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

Signed-off-by: Bugen Zhao <i@bugenzhao.com>

.buildkite/test_areas/distributed.yaml

@@ -208,7 +208,7 @@ steps:
   - tests/distributed/test_utils
   - tests/distributed/test_engram_dp_shard.py
   - vllm/config/engram.py
-  - vllm/models/deepseek_v4_1/
+  - vllm/models/deepseek_v41/
   commands:
   # https://github.com/NVIDIA/nccl/issues/1838
   - export NCCL_CUMEM_HOST_ENABLE=0

.github/mergify.yml

@@ -343,9 +343,8 @@ pull_request_rules:
   conditions:
     - label != stale
     - or:
-      # The 4.1 sources spell the model three ways: deepseek_v4_1 (models),
-      # deepseek_v41 (most) and deepseekv41 (tool parsers).
-      - files~=^vllm/models/deepseek_v4_1/
+      # The 4.1 sources use deepseek_v41 and deepseekv41 (tool parsers).
+      - files~=^vllm/models/deepseek_v41/
       - files~=^vllm/(?:parser|tokenizers|reasoning|transformers_utils/configs)/deepseek_v41.*\.py
       - files~=^vllm/tool_parsers/deepseekv41.*\.py
       - files~=^rust/src/(?:chat|parser)/src/.*deepseek_v41

tests/distributed/test_engram_dp_shard.py

@@ -39,9 +39,9 @@ from vllm.distributed.parallel_state import (
     initialize_model_parallel,
 )
 from vllm.forward_context import set_forward_context
-from vllm.models.deepseek_v4_1.common.engram import EngramLayout
-from vllm.models.deepseek_v4_1.nvidia import engram as engram_ops
-from vllm.models.deepseek_v4_1.nvidia.engram import (
+from vllm.models.deepseek_v41.common.engram import EngramLayout
+from vllm.models.deepseek_v41.nvidia import engram as engram_ops
+from vllm.models.deepseek_v41.nvidia.engram import (
     Engram,
     ParallelEngramEmbedding,
     engram_head_shard_rank,
@@ -473,8 +473,8 @@ def _check_dummy_hash_model_forward(
     vllm_config, engram, head_sizes, weight, scales, dp_shared_memory
 ):
     """A metadata-less replica joins DP sharding but skips a shared lookup."""
-    from vllm.models.deepseek_v4_1.common.engram import NgramHashState
-    from vllm.models.deepseek_v4_1.nvidia import model as model_ops
+    from vllm.models.deepseek_v41.common.engram import NgramHashState
+    from vllm.models.deepseek_v41.nvidia import model as model_ops
 
     dp_rank = vllm_config.parallel_config.data_parallel_rank
     if dp_shared_memory and dp_rank != 1:

tests/kernels/core/test_fused_q_kv_rmsnorm.py

@@ -125,7 +125,7 @@ def test_fused_q_kv_rmsnorm_quant_matches_separate(num_tokens, q_size, dtype):
     from vllm.model_executor.layers.quantization.utils.mxfp8_utils import (
         mxfp8_e4m3_quantize,
     )
-    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,
     )
 
@@ -198,15 +198,15 @@ def test_shared_query_quant_preserves_projection(
         mxfp8_e4m3_quantize,
     )
     from vllm.model_executor.layers.quantization.utils.quant_utils import kNvfp4Dynamic
-    from vllm.models.deepseek_v4_1.attention import (
+    from vllm.models.deepseek_v41.attention import (
         DeepseekV4Attention,
         DeepseekV4Indexer,
     )
-    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,
         fused_q_kv_rmsnorm_quant,
     )
-    from vllm.models.deepseek_v4_1.quant_config import DeepseekV4FP8Config
+    from vllm.models.deepseek_v41.quant_config import DeepseekV4FP8Config
 
     cls: (
         type[FlashInferCutedslMxfp8LinearKernel]

tests/kernels/test_compressor_kv_cache.py

@@ -65,7 +65,7 @@ def test_v41_fused_save_compress_and_insert(
     a Torch reference with BF16 rounding before RoPE. Replays change both
     inputs and history so stale scratch/cache values cannot satisfy the test.
     """
-    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,
     )
@@ -226,7 +226,7 @@ def test_v41_rope_insert_plain_row(compress_ratio: int, store_fp8: bool):
     Non-boundary positions, negative slots and untouched rows must keep their
     prior contents.
     """
-    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 (
         rope_quant_insert,
     )
 
