diff --git a/vllm/model_executor/models/deepseek_eagle3.py b/vllm/model_executor/models/deepseek_eagle3.py index 71ce971d53..faa2c26d9d 100644 --- a/vllm/model_executor/models/deepseek_eagle3.py +++ b/vllm/model_executor/models/deepseek_eagle3.py @@ -3,7 +3,6 @@ """Eagle3 speculative decoding model for DeepseekV2/V3 with MLP (no MoE).""" -import copy from collections.abc import Iterable import torch @@ -58,7 +57,6 @@ class DeepseekV2Eagle3DecoderLayer(nn.Module): quant_config = get_draft_quant_config(vllm_config) self.hidden_size = config.hidden_size - rope_scaling = getattr(config, "rope_scaling", None) max_position_embeddings = getattr(config, "max_position_embeddings", 8192) self.layer_idx = layer_idx @@ -68,13 +66,6 @@ class DeepseekV2Eagle3DecoderLayer(nn.Module): qk_rope_head_dim = getattr(config, "qk_rope_head_dim", 0) v_head_dim = getattr(config, "v_head_dim", 0) kv_lora_rank = getattr(config, "kv_lora_rank", 0) - config = copy.copy(config) - if rope_scaling: - rope_params = rope_scaling.copy() - rope_params["rope_type"] = "deepseek_yarn" - else: - rope_params = {"rope_type": "default"} - config.rope_parameters = rope_params self.self_attn = DeepseekV2MLAAttention( vllm_config=vllm_config, config=config,