weights re-quant for mlx-lm 0.31 KV-shared Gemma-4 arch: optiq_metadata.json

optiq_metadata.json

@@ -1,15 +1,15 @@
 {
   "method": "optiq_mixed_precision",
   "base_model": "google/gemma-4-e4b-it",
-  "reference": "bf16",
+  "reference": "uniform_4bit",
   "target_bpw": 5.0,
-  "achieved_bpw": 5.110966482264888,
-  "n_high_bits": 155,
-  "n_low_bits": 224,
+  "achieved_bpw": 5.180468492246783,
+  "n_high_bits": 219,
+  "n_low_bits": 124,
   "threshold": 0.0,
   "per_layer": {
     "language_model.model.per_layer_model_projection": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.41.per_layer_projection": {
@@ -36,14 +36,6 @@
       "bits": 8,
       "group_size": 64
     },
-    "language_model.model.layers.41.self_attn.v_proj": {
-      "bits": 4,
-      "group_size": 64
-    },
-    "language_model.model.layers.41.self_attn.k_proj": {
-      "bits": 4,
-      "group_size": 64
-    },
     "language_model.model.layers.41.self_attn.q_proj": {
       "bits": 8,
       "group_size": 64
@@ -61,7 +53,7 @@
       "group_size": 64
     },
     "language_model.model.layers.40.mlp.down_proj": {
-      "bits": 8,
+      "bits": 4,
       "group_size": 64
     },
     "language_model.model.layers.40.mlp.gate_proj": {
@@ -69,15 +61,7 @@
       "group_size": 64
     },
     "language_model.model.layers.40.self_attn.o_proj": {
-      "bits": 4,
-      "group_size": 64
-    },
-    "language_model.model.layers.40.self_attn.v_proj": {
-      "bits": 4,
-      "group_size": 64
-    },
-    "language_model.model.layers.40.self_attn.k_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.40.self_attn.q_proj": {
@@ -89,15 +73,15 @@
       "group_size": 64
     },
     "language_model.model.layers.39.per_layer_input_gate": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.39.mlp.up_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.39.mlp.down_proj": {
-      "bits": 8,
+      "bits": 4,
       "group_size": 64
     },
     "language_model.model.layers.39.mlp.gate_proj": {
@@ -105,15 +89,7 @@
       "group_size": 64
     },
     "language_model.model.layers.39.self_attn.o_proj": {
-      "bits": 4,
-      "group_size": 64
-    },
-    "language_model.model.layers.39.self_attn.v_proj": {
-      "bits": 4,
-      "group_size": 64
-    },
-    "language_model.model.layers.39.self_attn.k_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.39.self_attn.q_proj": {
@@ -125,7 +101,7 @@
       "group_size": 64
     },
     "language_model.model.layers.38.per_layer_input_gate": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.38.mlp.up_proj": {
@@ -133,7 +109,7 @@
       "group_size": 64
     },
     "language_model.model.layers.38.mlp.down_proj": {
-      "bits": 8,
+      "bits": 4,
       "group_size": 64
     },
     "language_model.model.layers.38.mlp.gate_proj": {
@@ -144,14 +120,6 @@
       "bits": 8,
       "group_size": 64
     },
-    "language_model.model.layers.38.self_attn.v_proj": {
-      "bits": 4,
-      "group_size": 64
-    },
-    "language_model.model.layers.38.self_attn.k_proj": {
-      "bits": 4,
-      "group_size": 64
-    },
     "language_model.model.layers.38.self_attn.q_proj": {
       "bits": 4,
       "group_size": 64
@@ -161,7 +129,7 @@
       "group_size": 64
     },
     "language_model.model.layers.37.per_layer_input_gate": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.37.mlp.up_proj": {
@@ -169,7 +137,7 @@
       "group_size": 64
     },
     "language_model.model.layers.37.mlp.down_proj": {
-      "bits": 8,
+      "bits": 4,
       "group_size": 64
     },
     "language_model.model.layers.37.mlp.gate_proj": {
@@ -177,15 +145,7 @@
       "group_size": 64
     },
     "language_model.model.layers.37.self_attn.o_proj": {
-      "bits": 4,
-      "group_size": 64
-    },
-    "language_model.model.layers.37.self_attn.v_proj": {
-      "bits": 4,
-      "group_size": 64
-    },
-    "language_model.model.layers.37.self_attn.k_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.37.self_attn.q_proj": {
@@ -197,15 +157,15 @@
       "group_size": 64
     },
     "language_model.model.layers.36.per_layer_input_gate": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.36.mlp.up_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.36.mlp.down_proj": {
