002f9879b2
Replaces cherry-picked upstream_ref with complete source trees. xllm/ — Iluvatar official C++ inference engine (15MB, 1470 files) Complete: kernels → layers → models → runtime → scheduler → api Excluded: .git, binary images, third_party submodule checkouts ds_vllm/ — Iluvatar official vllm fork (8MB, 703 files) Included: csrc/ (ALL CUDA kernels), fused_moe/, qwen3_5 model, _custom_ops Excluded: tests, benchmarks, docs, examples (not needed for reference) Critical call chains now fully traceable: MoE: moe_topk_softmax_kernels.cuh → ixformer.h → fused_moe.cpp → layer GDN: qwen3_gated_delta_net_base.cpp → qwen3_5_gated_delta_net.cpp Attention: ixformer.h → xllm_paged_attention → attention.cpp
How to compile xllm dynamic library
Run the following command in root directory:
python setup.py build --generate-so true
If you want to debug, it needs to set DEBUG environment variable.
export DEBUG=1
How to install dynamic library
Run installation script xllm/c_api/install.sh, headers and dynamic library will be installed in /usr/local/xllm directory.
cd xllm/c_api/tools
sh install.sh
You will see the following files in /usr/local/xllm directory:
[root@A03-R40-I189-101-4100046]# tree /usr/local/xllm
/usr/local/xllm
|-- include
| |-- llm.h
| |-- default.h
| |-- rec.h
| `-- types.h
`-- lib
`-- libxllm.so
3 directories, 5 files
How to compile c_api examples
GPU builds and NPU builds use different link commands. Replace
<example>.cpp and <example> with the example source file and output binary
name you want to build.
GPU
cd xllm/c_api/examples
g++ <example>.cpp -o <example> \
-std=c++17 \
-DUSE_CUDA \
-I/usr/local/xllm/include \
-L/usr/local/xllm/lib \
-lxllm \
-Wl,-rpath=/usr/local/xllm/lib
NPU
Before compiling or running examples, source the Ascend environment first:
source /usr/local/Ascend/ascend-toolkit/set_env.sh
Then compile with the extra custom op library used by the NPU build:
cd xllm/c_api/examples
g++ <example>.cpp -o <example> \
-std=c++17 \
-DUSE_NPU \
-I/usr/local/xllm/include \
-L/usr/local/xllm/lib \
-L/usr/local/Ascend/ascend-toolkit/latest/opp/vendors/xllm/op_api/lib \
-lxllm \
-lcust_opapi \
-Wl,-rpath=/usr/local/xllm/lib \
-Wl,-rpath=/usr/local/Ascend/ascend-toolkit/latest/opp/vendors/xllm/op_api/lib
If -lcust_opapi is missing from the NPU link command, the linker may report
undefined references to symbols such as aclnnBeamSearchGroup and
aclnnXAttention.
How to run c_api examples
Some examples, such as simple_rec_completions, support overriding the target
device from argv[1].
NPU
./simple_rec_completions npu:14
GPU
./simple_rec_completions cuda:0