[Bugfix][Spec Decode] Scope DSpark backend inheritance to DeepSeek V4 (#52809)

Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: Codex <noreply@openai.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
Misha Goin
2026-08-20 18:42:08 -07:00
committed by GitHub
co-authored by Codex OpenAI Codex
parent d29f7f5c92
commit 91a893de64
+34 -10
View File
@@ -3,13 +3,32 @@
import torch.nn as nn
from vllm.config import VllmConfig, replace
from vllm.config import ModelConfig, VllmConfig, replace
from vllm.distributed.parallel_state import get_pp_group
from vllm.model_executor.model_loader import get_model
from vllm.v1.worker.gpu.spec_decode.eagle.utils import (
_should_share,
get_target_lm_head,
)
from vllm.logger import init_logger
from vllm.v1.attention.backends.registry import AttentionBackendEnum
logger = init_logger(__name__)
def _resolve_dspark_attention_backend(
draft_model_config: ModelConfig,
draft_backend: AttentionBackendEnum | None,
target_backend: AttentionBackendEnum | None,
) -> AttentionBackendEnum | None:
if draft_backend is not None:
return draft_backend
# DeepSeek-V4 draft layers share the target's KV-cache layout. Other
# DSpark architectures may use a different attention kind.
if draft_model_config.hf_config.model_type == "deepseek_v4":
if target_backend is not None:
logger.info_once(
"Using the target model's %s attention backend for the "
"DeepSeek-V4 DSpark drafter.",
target_backend.name,
)
return target_backend
return None
def load_dspark_model(target_model: nn.Module, vllm_config: VllmConfig) -> nn.Module:
@@ -18,13 +37,18 @@ def load_dspark_model(target_model: nn.Module, vllm_config: VllmConfig) -> nn.Mo
draft_model_config = speculative_config.draft_model_config
from vllm.compilation.backends import set_model_tag
from vllm.model_executor.model_loader import get_model
from vllm.model_executor.models.qwen3_dflash import dflash_has_any_non_causal
from vllm.model_executor.models.utils import get_draft_quant_config
from vllm.v1.worker.gpu.spec_decode.eagle.utils import (
_should_share,
get_target_lm_head,
)
# None re-runs backend auto-selection for the draft, which can pick a
# different attention class than the target; fall back to the target's.
draft_attention_backend = (
speculative_config.attention_backend or vllm_config.attention_config.backend
draft_attention_backend = _resolve_dspark_attention_backend(
draft_model_config,
speculative_config.attention_backend,
vllm_config.attention_config.backend,
)
draft_vllm_config = replace(