[Mamba] Add FlashInfer ReplaySSM support for MTP (#52928)

Signed-off-by: Andrii Skliar <askliar@nvidia.com>
Co-authored-by: Andrii Skliar <askliar@nvidia.com>
This commit is contained in:
Andrii Skliar
2026-09-30 21:41:17 +00:00
committed by GitHub
co-authored by Andrii Skliar
parent 7e583e615c
commit deb148e730
18 changed files with 1029 additions and 113 deletions
+393 -5
View File
@@ -1,24 +1,30 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from types import SimpleNamespace
from unittest.mock import Mock
import pytest
import torch
from vllm.config.mamba import MambaBackendEnum, MambaConfig, MambaSSUAlgorithm
from vllm.model_executor.layers.mamba.mamba_mixer2 import MambaMixer2
from vllm.model_executor.layers.mamba.mamba_utils import MambaStateShapeCalculator
from vllm.model_executor.layers.mamba.ops.ssu_dispatch import (
FlashInferSSUBackend,
TritonSSUBackend,
commit_replayssm_ring_trackers,
get_mamba_ssu_backend,
initialize_mamba_ssu_backend,
reset_replayssm_ring_trackers,
selective_state_update,
selective_state_update_replayssm_flashinfer,
update_replayssm_ring_trackers,
)
from vllm.utils.torch_utils import set_random_seed
from vllm.v1.attention.backends.mamba2_attn import Mamba2AttentionMetadata
from vllm.v1.attention.backends.registry import MambaAttentionBackendEnum
from vllm.v1.attention.backends.utils import NULL_BLOCK_ID
from vllm.v1.kv_cache_interface import (
KVCacheConfig,
KVCacheGroupSpec,
@@ -32,6 +38,13 @@ try:
except ImportError:
HAS_FLASHINFER = False
try:
from flashinfer.mamba.checkpointing_ssu import CheckpointingSSURunner
HAS_FLASHINFER_CHECKPOINTING_SSU = CheckpointingSSURunner is not None
except ImportError:
HAS_FLASHINFER_CHECKPOINTING_SSU = False
@pytest.fixture(autouse=True)
def restore_backend_state():
@@ -47,6 +60,7 @@ def restore_backend_state():
def test_flashinfer_replayssm_ring_tracker_lifecycle():
ring_start = torch.zeros(2, dtype=torch.int32, device="cuda")
prev_num_accepted = torch.zeros(2, dtype=torch.int32, device="cuda")
prev_query_len = torch.zeros(2, dtype=torch.int32, device="cuda")
state_batch_indices = torch.tensor([1], dtype=torch.int32, device="cuda")
observed = []
@@ -54,6 +68,7 @@ def test_flashinfer_replayssm_ring_tracker_lifecycle():
update_replayssm_ring_trackers(
ring_start,
prev_num_accepted,
prev_query_len,
state_batch_indices,
logical_window=16,
ring_buffer_len=17,
@@ -69,9 +84,205 @@ def test_flashinfer_replayssm_ring_tracker_lifecycle():
reset_replayssm_ring_trackers(
ring_start,
prev_num_accepted,
prev_query_len,
state_batch_indices,
)
assert (ring_start[1].item(), prev_num_accepted[1].item()) == (0, 0)
assert (
ring_start[1].item(),
prev_num_accepted[1].item(),
prev_query_len[1].item(),
) == (0, 0, 0)
@pytest.mark.parametrize(
("accepted_sequence", "expected"),
[
pytest.param(
[4] * 22,
[
(0, 0, 4),
(0, 4, 4),
(0, 8, 4),
(0, 12, 4),
(0, 16, 4),
(16, 4, 4),
(16, 8, 4),
(16, 12, 4),
(16, 16, 4),
(12, 4, 4),
(12, 8, 4),
(12, 12, 4),
(12, 16, 4),
(8, 4, 4),
(8, 8, 4),
(8, 12, 4),
(8, 16, 4),
(4, 4, 4),
(4, 8, 4),
(4, 12, 4),
(4, 16, 4),
(0, 4, 4),
],
id="all-accepted",
),
pytest.param(
[4, 4, 0, 3, 4, 2, 4, 1],
[
(0, 0, 4),
(0, 4, 4),
(0, 4, 4),
(0, 7, 4),
(0, 11, 4),
(0, 13, 4),
(13, 4, 4),
(13, 5, 4),
],
id="mixed",
),
],
)
def test_replayssm_commit_tracker_acceptance_sequence(accepted_sequence, expected):
logical_window = 16
num_speculative_tokens = 3
query_len = 1 + num_speculative_tokens
ring_buffer_len = logical_window + 1 + num_speculative_tokens
ring_start = torch.zeros(2, dtype=torch.int32, device="cuda")
prev_num_accepted = torch.zeros(2, dtype=torch.int32, device="cuda")
prev_query_len = torch.zeros(2, dtype=torch.int32, device="cuda")
state_batch_indices = torch.tensor([1], dtype=torch.int32, device="cuda")
query_start_loc = torch.tensor([0, query_len], dtype=torch.int32, device="cuda")
observed = []
for accepted in accepted_sequence:
