From b7c5baf63d5f95d18493a7fae3c75fc4853732a7 Mon Sep 17 00:00:00 2001 From: Schwinn Saereesitthipitak <17022745+galletas1712@users.noreply.github.com> Date: Thu, 4 Jun 2026 15:30:29 -0700 Subject: [PATCH] fix: keep DeepSeek V4 RoPE cache on inv_freq device (#43926) Signed-off-by: Schwinn Saereesitthipitak Signed-off-by: Schwinn Saereesitthipitak <17022745+galletas1712@users.noreply.github.com> --- .../layers/rotary_embedding/deepseek_scaling_rope.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vllm/model_executor/layers/rotary_embedding/deepseek_scaling_rope.py b/vllm/model_executor/layers/rotary_embedding/deepseek_scaling_rope.py index 7362abcc8fb..bfaf81ad007 100644 --- a/vllm/model_executor/layers/rotary_embedding/deepseek_scaling_rope.py +++ b/vllm/model_executor/layers/rotary_embedding/deepseek_scaling_rope.py @@ -251,7 +251,7 @@ class DeepseekV4ScalingRotaryEmbedding(DeepseekScalingRotaryEmbedding): inv_freq = self._compute_inv_freq(self.scaling_factor) t = torch.arange( self.max_position_embeddings * self.scaling_factor, - device=current_platform.device_type, + device=inv_freq.device, dtype=torch.float32, ) freqs = torch.einsum("i,j -> ij", t, inv_freq)