docs(research): 更新调研报告至v2.0反映NVIDIA选型

- 调研报告从RouteLLM BERT切换为NVIDIA多头分类器作为推荐方案
- 新增选型变更记录、复杂度评分公式、测试结果
- 更新tx402技术对比表和演进路线
- nvidia_router.py添加use_safetensors=True兼容transformers 4.57
This commit is contained in:
2026-04-18 01:45:07 +08:00
parent a370061a96
commit 5a322e93a0
2 changed files with 198 additions and 251 deletions

View File

@@ -43,7 +43,8 @@ class NvidiaMultiHeadClassifier(nn.Module):
# DeBERTa backbone
self.backbone = DebertaV2Model.from_pretrained(
config.base_model,
ignore_mismatched_sizes=True
ignore_mismatched_sizes=True,
use_safetensors=True
)
hidden_size = 768 # DeBERTa-v3-base