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
26 lines
736 B
Python
26 lines
736 B
Python
import torch
|
|
import torch.nn as nn
|
|
from abc import ABC
|
|
from ixformer.train.speedformer.model_replacer_mapping import ModelMapping
|
|
|
|
# 外部接口
|
|
class SpeedFormer(ABC):
|
|
def __init__(self) -> None:
|
|
super().__init__()
|
|
self.replacer = None
|
|
|
|
|
|
def accelerate(self, model):
|
|
if model.config.model_type in ModelMapping:
|
|
self.replacer = ModelMapping[model.config.model_type]()
|
|
accelerate_model = self.replacer.accelerate(model)
|
|
else:
|
|
Warning(f"Warning: model '{model.config.model_type}' is not supported now.")
|
|
accelerate_model = model
|
|
|
|
return accelerate_model
|
|
|
|
|
|
def post_process(self, model):
|
|
self.replacer.post_process(model)
|