@@ -301,7 +301,7 @@ def test_v41_compressor_metadata_maps_tokens_to_their_ring():
     capacity`` and keep padding tokens at PAD."""
     from unittest.mock import MagicMock
 
-    from vllm.models.deepseek_v4_1.compressor import CompressorMetadataBuilder
+    from vllm.models.deepseek_v41.compressor import CompressorMetadataBuilder
     from vllm.v1.attention.backend import CommonAttentionMetadata
     from vllm.v1.kv_cache_interface import CircularBufferSpec
 
@@ -351,11 +351,11 @@ def test_v41_compressor_metadata_maps_tokens_to_their_ring():
 def test_v41_attention_joins_cache_writes_before_consumption(use_aux, use_graph):
     """Both reused-event joins must publish this forward's states and cache rows."""
     from vllm.forward_context import ForwardContext, override_forward_context
-    from vllm.models.deepseek_v4_1.attention import (
+    from vllm.models.deepseek_v41.attention import (
         DeepseekV4Attention,
         DeepseekV4Indexer,
     )
-    from vllm.models.deepseek_v4_1.compressor import DeepseekCompressor
+    from vllm.models.deepseek_v41.compressor import DeepseekCompressor
 
     torch.manual_seed(43)
     raw = torch.randn(19, 1024, device="cuda")

tests/kernels/test_engram.py

@@ -7,15 +7,15 @@ from types import SimpleNamespace
 import pytest
 import torch
 
-from vllm.models.deepseek_v4_1.common import engram as engram_ops
-from vllm.models.deepseek_v4_1.common.engram import (
+from vllm.models.deepseek_v41.common import engram as engram_ops
+from vllm.models.deepseek_v41.common.engram import (
     Engram as CommonEngram,
 )
-from vllm.models.deepseek_v4_1.common.engram import (
+from vllm.models.deepseek_v41.common.engram import (
     NgramHashState,
 )
-from vllm.models.deepseek_v4_1.nvidia import engram as nvidia_engram_ops
-from vllm.models.deepseek_v4_1.nvidia.engram import Engram, ParallelEngramEmbedding
+from vllm.models.deepseek_v41.nvidia import engram as nvidia_engram_ops
+from vllm.models.deepseek_v41.nvidia.engram import Engram, ParallelEngramEmbedding
 from vllm.platforms import current_platform
 
 
@@ -503,7 +503,7 @@ def test_lookback_window_reproduces_single_instance(runner, capture):
 @pytest.mark.skipif(not current_platform.is_cuda(), reason="CUDA required")
 def test_v2_model_state_gathers_lookback_window():
     """The window is gathered on device from the runner's token history."""
-    from vllm.models.deepseek_v4_1.nvidia.model_state import DeepseekV41ModelState
+    from vllm.models.deepseek_v41.nvidia.model_state import DeepseekV41ModelState
 
     depth, max_num_reqs, max_model_len = 3, 4, 16
     state = DeepseekV41ModelState.__new__(DeepseekV41ModelState)

tests/kernels/test_fused_indexer_q_rope_quant.py

@@ -359,7 +359,7 @@ def test_indexer_k_inserts_only_valid_groups_into_padded_pages(
     num_tokens, compress_ratio, use_fp4, cache_dtype
 ):
     """Insert group ends and preserve skipped slots, graph rows, and page padding."""
-    from vllm.models.deepseek_v4_1.common.ops.indexer_k_store import (
+    from vllm.models.deepseek_v41.common.ops.indexer_k_store import (
         indexer_k_norm_rope_store,
     )
 
