majia-yu opened a new issue #10455: Bug of group2ctxs for model parallelism URL: https://github.com/apache/incubator-mxnet/issues/10455 ## Description I use mx.mod.Module with group2ctxs to split a very larger softmax to multiple gpus. When I use less than 8 gpus, everything is fine, the input_grads from softmax are almost always same (for the same initilization and inputs), but when I use 8 gpus, the input_grads suddenly changes to a wrong value. I believe it's a bug. (A minimum code is provided below) ## Steps to reproduce Please run the code below. The result from my computer is: python magic.py --group2ctxs 1 --num_gpus 8 loss: 22.12132263183593750000000000000000 input_grads: 0.00453650718554854393005371093750 # ??? python magic.py --group2ctxs 1 --num_gpus 7 loss: 22.12132072448730468750000000000000 input_grads: 0.00030235946178436279296875000000 python magic.py --group2ctxs 1 --num_gpus 4 loss: 22.12132072448730468750000000000000 input_grads: 0.00030235989834181964397430419922 python magic.py --group2ctxs 1 --num_gpus 1 loss: 22.12132072448730468750000000000000 input_grads: 0.0003023596364073455333709716796 python magic.py --group2ctxs 0 --num_gpus 1 loss: 22.12132072448730468750000000000000 input_grads: 0.00030235992744565010070800781250 ``` import mxnet as mx import math from easydict import EasyDict import numpy as np import argparse parser = argparse.ArgumentParser() parser.add_argument('--num_gpus', type=int) parser.add_argument('--group2ctxs', type=int) args = parser.parse_args() np.random.seed(0) def softmax(input_shape, num_classes, gpus, s=64.0, m=0.5): embeddings = mx.sym.Variable(name='data', shape=input_shape) * s labels = mx.symbol.Variable('softmax_label') num_classes_gpu = [num_classes // len(gpus)] * len(gpus) num_classes_gpu[0] += num_classes - sum(num_classes_gpu) embedding_size = input_shape[1] numerators = [] numerators_index = [] numerators_mask = [] denominators = [] logits = [] for i, gpu in enumerate(gpus): with mx.AttrScope(ctx_group='dev_%d' % i): local_weights = mx.symbol.Variable( "softmax_%d_weight" % i, shape=(num_classes_gpu[i], embedding_size), lr_mult=1.0) local_weights = mx.symbol.L2Normalization(local_weights, mode='instance') local_logits = mx.sym.FullyConnected(data=embeddings, weight=local_weights, no_bias=True, num_hidden=num_classes_gpu[i]) local_numerators_mask = (labels >= sum(num_classes_gpu[: i])) * (labels < sum(num_classes_gpu[: i + 1])) local_labels = mx.sym.where(local_numerators_mask, labels - sum(num_classes_gpu[: i]), mx.sym.zeros_like(labels)) local_numerators_index = mx.sym.stack(mx.sym.arange(0, input_shape[0]), local_labels, axis=0) cos_t = mx.sym.pick(local_logits, local_labels, axis=1) / s sin_t = mx.sym.sqrt(1 - mx.sym.square(cos_t)) cos_m_t = cos_t * math.cos(m) - sin_t * math.sin(m) cos_m_t = mx.sym.where(cos_t > math.cos(math.pi - m), cos_m_t, cos_t - math.sin(math.pi - m) * m) cos_m_t_s = mx.sym.where(local_numerators_mask, cos_m_t * s, cos_t * s) local_numerators = mx.sym.where(local_numerators_mask, mx.sym.exp(cos_m_t_s), mx.sym.zeros_like(cos_m_t)) local_logits = local_logits + mx.sym.scatter_nd( cos_m_t_s - cos_t * s, local_numerators_index, (input_shape[0], num_classes_gpu[i])) local_denominators = mx.sym.sum(mx.sym.exp(local_logits), axis=1) numerators.append(local_numerators) numerators_index.append(local_numerators_index) numerators_mask.append(local_numerators_mask) denominators.append(local_denominators) logits.append(local_logits) numerators = mx.sym.add_n(*numerators) denominators = mx.sym.add_n(*denominators) fake_loss = [] for i, gpu in enumerate(gpus): with mx.AttrScope(ctx_group='dev_%d' % i): local_logits = logits[i] local_numerators_mask = numerators_mask[i] local_numerators_index = numerators_index[i] local_logits_grad = mx.sym.broadcast_div(mx.sym.exp(local_logits), mx.sym.expand_dims(denominators, axis=1)) local_logits_grad = local_logits_grad - mx.sym.scatter_nd( local_numerators_mask, local_numerators_index, (input_shape[0], num_classes_gpu[i])) local_logits_grad = mx.symbol.BlockGrad(local_logits_grad) local_fake_loss = mx.sym.sum(local_logits_grad * local_logits) / float(input_shape[0]) fake_loss.append(local_fake_loss) fake_loss = mx.sym.add_n(*fake_loss) return mx.sym.make_loss(fake_loss) gpus = list(range(args.num_gpus)) graph = mx.mod.Module( context=[mx.cpu(0)], symbol=softmax((512, 256), 500, gpus), group2ctxs={'dev_%d' % i:[mx.gpu(gpu)] for i, gpu in enumerate(gpus)} if args.group2ctxs else None ) graph.bind(data_shapes=[('data', (512, 256))], label_shapes=[('softmax_label',(512,))], inputs_need_grad=True) initializer = mx.init.Constant(1) graph.init_params(initializer=initializer, arg_params=None, aux_params=None) optimizer = mx.optimizer.SGD(learning_rate=0.1, momentum=0.9, wd=0.0005, rescale_grad=1) graph.init_optimizer(optimizer=optimizer) data = mx.nd.array(np.random.rand(512, 256)) data = mx.nd.L2Normalization(data, mode='instance') label = mx.nd.array(np.random.randint(0,500,(512,))) batch=EasyDict() batch.data=[data] batch.label=[label] graph.forward(batch, is_train=True) graph.backward() input_grads = graph.get_input_grads()[0].asnumpy() loss = graph.get_outputs()[0].asnumpy() print('loss: %.32f'%float(loss)) print('input_grads: %.32f'%np.std(input_grads)) ``` ## Environment info (Required) ----------Python Info---------- Version : 3.5.2 Compiler : GCC 5.4.0 20160609 Build : ('default', 'Sep 14 2017 22:51:06') Arch : ('64bit', 'ELF') ------------Pip Info----------- Version : 9.0.3 Directory : /mnt/ficusjordan/dfyu/Projects/SoftmaxFace/python/python3.5_ubuntu16.04/SoftmaxFace/lib/python3.5/site-packages/pip ----------MXNet Info----------- Version : 1.1.0 Directory : /mnt/ficusjordan/dfyu/Projects/SoftmaxFace/python/python3.5_ubuntu16.04/SoftmaxFace/lib/python3.5/site-packages/mxnet Commit Hash : 07a83a0325a3d782513a04f47d711710972cb144 ----------System Info---------- Platform : Linux-4.13.0-36-generic-x86_64-with-Ubuntu-16.04-xenial system : Linux node : WXRG0094 release : 4.13.0-36-generic version : #40~16.04.1-Ubuntu SMP Fri Feb 16 23:25:58 UTC 2018 ----------Hardware Info---------- machine : x86_64 processor : x86_64 Architecture: x86_64 CPU op-mode(s): 32-bit, 64-bit Byte Order: Little Endian CPU(s): 56 On-line CPU(s) list: 0-55 Thread(s) per core: 2 Core(s) per socket: 14 Socket(s): 2 NUMA node(s): 2 Vendor ID: GenuineIntel CPU family: 6 Model: 79 Model name: Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz Stepping: 1 CPU MHz: 2400.215 CPU max MHz: 3300.0000 CPU min MHz: 1200.0000 BogoMIPS: 4800.43 Virtualization: VT-x L1d cache: 32K L1i cache: 32K L2 cache: 256K L3 cache: 35840K NUMA node0 CPU(s): 0-13,28-41 NUMA node1 CPU(s): 14-27,42-55 Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 cdp_l3 invpcid_single pti retpoline intel_ppin intel_pt tpr_shadow vnmi flexpriority ept vpid fsgsbase tsc_adjust bmi1 hle avx2 smep bmi2 erms invpcid rtm cqm rdt_a rdseed adx smap xsaveopt cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local dtherm ida arat pln pts ----------Network Test---------- Setting timeout: 10 Timing for FashionMNIST: https://apache-mxnet.s3-accelerate.dualstack.amazonaws.com/gluon/dataset/fashion-mnist/train-labels-idx1-ubyte.gz, DNS: 0.1864 sec, LOAD: 1.3491 sec. Timing for MXNet: https://github.com/apache/incubator-mxnet, DNS: 0.0119 sec, LOAD: 3.8785 sec. Timing for Conda: https://repo.continuum.io/pkgs/free/, DNS: 0.0096 sec, LOAD: 0.4216 sec. Timing for Gluon Tutorial(cn): https://zh.gluon.ai, DNS: 0.1880 sec, LOAD: 0.8841 sec. Timing for Gluon Tutorial(en): http://gluon.mxnet.io, DNS: 0.2153 sec, LOAD: 0.4721 sec. Timing for PYPI: https://pypi.python.org/pypi/pip, DNS: 0.0118 sec, LOAD: 0.9239 sec.
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
