From 9d8ad5b408bf447e41a3629fc21a453720aaf52b Mon Sep 17 00:00:00 2001 From: Jee Jee Li Date: Wed, 29 Apr 2026 20:29:55 +0800 Subject: [PATCH] [Bugfix] Fix repeated DSv4 RoPE cache initialization (#41148) Signed-off-by: Jee Jee Li Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com> --- .../rotary_embedding/deepseek_scaling_rope.py | 13 +++++++++++-- vllm/model_executor/models/deepseek_v4.py | 1 - 2 files changed, 11 insertions(+), 3 deletions(-) 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 6cb9101a78..9a06eedd0f 100644 --- a/vllm/model_executor/layers/rotary_embedding/deepseek_scaling_rope.py +++ b/vllm/model_executor/layers/rotary_embedding/deepseek_scaling_rope.py @@ -45,6 +45,7 @@ class DeepseekScalingRotaryEmbedding(RotaryEmbeddingBase): beta_slow: int = 1, mscale: float = 1, mscale_all_dim: float = 0, + init_cache: bool = True, ) -> None: self.scaling_factor = scaling_factor self.extrapolation_factor = extrapolation_factor @@ -65,7 +66,13 @@ class DeepseekScalingRotaryEmbedding(RotaryEmbeddingBase): and head_size in [64, 128, 256, 512] ) super().__init__( - head_size, rotary_dim, max_position_embeddings, base, is_neox_style, dtype + head_size, + rotary_dim, + max_position_embeddings, + base, + is_neox_style, + dtype, + init_cache=init_cache, ) def _compute_inv_freq(self, scaling_factor: float) -> torch.Tensor: @@ -211,7 +218,9 @@ class DeepseekV4ScalingRotaryEmbedding(DeepseekScalingRotaryEmbedding): """ def __init__(self, *args, **kwargs): - super().__init__(*args, **kwargs) + # Avoid compute cache repeatedly + kwargs.pop("init_cache", None) + super().__init__(*args, **kwargs, init_cache=False) cache_fp32 = self._compute_cos_sin_cache() self.register_buffer("cos_sin_cache", cache_fp32, persistent=False) diff --git a/vllm/model_executor/models/deepseek_v4.py b/vllm/model_executor/models/deepseek_v4.py index d41a8b666d..baf28d0458 100644 --- a/vllm/model_executor/models/deepseek_v4.py +++ b/vllm/model_executor/models/deepseek_v4.py @@ -1027,7 +1027,6 @@ class DeepseekV4Attention(nn.Module): max_position=self.max_position_embeddings, rope_parameters=rope_parameters, is_neox_style=False, - dtype=config.torch_dtype, ) self.indexer = None