Fix gemma4_unified model type not supported (ml-explore#1349)

robertlangdonn · web-flow · commit 8239c72de5a0 · 2026-06-05T22:05:52.000+02:00
Add a MODEL_REMAPPING entry for gemma4_unified → gemma4, and extend
gemma4.sanitize() to skip the encoder-free vision_embedder weight
prefix so the text weights load cleanly.

Signed-off-by: robertlangdonn &lt;prasad@paperd.ink&gt;
diff --git a/mlx_lm/models/gemma4.py b/mlx_lm/models/gemma4.py
@@ -65,6 +65,7 @@ def sanitize(self, weights):
                     "audio_tower",
                     "embed_audio",
                     "embed_vision",
+                    "vision_embedder",  # gemma4_unified encoder-free vision
                 )
             ):
                 continue
diff --git a/mlx_lm/utils.py b/mlx_lm/utils.py
@@ -52,6 +52,7 @@
     "qwen2_5_vl": "qwen2_vl",
     "minimax_m2": "minimax",
     "iquestcoder": "llama",
+    "gemma4_unified": "gemma4",  # encoder-free multimodal variant; vision/audio weights stripped by sanitize()
 }
 
 MAX_FILE_SIZE_GB = 5

Original file line number	Diff line number	Diff line change
`@@ -65,6 +65,7 @@ def sanitize(self, weights):`
`65`	`65`	`"audio_tower",`
`66`	`66`	`"embed_audio",`
`67`	`67`	`"embed_vision",`
	`68`	`+ "vision_embedder", # gemma4_unified encoder-free vision`
`68`	`69`	`)`
`69`	`70`	`):`
`70`	`71`	`continue`
Original file line number	Diff line number	Diff line change
`@@ -52,6 +52,7 @@`
`52`	`52`	`"qwen2_5_vl": "qwen2_vl",`
`53`	`53`	`"minimax_m2": "minimax",`
`54`	`54`	`"iquestcoder": "llama",`
	`55`	`+ "gemma4_unified": "gemma4", # encoder-free multimodal variant; vision/audio weights stripped by sanitize()`
`55`	`56`	`}`
`56`	`57`
`57`	`58`	`MAX_FILE_SIZE_GB = 5`