From d9434fee4aab46766c8a0a1aa6a3a8dc9f09d8f0 Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Wed, 25 Feb 2026 10:29:05 +0000
Subject: [PATCH] fix: use raw github URL for vision.py patch + add VLM
 processor diagnostic logging

---
 setup.sh                                | 4 ++++
 studio/backend/core/training/trainer.py | 9 +++++++++
 2 files changed, 13 insertions(+)

diff --git a/setup.sh b/setup.sh
index 0d8821d50a..a911d9ed7a 100755
--- a/setup.sh
+++ b/setup.sh
@@ -169,6 +169,10 @@ if [ "$IS_COLAB" = true ]; then
     LLAMA_CPP_DST="$(pip show unsloth-zoo | grep -i '^Location:' | awk '{print $2}')/unsloth_zoo/llama_cpp.py"
     curl -sSL "https://raw.githubusercontent.com/unslothai/unsloth-zoo/refs/heads/main/unsloth_zoo/llama_cpp.py" \
         -o "$LLAMA_CPP_DST"
+    # Patch: override vision.py with fix from unsloth PR: https://github.com/unslothai/unsloth/pull/4091 until next pypi release
+    VISION_DST="$(pip show unsloth | grep -i '^Location:' | awk '{print $2}')/unsloth/vision.py"
+    curl -sSL "https://raw.githubusercontent.com/unslothai/unsloth/80e0108a684c882965a02a8ed851e3473c1145ab/unsloth/models/vision.py" \
+        -o "$VISION_DST"
     echo "   Installing studio dependencies..."
     run_quiet "pip install studio" pip install -r "$SCRIPT_DIR/studio/backend/requirements/studio.txt"
     echo "✅ Python dependencies installed"
diff --git a/studio/backend/core/training/trainer.py b/studio/backend/core/training/trainer.py
index 804baa33c0..c56b4a8a89 100644
--- a/studio/backend/core/training/trainer.py
+++ b/studio/backend/core/training/trainer.py
@@ -173,6 +173,15 @@ class UnslothTrainer:
                     token=hf_token,
                 )
                 logger.info("Loaded vision model")
+
+                # Diagnostic: check if FastVisionModel returned a real Processor or a raw tokenizer
+                from transformers import ProcessorMixin
+                tok = self.tokenizer
+                has_image_proc = isinstance(tok, ProcessorMixin) or hasattr(tok, "image_processor")
+                print(f"\n[VLM Diagnostic] FastVisionModel returned: {type(tok).__name__}")
+                print(f"[VLM Diagnostic] Is ProcessorMixin: {isinstance(tok, ProcessorMixin)}")
+                print(f"[VLM Diagnostic] Has image_processor: {hasattr(tok, 'image_processor')}")
+                print(f"[VLM Diagnostic] Usable as vision processor: {has_image_proc}\n")
             else:
                 # Load text model - returns (model, tokenizer)
                 self.model, self.tokenizer = FastLanguageModel.from_pretrained(