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
51 lines
1.7 KiB
Python
51 lines
1.7 KiB
Python
import ixformer._C as ops
|
|
import torch
|
|
|
|
__all__ = ["matmul", "ref_matmul"]
|
|
|
|
|
|
def ref_matmul(input, other, *, transa, transb, alpha):
|
|
if transa:
|
|
dims = list(range(input.ndim))
|
|
dims[-1], dims[-2] = dims[-2], dims[-1]
|
|
input = input.permute(*dims).contiguous()
|
|
|
|
if transb:
|
|
dims = list(range(other.ndim))
|
|
dims[-1], dims[-2] = dims[-2], dims[-1]
|
|
other = other.permute(*dims).contiguous()
|
|
|
|
return alpha * torch.matmul(input, other)
|
|
|
|
|
|
def matmul(
|
|
input: torch.Tensor,
|
|
other: torch.Tensor,
|
|
*,
|
|
transa: bool = False,
|
|
transb: bool = False,
|
|
alpha: float = 1.0,
|
|
) -> torch.Tensor:
|
|
"""
|
|
Args:
|
|
input: (...,m,k) or (...,k,m) torch.half
|
|
当transa为False shape : [...,m,k], 当transa为True shape : [...,k,m]
|
|
other: (...,k,n) or (...,n,k) torch.half
|
|
当transa为False shape : [...,m,k], 当transa为True shape : [...,k,m]
|
|
transa: bool
|
|
transb: bool
|
|
alpha: float
|
|
Returns:
|
|
Tensor: (..., m, n) torch.half
|
|
"""
|
|
if not input.is_contiguous():
|
|
input = input.contiguous()
|
|
|
|
if not other.is_contiguous():
|
|
if not other.transpose(-2, -1).is_contiguous():
|
|
other = other.contiguous()
|
|
|
|
return ops.train.matmul(
|
|
input, other, transa=transa, transb=transb, alpha=alpha, beta=0.0
|
|
)
|