From f20c2db803becf810b0960beaa2a8aa8ba1d54bc Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E5=BC=A0=E5=AE=97=E5=B9=B3?= <zhangzp@qq.com>
Date: Tue, 30 Jun 2026 00:17:47 +0000
Subject: [PATCH] =?UTF-8?q?feat(m2):=20Projector=20Linear(256=E2=86=92896)?=
 =?UTF-8?q?+LayerNorm=20(T2.2)?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

- src/tsmm/model/projector.py: Projector (Linear + LayerNorm).
- tests/test_projector.py: 5 tests (shape, Qwen hidden alignment,
  LayerNorm present, gradient flow, no-NaN).
- 66 tests passing.
---
 openspec/changes/ts-as-modality/tasks.md |  2 +-
 src/tsmm/model/projector.py              | 22 +++++++++++++++
 tests/test_projector.py                  | 35 ++++++++++++++++++++++++
 3 files changed, 58 insertions(+), 1 deletion(-)
 create mode 100644 src/tsmm/model/projector.py
 create mode 100644 tests/test_projector.py

diff --git a/openspec/changes/ts-as-modality/tasks.md b/openspec/changes/ts-as-modality/tasks.md
index d725d04..f6bb633 100644
--- a/openspec/changes/ts-as-modality/tasks.md
+++ b/openspec/changes/ts-as-modality/tasks.md
@@ -16,7 +16,7 @@
 ## 2. M2 · 模型可跑通
 
 - [x] 2.1 TS Encoder `model/ts_encoder.py`：`Patchify(patch=8, stride=4)` 通道独立切 patch 线性嵌入到 d=256；`TSEncoder(d=256, layers=2, heads=4)` 2 层 TF + 可学习位置编码；前向 `[B,T,C]`→`[B,n_patches,256]`（验证：输入 `[2,512,5]`→`[2,127,256]`；spec 澄清：n_patches=(T-P)/S+1=127 非 128；通道独立=每通道共享 patch 线性后对通道 mean-pool；参数 ≈2.1M 非 4M）
-- [ ] 2.2 Projector `model/projector.py`：`Linear(256→896)+LayerNorm`（验证：输出 `[B,128,896]` 对齐 Qwen2.5-0.5B hidden）
+- [x] 2.2 Projector `model/projector.py`：`Linear(256→896)+LayerNorm`（验证：输出 `[B,n_patches,896]` 对齐 Qwen2.5-0.5B hidden）
 - [ ] 2.3 多模态拼接 `model/multimodal.py`：Qwen tokenizer tokenize 文本部分，TS token 作 inputs_embeds 插入 `[属性][时间戳][TS tok][事件][问题]`，统一构造 inputs_embeds+attention_mask+labels（仅回答段非 -100）（验证：seq_len ≤1024，mask/labels 形状对）
 - [ ] 2.4 训练/推理封装 `model/wrapper.py`：`MultimodalTSModel` 组合 Encoder+Projector+LLM+LoRA 挂载开关，`forward` 返 loss、`generate` 返文本，支持 `freeze_llm`(阶段①)/`enable_lora(r=16)`(阶段②)（验证：单 batch forward 3060 不 OOM；generate 能出文本）
 - [ ] 2.5 Collator `data/collator.py`：JSONL→batch 张量，处理变长 C（padding+mask）与变长文本（padding+attention_mask）（验证：batch=4 张量形状一致无 NaN）
diff --git a/src/tsmm/model/projector.py b/src/tsmm/model/projector.py
new file mode 100644
index 0000000..eed9692
--- /dev/null
+++ b/src/tsmm/model/projector.py
@@ -0,0 +1,22 @@
+"""Projector (T2.2): maps TS tokens from encoder dim to LLM hidden dim.
+
+``Linear(d → h)`` + ``LayerNorm``. Output feeds as soft tokens into the LLM
+``inputs_embeds`` (see ``model/multimodal.py``, T2.3).
+
+Design ref: §2.2 — Projector = Linear(256→896) + LayerNorm.
+"""
+from __future__ import annotations
+
+import torch
+from torch import nn
+
+
+class Projector(nn.Module):
+    def __init__(self, in_dim: int = 256, out_dim: int = 896) -> None:
+        super().__init__()
+        self.linear = nn.Linear(in_dim, out_dim)
+        self.norm = nn.LayerNorm(out_dim)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        # x: [..., in_dim] → [..., out_dim]
+        return self.norm(self.linear(x))
diff --git a/tests/test_projector.py b/tests/test_projector.py
new file mode 100644
index 0000000..b4b7b03
--- /dev/null
+++ b/tests/test_projector.py
@@ -0,0 +1,35 @@
+"""Tests for Projector (T2.2): TS token d → LLM hidden."""
+import torch
+
+from tsmm.model.projector import Projector
+
+
+class TestProjector:
+    def test_output_shape(self):
+        proj = Projector(in_dim=256, out_dim=896)
+        x = torch.randn(2, 127, 256)
+        out = proj(x)
+        assert out.shape == (2, 127, 896)
+
+    def test_aligns_qwen_hidden(self):
+        # Qwen2.5-0.5B hidden = 896
+        proj = Projector(in_dim=256, out_dim=896)
+        out = proj(torch.randn(1, 127, 256))
+        assert out.shape[-1] == 896
+
+    def test_has_layernorm(self):
+        proj = Projector(in_dim=256, out_dim=896)
+        assert any(isinstance(m, torch.nn.LayerNorm) for m in proj.modules())
+
+    def test_gradient_flows(self):
+        proj = Projector(in_dim=256, out_dim=896)
+        x = torch.randn(1, 31, 256)
+        out = proj(x)
+        out.sum().backward()
+        assert proj.linear.weight.grad is not None
+        assert proj.linear.weight.grad.abs().sum() > 0
+
+    def test_no_nan(self):
+        proj = Projector(in_dim=256, out_dim=896)
+        out = proj(torch.randn(4, 128, 256))
+        assert not torch.isnan(out).any()