From 097989299222f94db6703ee0902972943f01aa7e Mon Sep 17 00:00:00 2001 From: Juan Calderon-Perez <835733+gaby@users.noreply.github.com> Date: Fri, 2 Oct 2026 11:25:37 -0400 Subject: [PATCH] [Bugfix][Multimodal] Fix GLM-5.3-Flash vision tower crashes on image input (#59126) Signed-off-by: Juan Calderon-Perez <835733+gaby@users.noreply.github.com> Co-authored-by: Claude Opus 5.5 Co-authored-by: Isotr0py --- vllm/models/glm5next/common/multimodal.py | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-) diff --git a/vllm/models/glm5next/common/multimodal.py b/vllm/models/glm5next/common/multimodal.py index 7384c8c316..23c90cce73 100644 --- a/vllm/models/glm5next/common/multimodal.py +++ b/vllm/models/glm5next/common/multimodal.py @@ -46,6 +46,7 @@ from vllm.model_executor.models.vision import ( ) from vllm.models.common.ops import fused_q_kv_rmsnorm from vllm.multimodal.parse import ImageSize, MultiModalDataItems +from vllm.utils.torch_utils import async_tensor_h2d from vllm.v1.attention.backends.registry import AttentionBackendEnum @@ -395,7 +396,7 @@ class Glm5NextVisionTransformer(nn.Module): head_dim = self.hidden_size // self.num_heads self.rotary_pos_emb = get_rope( head_size=head_dim, - max_position=8192, + max_position=text_config.max_position_embeddings, is_neox_style=True, rope_parameters={"partial_rotary_factor": 0.5}, ) @@ -479,7 +480,7 @@ class Glm5NextVisionTransformer(nn.Module): cos, sin = self.rotary_pos_emb.get_cos_sin(max_grid_size) - pos_ids = pos_ids.to(cos.device, non_blocking=True) + pos_ids = async_tensor_h2d(pos_ids, device=cos.device) cos_combined = cos[pos_ids].flatten(1) sin_combined = sin[pos_ids].flatten(1) return cos_combined, sin_combined, pos_ids @@ -590,7 +591,7 @@ class Glm5NextVisionTransformer(nn.Module): grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0] ).cumsum(dim=0, dtype=torch.int32) cu_seqlens = torch.cat([cu_seqlens.new_zeros(1), cu_seqlens]) - cu_seqlens = cu_seqlens.to(self.device, non_blocking=True) + cu_seqlens = async_tensor_h2d(cu_seqlens, device=self.device) max_seqlen = self.compute_attn_mask_seqlen(cu_seqlens) # transformers