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
57 lines
1.8 KiB
Python
57 lines
1.8 KiB
Python
from typing import List, Union
|
|
|
|
import ixformer._C as ops
|
|
import torch
|
|
|
|
__all__ = ["bnb_qgemm", "ref_bnb_qgemm"]
|
|
|
|
|
|
# qA : quant input shape : [bs, in_feature]
|
|
# qW : quant weight shape : [out_feature, in_feature]
|
|
# SA : scale vector of qA shape : [bs]
|
|
# SW : scale vector of qW shape : [out_feature]
|
|
|
|
|
|
def ref_bnb_qgemm(
|
|
qA: torch.Tensor,
|
|
qW: torch.Tensor,
|
|
SA: torch.Tensor,
|
|
SW: torch.Tensor,
|
|
training: bool = False,
|
|
scaleA: float = 127.0,
|
|
scaleW: float = 127.0,
|
|
):
|
|
y = torch.nn.functional.linear(qA.to(torch.float), qW.to(torch.float))
|
|
out = torch.empty(y.shape, dtype = SA.dtype, device = SA.device)
|
|
for i in range(qA.size(0)):
|
|
for j in range(qW.size(0)):
|
|
out[i][j] = y[i][j] * (SA[i].to(torch.float) / scaleA) * (SW[j].to(torch.float) / scaleW)
|
|
return out.to(SA.dtype)
|
|
|
|
|
|
def bnb_qgemm(
|
|
qA: torch.Tensor,
|
|
qW: torch.Tensor,
|
|
SA: torch.Tensor,
|
|
SW: torch.Tensor,
|
|
training: bool = False,
|
|
scaleA: float = 127.0,
|
|
scaleW: float = 127.0,
|
|
) -> torch.Tensor:
|
|
|
|
"""
|
|
Args:
|
|
qA: (bs, in_feature) torch.int8
|
|
qW: (out_feature, in_feature) torch.int8
|
|
SA: (bs) torch.half
|
|
scale vector of qA
|
|
SA: (out_feature) torch.half
|
|
scale vector of qW
|
|
training: bool
|
|
scaleA: float
|
|
scaleW: float
|
|
Returns:
|
|
Tensor: (bs, out_feature) torch.half
|
|
"""
|
|
return ops.infer.bnb_qgemm(qA, qW, SA, SW, scaleA, scaleW)
|