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
33 lines
935 B
Python
33 lines
935 B
Python
import os
|
|
from typing import Union
|
|
|
|
import ixformer._C as ops
|
|
import torch
|
|
|
|
__all__ = ["i8w8o32", "ref_i8w8o32"]
|
|
|
|
|
|
def ref_i8w8o32(input: torch.Tensor, weight: torch.Tensor):
|
|
output = torch.nn.functional.linear(input.float(), weight.float()).int()
|
|
return output
|
|
|
|
|
|
def i8w8o32(input: torch.Tensor, weight: torch.Tensor):
|
|
|
|
"""
|
|
Args:
|
|
input: (bs, ic) torch.int8
|
|
weight: (oc, ic) torch.int8
|
|
Returns:
|
|
Tensor: (bs, oc)) torch.int32
|
|
"""
|
|
if not torch.is_tensor(input):
|
|
raise RuntimeError("Not impl.")
|
|
output_shape = list(input.shape)
|
|
output_shape[-1] = weight.size(0)
|
|
output = torch.empty(output_shape, dtype=torch.int32, device=input.device)
|
|
ic_dim = input.size(-1)
|
|
input = input.view(-1, ic_dim)
|
|
ops.infer.linear_i8w8o32(input.view(-1, ic_dim), weight, output)
|
|
return output
|