[Bugfix][Model] Fix K3 DSpark config for 96-head drafts (#54005)
Signed-off-by: zixi-qi <zixi@inferact.ai> Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
@@ -1,10 +1,10 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
|
||||
from transformers import DeepseekV2Config
|
||||
from transformers import DeepseekV3Config
|
||||
|
||||
|
||||
class K3DSparkConfig(DeepseekV2Config):
|
||||
class K3DSparkConfig(DeepseekV3Config):
|
||||
"""Configuration for a dense MLA DSpark draft model."""
|
||||
|
||||
model_type = "k3_dspark"
|
||||
@@ -18,7 +18,7 @@ class K3DSparkConfig(DeepseekV2Config):
|
||||
rope_theta: float = 50000.0,
|
||||
**kwargs,
|
||||
) -> None:
|
||||
# DeepseekV2Config defaults to a MoE topology. Zero these fields so
|
||||
# DeepseekV3Config defaults to a MoE topology. Zero these fields so
|
||||
# generic vLLM config logic also recognizes this draft as dense.
|
||||
kwargs.setdefault("n_routed_experts", 0)
|
||||
kwargs.setdefault("n_shared_experts", 0)
|
||||
|
||||
Reference in New Issue
Block a user