# Owner(s): ["module: inductor"] import copy import functools import os import unittest from typing import Tuple import torch from torch import nn, Tensor from torch._dynamo.convert_frame import maybe_cprofile from torch._dynamo.test_case import run_tests, TestCase from torch._dynamo.testing import rand_strided, reduce_to_scalar_loss from torch._inductor import config, ir, metrics from torch._inductor.fx_passes import pad_mm as pad_mm_pass from torch._inductor.runtime.benchmarking import benchmarker from torch._inductor.utils import ceildiv, run_and_get_code from torch.testing._internal.common_utils import ( instantiate_parametrized_tests, parametrize, requires_cuda, serialTest, ) from torch.testing._internal.inductor_utils import HAS_CUDA DO_PERF_TEST = os.environ.get("DO_PERF_TEST") == "1" DO_ACC_TEST = os.environ.get("DO_ACC_TEST", "1") == "1" WITH_STACK = os.environ.get("WITH_STACK") == "1" USE_CUDA_GRAPHS = os.environ.get("USE_CUDA_GRAPHS", "1") == "1" try: import transformers # noqa: F401 HAS_TRANSFORMER = True except ImportError: HAS_TRANSFORMER = False def get_optim(m): return torch.optim.Adam(m.parameters(), lr=0.01, capturable=True, foreach=True) def gen_transformer_inputs(vocab_size, bs, seq_length): def geninp(): return torch.randint( 0, vocab_size, (bs, seq_length), dtype=torch.int64, requires_grad=False ) input_dict = {"input_ids": geninp(), "labels": geninp()} return input_dict class LinearAndSoftmax(nn.Module): """ It's very common that a transformer model will do a matmul and then softmax/log_softmax in the end. Creating this toy model to capture the pattern and make sure we do proper padding. """ def __init__(self, vocab_size=30523, bias=True): """ The default vocab size for BertForMaskedLM is 30522. We run a few test cases with good or bad vocab_size around Bert's default value. """ super().__init__() self.vocab_size = vocab_size self.linear = nn.Linear(768, vocab_size, bias=bias) self.ce = nn.CrossEntropyLoss() def forward(self, x, label): x = self.linear(x) return self.ce(x.view(-1, self.vocab_size), label.view(-1)) def get_example_inputs(self, batch_size=16): return torch.randn(batch_size, 512, 768), torch.randint( 0, self.vocab_size, (batch_size, 512) ) def forward_and_backward_pass(m, inputs): m(*inputs).sum().backward() @config.patch( { "benchmark_kernel": True, "triton.unique_kernel_names": True, "triton.cudagraphs": USE_CUDA_GRAPHS, } ) @requires_cuda class TestCaseBase(TestCase): @classmethod def setUpClass(cls): if HAS_CUDA: cls.prior_float32_matmul_precision = torch.get_float32_matmul_precision() cls.prior_default_device = torch.get_default_device() torch.set_float32_matmul_precision("high") torch.set_default_device("cuda") @classmethod def tearDownClass(cls): if HAS_CUDA: torch.set_float32_matmul_precision(cls.prior_float32_matmul_precision) torch.set_default_device(cls.prior_default_device) cls.prior_float32_matmul_precision = None cls.prior_default_device = None def check_close(self, ref, act, tol=1e-3): if type(ref).__name__ == "LongformerMaskedLMOutput": ref = ref.loss act = act.loss if type(ref).__name__ == "SequenceClassifierOutput": ref = ref.logits act = act.logits if isinstance(ref, dict) and "loss" in ref: ref = ref["loss"] act = act["loss"] self.assertTrue( torch.allclose(ref, act, atol=tol, rtol=tol), f"ref:\n{ref}\nact:\n{act}" ) def common_numeric_check(self, f, *args, tol=1e-3, **kwargs): ref = f(*args, **kwargs) opt_f = torch.compile(f) act = opt_f(*args, **kwargs) self.check_close(ref, act, tol) def do_profiling( self, f_lhs, f_rhs, tag_lhs="With padding", tag_rhs="Without padding", args=(), kwargs=None, ): if kwargs is None: kwargs = {} torch.cuda.synchronize() with torch.profiler.profile(with_stack=WITH_STACK) as p: niter = 3 for _ in range(niter): with torch.profiler.record_function(tag_lhs): f_lhs(*args, **kwargs) with torch.profiler.record_function(tag_rhs): f_rhs(*args, **kwargs) torch.cuda.synchronize() profile_path = "/tmp/chrome.json" p.export_chrome_trace(profile_path) print(f"Chrome