hack further

app.py

@@ -1,15 +1,24 @@
 import sys
 import torch
+import torch._dynamo.device_interface
+
 # Hack unsloth to work without GPU
+def fake_triton_check():
+    return False  # Prevents `bitsandbytes` from requiring GPU
 def fake_is_available():
     return True
 def fake_device_capability():
     return (8, 0)
 def fake_current_device():
     return 0
+class FakeDeviceProperties:
+    major = 7  # Mimic a compatible GPU
+    minor = 5
+
 sys.modules["torch"].cuda.is_available = fake_is_available
 sys.modules["torch"].cuda.get_device_capability = fake_device_capability
 sys.modules["torch"].cuda.current_device = fake_current_device
+sys.modules["torch._dynamo.device_interface"].Worker.get_device_properties = lambda: FakeDeviceProperties()
 
 import streamlit as st
 from unsloth.chat_templates import get_chat_template, CHAT_TEMPLATES

scratchpad.ipynb

@@ -262,7 +262,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 1,
+   "execution_count": 5,
    "metadata": {},
    "outputs": [
     {
@@ -271,7 +271,7 @@
        "False"
       ]
      },
-     "execution_count": 1,
+     "execution_count": 5,
      "metadata": {},
      "output_type": "execute_result"
     }
@@ -279,10 +279,11 @@
    "source": [
     "import os\n",
     "os.environ[\"CUDA_VISIBLE_DEVICES\"] = \"\"\n",
-    "\n",
+    "# import unsloth\n",
     "import torch\n",
-    "# torch.cuda.is_available = lambda: False\n",
-    "torch.cuda.is_available()"
+    "import torch.utils._triton\n",
+    "import sys\n",
+    "sys.modules[\"torch.utils._triton\"].has_triton()"
    ]
   },
   {