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
83 lines
2.4 KiB
Python
83 lines
2.4 KiB
Python
from typing import Union
|
|
|
|
import ixformer._C as ops
|
|
import torch
|
|
from torch.autograd.function import Function, FunctionCtx
|
|
|
|
import ixformer
|
|
|
|
__all__ = ["residual_bias"]
|
|
|
|
|
|
class ResidualBiasFunction(Function):
|
|
@staticmethod
|
|
def forward(
|
|
ctx,
|
|
input: torch.Tensor,
|
|
residual: torch.Tensor,
|
|
bias: torch.Tensor,
|
|
output: torch.Tensor,
|
|
alpha=1,
|
|
):
|
|
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)
|
|
ctx.has_bias = bias is not None
|
|
ctx.alpha = alpha
|
|
return output
|
|
|
|
@staticmethod
|
|
def backward(ctx: FunctionCtx, grad_output):
|
|
grad_input = torch.empty_like(grad_output)
|
|
grad_residual = torch.empty_like(grad_output)
|
|
if ctx.has_bias:
|
|
grad_bias = torch.empty(
|
|
[grad_output.size(-1)],
|
|
dtype=grad_output.dtype,
|
|
device=grad_output.device,
|
|
)
|
|
ops.train.add_residual_bias_backward(
|
|
grad_output,
|
|
grad_input,
|
|
grad_residual,
|
|
grad_bias,
|
|
ctx.alpha,
|
|
)
|
|
return (grad_input, grad_residual, grad_bias, None, None)
|
|
else:
|
|
ops.train.add_residual_bias_backward(
|
|
grad_output,
|
|
grad_input,
|
|
grad_residual,
|
|
ctx.alpha,
|
|
)
|
|
return (grad_input, grad_residual, None, None, None)
|
|
|
|
|
|
def residual_bias(
|
|
input: torch.Tensor,
|
|
residual: torch.Tensor,
|
|
bias: torch.Tensor = None,
|
|
output: torch.Tensor = None,
|
|
alpha=1,
|
|
):
|
|
"""
|
|
等价实现：
|
|
input = residual.float() * alpha + input.float() + bias.float()
|
|
|
|
参数说明:
|
|
Args:
|
|
input: shape:[batch_count, seq_len, hidden_size],dtype:[torch.half]
|
|
residual: shape:[batch_count, seq_len, hidden_size],dtype:[torch.half]
|
|
bias: shape:[hidden_size],dtype:[torch.half]
|
|
alpha: float
|
|
return:
|
|
output: shape:[batch_count, seq_len, hidden_size],dtype:[torch.half]
|
|
"""
|
|
return ResidualBiasFunction.apply(input, residual, bias, output, alpha)
|