commit_replayssm_ring_trackers(
ring_start,
prev_num_accepted,
prev_query_len,
state_batch_indices,
torch.tensor([accepted], dtype=torch.int32, device="cuda"),
query_start_loc,
logical_window,
ring_buffer_len,
)
snapshot = (
ring_start[1].item(),
prev_num_accepted[1].item(),
prev_query_len[1].item(),
)
observed.append(snapshot)
assert snapshot[1] + snapshot[2] <= ring_buffer_len
assert observed == expected
def test_replayssm_resume_resets_commit_history():
ring_start = torch.tensor([0, 13], dtype=torch.int32, device="cuda")
prev_num_accepted = torch.tensor([0, 13], dtype=torch.int32, device="cuda")
prev_query_len = torch.tensor([0, 4], dtype=torch.int32, device="cuda")
state_batch_indices = torch.tensor([1], dtype=torch.int32, device="cuda")
reset_replayssm_ring_trackers(
ring_start,
prev_num_accepted,
prev_query_len,
state_batch_indices,
)
assert (
ring_start[1].item(),
prev_num_accepted[1].item(),
prev_query_len[1].item(),
) == (0, 0, 0)
commit_replayssm_ring_trackers(
ring_start,
prev_num_accepted,
prev_query_len,
state_batch_indices,
torch.tensor([3], dtype=torch.int32, device="cuda"),
torch.tensor([0, 4], dtype=torch.int32, device="cuda"),
logical_window=16,
ring_buffer_len=20,
)
assert (
ring_start[1].item(),
prev_num_accepted[1].item(),
prev_query_len[1].item(),
) == (0, 0, 4)
def test_replayssm_commit_tracker_ragged_query_lengths():
ring_start = torch.zeros(3, dtype=torch.int32, device="cuda")
prev_num_accepted = torch.zeros(3, dtype=torch.int32, device="cuda")
prev_query_len = torch.zeros(3, dtype=torch.int32, device="cuda")
state_batch_indices = torch.tensor([1, 2], dtype=torch.int32, device="cuda")
query_start_loc = torch.tensor([0, 4, 6], dtype=torch.int32, device="cuda")
observed = []
for accepted in ([4, 2], [3, 1]):
commit_replayssm_ring_trackers(
ring_start,
prev_num_accepted,
prev_query_len,
state_batch_indices,
torch.tensor(accepted, dtype=torch.int32, device="cuda"),
query_start_loc,
logical_window=16,
ring_buffer_len=20,
)
observed.append(
[
(
ring_start[slot].item(),
prev_num_accepted[slot].item(),
prev_query_len[slot].item(),
)
for slot in (1, 2)
]
)
assert observed == [[(0, 0, 4), (0, 0, 2)], [(0, 3, 4), (0, 1, 2)]]
@pytest.mark.parametrize("operation", ["commit", "reset"])
def test_replayssm_tracker_kernels_mask_invalid_slots(operation):
num_states = 3
ring_start = torch.tensor([11, 2, 33], dtype=torch.int32, device="cuda")
prev_num_accepted = torch.tensor([11, 3, 33], dtype=torch.int32, device="cuda")
prev_query_len = torch.tensor([11, 4, 33], dtype=torch.int32, device="cuda")
state_batch_indices = torch.tensor(
[-1, num_states, NULL_BLOCK_ID, 1], dtype=torch.int32, device="cuda"
)
if operation == "commit":
commit_replayssm_ring_trackers(
ring_start,
prev_num_accepted,
prev_query_len,
state_batch_indices,
torch.tensor([4, 4, 4, 2], dtype=torch.int32, device="cuda"),
torch.tensor([0, 4, 8, 12, 16], dtype=torch.int32, device="cuda"),
logical_window=16,
ring_buffer_len=20,
)
expected_valid = (2, 5, 4)
else:
reset_replayssm_ring_trackers(
ring_start,
prev_num_accepted,
prev_query_len,
state_batch_indices,
)
expected_valid = (0, 0, 0)
assert (
ring_start.tolist(),
prev_num_accepted.tolist(),
prev_query_len.tolist(),
) == (
[11, expected_valid[0], 33],
[11, expected_valid[1], 33],
[11, expected_valid[2], 33],
)
def _kv_cache_config_with_ssu(
@@ -231,14 +442,190 @@ def test_triton_basic_call():
assert not torch.isnan(out).any()
@pytest.mark.parametrize("layout", ["packed", "dense"])
def test_replayssm_flashinfer_call_forwards_mtp_layout(monkeypatch, layout):
import vllm.model_executor.layers.mamba.ops.ssu_dispatch as mod
kernel = Mock(return_value=torch.empty(0))
monkeypatch.setattr(mod, "_flashinfer_replayssm_kernel", kernel)
batch, max_seqlen, nheads, dim, dstate, ngroups = 2, 4, 2, 4, 8, 1
state = torch.empty(2, nheads, dim, dstate)
x_shape: tuple[int, ...]
