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
62 lines
1.3 KiB
Python
62 lines
1.3 KiB
Python
import ixformer.inference.functions as ops
|
|
import torch
|
|
|
|
|
|
def fused_add_rmsnorm(
|
|
input: torch.Tensor, residual: torch.Tensor, weight: torch.Tensor, eps: float = 1e-6
|
|
):
|
|
r"""Fused add root mean square normalization.
|
|
|
|
Parameters
|
|
----------
|
|
input: torch.Tensor
|
|
Input tensor, shape (batch_size, hidden_size).
|
|
residual: torch.Tensor
|
|
Residual tensor, shape (batch_size, hidden_size).
|
|
weight: torch.Tensor
|
|
Weight tensor, shape (hidden_size,).
|
|
eps: float
|
|
Epsilon for numerical stability.
|
|
"""
|
|
return ops.residual_rms_norm(
|
|
input=input,
|
|
residual=residual,
|
|
weight=weight,
|
|
eps=eps,
|
|
)
|
|
|
|
|
|
def gemma_fused_add_rmsnorm():
|
|
pass
|
|
|
|
|
|
def gemma_rmsnorm():
|
|
pass
|
|
|
|
|
|
def rmsnorm(
|
|
input: torch.Tensor, weight: torch.Tensor, eps: float = 1e-6
|
|
) -> torch.Tensor:
|
|
r"""Root mean square normalization.
|
|
|
|
Parameters
|
|
----------
|
|
input: torch.Tensor
|
|
Input tensor, shape (batch_size, hidden_size).
|
|
weight: torch.Tensor
|
|
Weight tensor, shape (hidden_size,).
|
|
eps: float
|
|
Epsilon for numerical stability.
|
|
|
|
Returns
|
|
-------
|
|
output: torch.Tensor
|
|
Normalized tensor, shape (batch_size, hidden_size).
|
|
"""
|
|
|
|
return ops.rms_norm(
|
|
input=input,
|
|
weight=weight,
|
|
eps=eps,
|
|
)
|