[XPU] use rms_norm xpu kernel for context-key normalization (#58896)
Signed-off-by: Yan Ma <yan.ma@intel.com> Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
This commit is contained in:
@@ -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,
|
||||
|
||||
Reference in New Issue
Block a user