diff --git a/src/diffusers/models/attention_dispatch.py b/src/diffusers/models/attention_dispatch.py
index 8504504981..7face061b2 100644
--- a/src/diffusers/models/attention_dispatch.py
+++ b/src/diffusers/models/attention_dispatch.py
@@ -278,6 +278,7 @@ def attention_backend(backend: Union[str, AttentionBackendName] = AttentionBacke
 
     backend = AttentionBackendName(backend)
     _check_attention_backend_requirements(backend)
+    _maybe_download_kernel_for_backend(backend)
 
     old_backend = _AttentionBackendRegistry._active_backend
     _AttentionBackendRegistry._active_backend = backend
diff --git a/src/diffusers/pipelines/hunyuan_video1_5/image_processor.py b/src/diffusers/pipelines/hunyuan_video1_5/image_processor.py
index 6e3e818c7d..efeb0e2a5f 100644
--- a/src/diffusers/pipelines/hunyuan_video1_5/image_processor.py
+++ b/src/diffusers/pipelines/hunyuan_video1_5/image_processor.py
@@ -52,7 +52,7 @@ def get_closest_ratio(height: float, width: float, ratios: list, buckets: list):
     if aspect_ratio >= 1:
         indices = [(index, x) for index, x in enumerate(diff_ratios) if x <= 0]
     else:
-        indices = [(index, x) for index, x in enumerate(diff_ratios) if x > 0]
+        indices = [(index, x) for index, x in enumerate(diff_ratios) if x >= 0]
 
     closest_ratio_id = min(indices, key=lambda pair: abs(pair[1]))[0]
     closest_size = buckets[closest_ratio_id]
diff --git a/src/diffusers/pipelines/hunyuan_video1_5/pipeline_hunyuan_video1_5.py b/src/diffusers/pipelines/hunyuan_video1_5/pipeline_hunyuan_video1_5.py
index ee25031f60..7e23d2da3e 100644
--- a/src/diffusers/pipelines/hunyuan_video1_5/pipeline_hunyuan_video1_5.py
+++ b/src/diffusers/pipelines/hunyuan_video1_5/pipeline_hunyuan_video1_5.py
@@ -195,7 +195,7 @@ class HunyuanVideo15Pipeline(DiffusionPipeline):
             [ClassifierFreeGuidance]for classifier free guidance.
     """
 
-    model_cpu_offload_seq = "text_encoder->text_encoder_2->transformer->vae"
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
     _callback_tensor_inputs = ["latents", "prompt_embeds"]
 
     def __init__(