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
28 lines
923 B
Python
28 lines
923 B
Python
import torch
|
|
import torch.nn as nn
|
|
from torch.nn import LayerNorm
|
|
from types import ModuleType, MethodType
|
|
from abc import ABC
|
|
|
|
from ixformer.train.speedformer.models.gpt2.modeling_gpt2 import GPT2FlashAttention2
|
|
|
|
from ixformer.train.speedformer.layers.normalization import replace_layernorm_forward
|
|
from ixformer.train.speedformer.layers.gpt2.attention import replace_flash_attn_forward
|
|
|
|
|
|
class GPT2Replacer(ABC):
|
|
def __init__(self) -> None:
|
|
super().__init__()
|
|
|
|
@staticmethod
|
|
def accelerate(model):
|
|
# layer/kernel replace
|
|
for name, module in model.named_modules():
|
|
if isinstance(module, LayerNorm):
|
|
module.forward = MethodType(replace_layernorm_forward, module)
|
|
if isinstance(module, GPT2FlashAttention2):
|
|
module._flash_attention_forward = MethodType(
|
|
replace_flash_attn_forward, module)
|
|
|
|
return model
|