cf1584f373 · tduncan/Nvidia-Optimization-Stack-VLLM

.buildkite/test_areas/distributed.yaml

+1 -1

@@ -208,7 +208,7 @@ steps:

208 - tests/distributed/test_utils

209 - tests/distributed/test_engram_dp_shard.py

210 - vllm/config/engram.py

211 - - vllm/models/deepseek_v4_1/

211 + - vllm/models/deepseek_v41/

212 commands:

213 # https://github.com/NVIDIA/nccl/issues/1838

214 - export NCCL_CUMEM_HOST_ENABLE=0

.github/mergify.yml

+2 -3

@@ -343,9 +343,8 @@ pull_request_rules:

343 conditions:

344 - label != stale

345 - or:

346 - # The 4.1 sources spell the model three ways: deepseek_v4_1 (models),

347 - # deepseek_v41 (most) and deepseekv41 (tool parsers).

348 - - files~=^vllm/models/deepseek_v4_1/

346 + # The 4.1 sources use deepseek_v41 and deepseekv41 (tool parsers).

347 + - files~=^vllm/models/deepseek_v41/

348 - files~=^vllm/(?:parser|tokenizers|reasoning|transformers_utils/configs)/deepseek_v41.*\.py

349 - files~=^vllm/tool_parsers/deepseekv41.*\.py

350 - files~=^rust/src/(?:chat|parser)/src/.*deepseek_v41

tests/distributed/test_engram_dp_shard.py

+5 -5

@@ -39,9 +39,9 @@ from vllm.distributed.parallel_state import (

39 initialize_model_parallel,

40 )

41 from vllm.forward_context import set_forward_context

42 -from vllm.models.deepseek_v4_1.common.engram import EngramLayout

43 -from vllm.models.deepseek_v4_1.nvidia import engram as engram_ops

44 -from vllm.models.deepseek_v4_1.nvidia.engram import (

42 +from vllm.models.deepseek_v41.common.engram import EngramLayout

43 +from vllm.models.deepseek_v41.nvidia import engram as engram_ops

44 +from vllm.models.deepseek_v41.nvidia.engram import (

45 Engram,

46 ParallelEngramEmbedding,

47 engram_head_shard_rank,

@@ -473,8 +473,8 @@ def _check_dummy_hash_model_forward(

473 vllm_config, engram, head_sizes, weight, scales, dp_shared_memory

474 ):

475 """A metadata-less replica joins DP sharding but skips a shared lookup."""

476 - from vllm.models.deepseek_v4_1.common.engram import NgramHashState

477 - from vllm.models.deepseek_v4_1.nvidia import model as model_ops

476 + from vllm.models.deepseek_v41.common.engram import NgramHashState

477 + from vllm.models.deepseek_v41.nvidia import model as model_ops

478

479 dp_rank = vllm_config.parallel_config.data_parallel_rank

480 if dp_shared_memory and dp_rank != 1:

tests/kernels/core/test_fused_q_kv_rmsnorm.py

+4 -4

@@ -125,7 +125,7 @@ def test_fused_q_kv_rmsnorm_quant_matches_separate(num_tokens, q_size, dtype):

125 from vllm.model_executor.layers.quantization.utils.mxfp8_utils import (

126 mxfp8_e4m3_quantize,

127 )

128 - from vllm.models.deepseek_v4_1.common.ops.query_quant import (

128 + from vllm.models.deepseek_v41.common.ops.query_quant import (

129 fused_q_kv_rmsnorm_quant,

130 )

131

@@ -198,15 +198,15 @@ def test_shared_query_quant_preserves_projection(

198 mxfp8_e4m3_quantize,

199 )

200 from vllm.model_executor.layers.quantization.utils.quant_utils import kNvfp4Dynamic

201 - from vllm.models.deepseek_v4_1.attention import (

201 + from vllm.models.deepseek_v41.attention import (

202 DeepseekV4Attention,

203 DeepseekV4Indexer,

204 )

