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
48 lines
1.2 KiB
Python
48 lines
1.2 KiB
Python
from typing import List, Union
|
|
|
|
import ixformer._C as ops
|
|
import torch
|
|
from torch.autograd.function import Function, FunctionCtx
|
|
|
|
__all__ = [
|
|
"gelu",
|
|
]
|
|
|
|
|
|
class GeluFunction(Function):
|
|
@staticmethod
|
|
def forward(
|
|
ctx, input: torch.Tensor, in_place: bool = False, training: bool = False
|
|
):
|
|
if training:
|
|
if in_place:
|
|
ctx.save_for_backward(input.clone())
|
|
else:
|
|
ctx.save_for_backward(input)
|
|
if in_place:
|
|
return ops.train.gelu_forward(input, input)
|
|
else:
|
|
return ops.train.gelu_forward(input)
|
|
|
|
@staticmethod
|
|
def backward(ctx: FunctionCtx, grad_outputs):
|
|
input = ctx.saved_tensors[0]
|
|
grad_input = ops.train.gelu_backward(input, grad_outputs)
|
|
return grad_input, None, None
|
|
|
|
|
|
def gelu(
|
|
input: torch.Tensor, in_place: bool = False, training: bool = False
|
|
) -> torch.Tensor:
|
|
"""
|
|
等价实现：
|
|
torch.nn.functional.gelu
|
|
|
|
Args:
|
|
input: dtype:[torch.float, torch.half, torch.bfloat16]
|
|
in place: bool. Whether to operate directly on the original input data.
|
|
Returns:
|
|
output: dtype:[torch.float, torch.half, torch.bfloat16]
|
|
"""
|
|
return GeluFunction.apply(input, in_place, training)
|