Fix Gemma 4 for upcoming Transformers version (#49797)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2026-08-10 07:48:26 +00:00
committed by GitHub
parent 81840a172f
commit 70b84f0bcb
15 changed files with 569 additions and 105 deletions
+4 -15
View File
@@ -45,6 +45,7 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
VocabParallelEmbedding,
)
from vllm.sequence import IntermediateTensors
from vllm.transformers_utils.configs.gemma4 import gemma4_layer_config
from .gemma4 import Gemma4MLP, _get_text_config
from .utils import (
@@ -271,21 +272,9 @@ class Gemma4MTPDecoderLayer(nn.Module):
self.hidden_size = config.hidden_size
layer_idx = extract_layer_index(prefix)
layer_type = config.layer_types[layer_idx]
is_full_attention = layer_type == "full_attention"
head_dim = (
getattr(config, "global_head_dim", config.head_dim)
if is_full_attention
else config.head_dim
)
use_k_eq_v = is_full_attention and getattr(config, "attention_k_eq_v", False)
if use_k_eq_v:
num_kv_heads = getattr(
config, "num_global_key_value_heads", config.num_key_value_heads
)
else:
num_kv_heads = config.num_key_value_heads
layer_config = gemma4_layer_config(config, layer_idx)
head_dim = layer_config.head_dim
num_kv_heads = layer_config.num_key_value_heads
self.self_attn = Gemma4MTPAttention(
config=config,