87a19d2d00
来源:
1. Chranos/ixformer (GitHub) → ixformer_sdk/ (230 files, 70K lines)
- inference/functions/vllm.py: vllm_moe_topk_softmax 完整实现 (2033 lines)
- inference/functions/moe.py: MoE ops 完整实现 (1380 lines)
- contrib/vllm_flash_attn/: FA2 Python 接口 (1018 lines)
- contrib/tgi/fused_moe.py: TGI fused MoE (429 lines)
- csrc/include/ixformer/: C++ kernel headers + cmake
2. Deep-Spark/xllm (GitHub) → upstream_ref/xllm_latest/ (+15 files)
- npu_torch/qwen3_5_decoder_layer_impl.cpp/.h
- npu_torch/qwen3_5_gated_delta_net.cpp/.h
- npu_torch/qwen3_next_*.cpp/.h (6 files)
- npu_torch/attention.cpp/.h + fused_moe.cpp/.h + CMakeLists.txt
- models/llm/qwen3_5.h + qwen3_5_mtp.h + qwen3_next.h
- models/vlm/qwen3_5.h
调用链完整性:
ixformer_sdk/inference/functions/vllm.py
→ ops.infer.moe_topk_softmax() (C++ 层)
→ 这就是 base 镜像 libixformer.so 里的实现
upstream_ref/xllm_latest/core/layers/ilu/fused_moe.cpp
→ ixformer::infer::topk_softmax() (直接 C++ 调用)
→ ixformer::infer::group_gemm() → 完整 7-step MoE pipeline
38 lines
943 B
Python
38 lines
943 B
Python
from psutil import Process
|
|
|
|
|
|
def get_current_memory(pid=None):
|
|
return Process(pid).memory_full_info()
|
|
|
|
|
|
class MemoryMonitorContext(object):
|
|
def __init__(self, pid=None):
|
|
self._pid = pid
|
|
|
|
self.reset()
|
|
|
|
def __enter__(self):
|
|
self._enter_memory = self._get_used_memory()
|
|
return self
|
|
|
|
def __exit__(self, exc_type, exc_val, exc_tb):
|
|
self._exit_memory = self._get_used_memory()
|
|
return self
|
|
|
|
def _get_used_memory(self) -> float:
|
|
mem = get_current_memory(self._pid)
|
|
return mem.uss
|
|
|
|
def reset(self):
|
|
self._enter_memory = -1
|
|
self._exit_memory = -1
|
|
|
|
def delta(self) -> float:
|
|
if self._enter_memory < 0:
|
|
raise RuntimeError("Please using context manager to wrap your code.")
|
|
|
|
if self._exit_memory < 0:
|
|
return self._get_used_memory() - self._enter_memory
|
|
|
|
return self._exit_memory - self._enter_memory
|