[XPU] Fix incorrect context-key normalization for Qwen DFlash-based models (#56431)

Signed-off-by: Yan Ma <yan.ma@intel.com>
This commit is contained in:
Yan Ma
2026-09-18 07:52:01 +08:00
committed by GitHub
parent 092bdd6d57
commit 48f663c37f
@@ -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,