B_shape: tuple[int, ...]
expected_x_shape: tuple[int, ...]
expected_B_shape: tuple[int, ...]
if layout == "packed":
x_shape = (6, nheads, dim)
B_shape = (6, ngroups, dstate)
expected_x_shape = (1, 6, nheads, dim)
expected_B_shape = (1, 6, ngroups, dstate)
cu_seqlens = torch.tensor([0, 4, 6], dtype=torch.int32)
kernel_max_seqlen = max_seqlen
else:
x_shape = (batch, max_seqlen, nheads, dim)
B_shape = (batch, max_seqlen, ngroups, dstate)
expected_x_shape = x_shape
expected_B_shape = B_shape
cu_seqlens = None
kernel_max_seqlen = None
x = torch.empty(x_shape)
dt = torch.empty_like(x)
A = torch.empty(nheads, dim, dstate)
B = torch.empty(B_shape)
C = torch.empty_like(B)
out = torch.empty_like(x)
x_cache = torch.empty(2, nheads, 20, dim)
dt_cache = torch.empty(2, nheads, 20)
B_cache = torch.empty(2, ngroups, 20, dstate)
ring_start = torch.zeros(2, dtype=torch.int32)
prev_num_accepted = torch.zeros(2, dtype=torch.int32)
prev_query_len = torch.zeros(2, dtype=torch.int32)
selective_state_update_replayssm_flashinfer(
state,
x,
dt,
A,
B,
C,
out,
x_cache,
B_cache,
dt_cache,
ring_start,
prev_num_accepted,
prev_query_len,
logical_window=16,
state_batch_indices=torch.tensor([0, 1], dtype=torch.int32),
cu_seqlens=cu_seqlens,
max_seqlen=kernel_max_seqlen,
update_trackers=False,
)
args = kernel.call_args.args
assert args[6].shape == expected_x_shape
assert args[7].shape == expected_x_shape
assert args[9].shape == expected_B_shape
assert args[10].shape == expected_B_shape
assert args[11].shape == expected_x_shape
assert kernel.call_args.kwargs["cu_seqlens"] is cu_seqlens
assert kernel.call_args.kwargs["max_seqlen"] == kernel_max_seqlen
@pytest.mark.parametrize(
("backend", "expected_ring_len"),
("query_start_loc", "expected_shape", "expected_max_seqlen"),
[
(MambaBackendEnum.TRITON, 16),
(MambaBackendEnum.FLASHINFER, 17),
pytest.param([0, 4, 8], (2, 4, 2, 4), None, id="dense"),
pytest.param([0, 4, 6], (6, 2, 4), 4, id="packed"),
],
)
def test_replayssm_physical_ring_shape(backend, expected_ring_len):
def test_replayssm_mixer_selects_mtp_layout(
monkeypatch, query_start_loc, expected_shape, expected_max_seqlen
):
import vllm.model_executor.layers.mamba.mamba_mixer2 as mod
mixer = MambaMixer2.__new__(MambaMixer2)
torch.nn.Module.__init__(mixer)
mixer.prefix = "mixer"
mixer.tped_intermediate_size = 0
mixer.tped_conv_size = 1
mixer.tped_dt_size = 2
mixer.num_heads = 2
mixer.head_dim = 4
mixer.n_groups = mixer.tp_size = 1
mixer.ssm_state_size = 8
mixer.num_spec = 3
mixer.use_replayssm = True
mixer.replayssm_buffer_len = 16
mixer._commits_replayssm_trackers = True
mixer._updates_replayssm_trackers = False
mixer.mamba_config = MambaConfig(backend=MambaBackendEnum.FLASHINFER)
mixer.cache_config = SimpleNamespace(mamba_block_size=16, mamba_cache_mode="none")
mixer.conv_weights = torch.empty(0)
mixer.conv1d = SimpleNamespace(bias=None)
mixer.activation = "silu"
mixer.A = torch.empty(2)
mixer.dt_bias = torch.empty(2)
mixer.D = torch.empty(2)
mixer._replayssm_ring_start = torch.zeros(3, dtype=torch.int32)
mixer._replayssm_prev_num_accepted = torch.zeros(3, dtype=torch.int32)
mixer._replayssm_prev_query_len = torch.zeros(3, dtype=torch.int32)
mixer.kv_cache = (
torch.empty(3, 1),
torch.empty(3, 2, 4, 8),
torch.empty(3, 2, 20, 4),
torch.empty(3, 2, 20),
torch.empty(3, 1, 20, 8),
)
num_decode_tokens = query_start_loc[-1]
query_start_loc_d = torch.tensor(query_start_loc, dtype=torch.int32)
metadata = Mamba2AttentionMetadata(
num_prefills=0,
num_prefill_tokens=0,
num_decodes=2,
num_decode_tokens=num_decode_tokens,
num_reqs=2,
has_initial_states_p=None,
query_start_loc_p=None,
num_computed_tokens_p=None,
state_indices_tensor_p=None,
state_indices_tensor_d=torch.tensor([[1], [2]], dtype=torch.int32),
