Add 26B-A4B MoE config (hidden=2816, head_dim=256, 16 heads)

scripts/hf_job_selfdistill_train.py

@@ -88,6 +88,14 @@ GEMMA4_CONFIGS = {
         "num_key_value_heads": 2,
         "vocab_size": 262144,
     },
+    "26b": {
+        "hidden_size": 2816,
+        "intermediate_size": 2112,
+        "num_attention_heads": 16,
+        "num_key_value_heads": 8,
+        "vocab_size": 262144,
+        "head_dim": 256,
+    },
     "27b": {
         "hidden_size": 3584,
         "intermediate_size": 14336,