From 42d94df27953b1b3e5efc147c05416529b789095 Mon Sep 17 00:00:00 2001
From: Etherll <61019402+Etherll@users.noreply.github.com>
Date: Sat, 28 Feb 2026 05:58:32 +0200
Subject: [PATCH 1/3] Add Qwen3.5 to FORCE_FLOAT32

---
 unsloth/models/loader.py | 1 +
 1 file changed, 1 insertion(+)

diff --git a/unsloth/models/loader.py b/unsloth/models/loader.py
index 711476b759..58f5f4811c 100644
--- a/unsloth/models/loader.py
+++ b/unsloth/models/loader.py
@@ -106,6 +106,7 @@
     "gemma3text",  # Gemma3TextModel (EmbeddingGemma, standalone text-only Gemma3)
     "gemma3n",
     "gpt_oss",
+    "qwen3_5",  # Qwen3.5 RMSNorm uses (1+w) pattern like Gemma3, overflows float16
 ]
 
 global DISABLE_COMPILE_MODEL_NAMES

From 52769a6c9a6fcdf2447ed71463cb3bb2c6e91d79 Mon Sep 17 00:00:00 2001
From: Etherll <61019402+Etherll@users.noreply.github.com>
Date: Sat, 28 Feb 2026 06:03:12 +0200
Subject: [PATCH 2/3] fix vision encoder dtype mismatch

Add dtype casting for vision encoder when using FORCE_FLOAT32.
---
 unsloth/models/vision.py | 13 +++++++++++++
 1 file changed, 13 insertions(+)

diff --git a/unsloth/models/vision.py b/unsloth/models/vision.py
index 5aea471263..4c7cfb137d 100644
--- a/unsloth/models/vision.py
+++ b/unsloth/models/vision.py
@@ -1054,6 +1054,19 @@ def from_pretrained(
             do_forced_float32 = do_forced_float32,
             correct_dtype = correct_dtype,
         )
+
+        # FORCE_FLOAT32 converts all params to float16, but the vision encoder
+        # receives float32 pixel values. Cast it back to float32 to avoid dtype
+        # mismatches in nn.LayerNorm / nn.Linear (e.g. Qwen3.5, Gemma3).
+        if do_forced_float32:
+            _vision_attrs = ("visual", "vision_tower", "vision_model", "vision_encoder")
+            _inner = model.model if hasattr(model, "model") else model
+            for _va in _vision_attrs:
+                _ve = getattr(_inner, _va, None)
+                if _ve is not None:
+                    _ve.to(torch.float32)
+                    break
+
         try:
             model, tokenizer = patch_tokenizer(model, tokenizer)
         except Exception as _patch_err:

From 24f2ac51c207f788c9df8d82cbf84e75f4febf44 Mon Sep 17 00:00:00 2001
From: Etherll <61019402+Etherll@users.noreply.github.com>
Date: Mon, 2 Mar 2026 21:21:11 +0200
Subject: [PATCH 3/3] revert vision cast changes

---
 unsloth/models/vision.py | 12 ------------
 1 file changed, 12 deletions(-)

diff --git a/unsloth/models/vision.py b/unsloth/models/vision.py
index 4c7cfb137d..28c624c213 100644
--- a/unsloth/models/vision.py
+++ b/unsloth/models/vision.py
@@ -1055,18 +1055,6 @@ def from_pretrained(
             correct_dtype = correct_dtype,
         )
 
-        # FORCE_FLOAT32 converts all params to float16, but the vision encoder
-        # receives float32 pixel values. Cast it back to float32 to avoid dtype
-        # mismatches in nn.LayerNorm / nn.Linear (e.g. Qwen3.5, Gemma3).
-        if do_forced_float32:
-            _vision_attrs = ("visual", "vision_tower", "vision_model", "vision_encoder")
-            _inner = model.model if hasattr(model, "model") else model
-            for _va in _vision_attrs:
-                _ve = getattr(_inner, _va, None)
-                if _ve is not None:
-                    _ve.to(torch.float32)
-                    break
-
         try:
             model, tokenizer = patch_tokenizer(model, tokenizer)
         except Exception as _patch_err: