From cb1846cd4f1efa94e12f900f1768f92ffebcffc8 Mon Sep 17 00:00:00 2001
From: Chranos <826995883@qq.com>
Date: Tue, 10 Feb 2026 14:10:04 +0800
Subject: [PATCH] add gemma3

---
 .../vllm/model_executor/models/gemma3.py      | 29 ++++++++++---------
 1 file changed, 16 insertions(+), 13 deletions(-)

diff --git a/vllm-v0.6.2/vllm/model_executor/models/gemma3.py b/vllm-v0.6.2/vllm/model_executor/models/gemma3.py
index 81aa9a7..2c701ae 100644
--- a/vllm-v0.6.2/vllm/model_executor/models/gemma3.py
+++ b/vllm-v0.6.2/vllm/model_executor/models/gemma3.py
@@ -88,7 +88,6 @@ class Gemma3Attention(nn.Module):
                  num_kv_heads: int,
                  head_dim: int,
                  max_position_embeddings: int,
-                 rope_theta: float,
                  cache_config: Optional[CacheConfig] = None,
                  quant_config: Optional[QuantizationConfig] = None,
                  attn_logits_soft_cap: Optional[float] = None) -> None:
@@ -110,7 +109,22 @@ class Gemma3Attention(nn.Module):
         self.q_size = self.num_heads * self.head_dim
         self.kv_size = self.num_kv_heads * self.head_dim
         self.scaling = config.query_pre_attn_scalar**-0.5
-        self.rope_theta = rope_theta
+
+        # Extract rope_theta from config, compatible with both old-style
+        # (config.rope_theta) and new-style (config.rope_parameters dict).
+        rope_params = getattr(config, "rope_parameters", None)
+        if hasattr(config, "rope_theta"):
+            self.rope_theta = config.rope_theta
+        elif isinstance(rope_params, dict):
+            # Transformers v5: nested per layer_type
+            if "full_attention" in rope_params:
+                self.rope_theta = rope_params["full_attention"].get(
+                    "rope_theta", 10000.0)
+            else:
+                # Transformers v4: flat dict
+                self.rope_theta = rope_params.get("rope_theta", 10000.0)
+        else:
+            self.rope_theta = 10000.0
 
         self.qkv_proj = QKVParallelLinear(
             hidden_size,
@@ -236,16 +250,6 @@ class Gemma3DecoderLayer(nn.Module):
     ) -> None:
         super().__init__()
         self.hidden_size = config.hidden_size
-        # Extract rope_theta: try direct attribute first, then
-        # rope_parameters dict (newer transformers), fallback to 10000.0
-        rope_params = getattr(config, "rope_parameters", None)
-        if hasattr(config, "rope_theta"):
-            rope_theta = config.rope_theta
-        elif isinstance(rope_params, dict):
-            rope_theta = rope_params.get("rope_theta", 10000.0)
-        else:
-            rope_theta = 10000.0
-
         self.self_attn = Gemma3Attention(
             layer_idx=layer_idx,
             config=config,
@@ -254,7 +258,6 @@ class Gemma3DecoderLayer(nn.Module):
             num_kv_heads=config.num_key_value_heads,
             head_dim=config.head_dim,
             max_position_embeddings=config.max_position_embeddings,
-            rope_theta=rope_theta,
             cache_config=cache_config,
             quant_config=quant_config,
             # Gemma3 does not use attn logit softcapping