-      "bits": 8,
+      "bits": 4,
       "group_size": 64
     },
     "language_model.model.layers.36.mlp.gate_proj": {
@@ -213,15 +173,7 @@
       "group_size": 64
     },
     "language_model.model.layers.36.self_attn.o_proj": {
-      "bits": 4,
-      "group_size": 64
-    },
-    "language_model.model.layers.36.self_attn.v_proj": {
-      "bits": 4,
-      "group_size": 64
-    },
-    "language_model.model.layers.36.self_attn.k_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.36.self_attn.q_proj": {
@@ -233,7 +185,7 @@
       "group_size": 64
     },
     "language_model.model.layers.35.per_layer_input_gate": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.35.mlp.up_proj": {
@@ -252,16 +204,8 @@
       "bits": 8,
       "group_size": 64
     },
-    "language_model.model.layers.35.self_attn.v_proj": {
-      "bits": 4,
-      "group_size": 64
-    },
-    "language_model.model.layers.35.self_attn.k_proj": {
-      "bits": 4,
-      "group_size": 64
-    },
     "language_model.model.layers.35.self_attn.q_proj": {
-      "bits": 8,
+      "bits": 4,
       "group_size": 64
     },
     "language_model.model.layers.34.per_layer_projection": {
@@ -269,7 +213,7 @@
       "group_size": 64
     },
     "language_model.model.layers.34.per_layer_input_gate": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.34.mlp.up_proj": {
@@ -277,7 +221,7 @@
       "group_size": 64
     },
     "language_model.model.layers.34.mlp.down_proj": {
-      "bits": 8,
+      "bits": 4,
       "group_size": 64
     },
     "language_model.model.layers.34.mlp.gate_proj": {
@@ -285,15 +229,7 @@
       "group_size": 64
     },
     "language_model.model.layers.34.self_attn.o_proj": {
-      "bits": 4,
-      "group_size": 64
-    },
-    "language_model.model.layers.34.self_attn.v_proj": {
-      "bits": 4,
-      "group_size": 64
-    },
-    "language_model.model.layers.34.self_attn.k_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.34.self_attn.q_proj": {
@@ -305,15 +241,15 @@
       "group_size": 64
     },
     "language_model.model.layers.33.per_layer_input_gate": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.33.mlp.up_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.33.mlp.down_proj": {
-      "bits": 8,
+      "bits": 4,
       "group_size": 64
     },
     "language_model.model.layers.33.mlp.gate_proj": {
@@ -321,15 +257,7 @@
       "group_size": 64
     },
     "language_model.model.layers.33.self_attn.o_proj": {
-      "bits": 4,
-      "group_size": 64
-    },
-    "language_model.model.layers.33.self_attn.v_proj": {
-      "bits": 4,
-      "group_size": 64
-    },
-    "language_model.model.layers.33.self_attn.k_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.33.self_attn.q_proj": {
@@ -341,7 +269,7 @@
       "group_size": 64
     },
     "language_model.model.layers.32.per_layer_input_gate": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.32.mlp.up_proj": {
@@ -349,7 +277,7 @@
       "group_size": 64
     },
     "language_model.model.layers.32.mlp.down_proj": {
-      "bits": 8,
+      "bits": 4,
       "group_size": 64
     },
     "language_model.model.layers.32.mlp.gate_proj": {
@@ -360,14 +288,6 @@
       "bits": 8,
       "group_size": 64
     },
-    "language_model.model.layers.32.self_attn.v_proj": {
-      "bits": 4,
-      "group_size": 64
-    },
-    "language_model.model.layers.32.self_attn.k_proj": {
-      "bits": 4,
-      "group_size": 64
-    },
     "language_model.model.layers.32.self_attn.q_proj": {
       "bits": 4,
       "group_size": 64
@@ -377,7 +297,7 @@
       "group_size": 64
     },
     "language_model.model.layers.31.per_layer_input_gate": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.31.mlp.up_proj": {
@@ -385,7 +305,7 @@
       "group_size": 64
     },
     "language_model.model.layers.31.mlp.down_proj": {
-      "bits": 8,
+      "bits": 4,
       "group_size": 64
     },
     "language_model.model.layers.31.mlp.gate_proj": {
@@ -393,15 +313,7 @@
       "group_size": 64
     },
     "language_model.model.layers.31.self_attn.o_proj": {