205 - from vllm.models.deepseek_v4_1.common.ops.query_quant import (

205 + from vllm.models.deepseek_v41.common.ops.query_quant import (

206 can_fuse_query_quant,

207 fused_q_kv_rmsnorm_quant,

208 )

209 - from vllm.models.deepseek_v4_1.quant_config import DeepseekV4FP8Config

209 + from vllm.models.deepseek_v41.quant_config import DeepseekV4FP8Config

210

211 cls: (

212 type[FlashInferCutedslMxfp8LinearKernel]

tests/kernels/test_compressor_kv_cache.py

+5 -5

@@ -65,7 +65,7 @@ def test_v41_fused_save_compress_and_insert(

65 a Torch reference with BF16 rounding before RoPE. Replays change both

66 inputs and history so stale scratch/cache values cannot satisfy the test.

67 """

68 - from vllm.models.deepseek_v4_1.common.ops.fused_compress_quant_cache import (

68 + from vllm.models.deepseek_v41.common.ops.fused_compress_quant_cache import (

69 fused_save_compress_norm,

70 rope_quant_insert,

71 )

@@ -226,7 +226,7 @@ def test_v41_rope_insert_plain_row(compress_ratio: int, store_fp8: bool):

226 Non-boundary positions, negative slots and untouched rows must keep their

227 prior contents.

228 """

229 - from vllm.models.deepseek_v4_1.common.ops.fused_compress_quant_cache import (

229 + from vllm.models.deepseek_v41.common.ops.fused_compress_quant_cache import (

230 rope_quant_insert,

231 )

232

@@ -301,7 +301,7 @@ def test_v41_compressor_metadata_maps_tokens_to_their_ring():

301 capacity`` and keep padding tokens at PAD."""

302 from unittest.mock import MagicMock

303

304 - from vllm.models.deepseek_v4_1.compressor import CompressorMetadataBuilder

304 + from vllm.models.deepseek_v41.compressor import CompressorMetadataBuilder

305 from vllm.v1.attention.backend import CommonAttentionMetadata

306 from vllm.v1.kv_cache_interface import CircularBufferSpec

307

@@ -351,11 +351,11 @@ def test_v41_compressor_metadata_maps_tokens_to_their_ring():

351 def test_v41_attention_joins_cache_writes_before_consumption(use_aux, use_graph):

352 """Both reused-event joins must publish this forward's states and cache rows."""

353 from vllm.forward_context import ForwardContext, override_forward_context

354 - from vllm.models.deepseek_v4_1.attention import (

354 + from vllm.models.deepseek_v41.attention import (

355 DeepseekV4Attention,

356 DeepseekV4Indexer,

357 )

358 - from vllm.models.deepseek_v4_1.compressor import DeepseekCompressor

358 + from vllm.models.deepseek_v41.compressor import DeepseekCompressor

359

360 torch.manual_seed(43)

361 raw = torch.randn(19, 1024, device="cuda")

tests/kernels/test_engram.py

+6 -6

@@ -7,15 +7,15 @@ from types import SimpleNamespace

7 import pytest

8 import torch

9

10 -from vllm.models.deepseek_v4_1.common import engram as engram_ops

11 -from vllm.models.deepseek_v4_1.common.engram import (

10 +from vllm.models.deepseek_v41.common import engram as engram_ops

11 +from vllm.models.deepseek_v41.common.engram import (

12 Engram as CommonEngram,

13 )

14 -from vllm.models.deepseek_v4_1.common.engram import (

14 +from vllm.models.deepseek_v41.common.engram import (

15 NgramHashState,

16 )

17 -from vllm.models.deepseek_v4_1.nvidia import engram as nvidia_engram_ops

18 -from vllm.models.deepseek_v4_1.nvidia.engram import Engram, ParallelEngramEmbedding

17 +from vllm.models.deepseek_v41.nvidia import engram as nvidia_engram_ops

18 +from vllm.models.deepseek_v41.nvidia.engram import Engram, ParallelEngramEmbedding

19 from vllm.platforms import current_platform

20

21

@@ -503,7 +503,7 @@ def test_lookback_window_reproduces_single_instance(runner, capture):

503 @pytest.mark.skipif(not current_platform.is_cuda(), reason="CUDA required")

504 def test_v2_model_state_gathers_lookback_window():

505 """The window is gathered on device from the runner's token history."""

