[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
+5 -5
View File
@@ -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:
@@ -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]
+5 -5
View File
@@ -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")
+6 -6
View File
@@ -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)
@@ -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 (
+3 -3
View File
@@ -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)
+2 -2
View File
@@ -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
+1 -1
View File
@@ -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"
+6 -6
View File
@@ -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(
@@ -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