[Refactor] Normalize DeepSeek V4.1 model package naming (#56741)
Signed-off-by: Bugen Zhao <i@bugenzhao.com>
This commit is contained in:
@@ -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,
|
||||
)
|
||||
|
||||
|
||||
@@ -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"),
|
||||
|
||||
@@ -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,
|
||||
)
|
||||
+3
-3
@@ -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
|
||||
+3
-3
@@ -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,
|
||||
+3
-3
@@ -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,
|
||||
+3
-3
@@ -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
|
||||
@@ -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"
|
||||
|
||||
@@ -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"
|
||||
|
||||
|
||||
@@ -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"
|
||||
|
||||
Reference in New Issue
Block a user