506 - from vllm.models.deepseek_v4_1.nvidia.model_state import DeepseekV41ModelState

506 + from vllm.models.deepseek_v41.nvidia.model_state import DeepseekV41ModelState

507

508 depth, max_num_reqs, max_model_len = 3, 4, 16

509 state = DeepseekV41ModelState.__new__(DeepseekV41ModelState)

tests/kernels/test_fused_indexer_q_rope_quant.py

+3 -3

@@ -359,7 +359,7 @@ def test_indexer_k_inserts_only_valid_groups_into_padded_pages(

359 num_tokens, compress_ratio, use_fp4, cache_dtype

360 ):

361 """Insert group ends and preserve skipped slots, graph rows, and page padding."""

362 - from vllm.models.deepseek_v4_1.common.ops.indexer_k_store import (

362 + from vllm.models.deepseek_v41.common.ops.indexer_k_store import (

363 indexer_k_norm_rope_store,

364 )

365

@@ -442,7 +442,7 @@ def test_indexer_k_inserts_only_valid_groups_into_padded_pages(

442 @torch.inference_mode()

443 def test_indexer_k_cuda_graph_replay_reads_current_projection(use_fp4):

444 """Replay must consume updated keys and slot mapping through the captured API."""

445 - from vllm.models.deepseek_v4_1.common.ops.indexer_k_store import (

445 + from vllm.models.deepseek_v41.common.ops.indexer_k_store import (

446 indexer_k_norm_rope_store,

447 )

448

@@ -501,7 +501,7 @@ def test_indexer_k_store_roundtrips_through_rocm_gather(block_size, compress_rat

501 ``block_size > 1``, so a row-major store would feed the indexer permuted

502 key bytes and randomize its top-k. Assert the write/read pair is exact.

503 """

504 - from vllm.models.deepseek_v4_1.common.ops.indexer_k_store import (

504 + from vllm.models.deepseek_v41.common.ops.indexer_k_store import (

505 indexer_k_norm_rope_store,

506 )

507 from vllm.v1.attention.ops.rocm_aiter_mla_sparse import (

tests/kernels/test_mhc_kernels.py

+3 -3

@@ -32,7 +32,7 @@ from vllm.models.deepseek_v4.nvidia.model import (

32 DeepseekV4DecoderLayer,

33 DeepseekV4Model,

34 )

35 -from vllm.models.deepseek_v4_1.nvidia.model import (

35 +from vllm.models.deepseek_v41.nvidia.model import (

36 DeepseekV4DecoderLayer as DeepseekV41DecoderLayer,

37 )

38 from vllm.platforms import current_platform

@@ -104,7 +104,7 @@ def test_hc_collapse_custom_op_supports_compile():

104 @pytest.mark.parametrize("num_tokens", [1, 7])

105 def test_v41_dspark_head_collapses_with_last_ffn_mix(num_tokens, monkeypatch):

106 """Match reference DSparkBlock.forward_head's hc_pre before its RMSNorm."""

107 - from vllm.models.deepseek_v4_1.nvidia import dspark

107 + from vllm.models.deepseek_v41.nvidia import dspark

108

109 set_random_seed(0)

110 hidden_size, hc_mult = 5120, 4

@@ -296,7 +296,7 @@ def test_deepseek_v41_decoder_mixes_match_torch(

296 return post, res, decoder.attn_norm(collapsed), pre

297

298 monkeypatch.setattr(

299 - "vllm.models.deepseek_v4_1.nvidia.model.mhc_pre_delayed_tilelang",

299 + "vllm.models.deepseek_v41.nvidia.model.mhc_pre_delayed_tilelang",

300 reference,

301 )

