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
53 lines
1.4 KiB
Python
53 lines
1.4 KiB
Python
from typing import List, Union
|
|
|
|
import ixformer._C as ops
|
|
import torch
|
|
|
|
__all__ = ["bnb_rowcol_absmax", "ref_bnb_rowcol_absmax"]
|
|
|
|
|
|
# input : input shape : [row, col]
|
|
# threshold : abs of element exceeds threshold will be ignored
|
|
# type
|
|
# 0 : row absmax
|
|
def ref_bnb_rowcol_absmax(
|
|
input: torch.Tensor,
|
|
training: bool = False,
|
|
threshold: float = 0.0,
|
|
type: int = 0,
|
|
):
|
|
input = input.float()
|
|
if threshold ==0.0:
|
|
threshold = float('inf')
|
|
mask = (torch.abs(input) < threshold)
|
|
masked_input = mask * input
|
|
masked_input = masked_input.half()
|
|
if type == 0:
|
|
out = torch.amax(torch.abs(masked_input), dim=1)
|
|
|
|
else:
|
|
out = torch.amax(torch.abs(masked_input), dim=0)
|
|
return out
|
|
|
|
|
|
def bnb_rowcol_absmax(
|
|
input: torch.Tensor,
|
|
training: bool = False,
|
|
threshold: float = 0.0,
|
|
type: int = 0,
|
|
) -> torch.Tensor:
|
|
|
|
"""
|
|
Args:
|
|
input: (row, col) torch.half
|
|
目前col值必须满足col%2==0
|
|
training: bool
|
|
threshold: float
|
|
abs of element exceeds threshold will be ignored
|
|
type: int
|
|
row absmax, 目前只支持type=0
|
|
Returns:
|
|
Tensor: (row) torch.half
|
|
"""
|
|
return ops.infer.bnb_rowcol_absmax(input, threshold, type)
|