappa John Qin Cursor Hongxia Yang
parent 241e9391ed
commit 35dc273072
5 changed files with 390 additions and 5 deletions
@@ -30,7 +30,10 @@ def reset_aiter_mla_support_cache(monkeypatch: pytest.MonkeyPatch) -> None:
def _install_fake_aiter_modules(
monkeypatch: pytest.MonkeyPatch, *, supports_fp8: bool
monkeypatch: pytest.MonkeyPatch,
*,
supports_fp8: bool,
supports_causal: bool = False,
) -> None:
aiter_mod: Any = types.ModuleType("aiter")
mla_mod: Any = types.ModuleType("aiter.mla")
@@ -65,6 +68,14 @@ def _install_fake_aiter_modules(
mla_decode_fwd = mla_decode_fwd_without_fp8
if supports_causal:
inner = mla_decode_fwd
def mla_decode_fwd_with_causal(*args, causal=True, **kwargs):
return inner(*args, **kwargs)
mla_decode_fwd = mla_decode_fwd_with_causal
mla_mod.mla_decode_fwd = mla_decode_fwd
aiter_mod.mla = mla_mod
@@ -135,3 +146,107 @@ def test_aiter_mla_fp8_support_result_is_cached(monkeypatch):
assert _check_aiter_mla_fp8_support() is True
assert _check_aiter_mla_fp8_support() is True
assert signature_mock.call_count == 1
def test_missing_causal_arg_fails_closed(monkeypatch):
"""Old aiter wheels have no causal=; do not report that as a valid draft path."""
from vllm._aiter_ops import rocm_aiter_ops
from vllm.v1.attention.backends.mla.rocm_aiter_mla import AiterMLABackend
_install_fake_aiter_modules(monkeypatch, supports_fp8=True, supports_causal=False)
with pytest.raises(RuntimeError, match="causal-only"):
rocm_aiter_ops.mla_decode_supports_non_causal()
with pytest.raises(RuntimeError, match="causal-only"):
AiterMLABackend.supports_non_causal()
@pytest.mark.parametrize("has_kernels", [False, True])
def test_supports_non_causal_requires_arch_kernels(monkeypatch, has_kernels):
"""The Python argument alone does not guarantee an arch has the kernel."""
from vllm._aiter_ops import rocm_aiter_ops
from vllm.v1.attention.backends.mla import rocm_aiter_mla
from vllm.v1.attention.backends.mla.rocm_aiter_mla import AiterMLABackend
_install_fake_aiter_modules(monkeypatch, supports_fp8=True, supports_causal=True)
monkeypatch.setattr(
rocm_aiter_mla,
"_aiter_mla_non_causal_asm_kernels",
lambda: has_kernels,
)
assert rocm_aiter_ops.mla_decode_supports_non_causal() is True
assert AiterMLABackend.supports_non_causal() is has_kernels
@pytest.mark.parametrize("causal", [True, False])
def test_decode_fwd_forwards_causal_to_aiter(monkeypatch, causal):
"""The decode op must pass the block mask through; omitting it is causal-only."""
seen: list = []
def mla_decode_fwd(*args, causal=None, **kwargs):
seen.append(causal)
mla_mod: Any = types.ModuleType("aiter.mla")
mla_mod.mla_decode_fwd = mla_decode_fwd
aiter_mod: Any = types.ModuleType("aiter")
aiter_mod.mla = mla_mod
monkeypatch.setitem(sys.modules, "aiter", aiter_mod)
monkeypatch.setitem(sys.modules, "aiter.mla", mla_mod)
import torch
import vllm._aiter_ops as aiter_ops
aiter_ops._rocm_aiter_mla_decode_fwd_impl(
torch.zeros(2, 1, 8),
torch.zeros(2, 1, 1, 8),
torch.zeros(2, 1, 8),
torch.zeros(3, dtype=torch.int32),
1,
causal=causal,
)
assert seen == [causal]
def _aiter_mla_validate(
monkeypatch,
*,
dtype,
kv_cache_dtype,
use_non_causal,
):
from vllm.platforms.interface import DeviceCapability
from vllm.v1.attention.backend import AttentionType
from vllm.v1.attention.backends.mla import rocm_aiter_mla
from vllm.v1.attention.backends.mla.rocm_aiter_mla import AiterMLABackend
_install_fake_aiter_modules(monkeypatch, supports_fp8=True, supports_causal=True)
monkeypatch.setattr(
rocm_aiter_mla, "_aiter_mla_non_causal_asm_kernels", lambda: True
)
return AiterMLABackend.validate_configuration(
head_size=576,
dtype=dtype,
kv_cache_dtype=kv_cache_dtype,
block_size=1,
use_mla=True,
has_sink=False,
use_sparse=False,
use_mm_prefix=False,
use_per_head_quant_scales=False,
device_capability=DeviceCapability(9, 5),
attn_type=AttentionType.DECODER,
use_non_causal=use_non_causal,
)
def test_non_causal_fp16_auto_cache_is_rejected(monkeypatch):
"""Pinned AITER aborts fp16 Q; auto-select must fall through to Triton."""