-      "bits": 4,
-      "group_size": 64
-    },
-    "language_model.model.layers.31.self_attn.v_proj": {
-      "bits": 4,
-      "group_size": 64
-    },
-    "language_model.model.layers.31.self_attn.k_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.31.self_attn.q_proj": {
@@ -413,15 +325,15 @@
       "group_size": 64
     },
     "language_model.model.layers.30.per_layer_input_gate": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.30.mlp.up_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.30.mlp.down_proj": {
-      "bits": 8,
+      "bits": 4,
       "group_size": 64
     },
     "language_model.model.layers.30.mlp.gate_proj": {
@@ -429,15 +341,7 @@
       "group_size": 64
     },
     "language_model.model.layers.30.self_attn.o_proj": {
-      "bits": 4,
-      "group_size": 64
-    },
-    "language_model.model.layers.30.self_attn.v_proj": {
-      "bits": 4,
-      "group_size": 64
-    },
-    "language_model.model.layers.30.self_attn.k_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.30.self_attn.q_proj": {
@@ -449,7 +353,7 @@
       "group_size": 64
     },
     "language_model.model.layers.29.per_layer_input_gate": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.29.mlp.up_proj": {
@@ -468,16 +372,8 @@
       "bits": 8,
       "group_size": 64
     },
-    "language_model.model.layers.29.self_attn.v_proj": {
-      "bits": 4,
-      "group_size": 64
-    },
-    "language_model.model.layers.29.self_attn.k_proj": {
-      "bits": 4,
-      "group_size": 64
-    },
     "language_model.model.layers.29.self_attn.q_proj": {
-      "bits": 8,
+      "bits": 4,
       "group_size": 64
     },
     "language_model.model.layers.28.per_layer_projection": {
@@ -485,7 +381,7 @@
       "group_size": 64
     },
     "language_model.model.layers.28.per_layer_input_gate": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.28.mlp.up_proj": {
@@ -501,19 +397,11 @@
       "group_size": 64
     },
     "language_model.model.layers.28.self_attn.o_proj": {
-      "bits": 4,
-      "group_size": 64
-    },
-    "language_model.model.layers.28.self_attn.v_proj": {
-      "bits": 4,
-      "group_size": 64
-    },
-    "language_model.model.layers.28.self_attn.k_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.28.self_attn.q_proj": {
-      "bits": 8,
+      "bits": 4,
       "group_size": 64
     },
     "language_model.model.layers.27.per_layer_projection": {
@@ -521,15 +409,15 @@
       "group_size": 64
     },
     "language_model.model.layers.27.per_layer_input_gate": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.27.mlp.up_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.27.mlp.down_proj": {
-      "bits": 8,
+      "bits": 4,
       "group_size": 64
     },
     "language_model.model.layers.27.mlp.gate_proj": {
@@ -537,15 +425,7 @@
       "group_size": 64
     },
     "language_model.model.layers.27.self_attn.o_proj": {
-      "bits": 4,
-      "group_size": 64
-    },
-    "language_model.model.layers.27.self_attn.v_proj": {
-      "bits": 4,
-      "group_size": 64
-    },
-    "language_model.model.layers.27.self_attn.k_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.27.self_attn.q_proj": {
@@ -557,11 +437,11 @@
       "group_size": 64
     },
     "language_model.model.layers.26.per_layer_input_gate": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.26.mlp.up_proj": {
-      "bits": 8,
+      "bits": 4,
       "group_size": 64
     },
     "language_model.model.layers.26.mlp.down_proj": {
@@ -576,14 +456,6 @@
       "bits": 8,
       "group_size": 64
     },
-    "language_model.model.layers.26.self_attn.v_proj": {
-      "bits": 4,
-      "group_size": 64
-    },
-    "language_model.model.layers.26.self_attn.k_proj": {
-      "bits": 4,
-      "group_size": 64
-    },
     "language_model.model.layers.26.self_attn.q_proj": {
       "bits": 4,
       "group_size": 64
@@ -593,7 +465,7 @@
       "group_size": 64
     },
     "language_model.model.layers.25.per_layer_input_gate": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.25.mlp.up_proj": {
@@ -601,7 +473,7 @@
       "group_size": 64
     },