302 expected = decoder(x, positions, None, **kwargs)

tests/models/test_deepseek_v4_mega_moe.py

+2 -2

@@ -18,8 +18,8 @@ from vllm.models.deepseek_v4.nvidia.model import (

18 )

19 from vllm.models.deepseek_v4.nvidia.mtp import DeepSeekV4MTP

20 from vllm.models.deepseek_v4.nvidia.ops.prepare_megamoe import prepare_megamoe_inputs

21 -from vllm.models.deepseek_v4_1.common.mm_preprocess import IMAGE_SENTINEL_BASE_ID

22 -from vllm.models.deepseek_v4_1.nvidia.model import DeepseekV4MoE as DeepseekV41MoE

21 +from vllm.models.deepseek_v41.common.mm_preprocess import IMAGE_SENTINEL_BASE_ID

22 +from vllm.models.deepseek_v41.nvidia.model import DeepseekV4MoE as DeepseekV41MoE

23 from vllm.platforms import current_platform

24 from vllm.transformers_utils.configs.deepseek_v41 import DeepseekV41Config

25

tests/models/test_dspark_mla.py

+1 -1

@@ -216,7 +216,7 @@ def test_v41_dspark_loads_linear_scales(

216 """Checkpoint ``.scale`` maps to the quant method's scale parameter and

217 loads untouched. MXFP8 block-scale expansion lives in the KMxfp8Static

218 loader (see tests/quantization/test_modelopt.py), not in load_weights."""

219 - from vllm.models.deepseek_v4_1.nvidia import dspark

219 + from vllm.models.deepseek_v41.nvidia import dspark

220

221 mxfp8 = scale_dtype != torch.float32

222 scale_name = "weight_scale" if mxfp8 else "weight_scale_inv"

tests/quantization/test_fp8.py

+6 -6

@@ -77,8 +77,8 @@ def test_deepseek_v41_mxfp8_scale_loading(

77 MergedColumnParallelLinear,

78 RowParallelLinear,

79 )

80 - from vllm.models.deepseek_v4_1 import quant_config as quant_module

81 - from vllm.models.deepseek_v4_1.nvidia import model as model_module

80 + from vllm.models.deepseek_v41 import quant_config as quant_module

81 + from vllm.models.deepseek_v41.nvidia import model as model_module

82

83 default_vllm_config.model_config = SimpleNamespace(dtype=torch.bfloat16)

84 quant_config = quant_module.DeepseekV4FP8Config(

@@ -214,8 +214,8 @@ def test_deepseek_v41_vl_mapper_routes_linear_scales(

214 """The VL wrapper must map ``.scale`` keys to the parameter the linear

215 quant method registers, as the text model does. A hardcoded

216 ``weight_scale_inv`` raised KeyError for native MXFP8 checkpoints."""

217 - from vllm.models.deepseek_v4_1.nvidia import model as model_module

218 - from vllm.models.deepseek_v4_1.nvidia import vl_model as vl_module

217 + from vllm.models.deepseek_v41.nvidia import model as model_module

218 + from vllm.models.deepseek_v41.nvidia import vl_model as vl_module

219

220 vllm_config = SimpleNamespace(

221 quant_config=SimpleNamespace(weight_block_size=weight_block_size)

@@ -318,8 +318,8 @@ def test_mxfp8_bmm_loads_and_projects_grouped_weights(

318 compute_fp8_einsum_recipe,

319 deep_gemm_fp8_o_proj,

320 )

321 - from vllm.models.deepseek_v4_1.nvidia.model import DeepseekV4Model

322 - from vllm.models.deepseek_v4_1.quant_config import DeepseekV4FP8Config

321 + from vllm.models.deepseek_v41.nvidia.model import DeepseekV4Model

322 + from vllm.models.deepseek_v41.quant_config import DeepseekV4FP8Config

323 from vllm.utils.deep_gemm import is_deep_gemm_supported

324

325 if not is_deep_gemm_supported() or not current_platform.is_device_capability_family(

tests/v1/attention/test_indexer_deepseek_v4_slot_mapping.py

+1 -1

@@ -8,7 +8,7 @@ import torch

8

9 from tests.v1.attention.utils import create_vllm_config

10 from vllm.models.deepseek_v4.sparse_mla import DeepseekV4SparseMLABackend

11 -from vllm.models.deepseek_v4_1.sparse_mla import (

11 +from vllm.models.deepseek_v41.sparse_mla import (

12 DeepseekV4SparseMLABackend as DeepseekV41SparseMLABackend,

13 )

