JiaruiLi/DeepHealth
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Revert "Add nano and tiny model presets"

Browse Source 
This reverts commit 6f7b5be405.
This commit is contained in:
Jiarui Li
2026-07-23 16:03:29 +08:00
parent 6f7b5be405
commit 22faee7c51
3 changed files with 29 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files

29
Event_Trajectory_Shared_Reasoning_Backbone_设计方案.md
View File

@@ -1,7 +1,7 @@
# Event–Trajectory Shared Reasoning Backbone
> 状态：**Frozen implementation baseline**
> 架构标识：`event_trajectory_shared_v2`
> 架构标识：`event_trajectory_shared_v1`
> 固化日期：**2026-07-23**
## 1. 核心定义
@@ -16,12 +16,11 @@ self.reasoning_core = SharedEventTrajectoryCore(...)
禁止为不同推理轮创建独立 Transformer blocks。参数只保存一套，计算上顺序运行多轮。
模型规模固定为五档：
模型规模固定为四档：
| model_size | d_model | n_trajectory | trajectory_dim | traj_hidden |
|---|---:|---:|---:|---:|
| nano | 120 | 6 | 20 | 24 |
| tiny | 256 | 8 | 32 | 32 |
| nano | 256 | 8 | 32 | 32 |
| small | 512 | 8 | 64 | 32 |
| medium | 768 | 12 | 64 | 48 |
| huge | 1024 | 16 | 64 | 64 |
@@ -71,11 +70,10 @@ event_key_value = reasoning_core.project_event_memory(E)
## 3. Trajectory State
每个查询维护 `n_trajectory` 个显式 trajectory slots；nano 默认使用6个：
每个查询维护 `n_trajectory` 个显式 trajectory slots；nano 默认使用8个：
\[
S\in\mathbb{R}^{B\times Q\times n_{\mathrm{trajectory}}\times
d_{\mathrm{trajectory}}}.
S\in\mathbb{R}^{B\times Q\times8\times32}.
\]
其中：
@@ -86,7 +84,7 @@ d_{\mathrm{trajectory}}}.
定义可学习原型：
\[
P\in\mathbb{R}^{n_{\mathrm{trajectory}}\times d_{\mathrm{trajectory}}}.
P\in\mathbb{R}^{8\times32}.
\]
查询上下文经过投影并 reshape：
@@ -94,8 +92,7 @@ P\in\mathbb{R}^{n_{\mathrm{trajectory}}\times d_{\mathrm{trajectory}}}.
\[
C_Q
=\operatorname{QueryProjection}(\text{query features})
\in\mathbb{R}^{B\times Q\times n_{\mathrm{trajectory}}\times
d_{\mathrm{trajectory}}},
\in\mathbb{R}^{B\times Q\times8\times32},
\]
\[
@@ -343,12 +340,12 @@ next-token 的 query 位置全部并行，只有 reasoning rounds 顺序执行
训练配置必须写入：
```yaml
model_architecture: event_trajectory_shared_v2
model_architecture: event_trajectory_shared_v1
model_size: nano
d_model: 120
n_trajectory: 6
trajectory_dim: 20
traj_hidden: 24
d_model: 256
n_trajectory: 8
trajectory_dim: 32
traj_hidden: 32
n_reasoning_rounds: 12
model_parameter_count: <runtime count>
trainable_parameter_count: <runtime count>
@@ -357,7 +354,7 @@ trainable_parameter_count: <runtime count>
评估和导出入口必须同时验证：
1. `model_architecture` 完全匹配；
2. `model_size` 属于 `nano / tiny / small / medium / huge`；
2. `model_size` 属于 `nano / small / medium / huge`；
3. `d_model`、`n_trajectory`、`trajectory_dim` 和 `traj_hidden`
与对应规模预设完全匹配；
4. checkpoint 包含一套且仅一套 `reasoning_core` 关键参数；