     "language_model.model.layers.25.mlp.down_proj": {
-      "bits": 8,
+      "bits": 4,
       "group_size": 64
     },
     "language_model.model.layers.25.mlp.gate_proj": {
@@ -609,15 +481,7 @@
       "group_size": 64
     },
     "language_model.model.layers.25.self_attn.o_proj": {
-      "bits": 4,
-      "group_size": 64
-    },
-    "language_model.model.layers.25.self_attn.v_proj": {
-      "bits": 4,
-      "group_size": 64
-    },
-    "language_model.model.layers.25.self_attn.k_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.25.self_attn.q_proj": {
@@ -629,11 +493,11 @@
       "group_size": 64
     },
     "language_model.model.layers.24.per_layer_input_gate": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.24.mlp.up_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.24.mlp.down_proj": {
@@ -641,19 +505,11 @@
       "group_size": 64
     },
     "language_model.model.layers.24.mlp.gate_proj": {
-      "bits": 8,
-      "group_size": 64
-    },
-    "language_model.model.layers.24.self_attn.o_proj": {
-      "bits": 4,
-      "group_size": 64
-    },
-    "language_model.model.layers.24.self_attn.v_proj": {
       "bits": 4,
       "group_size": 64
     },
-    "language_model.model.layers.24.self_attn.k_proj": {
-      "bits": 4,
+    "language_model.model.layers.24.self_attn.o_proj": {
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.24.self_attn.q_proj": {
@@ -665,7 +521,7 @@
       "group_size": 64
     },
     "language_model.model.layers.23.per_layer_input_gate": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.23.mlp.up_proj": {
@@ -701,7 +557,7 @@
       "group_size": 64
     },
     "language_model.model.layers.22.per_layer_input_gate": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.22.mlp.up_proj": {
@@ -717,7 +573,7 @@
       "group_size": 64
     },
     "language_model.model.layers.22.self_attn.o_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.22.self_attn.v_proj": {
@@ -729,7 +585,7 @@
       "group_size": 64
     },
     "language_model.model.layers.22.self_attn.q_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.21.per_layer_projection": {
@@ -737,7 +593,7 @@
       "group_size": 64
     },
     "language_model.model.layers.21.per_layer_input_gate": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.21.mlp.up_proj": {
@@ -761,7 +617,7 @@
       "group_size": 64
     },
     "language_model.model.layers.21.self_attn.k_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.21.self_attn.q_proj": {
@@ -773,7 +629,7 @@
       "group_size": 64
     },
     "language_model.model.layers.20.per_layer_input_gate": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.20.mlp.up_proj": {
@@ -781,7 +637,7 @@
       "group_size": 64
     },
     "language_model.model.layers.20.mlp.down_proj": {
-      "bits": 8,
+      "bits": 4,
       "group_size": 64
     },
     "language_model.model.layers.20.mlp.gate_proj": {
@@ -797,7 +653,7 @@
       "group_size": 64
     },
     "language_model.model.layers.20.self_attn.k_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.20.self_attn.q_proj": {
@@ -833,7 +689,7 @@
       "group_size": 64
     },
     "language_model.model.layers.19.self_attn.k_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.19.self_attn.q_proj": {
@@ -861,7 +717,7 @@
       "group_size": 64
     },
     "language_model.model.layers.18.self_attn.o_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.18.self_attn.v_proj": {
@@ -869,11 +725,11 @@
       "group_size": 64
     },
     "language_model.model.layers.18.self_attn.k_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.18.self_attn.q_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.17.per_layer_projection": {
@@ -881,11 +737,11 @@
       "group_size": 64
     },
     "language_model.model.layers.17.per_layer_input_gate": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.17.mlp.up_proj": {
-      "bits": 8,
+      "bits": 4,
       "group_size": 64
     },
     "language_model.model.layers.17.mlp.down_proj": {
@@ -909,7 +765,7 @@
       "group_size": 64
     },
     "language_model.model.layers.17.self_attn.q_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.16.per_layer_projection": {
