[Misc] Replace mamba_type string literals with MambaAttentionBackendEnum (#41188)
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
This commit is contained in:
@@ -13,6 +13,7 @@ from vllm.model_executor.layers.mamba.ops.ssu_dispatch import (
|
||||
selective_state_update,
|
||||
)
|
||||
from vllm.utils.torch_utils import set_random_seed
|
||||
from vllm.v1.attention.backends.registry import MambaAttentionBackendEnum
|
||||
from vllm.v1.kv_cache_interface import (
|
||||
KVCacheConfig,
|
||||
KVCacheGroupSpec,
|
||||
@@ -27,7 +28,9 @@ except ImportError:
|
||||
HAS_FLASHINFER = False
|
||||
|
||||
|
||||
def _kv_cache_config_with_ssu(mamba_type: str = "mamba2") -> KVCacheConfig:
|
||||
def _kv_cache_config_with_ssu(
|
||||
mamba_type: MambaAttentionBackendEnum = MambaAttentionBackendEnum.MAMBA2,
|
||||
) -> KVCacheConfig:
|
||||
spec = MambaSpec(
|
||||
block_size=16,
|
||||
shapes=((16, 64),),
|
||||
@@ -77,7 +80,12 @@ def test_uninitialized_backend_raises():
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
"mamba_type", ["linear_attention", "gdn_attention", "short_conv"]
|
||||
"mamba_type",
|
||||
[
|
||||
MambaAttentionBackendEnum.LINEAR,
|
||||
MambaAttentionBackendEnum.GDN_ATTN,
|
||||
MambaAttentionBackendEnum.SHORT_CONV,
|
||||
],
|
||||
)
|
||||
def test_init_is_noop_for_non_ssu_mamba_type(mamba_type):
|
||||
import vllm.model_executor.layers.mamba.ops.ssu_dispatch as mod
|
||||
|
||||
@@ -13,6 +13,7 @@ from vllm.model_executor.models.minimax_text_01 import MiniMaxText01LinearAttent
|
||||
from vllm.v1.attention.backends.linear_attn import LinearAttentionBackend
|
||||
from vllm.v1.attention.backends.mamba1_attn import Mamba1AttentionBackend
|
||||
from vllm.v1.attention.backends.mamba2_attn import Mamba2AttentionBackend
|
||||
from vllm.v1.attention.backends.registry import MambaAttentionBackendEnum
|
||||
from vllm.v1.attention.backends.short_conv_attn import ShortConvAttentionBackend
|
||||
|
||||
|
||||
@@ -32,7 +33,7 @@ from vllm.v1.attention.backends.short_conv_attn import ShortConvAttentionBackend
|
||||
use_rms_norm=True,
|
||||
),
|
||||
Mamba1AttentionBackend,
|
||||
"mamba1",
|
||||
MambaAttentionBackendEnum.MAMBA1,
|
||||
),
|
||||
(
|
||||
MambaMixer2,
|
||||
@@ -48,7 +49,7 @@ from vllm.v1.attention.backends.short_conv_attn import ShortConvAttentionBackend
|
||||
head_dim=32,
|
||||
),
|
||||
Mamba2AttentionBackend,
|
||||
"mamba2",
|
||||
MambaAttentionBackendEnum.MAMBA2,
|
||||
),
|
||||
(
|
||||
MiniMaxText01LinearAttention,
|
||||
@@ -64,7 +65,7 @@ from vllm.v1.attention.backends.short_conv_attn import ShortConvAttentionBackend
|
||||
linear_layer_idx=0,
|
||||
),
|
||||
LinearAttentionBackend,
|
||||
"linear_attention",
|
||||
MambaAttentionBackendEnum.LINEAR,
|
||||
),
|
||||
(
|
||||
ShortConv,
|
||||
@@ -74,7 +75,7 @@ from vllm.v1.attention.backends.short_conv_attn import ShortConvAttentionBackend
|
||||
layer_idx=0,
|
||||
),
|
||||
ShortConvAttentionBackend,
|
||||
"short_conv",
|
||||
MambaAttentionBackendEnum.SHORT_CONV,
|
||||
),
|
||||
],
|
||||
)
|
||||
@@ -97,10 +98,14 @@ def test_mamba_layers_get_attn_backend(
|
||||
@pytest.mark.parametrize(
|
||||
"layer_class,expected_backend,expected_mamba_type",
|
||||
[
|
||||
(MambaMixer, Mamba1AttentionBackend, "mamba1"),
|
||||
(MambaMixer2, Mamba2AttentionBackend, "mamba2"),
|
||||
(MiniMaxText01LinearAttention, LinearAttentionBackend, "linear_attention"),
|
||||
(ShortConv, ShortConvAttentionBackend, "short_conv"),
|
||||
(MambaMixer, Mamba1AttentionBackend, MambaAttentionBackendEnum.MAMBA1),
|
||||
(MambaMixer2, Mamba2AttentionBackend, MambaAttentionBackendEnum.MAMBA2),
|
||||
(
|
||||
MiniMaxText01LinearAttention,
|
||||
LinearAttentionBackend,
|
||||
MambaAttentionBackendEnum.LINEAR,
|
||||
),
|
||||
(ShortConv, ShortConvAttentionBackend, MambaAttentionBackendEnum.SHORT_CONV),
|
||||
],
|
||||
)
|
||||
def test_mamba_layers_have_unified_interface(
|
||||
|
||||
Reference in New Issue
Block a user