From 97652d26b81f83fc9a3675be55ede7762fafb7bd Mon Sep 17 00:00:00 2001
From: contentis <lspindler@nvidia.com>
Date: Mon, 8 Sep 2025 21:08:18 +0200
Subject: [PATCH] Add explicit casting in apply_rope for Qwen VL (#9759)

---
 comfy/text_encoders/llama.py | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/comfy/text_encoders/llama.py b/comfy/text_encoders/llama.py
index 4c976058f..5e11956b5 100644
--- a/comfy/text_encoders/llama.py
+++ b/comfy/text_encoders/llama.py
@@ -128,11 +128,12 @@ def precompute_freqs_cis(head_dim, position_ids, theta, rope_dims=None, device=N
 
 
 def apply_rope(xq, xk, freqs_cis):
+    org_dtype = xq.dtype
     cos = freqs_cis[0]
     sin = freqs_cis[1]
     q_embed = (xq * cos) + (rotate_half(xq) * sin)
     k_embed = (xk * cos) + (rotate_half(xk) * sin)
-    return q_embed, k_embed
+    return q_embed.to(org_dtype), k_embed.to(org_dtype)
 
 
 class Attention(nn.Module):