ColossalAI/colossalai/nn/layer/fused_bias_gelu.py

35 lines
1.2 KiB
Python
Raw Normal View History

Develop/experiments (#59) * Add gradient accumulation, fix lr scheduler * fix FP16 optimizer and adapted torch amp with tensor parallel (#18) * fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes * fixed trainer * Revert "fixed trainer" This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097. * improved consistency between trainer, engine and schedule (#23) Co-authored-by: 1SAA <c2h214748@gmail.com> * Split conv2d, class token, positional embedding in 2d, Fix random number in ddp Fix convergence in cifar10, Imagenet1000 * Integrate 1d tensor parallel in Colossal-AI (#39) * fixed 1D and 2D convergence (#38) * optimized 2D operations * fixed 1D ViT convergence problem * Feature/ddp (#49) * remove redundancy func in setup (#19) (#20) * use env to control the language of doc (#24) (#25) * Support TP-compatible Torch AMP and Update trainer API (#27) * Add gradient accumulation, fix lr scheduler * fix FP16 optimizer and adapted torch amp with tensor parallel (#18) * fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes * fixed trainer * Revert "fixed trainer" This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097. * improved consistency between trainer, engine and schedule (#23) Co-authored-by: 1SAA <c2h214748@gmail.com> Co-authored-by: 1SAA <c2h214748@gmail.com> Co-authored-by: ver217 <lhx0217@gmail.com> * add an example of ViT-B/16 and remove w_norm clipping in LAMB (#29) * add explanation for ViT example (#35) (#36) * support torch ddp * fix loss accumulation * add log for ddp * change seed * modify timing hook Co-authored-by: Frank Lee <somerlee.9@gmail.com> Co-authored-by: 1SAA <c2h214748@gmail.com> Co-authored-by: binmakeswell <binmakeswell@gmail.com> * Feature/pipeline (#40) * remove redundancy func in setup (#19) (#20) * use env to control the language of doc (#24) (#25) * Support TP-compatible Torch AMP and Update trainer API (#27) * Add gradient accumulation, fix lr scheduler * fix FP16 optimizer and adapted torch amp with tensor parallel (#18) * fixed bugs in compatibility between torch amp and tensor parallel and performed some minor fixes * fixed trainer * Revert "fixed trainer" This reverts commit 2e0b0b76990e8d4e337add483d878c0f61cf5097. * improved consistency between trainer, engine and schedule (#23) Co-authored-by: 1SAA <c2h214748@gmail.com> Co-authored-by: 1SAA <c2h214748@gmail.com> Co-authored-by: ver217 <lhx0217@gmail.com> * add an example of ViT-B/16 and remove w_norm clipping in LAMB (#29) * add explanation for ViT example (#35) (#36) * optimize communication of pipeline parallel * fix grad clip for pipeline Co-authored-by: Frank Lee <somerlee.9@gmail.com> Co-authored-by: 1SAA <c2h214748@gmail.com> Co-authored-by: binmakeswell <binmakeswell@gmail.com> * optimized 3d layer to fix slow computation ; tested imagenet performance with 3d; reworked lr_scheduler config definition; fixed launch args; fixed some printing issues; simplified apis of 3d layers (#51) * Update 2.5d layer code to get a similar accuracy on imagenet-1k dataset * update api for better usability (#58) update api for better usability Co-authored-by: 1SAA <c2h214748@gmail.com> Co-authored-by: ver217 <lhx0217@gmail.com> Co-authored-by: puck_WCR <46049915+WANG-CR@users.noreply.github.com> Co-authored-by: binmakeswell <binmakeswell@gmail.com> Co-authored-by: アマデウス <kurisusnowdeng@users.noreply.github.com> Co-authored-by: BoxiangW <45734921+BoxiangW@users.noreply.github.com>
2021-12-09 07:08:29 +00:00
# adapted from Megatron-LM
# https://github.com/NVIDIA/Megatron-LM/blob/b31e1296354e979722627a6c4dedafe19b51fa97/megatron/model/fused_bias_gelu.py
import torch
@torch.jit.script
def bias_gelu(bias, y):
x = bias + y
return x * 0.5 * (1.0 + torch.tanh(0.79788456 * x * (1 + 0.044715 * x * x)))
# gradient of tanh approximation of gelu
# gradient of actual gelu is:
# 0.5 * (1. + torch.erf(x * 0.70710678)) + 0.3989423 * x * torch.exp(-0.5 * x * x)
@torch.jit.script
def bias_gelu_back(g, bias, y):
x = bias + y
tanh_out = torch.tanh(0.79788456 * x * (1 + 0.044715 * x * x))
# sqrt(2/pi) * 3 * 0.044715 -> 0.1070322243
ff = 0.5 * x * ((1 - tanh_out * tanh_out) * (0.79788456 + 0.1070322243 * x * x)) + 0.5 * (1 + tanh_out)
return ff*g
class GeLUFunction(torch.autograd.Function):
@staticmethod
# bias is an optional argument
def forward(ctx, input, bias):
ctx.save_for_backward(input, bias)
return bias_gelu(bias, input)
@staticmethod
def backward(ctx, grad_output):
input, bias = ctx.saved_tensors
tmp = bias_gelu_back(grad_output, bias, input)
return tmp, tmp
bias_gelu_impl = GeLUFunction.apply