eagle3: support Gemma4 eagle3 from RedHatAI

Author: ruixiang63

conversion/__init__.py

@@ -124,6 +124,9 @@
     "LlamaBidirectionalModel": "llama",
     "LlamaForCausalLM": "llama",
     "LlamaModel": "llama",
+    "Eagle3DraftModel": "llama",
+    "Eagle3Speculator": "llama",
+    "LlamaForCausalLMEagle3": "llama",
     "LlavaForConditionalGeneration": "llama",
     "LlavaStableLMEpochForCausalLM": "stablelm",
     "MPTForCausalLM": "mpt",

conversion/llama.py

@@ -63,6 +63,9 @@ def __init__(self, *args, **kwargs):
             with open(self.target_model_dir / "config.json", 'r', encoding='utf-8') as f:
                 target_config = json.load(f)
 
+            if "text_config" in target_config:
+                target_config = {**target_config, **target_config["text_config"]}
+
             # extract_layers: derived from target model layer count (low/mid/high)
             target_num_layers = target_config["num_hidden_layers"]
             extract_layers = [2, target_num_layers // 2, target_num_layers - 3]

src/models/gemma4.cpp

@@ -183,6 +183,8 @@ llama_model_gemma4::graph::graph(const llama_model & model, const llm_graph_para
         const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
         const int   n_rot_l      = hparams.n_rot(il);
 
+        res->t_layer_inp[il] = inpL;
+        
         // norm
         cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
         cb(cur, "attn_norm", il);