14 from vllm.v1.attention.backend import CommonAttentionMetadata

vllm/model_executor/layers/quantization/__init__.py

+1 -1

@@ -120,7 +120,7 @@ def get_quantization_config(quantization: str) -> type[QuantizationConfig]:

120 # The v4.1 class is the v4 one extended to accept model_type

121 # "deepseek_v41" and its 32x32 MXFP8 linear layout. V4.1 has

122 # platform-specific implementations for both CUDA and ROCm.

123 - from vllm.models.deepseek_v4_1 import (

123 + from vllm.models.deepseek_v41 import (

124 DeepseekV4FP8Config as DeepseekV41FP8Config,

125 )

126

vllm/model_executor/models/registry.py

+2 -2

@@ -377,7 +377,7 @@ _MULTIMODAL_MODELS = {

377 "DeepseekV4ForConditionalGeneration",

378 ),

379 "DeepseekV41ForCausalLM": (

380 - "vllm.models.deepseek_v4_1",

380 + "vllm.models.deepseek_v41",

381 "DeepseekV41ForCausalLM",

382 ),

383 "Dots3NoteForCausalLM": (

@@ -645,7 +645,7 @@ _SPECULATIVE_DECODING_MODELS = {

645 "DFlashMuseGlimmerAssistantModel": ("qwen3_dflash", "DFlashQwen3ForCausalLM"),

646 "DSparkDraftModel": ("vllm.models.deepseek_v4", "DSparkDeepseekV4ForCausalLM"),

647 "DSparkV41DraftModel": (

648 - "vllm.models.deepseek_v4_1",

648 + "vllm.models.deepseek_v41",

649 "DSparkDeepseekV4ForCausalLM",

650 ),

651 "Qwen3DSparkModel": ("qwen3_dspark", "Qwen3DSparkForCausalLM"),

vllm/models/deepseek_v41/__init__.py renamed

vllm/models/deepseek_v41/amd/__init__.py renamed

vllm/models/deepseek_v41/amd/dspark.py renamed

vllm/models/deepseek_v41/amd/model.py renamed

+2 -2

@@ -54,8 +54,8 @@ from vllm.models.common.ops.sequence_parallel import (

54 from vllm.models.deepseek_v4.amd.model import (

55 DeepseekV4MoE as DeepseekV4MoEBase,

56 )

57 -from vllm.models.deepseek_v4_1.amd.rocm import DeepseekV41ROCMAiterMLAAttention

58 -from vllm.models.deepseek_v4_1.attention import DeepseekV4Attention

57 +from vllm.models.deepseek_v41.amd.rocm import DeepseekV41ROCMAiterMLAAttention

58 +from vllm.models.deepseek_v41.attention import DeepseekV4Attention

59 from vllm.sequence import IntermediateTensors

60 from vllm.utils.math_utils import cdiv

61 from vllm.v1.attention.backends.registry import AttentionBackendEnum

vllm/models/deepseek_v41/amd/rocm.py renamed

+3 -3

@@ -13,9 +13,9 @@ from vllm.distributed import (

13 )

14 from vllm.forward_context import get_forward_context

15 from vllm.logger import init_logger

16 -from vllm.models.deepseek_v4_1.attention import DeepseekV4Attention

17 -from vllm.models.deepseek_v4_1.common.ops import dequantize_and_gather_k_cache

18 -from vllm.models.deepseek_v4_1.sparse_mla import (

16 +from vllm.models.deepseek_v41.attention import DeepseekV4Attention

17 +from vllm.models.deepseek_v41.common.ops import dequantize_and_gather_k_cache

18 +from vllm.models.deepseek_v41.sparse_mla import (

19 DeepseekV4FlashMLAMetadata,

20 DeepseekV4SparseMLABackend,

21 DeepseekV4SparseMLAMetadataBuilder,

vllm/models/deepseek_v41/amd/vl_model.py renamed

vllm/models/deepseek_v41/attention.py renamed

+7 -7

@@ -26,7 +26,7 @@ from vllm.model_executor.layers.linear import (

26 )

27 from vllm.model_executor.layers.sparse_attn_indexer import SparseAttnIndexer

28 from vllm.models.common.ops import fused_q_kv_rmsnorm

29 -from vllm.models.deepseek_v4_1.common.ops import (

29 +from vllm.models.deepseek_v41.common.ops import (

30 MXFP4_BLOCK_SIZE,

31 fused_indexer_q_rope_quant,

32 indexer_k_norm_rope_store,

@@ -49,8 +49,8 @@ from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase

49 from vllm.model_executor.layers.layernorm import RMSNorm

50 from vllm.model_executor.layers.quantization import QuantizationConfig

51 from vllm.model_executor.models.utils import extract_layer_index

52 -from vllm.models.deepseek_v4_1.common.rope import build_deepseek_v4_rope

53 -from vllm.models.deepseek_v4_1.compressor import DeepseekCompressor

52 +from vllm.models.deepseek_v41.common.rope import build_deepseek_v4_rope

53 +from vllm.models.deepseek_v41.compressor import DeepseekCompressor

54 from vllm.triton_utils import tl, triton

55 from vllm.utils.multi_stream_utils import (

56 execute_in_parallel,

@@ -156,7 +156,7 @@ class DeepseekV4Attention(nn.Module, AttentionLayerBase, ABC):

156 ``DeepseekV4FlashInferSM120Attention`` /

157 ``DeepseekV4FlashInferMLAAttention`` (CUDA) or

158 ``DeepseekV41ROCMAiterMLAAttention`` (ROCm) — selected by the platform-specific

159 - deepseek_v4_1 model module. The base is never instantiated directly.

159 + deepseek_v41 model module. The base is never instantiated directly.

160 """

161

162 # Provided by the platform subclass.

@@ -536,7 +536,7 @@ class DeepseekV4Attention(nn.Module, AttentionLayerBase, ABC):

536 "FLASHMLA_SPARSE_DSV41",

537 "ROCM_FLASHMLA_SPARSE_DSV4",

538 ):

539 - from vllm.models.deepseek_v4_1.common.ops.cache_utils import (

539 + from vllm.models.deepseek_v41.common.ops.cache_utils import (

540 _COMBINE_TOPK_SWA_INDICES_KERNEL,

541 )

542

@@ -581,7 +581,7 @@ class DeepseekV4Attention(nn.Module, AttentionLayerBase, ABC):

581

582 @cached_property

583 def _can_fuse_query_quant(self) -> bool:

584 - from vllm.models.deepseek_v4_1.common.ops.query_quant import (

584 + from vllm.models.deepseek_v41.common.ops.query_quant import (

585 can_fuse_query_quant,

586 )

587

@@ -600,7 +600,7 @@ class DeepseekV4Attention(nn.Module, AttentionLayerBase, ABC):

600 """

601 qr, kv = qr_kv.split([self.q_lora_rank, self.head_dim], dim=-1)

602 if self.q_lora_rank % 32 == 0 and self._can_fuse_query_quant:

603 - from vllm.models.deepseek_v4_1.common.ops.query_quant import (

603 + from vllm.models.deepseek_v41.common.ops.query_quant import (

604 fused_q_kv_rmsnorm_quant,

605 )

