From 87b08c5f6460cf487e47872c5fbc2595c97e74ef Mon Sep 17 00:00:00 2001 From: Woosuk Kwon Date: Mon, 18 May 2026 21:00:58 -0700 Subject: [PATCH] [Model Refactoring] Move DeepSeek V4 layers to `models/deepseek_v4/` [2/N] (#43039) Signed-off-by: Woosuk Kwon --- .github/CODEOWNERS | 3 +-- .../deepseek_v4/attention.py} | 2 +- .../deepseek_v4/compressor.py} | 0 vllm/models/deepseek_v4/nvidia/deepseek_v4.py | 10 +++++----- .../backends/mla/rocm_aiter_mla_sparse_dsv4.py | 4 +--- 5 files changed, 8 insertions(+), 11 deletions(-) rename vllm/{model_executor/layers/deepseek_v4_attention.py => models/deepseek_v4/attention.py} (99%) rename vllm/{model_executor/layers/deepseek_compressor.py => models/deepseek_v4/compressor.py} (100%) diff --git a/.github/CODEOWNERS b/.github/CODEOWNERS index 66041fcee5..3fc8818387 100644 --- a/.github/CODEOWNERS +++ b/.github/CODEOWNERS @@ -153,9 +153,8 @@ mkdocs.yaml @hmellor /vllm/model_executor/models/deepseek_mtp.py @luccafong # DeepseekV4-specific files +/vllm/models/deepseek_v4 @zyongye /vllm/v1/attention/ops/deepseek_v4_ops @zyongye -/vllm/model_executor/layers/deepseek_compressor.py @zyongye -/vllm/model_executor/layers/deepseek_v4_attention.py @zyongye /vllm/model_executor/layers/sparse_attn_indexer.py @zyongye # Mistral-specific files diff --git a/vllm/model_executor/layers/deepseek_v4_attention.py b/vllm/models/deepseek_v4/attention.py similarity index 99% rename from vllm/model_executor/layers/deepseek_v4_attention.py rename to vllm/models/deepseek_v4/attention.py index b2d2ac8164..68d0523414 100644 --- a/vllm/model_executor/layers/deepseek_v4_attention.py +++ b/vllm/models/deepseek_v4/attention.py @@ -46,7 +46,6 @@ from vllm.forward_context import ForwardContext, get_forward_context from vllm.logger import init_logger from vllm.model_executor.custom_op import PluggableLayer from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase -from vllm.model_executor.layers.deepseek_compressor import DeepseekCompressor from vllm.model_executor.layers.layernorm import RMSNorm from vllm.model_executor.layers.quantization import QuantizationConfig from vllm.model_executor.layers.quantization.input_quant_fp8 import ( @@ -55,6 +54,7 @@ from vllm.model_executor.layers.quantization.input_quant_fp8 import ( from vllm.model_executor.layers.quantization.utils.quant_utils import ( GroupShape, ) +from vllm.models.deepseek_v4.compressor import DeepseekCompressor from vllm.platforms import current_platform from vllm.utils.multi_stream_utils import ( execute_in_parallel, diff --git a/vllm/model_executor/layers/deepseek_compressor.py b/vllm/models/deepseek_v4/compressor.py similarity index 100% rename from vllm/model_executor/layers/deepseek_compressor.py rename to vllm/models/deepseek_v4/compressor.py diff --git a/vllm/models/deepseek_v4/nvidia/deepseek_v4.py b/vllm/models/deepseek_v4/nvidia/deepseek_v4.py index b08900df9e..395cd0201b 100644 --- a/vllm/models/deepseek_v4/nvidia/deepseek_v4.py +++ b/vllm/models/deepseek_v4/nvidia/deepseek_v4.py @@ -18,11 +18,6 @@ from vllm.distributed import ( ) from vllm.forward_context import get_forward_context from vllm.model_executor.layers.activation import SiluAndMul, SiluAndMulWithClamp -from vllm.model_executor.layers.deepseek_v4_attention import ( - DeepseekV4Indexer, - DeepseekV4MLAModules, - DeepseekV4MultiHeadLatentAttentionWrapper, -) from vllm.model_executor.layers.fused_moe import FusedMoE from vllm.model_executor.layers.fused_moe.router.fused_topk_bias_router import ( fused_topk_bias, @@ -61,6 +56,11 @@ from vllm.model_executor.models.utils import ( maybe_prefix, ) from vllm.model_executor.utils import set_weight_attrs +from vllm.models.deepseek_v4.attention import ( + DeepseekV4Indexer, + DeepseekV4MLAModules, + DeepseekV4MultiHeadLatentAttentionWrapper, +) from vllm.platforms import current_platform from vllm.sequence import IntermediateTensors from vllm.triton_utils import tl, triton diff --git a/vllm/v1/attention/backends/mla/rocm_aiter_mla_sparse_dsv4.py b/vllm/v1/attention/backends/mla/rocm_aiter_mla_sparse_dsv4.py index e7a82a1ee2..d225c8d785 100644 --- a/vllm/v1/attention/backends/mla/rocm_aiter_mla_sparse_dsv4.py +++ b/vllm/v1/attention/backends/mla/rocm_aiter_mla_sparse_dsv4.py @@ -31,9 +31,7 @@ from vllm.v1.attention.ops.rocm_aiter_mla_sparse import ( from vllm.v1.worker.workspace import current_workspace_manager if TYPE_CHECKING: - from vllm.model_executor.layers.deepseek_v4_attention import ( - DeepseekV4MLAAttention, - ) + from vllm.models.deepseek_v4.attention import DeepseekV4MLAAttention def _build_indptr_from_lengths(lengths: torch.Tensor) -> torch.Tensor: