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
+9 -13
View File
@@ -20,6 +20,7 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
VocabParallelEmbedding,
)
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
from vllm.transformers_utils.configs.gemma4 import gemma4_layer_config
from .gemma4_mtp import Gemma4MTPAttention, Gemma4MTPDecoderLayer
from .qwen3_dflash import DFlashQwen3Model, _dflash_layer_causal
@@ -38,18 +39,14 @@ class Gemma4DSparkAttention(Gemma4MTPAttention):
prefix: str,
) -> None:
layer_idx = extract_layer_index(prefix)
is_full = config.layer_types[layer_idx] == "full_attention"
head_dim = (
getattr(config, "global_head_dim", config.head_dim)
if is_full
else config.head_dim
)
use_k_eq_v = is_full and getattr(config, "attention_k_eq_v", False)
num_kv_heads = (
getattr(config, "num_global_key_value_heads", config.num_key_value_heads)
if use_k_eq_v
else config.num_key_value_heads
layer_type = config.layer_types[layer_idx]
self.use_k_eq_v = layer_type == "full_attention" and getattr(
config, "attention_k_eq_v", False
)
layer_config = gemma4_layer_config(config, layer_idx)
head_dim = layer_config.head_dim
num_kv_heads = layer_config.num_key_value_heads
super().__init__(
config=config,
hidden_size=config.hidden_size,
@@ -64,7 +61,6 @@ class Gemma4DSparkAttention(Gemma4MTPAttention):
)
self.is_kv_shared_layer = False
self.causal = _dflash_layer_causal(config, layer_idx)
self.use_k_eq_v = use_k_eq_v
self.kv_size = self.num_kv_heads * self.head_dim
attn_bias = getattr(config, "attention_bias", False)
self.k_proj = ColumnParallelLinear(
@@ -76,7 +72,7 @@ class Gemma4DSparkAttention(Gemma4MTPAttention):
)
self.v_proj = (
None
if use_k_eq_v
if self.use_k_eq_v
else ColumnParallelLinear(
config.hidden_size,
self.total_num_kv_heads * self.head_dim,