query_start_loc_d=query_start_loc_d,
num_accepted_tokens=torch.tensor([4, 2], dtype=torch.int32),
block_idx_last_scheduled_token=None,
block_idx_first_scheduled_token_p=None,
block_idx_last_computed_token=None,
block_idx_last_scheduled_token_prev_step=None,
seq_lens=torch.tensor([104, 102], dtype=torch.int32),
replayssm_scratch=(torch.empty(0), torch.empty(0), torch.empty(0)),
replayssm_state_indices_d=torch.tensor([1, 2], dtype=torch.int32),
)
def split_hidden_states_B_C(values):
tokens = values.size(0)
return (
torch.empty(tokens, 8),
torch.empty(tokens, 8),
torch.empty(tokens, 8),
)
mixer.split_hidden_states_B_C_fn = split_hidden_states_B_C
kernel = Mock()
monkeypatch.setattr(
mod,
"get_forward_context",
lambda: SimpleNamespace(attn_metadata={mixer.prefix: metadata}),
)
monkeypatch.setattr(mod, "commit_replayssm_ring_trackers", Mock())
monkeypatch.setattr(
mod, "causal_conv1d_update", lambda values, *args, **kwargs: values
)
monkeypatch.setattr(mod, "selective_state_update_replayssm_flashinfer", kernel)
mixer.conv_ssm_forward(
torch.empty(num_decode_tokens, 3), torch.empty(num_decode_tokens, 8)
)
assert kernel.call_args.args[1].shape == expected_shape
if expected_max_seqlen is None:
assert kernel.call_args.kwargs["cu_seqlens"] is None
else:
assert kernel.call_args.kwargs["cu_seqlens"] is query_start_loc_d
assert kernel.call_args.kwargs["max_seqlen"] == expected_max_seqlen
@pytest.mark.parametrize(
("backend", "num_speculative_tokens", "expected_ring_len"),
[
(MambaBackendEnum.TRITON, 0, 16),
(MambaBackendEnum.FLASHINFER, 0, 17),
(MambaBackendEnum.FLASHINFER, 3, 20),
],
)
def test_replayssm_physical_ring_shape(
backend, num_speculative_tokens, expected_ring_len
):
base_shapes = ((64, 3), (8, 4, 16))
shapes = MambaStateShapeCalculator.append_replayssm_ring(
@@ -247,6 +634,7 @@ def test_replayssm_physical_ring_shape(backend, expected_ring_len):
tp_world_size=2,
logical_window=16,
backend=backend,
num_speculative_tokens=num_speculative_tokens,
)
assert shapes[2:] == (
+19 -4
View File
@@ -67,8 +67,13 @@ def test_replayssm_autotune_decode_kwargs(runner_kwargs, expected_num_reqs):
with patch.object(
warmup, "flashinfer_replayssm_autotune_supported", return_value=True
):
result = warmup._replayssm_autotune_kwargs(_autotune_runner(**runner_kwargs))
result = warmup._replayssm_autotune_kwargs(
_autotune_runner(**runner_kwargs), PREFILL_KWARGS
)
assert result is not None
max_num_reqs, decode_kwargs = result
assert max_num_reqs == expected_num_reqs
expected_kwargs = {
**PREFILL_KWARGS,
"num_tokens": expected_num_reqs * query_len,
@@ -82,7 +87,7 @@ def test_replayssm_autotune_decode_kwargs(runner_kwargs, expected_num_reqs):
force_attention=True,
profile_seq_lens=query_len + 1,
)
assert result == (expected_num_reqs, expected_kwargs)
assert decode_kwargs == expected_kwargs
@pytest.mark.parametrize(
@@ -91,8 +96,14 @@ def test_replayssm_autotune_decode_kwargs(runner_kwargs, expected_num_reqs):
(dict(use_replayssm=False), True),
(dict(backend=MambaBackendEnum.TRITON), True),
({}, False),
(dict(num_blocks=1), True),
],
ids=[
"replayssm_disabled",
"non_flashinfer_backend",
"kernel_unavailable",
"zero_non_padding_slots",
],
ids=["replayssm_disabled", "non_flashinfer_backend", "kernel_unavailable"],
)
def test_replayssm_autotune_kwargs_skipped(runner_kwargs, flashinfer_supported):
with patch.object(
@@ -100,7 +111,9 @@ def test_replayssm_autotune_kwargs_skipped(runner_kwargs, flashinfer_supported):
"flashinfer_replayssm_autotune_supported",
return_value=flashinfer_supported,
):
result = warmup._replayssm_autotune_kwargs(_autotune_runner(**runner_kwargs))
result = warmup._replayssm_autotune_kwargs(
_autotune_runner(**runner_kwargs), PREFILL_KWARGS
)