import torch
reasons = _aiter_mla_validate(
monkeypatch,
dtype=torch.float16,
kv_cache_dtype="auto",
use_non_causal=True,
)
assert any("non-causal fp16" in reason for reason in reasons)
@@ -176,3 +176,8 @@ def test_pad_unpad_round_trip_preserves_head_order(num_heads):
assert unpadded.shape == q.shape
torch.testing.assert_close(unpadded, q)
def test_a_non_causal_block_never_routes_to_gluon(gluon_available):
"""A small-head block that normally uses Gluon must use ASM when non-causal."""
assert not AiterMLAHelper.use_gluon_verify(12, 8, "auto", causal=False)
@@ -130,6 +130,7 @@ def _builder(
_mla_q_dtype=torch.bfloat16,
_mla_kv_dtype=torch.bfloat16,
decode_attn_out_dtype=torch.bfloat16,
_decode_causal=True,
)
# Bound to the stub rather than faked: the verify flatten's per-row view is
# part of what _build_decode is being tested for.
@@ -155,6 +156,25 @@ def test_backend_declares_uniform_batch_support():
)
@pytest.mark.parametrize("causal", [True, False])
def test_build_hands_common_causal_flag_to_aiter_builder(monkeypatch, causal):
seen = []
def fake_common_build(self, *args):
seen.append(self._decode_causal)
return SimpleNamespace(decode=None, prefill=None)
monkeypatch.setattr(
rocm_aiter_mla.MLACommonMetadataBuilder, "build", fake_common_build
)
builder = object.__new__(AiterMLAMetadataBuilder)
builder._fp8_prefill_enabled = False
builder.build(0, SimpleNamespace(causal=causal))
assert seen == [causal]
def test_dcp_verify_row_view_is_causal_per_row():
"""Row t must cover this rank's share of positions [0, seq_len - qlen + t].
@@ -949,3 +969,81 @@ def test_persistent_metadata_gate_without_gluon_build(
assert metadata.has_persistent_metadata is expect_persistent
assert get_mla_metadata_v1.called is expect_persistent
def _build_non_causal(monkeypatch, *, num_heads, kv_cache_dtype, qlen, mtp_qlen):
"""Drive _build_decode with causal=False and hand back the aiter mock."""
get_mla_metadata_v1 = mock.MagicMock()
monkeypatch.setitem(
sys.modules,
"aiter",
SimpleNamespace(get_mla_metadata_v1=get_mla_metadata_v1),
)
monkeypatch.setattr(
rocm_aiter_mla, "_expand_page_indices_kernel", _NoOpTritonKernel()
)
builder = _builder(
mtp_decode_qlen=mtp_qlen,
kv_cache_dtype=kv_cache_dtype,
num_heads=num_heads,
)
builder._decode_causal = False
metadata = AiterMLAMetadataBuilder._build_decode(
builder,
block_table_tensor=torch.arange(16, dtype=torch.int32).view(2, 8),
seq_lens_device=torch.tensor([7, 5], dtype=torch.int32),
max_seq_len=7,
query_start_loc_cpu=torch.tensor([0, qlen, 2 * qlen], dtype=torch.int32),
query_start_loc_device=torch.tensor([0, qlen, 2 * qlen], dtype=torch.int32),
num_decode_tokens=2 * qlen,
dcp_tot_seq_lens_device=None,
)
return metadata, get_mla_metadata_v1
def test_non_causal_build_hands_the_mask_to_aiter(monkeypatch):
"""The schedule carries the mask, so is_causal must follow the block."""
metadata, get_mla_metadata_v1 = _build_non_causal(
monkeypatch, num_heads=12, kv_cache_dtype="auto", qlen=8, mtp_qlen=8
)
assert metadata.has_persistent_metadata
# 6th positional arg of get_mla_metadata_v1 is is_causal.
assert get_mla_metadata_v1.call_args.args[5] is False
@pytest.mark.parametrize("num_heads", [12, 16, 48])
def test_a_two_token_fp8_block_is_refused_without_a_fold(monkeypatch, num_heads):
"""16-head (and padded-to-16) qlen-2 fp8 has no non-causal kernel.
48/80/112 (and 33 padded to 48) fold to H16 while keeping Q2, so they
must be refused the same way as native 16-head qlen 2.
"""
with pytest.raises(ValueError, match="2-token"):
_build_non_causal(
monkeypatch,
num_heads=num_heads,
kv_cache_dtype="fp8",
qlen=2,
mtp_qlen=8,
)
def test_a_two_token_fp8_block_is_allowed_when_folded(monkeypatch):
"""32 heads at qlen 2 folds onto the 16-head / 4-token kernel."""
metadata, get_mla_metadata_v1 = _build_non_causal(
monkeypatch,
num_heads=32,
kv_cache_dtype="fp8",
qlen=2,
mtp_qlen=8,
)
assert metadata.has_persistent_metadata
assert get_mla_metadata_v1.called
def test_a_two_token_bf16_block_is_allowed(monkeypatch):
"""The bf16 fold has no such hole, so the guard must not fire there."""
metadata, _ = _build_non_causal(
monkeypatch, num_heads=12, kv_cache_dtype="auto", qlen=2, mtp_qlen=8
)
assert metadata.has_persistent_metadata