A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 9, 03:45 UTC
Fix DeepSeek V2 RoPE initialization error (#33501)
Signed-off-by: Eduardo Salinas <edus@microsoft.com> Signed-off-by: catswe <212922539+catswe@users.noreply.github.com> Co-authored-by: Eduardo Salinas <edus@microsoft.com>
will b. committed
Feb 1, 2026 at 15:00 UTC
46b4a02794223a75987d07b2ceaaa1c1010dfbf0
1 file changed
-2
vllm/model_executor/layers/rotary_embedding/deepseek_scaling_rope.py
-2
@@ -75,7 +75,6 @@ class DeepseekScalingRotaryEmbedding(RotaryEmbeddingBase):
75
self.rotary_dim,
76
2,
77
dtype=torch.float,
78
- device=current_platform.device_type,
78
)
79
/ self.rotary_dim
80
)
@@ -104,7 +103,6 @@ class DeepseekScalingRotaryEmbedding(RotaryEmbeddingBase):
103
inv_freq = self._compute_inv_freq(self.scaling_factor)
104
t = torch.arange(
105
self.max_position_embeddings * self.scaling_factor,
107
- device=current_platform.device_type,
106
dtype=torch.float32,
107
)
108
freqs = torch.einsum("i,j -> ij", t, inv_freq)