assert result is None
@@ -116,10 +129,12 @@ def test_replayssm_autotune_slots_restore_state_and_trackers():
)
mixer._replayssm_ring_start = torch.full((4,), 3, dtype=torch.int32)
mixer._replayssm_prev_num_accepted = torch.full((4,), 3, dtype=torch.int32)
mixer._replayssm_prev_query_len = torch.full((4,), 3, dtype=torch.int32)
tracked = (
*mixer.kv_cache,
mixer._replayssm_ring_start,
mixer._replayssm_prev_num_accepted,
mixer._replayssm_prev_query_len,
)
block_ids = np.arange(10, 14, dtype=np.int32).reshape(4, 1)
+4 -2
View File
@@ -153,6 +153,7 @@ def test_kda_recoverssm_derivation_is_revalidated():
use_replayssm=True,
use_kda_recoverssm=False,
mamba_cache_mode="none",
replayssm_buffer_len=16,
),
num_speculative_tokens=3,
model_config=SimpleNamespace(
@@ -184,8 +185,9 @@ def test_kda_recoverssm_derivation_is_revalidated():
config.cache_config.mamba_cache_mode = "none"
config.model_config.architecture = "NemotronHForCausalLM"
with pytest.raises(ValueError, match="only supported for Kimi-K3 KDA"):
VllmConfig.validate_mamba_cached_kernel(config)
config.mamba_config.backend = MambaBackendEnum.FLASHINFER
VllmConfig.validate_mamba_cached_kernel(config)
assert not config.cache_config.use_kda_recoverssm
config.model_config.architecture = "KimiLinearForCausalLM"
config.parallel_config.pipeline_parallel_size = 2
@@ -5,7 +5,9 @@
from types import SimpleNamespace
import pytest
import torch
from vllm.model_executor.layers.mamba.abstract import MambaBase
from vllm.model_executor.layers.mamba.linear.minimax_linear_attn import (
MiniMaxText01LinearAttention,
)
@@ -19,6 +21,32 @@ from vllm.v1.attention.backends.registry import MambaAttentionBackendEnum
from vllm.v1.attention.backends.short_conv_attn import ShortConvAttentionBackend
@pytest.mark.parametrize(("use_replayssm", "expected_blocks"), [(False, 3), (True, 0)])
def test_replayssm_does_not_reserve_speculative_state_blocks(
use_replayssm, expected_blocks
):
layer = SimpleNamespace(
get_state_shape=lambda: ((2,),),
get_state_dtype=lambda: (torch.float32,),
mamba_type=MambaAttentionBackendEnum.MAMBA2,
is_kv_cache_tp_replicated=False,
)
vllm_config = SimpleNamespace(
cache_config=SimpleNamespace(
mamba_block_size=1,
mamba_page_size_padded=None,
mamba_cache_mode="none",
use_replayssm=use_replayssm,
),
num_speculative_tokens=3,
)
spec = MambaBase.get_kv_cache_spec(layer, vllm_config)
assert spec is not None
assert spec.num_speculative_blocks == expected_blocks
@pytest.mark.parametrize(
"layer_class, init_kwargs, expected_backend, expected_mamba_type",
[
@@ -16,6 +16,8 @@ from tests.v1.attention.utils import (
create_common_attn_metadata,
create_vllm_config,
)
from vllm.config import SpeculativeConfig
from vllm.config.compilation import CUDAGraphMode
from vllm.config.mamba import MambaBackendEnum
from vllm.v1.kv_cache_interface import MambaSpec
@@ -191,9 +193,12 @@ REPLAYSSM_BUILD_CASES = {
def _make_mamba_spec(
buffer_len: int,
mamba_backend: MambaBackendEnum,
num_speculative_tokens: int = 0,
) -> MambaSpec:
ring_buffer_len = buffer_len + (
1 if mamba_backend == MambaBackendEnum.FLASHINFER else 0
1 + num_speculative_tokens
if mamba_backend == MambaBackendEnum.FLASHINFER
else 0
)
shapes = (
(1, 1),
@@ -214,6 +219,7 @@ def _create_replayssm_builder(
mamba_cache_mode: str = "none",
*,
mamba_backend: MambaBackendEnum = MambaBackendEnum.TRITON,
num_speculative_tokens: int = 0,
) -> MockMambaBuilder:
vllm_config = create_vllm_config(
model_name="Qwen/Qwen3.5-0.8B", block_size=BLOCK_SIZE
@@ -224,21 +230,32 @@ def _create_replayssm_builder(
vllm_config.cache_config.replayssm_buffer_len = buffer_len
vllm_config.cache_config.mamba_cache_mode = mamba_cache_mode
vllm_config.mamba_config.backend = mamba_backend
if num_speculative_tokens > 0:
