A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 3, 09:15 UTC
[Model Refactoring] Move DeepSeek V4 layers to `models/deepseek_v4/` [2/N] (#43039)
Signed-off-by: Woosuk Kwon <woosuk@inferact.ai>
Woosuk Kwon committed
May 18, 2026 at 21:00 UTC
87b08c5f6460cf487e47872c5fbc2595c97e74ef
5 files changed
+8
-11
.github/CODEOWNERS
+1
-2
@@ -153,9 +153,8 @@ mkdocs.yaml @hmellor
153
/vllm/model_executor/models/deepseek_mtp.py @luccafong
154
155
# DeepseekV4-specific files
156
+/vllm/models/deepseek_v4 @zyongye
157
/vllm/v1/attention/ops/deepseek_v4_ops @zyongye
157
-/vllm/model_executor/layers/deepseek_compressor.py @zyongye
158
-/vllm/model_executor/layers/deepseek_v4_attention.py @zyongye
158
/vllm/model_executor/layers/sparse_attn_indexer.py @zyongye
159
160
# Mistral-specific files
vllm/models/deepseek_v4/attention.py
renamed
+1
-1
@@ -46,7 +46,6 @@ from vllm.forward_context import ForwardContext, get_forward_context
46
from vllm.logger import init_logger
47
from vllm.model_executor.custom_op import PluggableLayer
48
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
49
-from vllm.model_executor.layers.deepseek_compressor import DeepseekCompressor
49
from vllm.model_executor.layers.layernorm import RMSNorm
50
from vllm.model_executor.layers.quantization import QuantizationConfig
51
from vllm.model_executor.layers.quantization.input_quant_fp8 import (
@@ -55,6 +54,7 @@ from vllm.model_executor.layers.quantization.input_quant_fp8 import (
54
from vllm.model_executor.layers.quantization.utils.quant_utils import (
55
GroupShape,
56
)
57
+from vllm.models.deepseek_v4.compressor import DeepseekCompressor
58
from vllm.platforms import current_platform
59
from vllm.utils.multi_stream_utils import (
60
execute_in_parallel,
vllm/models/deepseek_v4/compressor.py
renamed
vllm/models/deepseek_v4/nvidia/deepseek_v4.py
+5
-5
@@ -18,11 +18,6 @@ from vllm.distributed import (
18
)
19
from vllm.forward_context import get_forward_context
20
from vllm.model_executor.layers.activation import SiluAndMul, SiluAndMulWithClamp
21
-from vllm.model_executor.layers.deepseek_v4_attention import (
22
- DeepseekV4Indexer,
23
- DeepseekV4MLAModules,
24
- DeepseekV4MultiHeadLatentAttentionWrapper,
25
-)
21
from vllm.model_executor.layers.fused_moe import FusedMoE
22
from vllm.model_executor.layers.fused_moe.router.fused_topk_bias_router import (
23
fused_topk_bias,
@@ -61,6 +56,11 @@ from vllm.model_executor.models.utils import (
56
maybe_prefix,
57
)
58
from vllm.model_executor.utils import set_weight_attrs
59
+from vllm.models.deepseek_v4.attention import (
60
+ DeepseekV4Indexer,
61
+ DeepseekV4MLAModules,
62
+ DeepseekV4MultiHeadLatentAttentionWrapper,
63
+)
64
from vllm.platforms import current_platform
65
from vllm.sequence import IntermediateTensors
66
from vllm.triton_utils import tl, triton
vllm/v1/attention/backends/mla/rocm_aiter_mla_sparse_dsv4.py
+1
-3
@@ -31,9 +31,7 @@ from vllm.v1.attention.ops.rocm_aiter_mla_sparse import (
31
from vllm.v1.worker.workspace import current_workspace_manager
32
33
if TYPE_CHECKING:
34
- from vllm.model_executor.layers.deepseek_v4_attention import (
35
- DeepseekV4MLAAttention,
36
- )
34
+ from vllm.models.deepseek_v4.attention import DeepseekV4MLAAttention
35
36
37
def _build_indptr_from_lengths(lengths: torch.Tensor) -> torch.Tensor: