Serendipity/CTI-Inference-Opt
1
1
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

chore: 移除 emb_fp16 开关，暂不启用 Embedding FP16

Browse Source 
Co-Authored-By: Claude <noreply@anthropic.com>
This commit is contained in:
刘航宇
2026-06-15 17:33:54 +08:00
parent af1795d371
commit 887a8cff86
1 changed files with 0 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
代码/code/infer.py
-1
View File
@@ -34,7 +34,6 @@ except Exception:
# ============================================================ # ============================================================
CONFIG = { CONFIG = {
"fp16": True, # True=半精度推理；False=FP32 参考跑（确立 AUC 天花板） "fp16": True, # True=半精度推理；False=FP32 参考跑（确立 AUC 天花板）
"emb_fp16": False, # True=Embedding 也 FP16（省 ~10GB 显存带宽，AUC 可能微降）
"keep_fp32_modules": (), # fp16 下仍保留 FP32 的子模块名前缀，如 ("linear",) "keep_fp32_modules": (), # fp16 下仍保留 FP32 的子模块名前缀，如 ("linear",)
"expert_merge": True, # 是否做 expert 权重相似度合并 "expert_merge": True, # 是否做 expert 权重相似度合并
"merge_threshold": 0.90, # 合并的余弦相似度阈值 "merge_threshold": 0.90, # 合并的余弦相似度阈值