From 754aa29c006ddfbc518ae7d89e30414c6866d890 Mon Sep 17 00:00:00 2001 From: Yan Ma Date: Tue, 29 Sep 2026 07:09:58 +0800 Subject: [PATCH] [XPU] use rms_norm xpu kernel for context-key normalization (#58896) Signed-off-by: Yan Ma Co-authored-by: Kunshang Ji --- vllm/model_executor/models/qwen3_dflash.py | 10 ---------- 1 file changed, 10 deletions(-) diff --git a/vllm/model_executor/models/qwen3_dflash.py b/vllm/model_executor/models/qwen3_dflash.py index 36a6c4c2c1..3161cf2261 100644 --- a/vllm/model_executor/models/qwen3_dflash.py +++ b/vllm/model_executor/models/qwen3_dflash.py @@ -34,7 +34,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( VocabParallelEmbedding, ) from vllm.multimodal.inputs import NestedTensors -from vllm.platforms import current_platform from vllm.transformers_utils.config import set_default_rope_theta from vllm.transformers_utils.repo_utils import get_hf_file_bytes from vllm.v1.attention.backend import AttentionType @@ -622,15 +621,6 @@ class DFlashQwen3Model(nn.Module): # --- Grouped RMSNorm K across all layers ([L, num_ctx, nkv, hd]) --- # The weight is selected per layer by the outermost (layer) index. all_k_normed = torch.empty_like(all_k) - if current_platform.is_xpu(): - for layer_idx in range(all_k.shape[0]): - ops.rms_norm( - all_k_normed[layer_idx], - all_k[layer_idx], - self._k_norm_weights[layer_idx], - self._rms_norm_eps, - ) - return all_k_normed ops.rms_norm( all_k_normed, all_k,