[Attention] Mamba attention module refactor (#41126)

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
This commit is contained in:
wangxiyuan
2026-05-22 17:13:12 +08:00
committed by GitHub
parent 65b7a812a2
commit 7e1b45a092
10 changed files with 765 additions and 774 deletions
+13 -27
View File
@@ -7,7 +7,7 @@ import torch
from torch import nn
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, ModelConfig, ParallelConfig, VllmConfig
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import (
get_pp_group,
get_tensor_model_parallel_world_size,
@@ -18,7 +18,6 @@ from vllm.model_executor.layers.fused_moe import (
FusedMoE,
fused_moe_make_expert_params_mapping,
)
from vllm.model_executor.layers.kda import KimiDeltaAttention
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
ColumnParallelLinear,
@@ -27,6 +26,9 @@ from vllm.model_executor.layers.linear import (
RowParallelLinear,
)
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.mamba.gdn.kimi_gdn_linear_attn import (
KimiGatedDeltaNetAttention,
)
from vllm.model_executor.layers.mamba.mamba_utils import (
MambaStateCopyFunc,
MambaStateCopyFuncCalculator,
@@ -286,26 +288,22 @@ class KimiDecoderLayer(nn.Module):
def __init__(
self,
config: KimiLinearConfig,
layer_idx: int,
cache_config: CacheConfig | None = None,
quant_config: QuantizationConfig | None = None,
parallel_config: ParallelConfig | None = None,
model_config: ModelConfig | None = None,
vllm_config: VllmConfig,
prefix: str = "",
**kwargs,
) -> None:
super().__init__()
self.hidden_size = config.hidden_size
self.is_moe = config.is_moe
layer_idx = int(prefix.rsplit(".", 1)[1])
model_config = vllm_config.model_config
cache_config = vllm_config.cache_config
quant_config = vllm_config.quant_config
if config.is_kda_layer(layer_idx):
self.self_attn = KimiDeltaAttention(
layer_idx=layer_idx,
hidden_size=config.hidden_size,
quant_config=quant_config,
cache_config=cache_config,
model_config=config,
self.self_attn = KimiGatedDeltaNetAttention(
config,
vllm_config,
prefix=f"{prefix}.self_attn",
)
else:
@@ -385,10 +383,6 @@ class KimiLinearModel(nn.Module):
super().__init__()
config = vllm_config.model_config.hf_text_config
model_config = vllm_config.model_config
cache_config = vllm_config.cache_config
quant_config = vllm_config.quant_config
parallel_config = vllm_config.parallel_config
self.config = config
self.vocab_size = config.vocab_size
@@ -402,19 +396,11 @@ class KimiLinearModel(nn.Module):
else:
self.embed_tokens = PPMissingLayer()
extra_kwargs = {}
def get_layer(prefix: str):
layer_idx = int(prefix.rsplit(".", 1)[1])
return KimiDecoderLayer(
config,
layer_idx,
cache_config,
quant_config,
parallel_config,
model_config,
vllm_config,
prefix,
**extra_kwargs,
)
self.start_layer, self.end_layer, self.layers = make_layers(