sxjscience edited a comment on pull request #6699: URL: https://github.com/apache/incubator-tvm/pull/6699#issuecomment-711111989
I've verified the TVM integration with 5 NLP backbones in GluonNLP: BERT, ALBERT, ELECTRA, RoBERTA, and BART ```python import mxnet as mx import numpy as np import gluonnlp from gluonnlp.models import get_backbone import numpy.testing as npt import tvm from tvm import relay import tvm.contrib.graph_runtime as runtime mx.npx.set_np() instance_info = { 'g4': {'target': "cuda -model=t4", 'use_gpu': True}, 'c4': {'target': 'llvm -mcpu=core-avx2 -libs=cblas', 'use_gpu': False}, 'c5': {'target': 'llvm -mcpu=skylake-avx512 -libs=cblas', 'use_gpu': False}, 'p3': {'target': 'cuda -model=v100', 'use_gpu': True} } def test_backbone(model_name, batch_size=2, seq_length=128, instance='g4', required_pass=None, opt_level=3): if required_pass is None: required_pass = ["FastMath"] model_cls, cfg, tokenizer, backbone_param_path, _ = get_backbone(model_name) model = model_cls.from_cfg(cfg) model.load_parameters(backbone_param_path) model.hybridize() token_ids = mx.np.random.randint(0, cfg.MODEL.vocab_size, (batch_size, seq_length), dtype=np.int32) token_types = mx.np.random.randint(0, 2, (batch_size, seq_length), dtype=np.int32) valid_length = mx.np.random.randint(seq_length // 2, seq_length, (batch_size,), dtype=np.int32) if 'bart' in model_name: mx_out = model(token_ids, valid_length, token_ids, valid_length) shape_dict = { 'data0': token_ids.shape, 'data1': valid_length.shape, 'data2': token_ids.shape, 'data3': valid_length.shape, } dtype_dict = { 'data0': token_ids.dtype.name, 'data1': valid_length.dtype.name, 'data2': token_ids.dtype.name, 'data3': valid_length.dtype.name, } elif 'roberta' in model_name or 'xlmr' in model_name: mx_out = model(token_ids, valid_length) shape_dict = { 'data0': token_ids.shape, 'data1': valid_length.shape, } dtype_dict = { 'data0': token_ids.dtype.name, 'data1': valid_length.dtype.name, } else: mx_out = model(token_ids, token_types, valid_length) shape_dict = { 'data0': token_ids.shape, 'data1': token_types.shape, 'data2': valid_length.shape } dtype_dict = { 'data0': token_ids.dtype.name, 'data1': token_types.dtype.name, 'data2': valid_length.dtype.name } sym = model._cached_graph[1] params = {} for k, v in model.collect_params().items(): params[v._var_name] = tvm.nd.array(v.data().asnumpy()) mod, params = relay.frontend.from_mxnet(sym, shape=shape_dict, dtype=dtype_dict, arg_params=params) target = instance_info[instance]['target'] use_gpu = instance_info[instance]['use_gpu'] with relay.build_config(opt_level=opt_level, required_pass=required_pass): graph, lib, cparams = relay.build(mod, target, params=params) if use_gpu: ctx = tvm.gpu() else: ctx = tvm.cpu() rt = runtime.create(graph, lib, ctx) rt.set_input(**cparams) if 'bart' in model_name: rt.set_input(data0=token_ids, data1=valid_length, data2=token_ids, data3=valid_length) elif 'roberta' in model_name: rt.set_input(data0=token_ids, data1=valid_length) else: rt.set_input(data0=token_ids, data1=token_types, data2=valid_length) rt.run() for i in range(rt.get_num_outputs()): out = rt.get_output(i) if rt.get_num_outputs() == 1: mx_out_gt = mx_out.asnumpy() else: mx_out_gt = mx_out[i].asnumpy() if 'mobilebert' in model_name and len(out.shape) == 3: npt.assert_allclose(out.asnumpy()[:, 1:, :], mx_out[i].asnumpy()[:, 1:, :], rtol=6e-2, atol=6e-2) else: npt.assert_allclose(out.asnumpy(), mx_out_gt, rtol=6e-2, atol=6e-2) # test_backbone('google_en_cased_bert_base', instance='g4') test_model_names = ['google_albert_base_v2', 'google_en_cased_bert_base', 'google_electra_small', 'fairseq_roberta_base', 'fairseq_bart_base'] for model_name in test_model_names: test_backbone(model_name, instance='g4') ``` ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to 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