From 48f663c37f8c77803d9281778d49ec0727abcacf Mon Sep 17 00:00:00 2001 From: Yan Ma Date: Fri, 18 Sep 2026 07:52:01 +0800 Subject: [PATCH] [XPU] Fix incorrect context-key normalization for Qwen DFlash-based models (#56431) Signed-off-by: Yan Ma --- vllm/model_executor/models/qwen3_dflash.py | 10 ++++++++++ 1 file changed, 10 insertions(+) diff --git a/vllm/model_executor/models/qwen3_dflash.py b/vllm/model_executor/models/qwen3_dflash.py index ce3884a10e..5504304c73 100644 --- a/vllm/model_executor/models/qwen3_dflash.py +++ b/vllm/model_executor/models/qwen3_dflash.py @@ -32,6 +32,7 @@ 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 @@ -561,6 +562,15 @@ 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,