Add CogVLM2 support for Windows

Author: jhc13

requirements.txt

@@ -7,11 +7,6 @@ pyparsing==3.2.1
 PySide6==6.8.2.1
 transformers==4.48.3
 
-# PyTorch
-torch==2.6.0; platform_system != "Windows"
-https://download.pytorch.org/whl/cu124/torch-2.6.0%2Bcu124-cp312-cp312-win_amd64.whl; platform_system == "Windows" and python_version == "3.12"
-https://download.pytorch.org/whl/cu124/torch-2.6.0%2Bcu124-cp311-cp311-win_amd64.whl; platform_system == "Windows" and python_version == "3.11"
-
 # CogAgent
 timm==1.0.14
 
@@ -30,7 +25,17 @@ numpy==2.2.3
 huggingface-hub==0.29.1
 onnxruntime==1.20.1
 
+# PyTorch
+torch==2.6.0; platform_system != "Windows"
+https://download.pytorch.org/whl/cu124/torch-2.6.0%2Bcu124-cp312-cp312-win_amd64.whl; platform_system == "Windows" and python_version == "3.12"
+https://download.pytorch.org/whl/cu124/torch-2.6.0%2Bcu124-cp311-cp311-win_amd64.whl; platform_system == "Windows" and python_version == "3.11"
+
 # FlashAttention (Florence-2, Phi-3-Vision)
 flash-attn==2.7.4.post1; platform_system == "Linux"
 https://github.com/kingbri1/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu124torch2.6.0cxx11abiFALSE-cp312-cp312-win_amd64.whl; platform_system == "Windows" and python_version == "3.12"
 https://github.com/kingbri1/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu124torch2.6.0cxx11abiFALSE-cp311-cp311-win_amd64.whl; platform_system == "Windows" and python_version == "3.11"
+
+# Triton (CogVLM2)
+triton==3.2.0; platform_system == "Linux"
+https://github.com/woct0rdho/triton-windows/releases/download/v3.2.0-windows.post10/triton-3.2.0-cp312-cp312-win_amd64.whl; platform_system == "Windows" and python_version == "3.12"
+https://github.com/woct0rdho/triton-windows/releases/download/v3.2.0-windows.post10/triton-3.2.0-cp311-cp311-win_amd64.whl; platform_system == "Windows" and python_version == "3.11"

taggui-windows.spec

@@ -3,6 +3,7 @@ from PyInstaller.utils.hooks import collect_data_files
 
 datas = [('clip-vit-base-patch32', 'clip-vit-base-patch32'),
          ('images/icon.ico', 'images')]
+datas += collect_data_files('triton')
 datas += collect_data_files('xformers')
 hiddenimports = [
     'timm.models.layers',
@@ -27,6 +28,7 @@ a = Analysis(
     cipher=block_cipher,
     noarchive=False,
     module_collection_mode={
+        'triton': 'py',
         'xformers': 'pyz+py',
     },
 )

taggui/auto_captioning/models/cogvlm2.py

@@ -19,8 +19,8 @@ class Cogvlm2(AutoCaptioningModel):
     def get_additional_error_message(self) -> str | None:
         if not importlib.util.find_spec('triton'):
             return ('This model requires the `triton` package, which is only '
-                    'available on Linux. Therefore, this model cannot be run '
-                    'on this system.')
+                    'available for Linux and Windows. Therefore, this model '
+                    'cannot be run on this system.')
         is_4_bit_model = 'int4' in self.model_id
         if is_4_bit_model:
             if self.device_setting == CaptionDevice.CPU: