From c771b3a75a6ebbfbfc398a028a477246b0799cf0 Mon Sep 17 00:00:00 2001 From: Tim Dettmers Date: Fri, 22 Jul 2022 14:41:05 -0700 Subject: Most tests passing. --- tests/test_optim.py | 87 ++++++++++++----------------------------------------- 1 file changed, 19 insertions(+), 68 deletions(-) (limited to 'tests/test_optim.py') diff --git a/tests/test_optim.py b/tests/test_optim.py index c80fe51..b173eaa 100644 --- a/tests/test_optim.py +++ b/tests/test_optim.py @@ -1,12 +1,9 @@ -# Copyright (c) Facebook, Inc. and its affiliates. -# -# This source code is licensed under the MIT license found in the -# LICENSE file in the root directory of this source tree. import os import time import shutil import uuid import pytest +import ctypes import torch import bitsandbytes as bnb import bitsandbytes.functional as F @@ -14,7 +11,9 @@ import bitsandbytes.functional as F from os.path import join from itertools import product -import apex +#import apex + +k = 20 def get_temp_dir(): path = '/tmp/autoswap/{0}'.format(str(uuid.uuid4())) @@ -26,55 +25,47 @@ def rm_path(path): str2optimizers = {} str2optimizers['adam_pytorch'] = (None, torch.optim.Adam, bnb.optim.Adam) -str2optimizers['adam_apex'] = (None, apex.optimizers.FusedAdam, bnb.optim.Adam) -str2optimizers['momentum_apex'] = (None, lambda pxx: apex.optimizers.FusedSGD(pxx, 0.01, 0.9), bnb.optim.Adam) +#str2optimizers['adam_apex'] = (None, apex.optimizers.FusedAdam, bnb.optim.Adam) +#str2optimizers['momentum_apex'] = (None, lambda pxx: apex.optimizers.FusedSGD(pxx, 0.01, 0.9), bnb.optim.Adam) str2optimizers['momentum_pytorch'] = (None, lambda pxx: torch.optim.SGD(pxx, 0.01, 0.9), bnb.optim.Adam) -str2optimizers['lamb_apex'] = (None, lambda pxx: apex.optimizers.FusedLAMB(pxx, weight_decay=0.00, use_nvlamb=True), bnb.optim.Adam) -str2optimizers['lars_apex'] = (None, lambda pxx: apex.parallel.LARC.LARC(apex.optimizers.FusedSGD(pxx, 0.01, 0.9)), bnb.optim.Adam) +#str2optimizers['lamb_apex'] = (None, lambda pxx: apex.optimizers.FusedLAMB(pxx, weight_decay=0.00, use_nvlamb=True), bnb.optim.Adam) +#str2optimizers['lars_apex'] = (None, lambda pxx: apex.parallel.LARC.LARC(apex.optimizers.FusedSGD(pxx, 0.01, 0.9)), bnb.optim.Adam) str2optimizers['adam'] = (torch.optim.Adam, bnb.optim.Adam) -str2optimizers['adamw'] = (torch.optim.AdamW, bnb.optim.AdamW) -str2optimizers['fused_adam'] = (apex.optimizers.FusedAdam, bnb.optim.Adam) +#str2optimizers['fused_adam'] = (apex.optimizers.FusedAdam, bnb.optim.Adam) str2optimizers['momentum'] = (lambda pxx: torch.optim.SGD(pxx, 0.01, 0.9), lambda pxx: bnb.optim.SGD(pxx, 0.01, 0.9, block_wise=False)) str2optimizers['lars'] = (lambda pxx: bnb.optim.PytorchLARS(pxx, 0.01, 0.9), lambda pxx: bnb.optim.LARS(pxx, 0.01, 0.9)) -str2optimizers['lamb'] = (lambda pxx: apex.optimizers.FusedLAMB(pxx, weight_decay=0.0, max_grad_norm=10000.0, eps=1e-8, use_nvlamb=True), bnb.optim.LAMB) +#str2optimizers['lamb'] = (lambda pxx: apex.optimizers.FusedLAMB(pxx, weight_decay=0.0, max_grad_norm=10000.0, eps=1e-8, use_nvlamb=True), bnb.optim.LAMB) str2optimizers['rmsprop'] = (lambda pxx: torch.optim.RMSprop(pxx, 0.01, 0.9), lambda pxx: bnb.optim.RMSprop(pxx, 0.01, 0.9, block_wise=False)) -str2optimizers['adagrad'] = (lambda pxx: torch.optim.Adagrad(pxx, 0.01), lambda pxx: bnb.optim.Adagrad(pxx, 0.01, block_wise=False)) str2optimizers['adam8bit'] = (torch.optim.Adam, lambda pxx: bnb.optim.Adam8bit(pxx, block_wise=False)) str2optimizers['momentum8bit'] = (lambda pxx: torch.optim.SGD(pxx, 0.01, 0.9), lambda pxx: bnb.optim.SGD8bit(pxx, 0.01, 0.9, block_wise=False)) str2optimizers['rmsprop8bit'] = (lambda pxx: torch.optim.RMSprop(pxx, 0.01, 0.9), lambda pxx: bnb.optim.RMSprop8bit(pxx, 0.01, 0.9, block_wise=False)) -str2optimizers['lamb8bit'] = (lambda pxx: apex.optimizers.FusedLAMB(pxx, weight_decay=0.0, max_grad_norm=10000.0, eps=1e-8, use_nvlamb=True), bnb.optim.LAMB8bit) +#str2optimizers['lamb8bit'] = (lambda pxx: apex.optimizers.FusedLAMB(pxx, weight_decay=0.0, max_grad_norm=10000.0, eps=1e-8, use_nvlamb=True), bnb.optim.LAMB8bit) str2optimizers['lars8bit'] = (lambda pxx: bnb.optim.PytorchLARS(pxx, 0.01, 0.9), lambda pxx: bnb.optim.LARS8bit(pxx, 0.01, 0.9)) str2optimizers['adam8bit_blockwise'] = (torch.optim.Adam, lambda pxx: bnb.optim.Adam8bit(pxx, block_wise=True)) -str2optimizers['adamw8bit_blockwise'] = (torch.optim.Adam, lambda pxx: bnb.optim.AdamW8bit(pxx, block_wise=True)) str2optimizers['momentum8bit_blockwise'] = (lambda pxx: torch.optim.SGD(pxx, 0.01, 0.9), lambda pxx: bnb.optim.SGD8bit(pxx, 0.01, 0.9, block_wise=True)) str2optimizers['rmsprop8bit_blockwise'] = (lambda pxx: torch.optim.RMSprop(pxx, 0.01, 0.9), lambda pxx: bnb.optim.RMSprop8bit(pxx, 0.01, 0.9, block_wise=True)) -str2optimizers['adagrad8bit_blockwise'] = (lambda pxx: torch.optim.Adagrad(pxx, 0.01), lambda pxx: bnb.optim.Adagrad8bit(pxx, 0.01, block_wise=True)) str2statenames = {} str2statenames['adam'] = [('exp_avg', 'state1'), ('exp_avg_sq', 'state2')] -str2statenames['adamw'] = [('exp_avg', 'state1'), ('exp_avg_sq', 'state2')] str2statenames['momentum'] = [('momentum_buffer', 'state1')] str2statenames['lars'] = [('momentum_buffer', 'state1')] str2statenames['lamb'] = [('exp_avg', 'state1'), ('exp_avg_sq', 'state2')] str2statenames['rmsprop'] = [('square_avg', 'state1')] -str2statenames['adagrad'] = [('sum', 'state1')] str2statenames['adam8bit'] = [('exp_avg', 'state1', 'qmap1', 'max1'), ('exp_avg_sq', 'state2', 'qmap2', 'max2')] str2statenames['lamb8bit'] = [('exp_avg', 'state1', 'qmap1', 'max1'), ('exp_avg_sq', 'state2', 'qmap2', 'max2')] str2statenames['adam8bit_blockwise'] = [('exp_avg', 'state1', 'qmap1', 'absmax1'), ('exp_avg_sq', 'state2', 'qmap2', 'absmax2')] -str2statenames['adamw8bit_blockwise'] = [('exp_avg', 'state1', 'qmap1', 'absmax1'), ('exp_avg_sq', 'state2', 'qmap2', 'absmax2')] str2statenames['momentum8bit'] = [('momentum_buffer', 'state1', 'qmap1', 'max1')] str2statenames['momentum8bit_blockwise'] = [('momentum_buffer', 'state1', 'qmap1', 'absmax1')] str2statenames['lars8bit'] = [('momentum_buffer', 'state1', 'qmap1', 'max1')] str2statenames['rmsprop8bit'] = [('square_avg', 'state1', 'qmap1', 'max1')] str2statenames['rmsprop8bit_blockwise'] = [('square_avg', 'state1', 'qmap1', 'absmax1')] -str2statenames['adagrad8bit_blockwise'] = [('sum', 'state1', 'qmap1', 'absmax1')] dim1 = [1024] dim2 = [32, 1024, 4097, 1] gtype = [torch.float32, torch.float16] -optimizer_names = ['adam', 'adamw', 'momentum', 'rmsprop', 'lars', 'lamb', 'adagrad'] +optimizer_names = ['adam', 'momentum', 'rmsprop', 'lars', 'lamb'] values = list(product(dim1,dim2, gtype, optimizer_names)) names = ['dim1_{0}_dim2_{1}_gtype_{2}_optim_{3}'.format(*vals) for vals in values] @pytest.mark.parametrize("dim1, dim2, gtype, optim_name", values, ids=names) @@ -89,12 +80,12 @@ def test_optimizer32bit(dim1, dim2, gtype, optim_name): bnb_optimizer = str2optimizers[optim_name][1]([p2]) if gtype == torch.float32: - atol, rtol = 2e-6, 1e-5 + atol, rtol = 1e-6, 1e-5 else: atol, rtol = 1e-4, 1e-3 - for i in range(50): + for i in range(k): g = torch.randn(dim1,dim2, device='cuda', dtype=gtype)*0.01 p1.grad = g.clone().float() p2.grad = g.clone() @@ -107,7 +98,7 @@ def test_optimizer32bit(dim1, dim2, gtype, optim_name): torch.testing.assert_allclose(p1, p2.float(), atol=atol, rtol=rtol) - if