Eagle3 DP attention for Qwen3 MoE (#12002)

This commit is contained in:
Rain H
2025-10-29 20:25:17 +08:00
committed by GitHub
parent 42f8ea4030
commit 750940ae36
9 changed files with 219 additions and 27 deletions
+2
View File
@@ -84,6 +84,8 @@ DEFAULT_MODEL_NAME_FOR_TEST_AWQ_INT4 = (
DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST = "meta-llama/Llama-2-7b-chat-hf"
DEFAULT_EAGLE_DRAFT_MODEL_FOR_TEST = "lmsys/sglang-EAGLE-llama2-chat-7B"
DEFAULT_EAGLE_TARGET_MODEL_FOR_TEST_EAGLE3 = "meta-llama/Llama-3.1-8B-Instruct"
DEFAULT_EAGLE_DP_ATTENTION_TARGET_MODEL_FOR_TEST = "Qwen/Qwen3-30B-A3B"
DEFAULT_EAGLE_DP_ATTENTION_DRAFT_MODEL_FOR_TEST = "Tengyunw/qwen3_30b_moe_eagle3"
DEFAULT_MODEL_NAME_FOR_TEST_EAGLE3 = "lmsys/sglang-EAGLE3-LLaMA3.1-Instruct-8B"
DEFAULT_STANDALONE_SPECULATIVE_TARGET_MODEL_FOR_TEST = (
"meta-llama/Llama-3.1-8B-Instruct"