From 465e74f49e144b1d256a04726943426d63a32766 Mon Sep 17 00:00:00 2001
From: marauder-actual <marauder@saiden.dev>
Date: Mon, 1 Jun 2026 04:26:15 +0200
Subject: [PATCH] fix: patch is_mllm_model for Qwen3.6 text-only model

AutoRound misidentifies Qwen3_5ForConditionalGeneration as a VLM
and tries to load a vision processor. Patch to force LLM mode.
---
 quantize_awq.py | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/quantize_awq.py b/quantize_awq.py
index 746c43d..c065160 100644
--- a/quantize_awq.py
+++ b/quantize_awq.py
@@ -9,6 +9,12 @@ Uses domain calibration data from substrate_v5.jsonl.
 import json
 import torch
 from transformers import AutoModelForCausalLM, AutoTokenizer
+
+# Qwen3.6-27B uses Qwen3_5ForConditionalGeneration which AutoRound
+# misidentifies as a multimodal model. Patch before importing AutoRound.
+import auto_round.autoround as _ar
+_ar.is_mllm_model = lambda *a, **kw: False
+
 from auto_round import AutoRound
 
 MODEL = "/workspace/substrate-qwen36-27b-merged"