[Refactor] Normalize DeepSeek V4.1 model package naming (#56741)

Signed-off-by: Bugen Zhao <i@bugenzhao.com>
This commit is contained in:
Bugen Zhao
2026-09-13 23:17:50 -07:00
committed by GitHub
parent 10e8d5614c
commit cf1584f373
44 changed files with 71 additions and 72 deletions
@@ -54,8 +54,8 @@ from vllm.models.common.ops.sequence_parallel import (
from vllm.models.deepseek_v4.amd.model import (
DeepseekV4MoE as DeepseekV4MoEBase,
)
from vllm.models.deepseek_v4_1.amd.rocm import DeepseekV41ROCMAiterMLAAttention
from vllm.models.deepseek_v4_1.attention import DeepseekV4Attention
from vllm.models.deepseek_v41.amd.rocm import DeepseekV41ROCMAiterMLAAttention
from vllm.models.deepseek_v41.attention import DeepseekV4Attention
from vllm.sequence import IntermediateTensors
from vllm.utils.math_utils import cdiv
from vllm.v1.attention.backends.registry import AttentionBackendEnum
@@ -13,9 +13,9 @@ from vllm.distributed import (
)
from vllm.forward_context import get_forward_context
from vllm.logger import init_logger
from vllm.models.deepseek_v4_1.attention import DeepseekV4Attention
from vllm.models.deepseek_v4_1.common.ops import dequantize_and_gather_k_cache
from vllm.models.deepseek_v4_1.sparse_mla import (
from vllm.models.deepseek_v41.attention import DeepseekV4Attention
from vllm.models.deepseek_v41.common.ops import dequantize_and_gather_k_cache
from vllm.models.deepseek_v41.sparse_mla import (
DeepseekV4FlashMLAMetadata,
DeepseekV4SparseMLABackend,
DeepseekV4SparseMLAMetadataBuilder,
@@ -26,7 +26,7 @@ from vllm.model_executor.layers.linear import (
)
from vllm.model_executor.layers.sparse_attn_indexer import SparseAttnIndexer
from vllm.models.common.ops import fused_q_kv_rmsnorm
from vllm.models.deepseek_v4_1.common.ops import (
from vllm.models.deepseek_v41.common.ops import (
MXFP4_BLOCK_SIZE,
fused_indexer_q_rope_quant,
indexer_k_norm_rope_store,
@@ -49,8 +49,8 @@ from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.models.utils import extract_layer_index
from vllm.models.deepseek_v4_1.common.rope import build_deepseek_v4_rope
from vllm.models.deepseek_v4_1.compressor import DeepseekCompressor
from vllm.models.deepseek_v41.common.rope import build_deepseek_v4_rope
from vllm.models.deepseek_v41.compressor import DeepseekCompressor
from vllm.triton_utils import tl, triton
from vllm.utils.multi_stream_utils import (
execute_in_parallel,
@@ -156,7 +156,7 @@ class DeepseekV4Attention(nn.Module, AttentionLayerBase, ABC):
``DeepseekV4FlashInferSM120Attention`` /
``DeepseekV4FlashInferMLAAttention`` (CUDA) or
``DeepseekV41ROCMAiterMLAAttention`` (ROCm) — selected by the platform-specific
deepseek_v4_1 model module. The base is never instantiated directly.
deepseek_v41 model module. The base is never instantiated directly.
"""