# Triton ReplaySSM requires exact synchronous CPU metadata.
vllm_config.scheduler_config.async_scheduling = False
vllm_config.speculative_config = SpeculativeConfig(
method="ngram",
num_speculative_tokens=num_speculative_tokens,
)
return MockMambaBuilder(
_make_mamba_spec(buffer_len, mamba_backend),
_make_mamba_spec(buffer_len, mamba_backend, num_speculative_tokens),
["layer0"],
vllm_config,
DEVICE,
)
def _build(builder: MockMambaBuilder, case: ReplaySSMBuildCase):
def _build(
builder: MockMambaBuilder,
case: ReplaySSMBuildCase,
num_accepted_tokens: torch.Tensor | None = None,
):
batch = BatchSpec(seq_lens=case.seq_lens, query_lens=case.query_lens)
common = create_common_attn_metadata(batch, BLOCK_SIZE, DEVICE).replace(
is_prefilling=torch.tensor(case.is_prefilling, dtype=torch.bool),
replayssm_decode_base_cpu=torch.tensor(case.decode_base, dtype=torch.int32),
)
return builder.build(0, common)
return builder.build(0, common, num_accepted_tokens=num_accepted_tokens)
@pytest.mark.parametrize(
@@ -270,21 +287,60 @@ def test_resumed_request_differs_from_fresh():
assert meta.is_flush_d.tolist()[:2] == [0, 0]
def test_flashinfer_replayssm_scratch_metadata_fresh_decode():
checkpointing_ssu = pytest.importorskip("flashinfer.mamba.checkpointing_ssu")
if not hasattr(checkpointing_ssu, "allocate_checkpointing_ssu_scratch"):
pytest.skip("FlashInfer does not expose ReplaySSM scratch allocation")
def test_spec_decode_single_token_chunk_synthesizes_acceptance_metadata():
builder = _create_replayssm_builder(16, num_speculative_tokens=3)
case = REPLAYSSM_BUILD_CASES["leftover_prompt_one_token_flush"]
builder = _create_replayssm_builder(16, mamba_backend=MambaBackendEnum.FLASHINFER)
case = REPLAYSSM_BUILD_CASES["fresh_decode"]
meta = _build(builder, case)
assert meta.write_pos_d is None
assert meta.is_flush_d is None
assert meta.bc_pre_scratch is None
assert meta.replayssm_scratch is not None
assert [tensor.shape for tensor in meta.replayssm_scratch] == [
(1, 1, 32, 8),
(1, 1, 16),
(1, 1, 32, 8),
]
assert meta.query_start_loc_d is not None
assert meta.query_start_loc_d.tolist() == [0, 1]
assert meta.num_accepted_tokens is not None
assert meta.num_accepted_tokens.tolist() == [1]
def test_flashinfer_replayssm_state_indices_are_stable_for_full_cudagraph():
checkpointing_ssu = pytest.importorskip("flashinfer.mamba.checkpointing_ssu")
if not hasattr(checkpointing_ssu, "allocate_checkpointing_ssu_scratch"):
pytest.skip("requires FlashInfer ReplaySSM autotuning support")
builder = _create_replayssm_builder(
16,
mamba_backend=MambaBackendEnum.FLASHINFER,
num_speculative_tokens=3,
)
builder.compilation_config.cudagraph_mode = CUDAGraphMode.FULL
first = _build(
builder,
ReplaySSMBuildCase(
seq_lens=[106, 106],
query_lens=[1, 1],
is_prefilling=[False, False],
decode_base=[100, 100],
buffer_len=16,
expected_write_pos=[],
expected_is_flush=[],
),
)
first_indices = first.replayssm_state_indices_d
assert first_indices is not None
assert first_indices.is_contiguous()
first_ptr = first_indices.data_ptr()
second = _build(
builder,
ReplaySSMBuildCase(
seq_lens=[122, 122],
query_lens=[1, 1],
is_prefilling=[False, False],
decode_base=[116, 116],
buffer_len=16,
expected_write_pos=[],
expected_is_flush=[],
),
)
second_indices = second.replayssm_state_indices_d
assert second_indices is not None
assert second_indices.data_ptr() == first_ptr
assert torch.equal(second_indices, second.state_indices_tensor_d[:, 0])
+120 -26
View File
@@ -10,11 +10,16 @@ from vllm.v1.metrics.reader import Counter
from ...models.utils import check_logprobs_close
from ...utils import large_gpu_mark, multi_gpu_test
# Mamba2 (Nemotron-3) hybrid.
# Mamba2 (Nemotron-3) hybrids.
MAMBA2_MODEL = "nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16"
# FlashInfer 0.7.0's standard SSU cannot compile Nano's DIM=80 configuration.
MAMBA2_FLASHINFER_MODEL = "nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4"
MODELS = [
pytest.param(MAMBA2_MODEL, marks=large_gpu_mark(min_gb=40)),
]
FLASHINFER_MODELS = [
pytest.param(MAMBA2_FLASHINFER_MODEL, marks=large_gpu_mark(min_gb=40)),
]
PROMPTS = [
"The capital of France is",
@@ -22,6 +27,16 @@ PROMPTS = [
]
@pytest.fixture(autouse=True)
def _use_v1_model_runner_by_default(monkeypatch):
# Triton ReplaySSM is V1-only. FlashInfer V2 tests override this locally.
with monkeypatch.context() as patch:
patch.setenv("VLLM_USE_V2_MODEL_RUNNER", "0")
envs.disable_envs_cache()
yield
envs.disable_envs_cache()
def _check_replayssm_parity(
vllm_runner,
model_name,
@@ -29,17 +44,11 @@ def _check_replayssm_parity(
tensor_parallel_size=1,
mamba_backend: str = "triton",
name_1: str = "replayssm",
require_v2: bool = False,
monkeypatch: pytest.MonkeyPatch | None = None,
expected_v2: bool | None = None,
):
# Compare logprobs, not greedy ids: ReplaySSM's fp arithmetic can flip a
# near-tie. Baseline and ReplaySSM run at the same TP, so TP numerics are
# common-mode and only ReplaySSM varies.
if require_v2:
assert monkeypatch is not None
monkeypatch.setenv("VLLM_USE_V2_MODEL_RUNNER", "1")
envs.disable_envs_cache()
# near-tie. Baseline and ReplaySSM use the same backend so only ReplaySSM
# behavior varies.
common = dict(
max_model_len=1024,
trust_remote_code=True,
@@ -49,14 +58,17 @@ def _check_replayssm_parity(
mamba_backend=mamba_backend,
)
with vllm_runner(model_name, **common) as llm:
if require_v2:
assert llm.llm.llm_engine.vllm_config.use_v2_model_runner
if expected_v2 is not None:
assert llm.llm.llm_engine.vllm_config.use_v2_model_runner is expected_v2
baseline = llm.generate_greedy_logprobs(PROMPTS, max_tokens=32, num_logprobs=5)
with vllm_runner(
model_name, use_replayssm=True, replayssm_buffer_len=16, **common
model_name,
use_replayssm=True,
replayssm_buffer_len=16,
**common,
) as llm:
if require_v2:
assert llm.llm.llm_engine.vllm_config.use_v2_model_runner
if expected_v2 is not None:
assert llm.llm.llm_engine.vllm_config.use_v2_model_runner is expected_v2
replay = llm.generate_greedy_logprobs(PROMPTS, max_tokens=32, num_logprobs=5)
check_logprobs_close(
@@ -80,19 +92,101 @@ def test_replayssm_decode_matches_baseline_tp2(vllm_runner, model_name):
_check_replayssm_parity(vllm_runner, model_name, tensor_parallel_size=2)
@pytest.mark.parametrize("model_name", MODELS)
def test_replayssm_flashinfer_decode_matches_baseline_v2(
vllm_runner, model_name, monkeypatch
@pytest.mark.parametrize("model_name", FLASHINFER_MODELS)
@pytest.mark.parametrize("use_v2_model_runner", [False, True], ids=["v1", "v2"])
def test_replayssm_flashinfer_decode_matches_baseline(
vllm_runner, model_name, monkeypatch, use_v2_model_runner
):
pytest.importorskip("flashinfer.mamba.checkpointing_ssu")
_check_replayssm_parity(
vllm_runner,
model_name,
mamba_backend="flashinfer",
name_1="replayssm_flashinfer_v2",
require_v2=True,
monkeypatch=monkeypatch,
try:
with monkeypatch.context() as patch:
patch.setenv("VLLM_USE_V2_MODEL_RUNNER", str(int(use_v2_model_runner)))
envs.disable_envs_cache()
_check_replayssm_parity(
vllm_runner,
model_name,
mamba_backend="flashinfer",
name_1="replayssm_flashinfer",
expected_v2=use_v2_model_runner,
)
finally:
# The context restores the environment before the final cache reset.
envs.disable_envs_cache()
@pytest.mark.parametrize("model_name", FLASHINFER_MODELS)
def test_replayssm_flashinfer_spec_decode_matches_baseline(vllm_runner, model_name):
common = dict(
max_model_len=1024,
trust_remote_code=True,
enable_prefix_caching=False,
mamba_cache_mode="none",
speculative_config={
"method": "ngram",
"num_speculative_tokens": 3,
"prompt_lookup_max": 3,
},
)
with vllm_runner(model_name, mamba_backend="flashinfer", **common) as llm:
baseline = llm.generate_greedy_logprobs(PROMPTS, max_tokens=32, num_logprobs=5)
with vllm_runner(
model_name,
use_replayssm=True,
replayssm_buffer_len=16,
mamba_backend="flashinfer",
**common,
) as llm:
replay = llm.generate_greedy_logprobs(PROMPTS, max_tokens=32, num_logprobs=5)
check_logprobs_close(
outputs_0_lst=baseline,
outputs_1_lst=replay,
name_0="baseline_spec",
name_1="replayssm_flashinfer_spec",
)
@multi_gpu_test(num_gpus=2)
@large_gpu_mark(min_gb=40)
@pytest.mark.parametrize("use_v2_model_runner", [False, True], ids=["v1", "v2"])
def test_replayssm_flashinfer_mtp(vllm_runner, monkeypatch, use_v2_model_runner):
common = dict(
max_model_len=1024,
trust_remote_code=True,
enable_prefix_caching=False,
mamba_cache_mode="none",
mamba_backend="flashinfer",
tensor_parallel_size=2,
disable_log_stats=False,
speculative_config={"method": "mtp", "num_speculative_tokens": 3},
)
try:
with monkeypatch.context() as patch:
patch.setenv("VLLM_USE_V2_MODEL_RUNNER", str(int(use_v2_model_runner)))
envs.disable_envs_cache()
with vllm_runner(
MAMBA2_FLASHINFER_MODEL,
use_replayssm=True,
replayssm_buffer_len=16,
**common,
) as llm:
assert (
llm.llm.llm_engine.vllm_config.use_v2_model_runner
is use_v2_model_runner
)
outputs = llm.generate_greedy(PROMPTS, max_tokens=32)
draft_count = sum(
metric.value
for metric in llm.llm.get_metrics()
if isinstance(metric, Counter)
and metric.name == "vllm:spec_decode_num_drafts"
)
finally:
envs.disable_envs_cache()
# At least one request must run past the 16-token replay window; another
# may legitimately stop early on EOS.
assert any(len(token_ids) > 16 for token_ids, _ in outputs)
assert draft_count > 0
# Prefix spans several mamba blocks; prefix caching only reuses full blocks.
+14 -12
View File
@@ -1260,6 +1260,8 @@ class _TestReplaySSMMixer(MambaMixer2):
self.mamba_config = MambaConfig(backend=MambaBackendEnum.FLASHINFER)
self._replayssm_ring_start = torch.empty(0, dtype=torch.int32)
self._replayssm_prev_num_accepted = torch.empty(0, dtype=torch.int32)
self._replayssm_prev_query_len = torch.empty(0, dtype=torch.int32)
self._commits_replayssm_trackers = True
self._updates_replayssm_trackers = True
def get_state_shape(self) -> tuple[tuple[int, ...], ...]:
@@ -1294,19 +1296,19 @@ def test_bind_kv_cache_shares_replayssm_trackers_by_cache_group(layers_only):
else:
bind_kv_cache(kv_cache, ctx, [], kv_cache_groups=kv_cache_groups)
assert (
mixers[0]._replayssm_ring_start.data_ptr()
== mixers[2]._replayssm_ring_start.data_ptr()
tracker_names = (
"_replayssm_ring_start",
"_replayssm_prev_num_accepted",
"_replayssm_prev_query_len",
)
assert (
mixers[0]._replayssm_prev_num_accepted.data_ptr()
== mixers[2]._replayssm_prev_num_accepted.data_ptr()
)
assert (
mixers[1]._replayssm_ring_start.data_ptr()
!= mixers[0]._replayssm_ring_start.data_ptr()
)
# Group {0, 2} shares trackers; layer 2 (not 0) updates after both run.
for tracker_name in tracker_names:
group_tracker = getattr(mixers[0], tracker_name)
assert group_tracker.data_ptr() == getattr(mixers[2], tracker_name).data_ptr()
assert group_tracker.data_ptr() != getattr(mixers[1], tracker_name).data_ptr()
assert group_tracker.shape == (4,)
assert torch.count_nonzero(group_tracker) == 0
assert [m._commits_replayssm_trackers for m in mixers] == [True, True, False]
assert [m._updates_replayssm_trackers for m in mixers] == [False, True, True]