@@ -442,7 +442,7 @@ def test_indexer_k_inserts_only_valid_groups_into_padded_pages(
 @torch.inference_mode()
 def test_indexer_k_cuda_graph_replay_reads_current_projection(use_fp4):
     """Replay must consume updated keys and slot mapping through the captured API."""
-    from vllm.models.deepseek_v4_1.common.ops.indexer_k_store import (
+    from vllm.models.deepseek_v41.common.ops.indexer_k_store import (
         indexer_k_norm_rope_store,
     )
 
@@ -501,7 +501,7 @@ def test_indexer_k_store_roundtrips_through_rocm_gather(block_size, compress_rat
     ``block_size > 1``, so a row-major store would feed the indexer permuted
     key bytes and randomize its top-k. Assert the write/read pair is exact.
     """
-    from vllm.models.deepseek_v4_1.common.ops.indexer_k_store import (
+    from vllm.models.deepseek_v41.common.ops.indexer_k_store import (
         indexer_k_norm_rope_store,
     )
     from vllm.v1.attention.ops.rocm_aiter_mla_sparse import (

tests/kernels/test_mhc_kernels.py

@@ -32,7 +32,7 @@ from vllm.models.deepseek_v4.nvidia.model import (
     DeepseekV4DecoderLayer,
     DeepseekV4Model,
 )
-from vllm.models.deepseek_v4_1.nvidia.model import (
+from vllm.models.deepseek_v41.nvidia.model import (
     DeepseekV4DecoderLayer as DeepseekV41DecoderLayer,
 )
 from vllm.platforms import current_platform
@@ -104,7 +104,7 @@ def test_hc_collapse_custom_op_supports_compile():
 @pytest.mark.parametrize("num_tokens", [1, 7])
 def test_v41_dspark_head_collapses_with_last_ffn_mix(num_tokens, monkeypatch):
     """Match reference DSparkBlock.forward_head's hc_pre before its RMSNorm."""
-    from vllm.models.deepseek_v4_1.nvidia import dspark
+    from vllm.models.deepseek_v41.nvidia import dspark
 
     set_random_seed(0)
     hidden_size, hc_mult = 5120, 4
@@ -296,7 +296,7 @@ def test_deepseek_v41_decoder_mixes_match_torch(
         return post, res, decoder.attn_norm(collapsed), pre
 
     monkeypatch.setattr(
-        "vllm.models.deepseek_v4_1.nvidia.model.mhc_pre_delayed_tilelang",
+        "vllm.models.deepseek_v41.nvidia.model.mhc_pre_delayed_tilelang",
         reference,
     )
     expected = decoder(x, positions, None, **kwargs)

tests/models/test_deepseek_v4_mega_moe.py

@@ -18,8 +18,8 @@ from vllm.models.deepseek_v4.nvidia.model import (
 )
 from vllm.models.deepseek_v4.nvidia.mtp import DeepSeekV4MTP
 from vllm.models.deepseek_v4.nvidia.ops.prepare_megamoe import prepare_megamoe_inputs
-from vllm.models.deepseek_v4_1.common.mm_preprocess import IMAGE_SENTINEL_BASE_ID
-from vllm.models.deepseek_v4_1.nvidia.model import DeepseekV4MoE as DeepseekV41MoE
+from vllm.models.deepseek_v41.common.mm_preprocess import IMAGE_SENTINEL_BASE_ID
+from vllm.models.deepseek_v41.nvidia.model import DeepseekV4MoE as DeepseekV41MoE
 from vllm.platforms import current_platform
 from vllm.transformers_utils.configs.deepseek_v41 import DeepseekV41Config
 