i % 10 == 0 and i > 0: + if i % (k//5) == 0 and i > 0: path = get_temp_dir() torch.save(bnb_optimizer.state_dict(),join(path, 'opt.pt')) del bnb_optimizer @@ -148,7 +139,6 @@ def test_global_config(dim1, dim2, gtype): eps = 1e-8 bnb.optim.GlobalOptimManager.get_instance().initialize() - bnb.optim.GlobalOptimManager.get_instance().override_config(p2, 'skip_zeros', True) bnb.optim.GlobalOptimManager.get_instance().override_config(p3, 'optim_bits', 8) bnb.optim.GlobalOptimManager.get_instance().register_parameters([p1, p2, p3]) @@ -163,8 +153,6 @@ def test_global_config(dim1, dim2, gtype): else: atol, rtol = 1e-4, 1e-3 - original_p2 = p2[mask].clone() - for i in range(50): g1 = torch.randn(dim1,dim2, device='cuda', dtype=gtype)*0.1 + 0.001 g2 = torch.randn(dim1,dim2, device='cuda', dtype=gtype)*0.1 + 0.001 @@ -173,38 +161,17 @@ def test_global_config(dim1, dim2, gtype): p2.grad = g2 p3.grad = g3 - if i > 30 and i % 10 == 0: - g1.data[mask] = 0.0 - g2.data[mask] = 0.0 - p1.grad = g1 - p2.grad = g2 - original_p1 = p1[mask].clone() - original_p2 = p2[mask].clone() - og_s1 = adam2.state[p2]['state1'][mask].clone() - og_s2 = adam2.state[p2]['state2'][mask].clone() - og_s11 = adam2.state[p1]['state1'][mask].clone() - og_s21 = adam2.state[p1]['state2'][mask].clone() - adam2.step() assert adam2.state[p3]['state1'].dtype == torch.uint8 assert adam2.state[p3]['state2'].dtype == torch.uint8 - if i > 30 and i % 10 == 0: - torch.testing.assert_allclose(original_p2, p2[mask]) - torch.testing.assert_allclose(adam2.state[p2]['state1'][mask], og_s1) - torch.testing.assert_allclose(adam2.state[p2]['state2'][mask], og_s2) - assert ((p1[mask]- original_p1)==0.0).sum() < p1.numel() - assert ((adam2.state[p1]['state1'][mask]- og_s11)==0.0).sum() == 0.0 - assert ((adam2.state[p1]['state2'][mask]- og_s21)==0.0).sum() == 0.0 - - dim1 = [1024] dim2 = [32, 1024, 4097] gtype = [torch.float32, torch.float16] -optimizer_names = ['adam8bit', 'momentum8bit', 'rmsprop8bit', 'adam8bit_blockwise', 'adamw8bit_blockwise', 'lamb8bit', 'lars8bit', 'momentum8bit_blockwise', 'rmsprop8bit_blockwise', 'adagrad8bit_blockwise'] +optimizer_names = ['adam8bit', 'momentum8bit', 'rmsprop8bit', 'adam8bit_blockwise', 'lamb8bit', 'lars8bit', 'momentum8bit_blockwise', 'rmsprop8bit_blockwise'] values = list(product(dim1,dim2, gtype, optimizer_names)) names = ['dim1_{0}_dim2_{1}_gtype_{2}_optim_{3}'.format(*vals) for vals in values] @pytest.mark.parametrize("dim1, dim2, gtype, optim_name", values, ids=names) @@ -370,13 +337,12 @@ def test_benchmark_blockwise(dim1, dim2, gtype, optim_name): if dim1 == 1 and dim2 == 1: return p1 = torch.randn(dim1,dim2, device='cuda', dtype=gtype)*0.1 - bnb_optimizer = str2optimizers[optim_name][1]([p1]) g = torch.randn(dim1,dim2, device='cuda', dtype=gtype)*0.01 p1.grad = g - for i in range(5000): - if i == 500: + for i in range(k): + if i == k//5: # 100 iterations for burn-in torch.cuda.synchronize() t0 = time.time() @@ -386,23 +352,8 @@ def test_benchmark_blockwise(dim1, dim2, gtype, optim_name): torch.cuda.synchronize() s = time.time()-t0 print('') - params = 4500*4096*4096 + params = (k-k//5)*dim1*dim2 print(optim_name, gtype, s/params) #assert s < 3.9 - -def test_str_betas(): - betas = (0.80, 0.95) - strbetas = '(0.80, 0.95)' - - layer = torch.nn.Linear(10, 10) - - base = bnb.optim.Adam(layer.parameters(), betas=betas) - strbase = bnb.optim.Adam(layer.parameters(), betas=strbetas) - assert base.defaults['betas'][0] == 0.8 - assert base.defaults['betas'][1] == 0.95 - assert strbase.defaults['betas'][0] == 0.8 - assert strbase.defaults['betas'][1] == 0.95 - - -- cgit v1.2.3