apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-09-26 20:41:53 +08:00
Code Issues Actions 5 Packages Projects Releases Wiki Activity

[XPU] Enable XPU V1 mode based on environment variable (#4213)

Browse Source 
* Enable XPU V1 mode based on environment variable
* add default param to xft_moe_fc_block_eb for latest xvllm compatibility; update run_ci_xpu to use latest xvllm
This commit is contained in:
yyssys
2025-09-24 10:29:48 +08:00
committed by GitHub
parent ec99474e71
commit d6e59447f5
4 changed files with 5 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
custom_ops/xpu_ops/src/ops/moe_expert_ffn.cc
View File

@@ -72,6 +72,7 @@ void MoeExpertFFNImpl(xftblock::Tensor* ffn_in,
is_padding_input ? token_num_info : nullptr,
expert_num,
1, // moe_topk
0, // group_size
ffn1_out_shape.size() == 2 ? xftblock::MoeFCInputMode::DENSE
: xftblock::MoeFCInputMode::SPARSE);
PD_CHECK(ret == 0);
@@ -134,6 +135,7 @@ void MoeExpertFFNImpl(xftblock::Tensor* ffn_in,
is_padding_input ? token_num_info : nullptr,
expert_num,
1, // moe_topk
0, // group_size
ffn1_out_shape.size() == 2
? xftblock::MoeFCInputMode::DENSE
: xftblock::MoeFCInputMode::SPARSE); // bias_mode

2
fastdeploy/engine/args_utils.py
View File

@@ -424,7 +424,7 @@ class EngineArgs:
envs.ENABLE_V1_KVCACHE_SCHEDULER = 0
if self.splitwise_role != "mixed" and self.cache_transfer_protocol != "rdma":
envs.ENABLE_V1_KVCACHE_SCHEDULER = 0
if not current_platform.is_cuda():
if not current_platform.is_cuda() and not current_platform.is_xpu():
envs.ENABLE_V1_KVCACHE_SCHEDULER = 0
if self.guided_decoding_backend != "off":
envs.ENABLE_V1_KVCACHE_SCHEDULER = 0

2
fastdeploy/worker/worker_process.py
View File

@@ -778,7 +778,7 @@ def initialize_fd_config(args, ranks: int = 1, local_rank: int = 0) -> FDConfig:
envs.ENABLE_V1_KVCACHE_SCHEDULER = 0
if args.splitwise_role != "mixed" and args.cache_transfer_protocol != "rdma":
envs.ENABLE_V1_KVCACHE_SCHEDULER = 0
if not current_platform.is_cuda():
if not current_platform.is_cuda() and not current_platform.is_xpu():
logger.info("Set ENABLE_V1_KVCACHE_SCHEDULER to 0 due to not supported.")
envs.ENABLE_V1_KVCACHE_SCHEDULER = 0
if parallel_config.guided_decoding_backend != "off":

5
scripts/run_ci_xpu.sh
View File

@@ -24,10 +24,7 @@ python -m pip install paddlepaddle-xpu -i https://www.paddlepaddle.org.cn/packag
echo "build whl"
bash custom_ops/xpu_ops/download_dependencies.sh develop
export CLANG_PATH=$(pwd)/custom_ops/xpu_ops/third_party/xtdk
# export XVLLM_PATH=$(pwd)/custom_ops/xpu_ops/third_party/xvllm
# 由于xvllm更新导致编译报错暂时锁定xvllm版本
wget https://klx-sdk-release-public.su.bcebos.com/xinfer/daily/eb/20250921/output.tar.gz --no-proxy && tar xf output.tar.gz && mv output xvllm
export XVLLM_PATH=${PWD}/xvllm
export XVLLM_PATH=$(pwd)/custom_ops/xpu_ops/third_party/xvllm
bash build.sh || exit 1
echo "pip others"