@@ -945,7 +801,7 @@
       "group_size": 64
     },
     "language_model.model.layers.16.self_attn.q_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.15.per_layer_projection": {
@@ -977,15 +833,15 @@
       "group_size": 64
     },
     "language_model.model.layers.15.self_attn.k_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.15.self_attn.q_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.14.per_layer_projection": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.14.per_layer_input_gate": {
@@ -997,7 +853,7 @@
       "group_size": 64
     },
     "language_model.model.layers.14.mlp.down_proj": {
-      "bits": 8,
+      "bits": 4,
       "group_size": 64
     },
     "language_model.model.layers.14.mlp.gate_proj": {
@@ -1009,15 +865,15 @@
       "group_size": 64
     },
     "language_model.model.layers.14.self_attn.v_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.14.self_attn.k_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.14.self_attn.q_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.13.per_layer_projection": {
@@ -1053,7 +909,7 @@
       "group_size": 64
     },
     "language_model.model.layers.13.self_attn.q_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.12.per_layer_projection": {
@@ -1077,7 +933,7 @@
       "group_size": 64
     },
     "language_model.model.layers.12.self_attn.o_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.12.self_attn.v_proj": {
@@ -1089,7 +945,7 @@
       "group_size": 64
     },
     "language_model.model.layers.12.self_attn.q_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.11.per_layer_projection": {
@@ -1097,7 +953,7 @@
       "group_size": 64
     },
     "language_model.model.layers.11.per_layer_input_gate": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.11.mlp.up_proj": {
@@ -1109,7 +965,7 @@
       "group_size": 64
     },
     "language_model.model.layers.11.mlp.gate_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.11.self_attn.o_proj": {
@@ -1125,7 +981,7 @@
       "group_size": 64
     },
     "language_model.model.layers.11.self_attn.q_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.10.per_layer_projection": {
@@ -1137,11 +993,11 @@
       "group_size": 64
     },
     "language_model.model.layers.10.mlp.up_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.10.mlp.down_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.10.mlp.gate_proj": {
@@ -1185,7 +1041,7 @@
       "group_size": 64
     },
     "language_model.model.layers.9.self_attn.o_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.9.self_attn.v_proj": {
@@ -1193,7 +1049,7 @@
       "group_size": 64
     },
     "language_model.model.layers.9.self_attn.k_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.9.self_attn.q_proj": {
@@ -1209,7 +1065,7 @@
       "group_size": 64
     },
     "language_model.model.layers.8.mlp.up_proj": {
-      "bits": 8,
+      "bits": 4,
       "group_size": 64
     },
     "language_model.model.layers.8.mlp.down_proj": {
@@ -1217,11 +1073,11 @@
       "group_size": 64
     },
     "language_model.model.layers.8.mlp.gate_proj": {
-      "bits": 8,
+      "bits": 4,
       "group_size": 64
     },
     "language_model.model.layers.8.self_attn.o_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.8.self_attn.v_proj": {
@@ -1229,11 +1085,11 @@
       "group_size": 64
     },
     "language_model.model.layers.8.self_attn.k_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.8.self_attn.q_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.7.per_layer_projection": {
@@ -1249,7 +1105,7 @@
       "group_size": 64
     },
     "language_model.model.layers.7.mlp.down_proj": {
-      "bits": 8,
+      "bits": 4,
       "group_size": 64
     },
     "language_model.model.layers.7.mlp.gate_proj": {
@@ -1257,7 +1113,7 @@
       "group_size": 64
     },
     "language_model.model.layers.7.self_attn.o_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.7.self_attn.v_proj": {
@@ -1269,11 +1125,11 @@
       "group_size": 64
     },
     "language_model.model.layers.7.self_attn.q_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.6.per_layer_projection": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.6.per_layer_input_gate": {
