From 5555a2d8f7116b2b4877f66516f4128176cf722a Mon Sep 17 00:00:00 2001 From: Mook <68294499+Godmook@users.noreply.github.com> Date: Tue, 13 Oct 2026 14:02:12 +0900 Subject: [PATCH] [Bugfix][Model] Fix EXAONE-4.5 image inputs: ViT received un-normalised pixels (#61104) Signed-off-by: godmook Co-authored-by: cjackal <44624812+cjackal@users.noreply.github.com> --- vllm/model_executor/models/exaone4_5.py | 6 ++++++ vllm/model_executor/models/exaone4_5_mtp.py | 1 + 2 files changed, 7 insertions(+) diff --git a/vllm/model_executor/models/exaone4_5.py b/vllm/model_executor/models/exaone4_5.py index 5a53e60b7d..ad4049c8ee 100644 --- a/vllm/model_executor/models/exaone4_5.py +++ b/vllm/model_executor/models/exaone4_5.py @@ -36,6 +36,7 @@ from vllm.distributed import parallel_state from vllm.distributed import utils as dist_utils from vllm.logger import init_logger from vllm.model_executor.layers.attention.mm_encoder_attention import MMEncoderAttention +from vllm.model_executor.layers.fusion.mm_input_norm import build_mm_input_norm from vllm.model_executor.layers.layernorm import RMSNorm from vllm.model_executor.layers.linear import QKVParallelLinear, RowParallelLinear from vllm.model_executor.layers.quantization import QuantizationConfig @@ -272,6 +273,7 @@ class EXAONE4_5_VisionTransformer(Qwen2_5_VisionTransformer): vision_config: Exaone4_5_VisionConfig, norm_eps: float = 1e-6, quant_config: QuantizationConfig | None = None, + input_norm: nn.Module | None = None, prefix: str = "", use_data_parallel: bool = False, ) -> None: @@ -279,6 +281,7 @@ class EXAONE4_5_VisionTransformer(Qwen2_5_VisionTransformer): vision_config=vision_config, norm_eps=norm_eps, quant_config=quant_config, + input_norm=input_norm, prefix=prefix, ) depth = vision_config.depth @@ -335,7 +338,9 @@ class Exaone4_5_ForConditionalGeneration(Qwen2_5_VLForConditionalGeneration): self.use_data_parallel = multimodal_config.mm_encoder_tp_mode == "data" self.config = config + self.model_config = vllm_config.model_config self.multimodal_config = multimodal_config + self.video_pruning_rate = multimodal_config.video_pruning_rate self.is_multimodal_pruning_enabled = ( multimodal_config.is_multimodal_pruning_enabled() ) @@ -345,6 +350,7 @@ class Exaone4_5_ForConditionalGeneration(Qwen2_5_VLForConditionalGeneration): config.vision_config, norm_eps=getattr(config, "rms_norm_eps", 1e-6), quant_config=self.quant_config, + input_norm=build_mm_input_norm(self.model_config), prefix=maybe_prefix(prefix, "visual"), use_data_parallel=self.use_data_parallel, ) diff --git a/vllm/model_executor/models/exaone4_5_mtp.py b/vllm/model_executor/models/exaone4_5_mtp.py index d9c7960010..4123592986 100644 --- a/vllm/model_executor/models/exaone4_5_mtp.py +++ b/vllm/model_executor/models/exaone4_5_mtp.py @@ -75,6 +75,7 @@ class Exaone4_5MultiTokenPredictor(ExaoneMoeMultiTokenPredictor): self.layers = nn.ModuleList( Exaone4DecoderLayer( text_config, + cache_config=vllm_config.cache_config, quant_config=quant_config, prefix=f"{prefix}.layers.{idx}", )