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
48 lines
1.4 KiB
Python
48 lines
1.4 KiB
Python
from typing import Union
|
|
|
|
import ixformer._C as ops
|
|
import torch
|
|
|
|
__all__ = ["residual_bias", "ref_residual_bias"]
|
|
|
|
|
|
def ref_residual_bias(
|
|
input: torch.Tensor,
|
|
residual: torch.Tensor,
|
|
bias: torch.Tensor = None,
|
|
alpha: float = 1,
|
|
):
|
|
if bias is not None:
|
|
output = residual.float() * alpha + input.float() + bias.float()
|
|
else:
|
|
output = residual.float() * alpha + input.float()
|
|
|
|
return output.to(residual.dtype)
|
|
|
|
|
|
def residual_bias(
|
|
input: torch.Tensor,
|
|
residual: torch.Tensor,
|
|
bias: torch.Tensor = None,
|
|
alpha: float = 1,
|
|
output: torch.Tensor = None
|
|
):
|
|
"""
|
|
Args:
|
|
input: [batch_count, seq_len, hidden_size] or [batch_tokens, hidden_size] torch.half
|
|
residual: [batch_count, seq_len, hidden_size] or [batch_tokens, hidden_size] torch.half
|
|
bias: [hidden_size] torch.half
|
|
alpha: float
|
|
Returns:
|
|
output: [batch_count, seq_len, hidden_size] or [batch_tokens, hidden_size] torch.half
|
|
"""
|
|
if output is None:
|
|
output = torch.empty_like(input)
|
|
if alpha is None:
|
|
alpha = 1
|
|
if bias is not None:
|
|
ops.train.add_residual_bias_forward(input, residual, bias, alpha, output)
|
|
else:
|
|
ops.train.add_residual_bias_forward(input, residual, alpha, output)
|
|
return output
|