606

vllm/models/deepseek_v41/common/__init__.py renamed

vllm/models/deepseek_v41/common/engram.py renamed

vllm/models/deepseek_v41/common/mm_preprocess.py renamed

vllm/models/deepseek_v41/common/ops/__init__.py renamed

vllm/models/deepseek_v41/common/ops/cache_utils.py renamed

vllm/models/deepseek_v41/common/ops/fused_compress_quant_cache.py renamed

vllm/models/deepseek_v41/common/ops/indexer_k_store.py renamed

vllm/models/deepseek_v41/common/ops/query_quant.py renamed

vllm/models/deepseek_v41/common/rope.py renamed

vllm/models/deepseek_v41/compressor.py renamed

+1 -1

@@ -12,7 +12,7 @@ from vllm.forward_context import get_forward_context

12 from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase

13 from vllm.model_executor.layers.layernorm import RMSNorm

14 from vllm.model_executor.layers.linear import MergedColumnParallelLinear

15 -from vllm.models.deepseek_v4_1.common.ops.fused_compress_quant_cache import (

15 +from vllm.models.deepseek_v41.common.ops.fused_compress_quant_cache import (

16 fused_save_compress_norm,

17 rope_quant_insert,

18 )

vllm/models/deepseek_v41/nvidia/__init__.py renamed

vllm/models/deepseek_v41/nvidia/dspark.py renamed

vllm/models/deepseek_v41/nvidia/engram.py renamed

+3 -3

@@ -23,16 +23,16 @@ from vllm.distributed.parallel_state import GroupCoordinator

23 from vllm.forward_context import get_forward_context

24 from vllm.logger import init_logger

25 from vllm.model_executor.utils import set_weight_attrs

26 -from vllm.models.deepseek_v4_1.common.engram import (

26 +from vllm.models.deepseek_v41.common.engram import (

27 DEAD_ID,

28 EngramLayout,

29 _engram_head_shard_weight_loader,

30 _engram_select_rows,

31 )

32 -from vllm.models.deepseek_v4_1.common.engram import (

32 +from vllm.models.deepseek_v41.common.engram import (

33 Engram as BaseEngram,

34 )

35 -from vllm.models.deepseek_v4_1.common.engram import (

35 +from vllm.models.deepseek_v41.common.engram import (

36 ParallelEngramEmbedding as BaseParallelEngramEmbedding,

37 )

38 from vllm.utils.platform_utils import is_uva_available

vllm/models/deepseek_v41/nvidia/flashinfer_sparse.py renamed

+3 -3

@@ -13,12 +13,12 @@ from vllm.models.deepseek_v4.nvidia.ops.o_proj import (

13 compute_fp8_einsum_recipe,

14 deep_gemm_fp8_o_proj,

15 )

16 -from vllm.models.deepseek_v4_1.attention import DeepseekV4Attention

17 -from vllm.models.deepseek_v4_1.common.ops import (

16 +from vllm.models.deepseek_v41.attention import DeepseekV4Attention

17 +from vllm.models.deepseek_v41.common.ops import (

18 build_flashinfer_mixed_sparse_indices,

19 compute_global_topk_indices_and_lens,

20 )

21 -from vllm.models.deepseek_v4_1.sparse_mla import (

21 +from vllm.models.deepseek_v41.sparse_mla import (

22 DeepseekV4FlashMLAMetadata,

23 DeepseekV4SparseMLABackend,

24 DeepseekV4SparseMLAMetadataBuilder,

vllm/models/deepseek_v41/nvidia/flashmla.py renamed

+3 -3

@@ -10,13 +10,13 @@ from vllm.models.deepseek_v4.nvidia.ops.o_proj import (

10 compute_fp8_einsum_recipe,

11 deep_gemm_fp8_o_proj,

12 )

13 -from vllm.models.deepseek_v4_1.attention import DeepseekV4Attention

14 -from vllm.models.deepseek_v4_1.common.ops import (

13 +from vllm.models.deepseek_v41.attention import DeepseekV4Attention

14 +from vllm.models.deepseek_v41.common.ops import (

15 combine_topk_swa_indices,

16 compute_global_topk_indices_and_lens,

17 dequantize_and_gather_k_cache,

18 )