# Provided by the platform subclass.
@@ -536,7 +536,7 @@ class DeepseekV4Attention(nn.Module, AttentionLayerBase, ABC):
"FLASHMLA_SPARSE_DSV41",
"ROCM_FLASHMLA_SPARSE_DSV4",
):
from vllm.models.deepseek_v4_1.common.ops.cache_utils import (
from vllm.models.deepseek_v41.common.ops.cache_utils import (
_COMBINE_TOPK_SWA_INDICES_KERNEL,
)
@@ -581,7 +581,7 @@ class DeepseekV4Attention(nn.Module, AttentionLayerBase, ABC):
@cached_property
def _can_fuse_query_quant(self) -> bool:
from vllm.models.deepseek_v4_1.common.ops.query_quant import (
from vllm.models.deepseek_v41.common.ops.query_quant import (
can_fuse_query_quant,
)
@@ -600,7 +600,7 @@ class DeepseekV4Attention(nn.Module, AttentionLayerBase, ABC):
"""
qr, kv = qr_kv.split([self.q_lora_rank, self.head_dim], dim=-1)
if self.q_lora_rank % 32 == 0 and self._can_fuse_query_quant:
from vllm.models.deepseek_v4_1.common.ops.query_quant import (
from vllm.models.deepseek_v41.common.ops.query_quant import (
fused_q_kv_rmsnorm_quant,
)
@@ -12,7 +12,7 @@ from vllm.forward_context import get_forward_context
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import MergedColumnParallelLinear
from vllm.models.deepseek_v4_1.common.ops.fused_compress_quant_cache import (
from vllm.models.deepseek_v41.common.ops.fused_compress_quant_cache import (
fused_save_compress_norm,
rope_quant_insert,
)
@@ -23,16 +23,16 @@ from vllm.distributed.parallel_state import GroupCoordinator
from vllm.forward_context import get_forward_context
from vllm.logger import init_logger
from vllm.model_executor.utils import set_weight_attrs
from vllm.models.deepseek_v4_1.common.engram import (
from vllm.models.deepseek_v41.common.engram import (
DEAD_ID,
EngramLayout,
_engram_head_shard_weight_loader,
_engram_select_rows,
)
from vllm.models.deepseek_v4_1.common.engram import (
from vllm.models.deepseek_v41.common.engram import (
Engram as BaseEngram,
)
from vllm.models.deepseek_v4_1.common.engram import (
from vllm.models.deepseek_v41.common.engram import (
ParallelEngramEmbedding as BaseParallelEngramEmbedding,
)
from vllm.utils.platform_utils import is_uva_available
@@ -13,12 +13,12 @@ from vllm.models.deepseek_v4.nvidia.ops.o_proj import (
compute_fp8_einsum_recipe,
deep_gemm_fp8_o_proj,
)
from vllm.models.deepseek_v4_1.attention import DeepseekV4Attention
from vllm.models.deepseek_v4_1.common.ops import (
from vllm.models.deepseek_v41.attention import DeepseekV4Attention
from vllm.models.deepseek_v41.common.ops import (
build_flashinfer_mixed_sparse_indices,
compute_global_topk_indices_and_lens,
)
from vllm.models.deepseek_v4_1.sparse_mla import (
from vllm.models.deepseek_v41.sparse_mla import (
DeepseekV4FlashMLAMetadata,
DeepseekV4SparseMLABackend,
DeepseekV4SparseMLAMetadataBuilder,
@@ -10,13 +10,13 @@ from vllm.models.deepseek_v4.nvidia.ops.o_proj import (
compute_fp8_einsum_recipe,
deep_gemm_fp8_o_proj,
)
from vllm.models.deepseek_v4_1.attention import DeepseekV4Attention
from vllm.models.deepseek_v4_1.common.ops import (
from vllm.models.deepseek_v41.attention import DeepseekV4Attention
from vllm.models.deepseek_v41.common.ops import (
combine_topk_swa_indices,
compute_global_topk_indices_and_lens,
dequantize_and_gather_k_cache,
)
from vllm.models.deepseek_v4_1.sparse_mla import (
from vllm.models.deepseek_v41.sparse_mla import (
DeepseekV4FlashMLABackend,
DeepseekV4FlashMLAMetadata,
DeepseekV41SparseSWAMetadataBuilder,
@@ -63,12 +63,12 @@ from vllm.models.deepseek_v4.nvidia.model import (
from vllm.models.deepseek_v4.nvidia.model import (
make_deepseek_v4_expert_params_mapping,
)
from vllm.models.deepseek_v4_1.attention import DeepseekV4Attention
from vllm.models.deepseek_v4_1.nvidia.flashinfer_sparse import (
from vllm.models.deepseek_v41.attention import DeepseekV4Attention
from vllm.models.deepseek_v41.nvidia.flashinfer_sparse import (
DeepseekV4FlashInferMLAAttention,
DeepseekV4FlashInferSM120Attention,
)
from vllm.models.deepseek_v4_1.nvidia.flashmla import DeepseekV4FlashMLAAttention
from vllm.models.deepseek_v41.nvidia.flashmla import DeepseekV4FlashMLAAttention
from vllm.platforms import current_platform
from vllm.sequence import IntermediateTensors
from vllm.utils.math_utils import cdiv