tests/models/test_dspark_mla.py

@@ -216,7 +216,7 @@ def test_v41_dspark_loads_linear_scales(
     """Checkpoint ``.scale`` maps to the quant method's scale parameter and
     loads untouched. MXFP8 block-scale expansion lives in the KMxfp8Static
     loader (see tests/quantization/test_modelopt.py), not in load_weights."""
-    from vllm.models.deepseek_v4_1.nvidia import dspark
+    from vllm.models.deepseek_v41.nvidia import dspark
 
     mxfp8 = scale_dtype != torch.float32
     scale_name = "weight_scale" if mxfp8 else "weight_scale_inv"

tests/quantization/test_fp8.py

@@ -77,8 +77,8 @@ def test_deepseek_v41_mxfp8_scale_loading(
         MergedColumnParallelLinear,
         RowParallelLinear,
     )
-    from vllm.models.deepseek_v4_1 import quant_config as quant_module
-    from vllm.models.deepseek_v4_1.nvidia import model as model_module
+    from vllm.models.deepseek_v41 import quant_config as quant_module
+    from vllm.models.deepseek_v41.nvidia import model as model_module
 
     default_vllm_config.model_config = SimpleNamespace(dtype=torch.bfloat16)
     quant_config = quant_module.DeepseekV4FP8Config(
@@ -214,8 +214,8 @@ def test_deepseek_v41_vl_mapper_routes_linear_scales(
     """The VL wrapper must map ``.scale`` keys to the parameter the linear
     quant method registers, as the text model does. A hardcoded
     ``weight_scale_inv`` raised KeyError for native MXFP8 checkpoints."""
-    from vllm.models.deepseek_v4_1.nvidia import model as model_module
-    from vllm.models.deepseek_v4_1.nvidia import vl_model as vl_module
+    from vllm.models.deepseek_v41.nvidia import model as model_module
+    from vllm.models.deepseek_v41.nvidia import vl_model as vl_module
 
     vllm_config = SimpleNamespace(
         quant_config=SimpleNamespace(weight_block_size=weight_block_size)
@@ -318,8 +318,8 @@ def test_mxfp8_bmm_loads_and_projects_grouped_weights(
         compute_fp8_einsum_recipe,
         deep_gemm_fp8_o_proj,
     )
-    from vllm.models.deepseek_v4_1.nvidia.model import DeepseekV4Model
-    from vllm.models.deepseek_v4_1.quant_config import DeepseekV4FP8Config
+    from vllm.models.deepseek_v41.nvidia.model import DeepseekV4Model
+    from vllm.models.deepseek_v41.quant_config import DeepseekV4FP8Config
     from vllm.utils.deep_gemm import is_deep_gemm_supported
 
     if not is_deep_gemm_supported() or not current_platform.is_device_capability_family(

tests/v1/attention/test_indexer_deepseek_v4_slot_mapping.py

@@ -8,7 +8,7 @@ import torch
 
 from tests.v1.attention.utils import create_vllm_config
 from vllm.models.deepseek_v4.sparse_mla import DeepseekV4SparseMLABackend
-from vllm.models.deepseek_v4_1.sparse_mla import (
+from vllm.models.deepseek_v41.sparse_mla import (
     DeepseekV4SparseMLABackend as DeepseekV41SparseMLABackend,
 )
 from vllm.v1.attention.backend import CommonAttentionMetadata

vllm/model_executor/layers/quantization/__init__.py

@@ -120,7 +120,7 @@ def get_quantization_config(quantization: str) -> type[QuantizationConfig]:
         # The v4.1 class is the v4 one extended to accept model_type
         # "deepseek_v41" and its 32x32 MXFP8 linear layout. V4.1 has
         # platform-specific implementations for both CUDA and ROCm.
-        from vllm.models.deepseek_v4_1 import (
+        from vllm.models.deepseek_v41 import (
             DeepseekV4FP8Config as DeepseekV41FP8Config,
         )
 

vllm/model_executor/models/registry.py

@@ -377,7 +377,7 @@ _MULTIMODAL_MODELS = {
         "DeepseekV4ForConditionalGeneration",
     ),
     "DeepseekV41ForCausalLM": (
-        "vllm.models.deepseek_v4_1",
+        "vllm.models.deepseek_v41",
         "DeepseekV41ForCausalLM",
     ),
     "Dots3NoteForCausalLM": (
@@ -645,7 +645,7 @@ _SPECULATIVE_DECODING_MODELS = {
     "DFlashMuseGlimmerAssistantModel": ("qwen3_dflash", "DFlashQwen3ForCausalLM"),
     "DSparkDraftModel": ("vllm.models.deepseek_v4", "DSparkDeepseekV4ForCausalLM"),
     "DSparkV41DraftModel": (
-        "vllm.models.deepseek_v4_1",
+        "vllm.models.deepseek_v41",
         "DSparkDeepseekV4ForCausalLM",
     ),
     "Qwen3DSparkModel": ("qwen3_dspark", "Qwen3DSparkForCausalLM"),

vllm/models/deepseek_v41/amd/model.py

@@ -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

vllm/models/deepseek_v41/amd/rocm.py

@@ -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,

vllm/models/deepseek_v41/attention.py

@@ -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,
             )
 

vllm/models/deepseek_v41/compressor.py

@@ -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,
 )

vllm/models/deepseek_v41/nvidia/engram.py

@@ -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

vllm/models/deepseek_v41/nvidia/flashinfer_sparse.py

@@ -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,

vllm/models/deepseek_v41/nvidia/flashmla.py

@@ -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,

vllm/models/deepseek_v41/nvidia/model.py

@@ -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

vllm/transformers_utils/configs/deepseek_v41.py

@@ -12,7 +12,7 @@ class DeepseekV41Config(PretrainedConfig):
     tower under ``vision_config``. The vLLM model code (ported from
     ``deepseek_v4``) reads flat attributes, so both sub-configs are flattened
     onto the top level here: text fields are exposed as-is, vision fields
-    with the ``vision_*`` naming used by ``deepseek_v4_1.common.vision``.
+    with the ``vision_*`` naming used by ``deepseek_v41.common.vision``.
     """
 
     model_type = "deepseek_v41"

vllm/v1/attention/backends/mla/sparse_swa.py

@@ -49,7 +49,7 @@ _LAYER_TYPE_C128A = "c128a"
 # v4.1 ratio-1 / ratio-2 indexer layers: same indexer-path shape as C4A, but
 # their compressed page block is block_size // ratio, so each gets its own plan.
 # v4.1 builders classify layer types themselves (ratio 0 is SWA-only there); see
-# deepseek_v4_1/sparse_mla.py.
+# deepseek_v41/sparse_mla.py.
 _LAYER_TYPE_C1A = "c1a"
 _LAYER_TYPE_C2A = "c2a"
 

vllm/v1/attention/backends/registry.py

@@ -108,10 +108,10 @@ class AttentionBackendEnum(Enum, metaclass=_AttentionBackendEnumMeta):
     # layer). Separate names from DSV4 so a V4.1 model never resolves the
     # V4.0 backend classes through this enum.
     FLASHMLA_SPARSE_DSV41 = (
-        "vllm.models.deepseek_v4_1.sparse_mla.DeepseekV4FlashMLABackend"
+        "vllm.models.deepseek_v41.sparse_mla.DeepseekV4FlashMLABackend"
     )
     FLASHINFER_MLA_SPARSE_DSV41 = (
-        "vllm.models.deepseek_v4_1.nvidia.flashinfer_sparse."
+        "vllm.models.deepseek_v41.nvidia.flashinfer_sparse."
         "DeepseekV4FlashInferMLASparseBackend"
     )
     B12X = "vllm.v1.attention.backends.b12x.B12xPagedAttentionBackend"