19 -from vllm.models.deepseek_v4_1.sparse_mla import (

19 +from vllm.models.deepseek_v41.sparse_mla import (

20 DeepseekV4FlashMLABackend,

21 DeepseekV4FlashMLAMetadata,

22 DeepseekV41SparseSWAMetadataBuilder,

vllm/models/deepseek_v41/nvidia/model.py renamed

+3 -3

@@ -63,12 +63,12 @@ from vllm.models.deepseek_v4.nvidia.model import (

63 from vllm.models.deepseek_v4.nvidia.model import (

64 make_deepseek_v4_expert_params_mapping,

65 )

66 -from vllm.models.deepseek_v4_1.attention import DeepseekV4Attention

67 -from vllm.models.deepseek_v4_1.nvidia.flashinfer_sparse import (

66 +from vllm.models.deepseek_v41.attention import DeepseekV4Attention

67 +from vllm.models.deepseek_v41.nvidia.flashinfer_sparse import (

68 DeepseekV4FlashInferMLAAttention,

69 DeepseekV4FlashInferSM120Attention,

70 )

71 -from vllm.models.deepseek_v4_1.nvidia.flashmla import DeepseekV4FlashMLAAttention

71 +from vllm.models.deepseek_v41.nvidia.flashmla import DeepseekV4FlashMLAAttention

72 from vllm.platforms import current_platform

73 from vllm.sequence import IntermediateTensors

74 from vllm.utils.math_utils import cdiv

vllm/models/deepseek_v41/nvidia/model_state.py renamed

vllm/models/deepseek_v41/nvidia/vl_model.py renamed

vllm/models/deepseek_v41/quant_config.py renamed

vllm/models/deepseek_v41/sparse_mla.py renamed

vllm/transformers_utils/configs/deepseek_v41.py

+1 -1

@@ -12,7 +12,7 @@ class DeepseekV41Config(PretrainedConfig):

12 tower under ``vision_config``. The vLLM model code (ported from

13 ``deepseek_v4``) reads flat attributes, so both sub-configs are flattened

14 onto the top level here: text fields are exposed as-is, vision fields

15 - with the ``vision_*`` naming used by ``deepseek_v4_1.common.vision``.

15 + with the ``vision_*`` naming used by ``deepseek_v41.common.vision``.

16 """

17

18 model_type = "deepseek_v41"

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

+1 -1

@@ -49,7 +49,7 @@ _LAYER_TYPE_C128A = "c128a"

49 # v4.1 ratio-1 / ratio-2 indexer layers: same indexer-path shape as C4A, but

50 # their compressed page block is block_size // ratio, so each gets its own plan.

51 # v4.1 builders classify layer types themselves (ratio 0 is SWA-only there); see

52 -# deepseek_v4_1/sparse_mla.py.

52 +# deepseek_v41/sparse_mla.py.

53 _LAYER_TYPE_C1A = "c1a"

54 _LAYER_TYPE_C2A = "c2a"

55

vllm/v1/attention/backends/registry.py

+2 -2

@@ -108,10 +108,10 @@ class AttentionBackendEnum(Enum, metaclass=_AttentionBackendEnumMeta):

108 # layer). Separate names from DSV4 so a V4.1 model never resolves the

109 # V4.0 backend classes through this enum.

110 FLASHMLA_SPARSE_DSV41 = (

111 - "vllm.models.deepseek_v4_1.sparse_mla.DeepseekV4FlashMLABackend"

111 + "vllm.models.deepseek_v41.sparse_mla.DeepseekV4FlashMLABackend"

112 )

113 FLASHINFER_MLA_SPARSE_DSV41 = (

114 - "vllm.models.deepseek_v4_1.nvidia.flashinfer_sparse."

114 + "vllm.models.deepseek_v41.nvidia.flashinfer_sparse."

115 "DeepseekV4FlashInferMLASparseBackend"

116 )

117 B12X = "vllm.v1.attention.backends.b12x.B12xPagedAttentionBackend"

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