From 253b01b9a552b9745aaea827bd8944c87ea7b09a Mon Sep 17 00:00:00 2001 From: Wang Yixuan <88923622+hust17yixuan@users.noreply.github.com> Date: Tue, 2 Sep 2025 17:21:56 +0800 Subject: [PATCH] [7/N][refactor]fix torchair rope ops (#2683) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit ### What this PR does / why we need it? Due to the registration mechanism, torchair ops can not take effect, so have to patch the Ascend ops to adapt torchair ### Does this PR introduce _any_ user-facing change? No ### How was this patch tested? vLLM version: main vLLM main: https://github.com/vllm-project/vllm/commit/7ea22e42d5f666a26b3ce4117724dadfdb4d3887 - vLLM version: main - vLLM main: https://github.com/vllm-project/vllm/commit/7ea22e42d5f666a26b3ce4117724dadfdb4d3887 Signed-off-by: hust17yixuan <303660421@qq.com> --- vllm_ascend/torchair/utils.py | 13 ++++++------- 1 file changed, 6 insertions(+), 7 deletions(-) diff --git a/vllm_ascend/torchair/utils.py b/vllm_ascend/torchair/utils.py index a9bbdae..8dd1e3f 100644 --- a/vllm_ascend/torchair/utils.py +++ b/vllm_ascend/torchair/utils.py @@ -185,15 +185,14 @@ def torchair_quant_method_register(): def torchair_ops_patch(): - from vllm.model_executor.layers.rotary_embedding import ( - DeepseekScalingRotaryEmbedding, RotaryEmbedding) - + from vllm_ascend.ops.rotary_embedding import ( + AscendDeepseekScalingRotaryEmbedding, AscendRotaryEmbedding) from vllm_ascend.torchair.ops.torchair_rotary_embedding import ( deepseek_rope_init_func, native_rope_deepseek_forward, qwen_rope_init_func, rope_forward) - RotaryEmbedding.__init__ = qwen_rope_init_func - RotaryEmbedding.forward_oot = rope_forward + AscendRotaryEmbedding.__init__ = qwen_rope_init_func # type: ignore[method-assign] + AscendRotaryEmbedding.forward_oot = rope_forward # type: ignore[method-assign] - DeepseekScalingRotaryEmbedding.__init__ = deepseek_rope_init_func - DeepseekScalingRotaryEmbedding.forward = native_rope_deepseek_forward + AscendDeepseekScalingRotaryEmbedding.__init__ = deepseek_rope_init_func # type: ignore[method-assign] + AscendDeepseekScalingRotaryEmbedding.forward = native_rope_deepseek_forward # type: ignore[method-assign]