@@ -1281,7 +1137,7 @@
       "group_size": 64
     },
     "language_model.model.layers.6.mlp.up_proj": {
-      "bits": 8,
+      "bits": 4,
       "group_size": 64
     },
     "language_model.model.layers.6.mlp.down_proj": {
@@ -1289,11 +1145,11 @@
       "group_size": 64
     },
     "language_model.model.layers.6.mlp.gate_proj": {
-      "bits": 8,
+      "bits": 4,
       "group_size": 64
     },
     "language_model.model.layers.6.self_attn.o_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.6.self_attn.v_proj": {
@@ -1305,7 +1161,7 @@
       "group_size": 64
     },
     "language_model.model.layers.6.self_attn.q_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.5.per_layer_projection": {
@@ -1321,11 +1177,11 @@
       "group_size": 64
     },
     "language_model.model.layers.5.mlp.down_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.5.mlp.gate_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.5.self_attn.o_proj": {
@@ -1341,7 +1197,7 @@
       "group_size": 64
     },
     "language_model.model.layers.5.self_attn.q_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.4.per_layer_projection": {
@@ -1357,15 +1213,15 @@
       "group_size": 64
     },
     "language_model.model.layers.4.mlp.down_proj": {
-      "bits": 8,
+      "bits": 4,
       "group_size": 64
     },
     "language_model.model.layers.4.mlp.gate_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.4.self_attn.o_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.4.self_attn.v_proj": {
@@ -1397,7 +1253,7 @@
       "group_size": 64
     },
     "language_model.model.layers.3.mlp.gate_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.3.self_attn.o_proj": {
@@ -1409,15 +1265,15 @@
       "group_size": 64
     },
     "language_model.model.layers.3.self_attn.k_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.3.self_attn.q_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.2.per_layer_projection": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.2.per_layer_input_gate": {
@@ -1429,7 +1285,7 @@
       "group_size": 64
     },
     "language_model.model.layers.2.mlp.down_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.2.mlp.gate_proj": {
@@ -1437,7 +1293,7 @@
       "group_size": 64
     },
     "language_model.model.layers.2.self_attn.o_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.2.self_attn.v_proj": {
@@ -1445,7 +1301,7 @@
       "group_size": 64
     },
     "language_model.model.layers.2.self_attn.k_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.2.self_attn.q_proj": {
@@ -1453,7 +1309,7 @@
       "group_size": 64
     },
     "language_model.model.layers.1.per_layer_projection": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.1.per_layer_input_gate": {
@@ -1469,7 +1325,7 @@
       "group_size": 64
     },
     "language_model.model.layers.1.mlp.gate_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.1.self_attn.o_proj": {
@@ -1485,7 +1341,7 @@
       "group_size": 64
     },
     "language_model.model.layers.1.self_attn.q_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.0.per_layer_projection": {
@@ -1497,7 +1353,7 @@
       "group_size": 64
     },
     "language_model.model.layers.0.mlp.up_proj": {
-      "bits": 4,
+      "bits": 8,
       "group_size": 64
     },
     "language_model.model.layers.0.mlp.down_proj": {
@@ -1524,33 +1380,5 @@
       "bits": 8,
       "group_size": 64
     }
-  },
-  "post_processing": [
-    {
-      "op": "strip_multimodal_metadata",
-      "architectures": {
-        "from": [
-          "Gemma4ForConditionalGeneration"
-        ],
-        "to": [
-          "Gemma4ForCausalLM"
-        ]
-      },
-      "model_type": null,
-      "flattened_text_config": false,
-      "dropped_keys": [
-        "audio_config",
-        "audio_token_id",
-        "boa_token_id",
-        "boi_token_id",
-        "eoa_token_id",
-        "eoa_token_index",
-        "eoi_token_id",
-        "image_token_id",
-        "video_token_id",
-        "vision_soft_tokens_per_image"
-      ],
-      "dropped_mrope_keys": []
-    }
-  ]
-}
+  }
+}