From 841f47d4ef6d6519473ed2a685455a934a8637d2 Mon Sep 17 00:00:00 2001
From: xinyanghuang7 <hsinyanghuang7@gmail.com>
Date: Sat, 2 Nov 2024 05:56:42 -0400
Subject: [PATCH] speedup and siglip

---
 3-eval_chat.py                             |   5 +++--
 model/__pycache__/LMConfig.cpython-310.pyc | Bin 1601 -> 1601 bytes
 model/__pycache__/model.cpython-310.pyc    | Bin 16442 -> 16464 bytes
 model/model.py                             |   5 +++++
 4 files changed, 8 insertions(+), 2 deletions(-)

diff --git a/3-eval_chat.py b/3-eval_chat.py
index 84fa53f..f505979 100644
--- a/3-eval_chat.py
+++ b/3-eval_chat.py
@@ -41,7 +41,7 @@ def init_model(lm_config, device, multi):
     model = model.to(device)
     print(f'模型参数: {count_parameters(model) / 1e6} 百万 = {count_parameters(model) / 1e9} B (Billion)')
 
-    vision_model, preprocess = get_vision_model()
+    vision_model, preprocess = get_vision_model(encoder_type="clip")
     vision_model = vision_model.to(device)
     return model, tokenizer, vision_model, preprocess
 
@@ -66,7 +66,8 @@ if __name__ == "__main__":
     device = 'cuda:0' if torch.cuda.is_available() else 'cpu'
     dtype = 'bfloat16'
     max_seq_len = 1024
-    lm_config = LMConfig()
+    # lm_config = LMConfig()
+    lm_config = LMConfig(image_special_token='<'*2+'>'*2, image_ids=[30]*2+[32]*2)
     lm_config.max_seq_len = max_seq_len
     model, tokenizer, vision_model, preprocess = init_model(lm_config, device, multi)
     model.eval()
diff --git a/model/__pycache__/LMConfig.cpython-310.pyc b/model/__pycache__/LMConfig.cpython-310.pyc
index fdc47c4d11b047d042095a122eaa8429a037187e..098838d25eb7fb5387a55a112037fe3c1a605a93 100644
GIT binary patch
delta 19
ZcmX@ebC8ECpO=@50SI!xZR9d%0{}1F1d{*&

delta 19
ZcmX@ebC8ECpO=@50SGcrZR9d%0{|}Q1Y-aI

diff --git a/model/__pycache__/model.cpython-310.pyc b/model/__pycache__/model.cpython-310.pyc
index 7d7909d9fe156ec64a8cd4cd78e18e83875c6e47..7a19d9a61c3b447e080a45cf372f555e8cd6393c 100644
GIT binary patch
delta 204
zcmdnhz<8m7kuRT@mx}=i8osHfPY&G3ch-n;$>cjmvQ`}oH4F<_Qy3RAX0sHDr7#48
z_-rYR9SlG*=33?&<{GAKrY!aprWEF0##)w4hFaDV4v?P7ipJWE-jjWe#aZ=gSc?KC
zXBsbO?3gTQQpn0uoLQARxzR*;@+^~X9x0%`LW~@Ye9SCNlXXpN7<neIGF{0iwmHNs
vnhD6+U{TAsY_p!_dJyNH^+m>0lQ-Dh0kV5-{TM$^zGM5F(RlL>yCg;c?HWAL

delta 196
zcmcc6z__b{kuRT@mx}=i9$!>V?+Dn)ch-n;(&RfvvKloE3s_SaI~Y<J7c$o}*D%*G
zWiw^5r7)#1_cGS9WHQvUmaqfWa7;Ef)@F2?9B(Yns#?QZ<UYCDcsXP9WMz{=akkRD
zwEUcuqAH;IMSFq7EtcZUs?^D=O?o)QfLtL)4#vsOrZtSblMk4#WE9+-Wfsi@<Q%i8
pWn8$~#d1A}^Tqli<I%~-Z0-Qri*5ZF-%S2w`<qd3^E<mFMgSTJJ)8gl

diff --git a/model/model.py b/model/model.py
index 6f8fdbb..bfa4c65 100644
--- a/model/model.py
+++ b/model/model.py
@@ -374,6 +374,11 @@ class Transformer(PreTrainedModel):
             image_ids_tensor = torch.tensor(image_ids).to(tokens.device)
             len_image_ids = len(image_ids)
 
+            # .generate时，在初始化后直接跳过
+            if len_image_ids > tokens.size(1):
+                # print(f"len_image_ids ({len_image_ids}) is greater than sequence length ({tokens.size(1)}), skipping.")
+                return None
+            
             # 使用view来创建一个视图，便于处理滑动窗口
             tokens_view = tokens.unfold(1, len_image_ids, 1)  # 在第二维度创建滑动窗口
             # 检查每个滑动窗口是否与image_ids_tensor相等