trace is written to {profile_path}") class PerfTestBetweenGoodAndBadShape(TestCaseBase): @unittest.skipIf(not DO_PERF_TEST, "Perf test not enabled") def test_nobias_LinearAndSoftmax_both_shapes(self): self.test_LinearAndSoftmax_both_shapes(bias=False) @unittest.skipIf(not DO_PERF_TEST, "Perf test not enabled") def test_LinearAndSoftmax_both_shapes(self, bias=True): """ Compare the perf with good and bad shape. """ m_bad_shape = LinearAndSoftmax(vocab_size=30523, bias=bias) inptus_bad_shape = m_bad_shape.get_example_inputs() m_good_shape = LinearAndSoftmax(vocab_size=30528, bias=bias) inputs_good_shape = m_good_shape.get_example_inputs() m_bad_shape_opt = torch.compile(m_bad_shape) m_good_shape_opt = torch.compile(m_good_shape) latency_good_shape = benchmarker.benchmark_gpu( lambda: forward_and_backward_pass(m_good_shape_opt, inputs_good_shape) ) latency_bad_shape = benchmarker.benchmark_gpu( lambda: forward_and_backward_pass(m_bad_shape_opt, inptus_bad_shape) ) print( f"Latency for good shape v.s. bad shape: {latency_good_shape:.3f}ms v.s. {latency_bad_shape:.3f}ms" ) @unittest.skipIf(not DO_PERF_TEST or not HAS_TRANSFORMER, "Perf test not enabled") def test_BertForMaskedLM(self, num_layers=1): """ Compare the perf between doing padding and good shape. """ from transformers import BertForMaskedLM config_cls = BertForMaskedLM.config_class bs = 16 seq_length = 512 def create_model(vocab_size): config = config_cls() config.num_hidden_layers = num_layers config.vocab_size = vocab_size inputs = gen_transformer_inputs(config.vocab_size, bs, seq_length) model = BertForMaskedLM(config) optim = get_optim(model) def f(**inputs): optim.zero_grad(True) with torch.cuda.amp.autocast(): pred = model(**inputs) loss = pred[0] loss.backward() optim.step() return torch.compile(f), inputs f_good_shape, inputs_good_shape = create_model(30528) f_bad_shape, inputs_bad_shape = create_model(30522) print("benchmark for good shape") latency_good_shape = benchmarker.benchmark_gpu( lambda: f_good_shape(**inputs_good_shape) ) print("benchmark for bad shape") latency_bad_shape = benchmarker.benchmark_gpu( lambda: f_bad_shape(**inputs_bad_shape) ) print( f"Latency with good and bad shape: {latency_good_shape:.3f} v.s. {latency_bad_shape:.3f}" ) self.do_profiling( lambda: f_good_shape(**inputs_good_shape), lambda: f_bad_shape(**inputs_bad_shape), tag_lhs="With good shape", tag_rhs="With bad shape", ) class PerfTestWithAndWithoutPadding(TestCaseBase): @maybe_cprofile def run_acc_and_perf_test(self, model, inputs, perf_inputs=None, tol=1e-3): """ Run accuracy test. Also compare the perf with and without the comprehensive padding if DO_PERF_TEST is true. """ if perf_inputs is None: perf_inputs = inputs def _process_inputs(x): """ return args and kwargs """ if isinstance(x, dict): return [], x if not isinstance(inputs, (tuple, list)): x = [x] return x, {} args, kwargs = _process_inputs(inputs) perf_args, perf_kwargs = _process_inputs(perf_inputs) if DO_ACC_TEST: model.eval() self.common_numeric_check(model, *args, **kwargs, tol=tol) else: print("Accuracy test skipped") model.train() if DO_PERF_TEST: print("Do performance test") def get_f(m, optim): def f(*args, **kwargs): optim.zero_grad(True) with torch.cuda.amp.autocast(): pred = m(*args, **kwargs) loss = reduce_to_scalar_loss(pred) loss.backward() optim.step() return f latency_with_padding = None print("Benchmark with padding") with config.patch(comprehensive_padding=True): m_copy_with_padding = copy.deepcopy(model) optim_with_padding = get_optim(m_copy_with_padding) opt_f_with_padding = torch.compile( get_f(m_copy_with_padding, optim_with_padding) ) latency_with_padding = benchmarker.benchmark_gpu( lambda: opt_f_with_padding(*perf_args, **perf_kwargs) ) latency_without_padding = None print("bencmark without padding") with config.patch(comprehensive_padding=False): m_copy_without_padding = copy.deepcopy(model) optim_without_padding = get_optim(m_copy_without_padding) opt_f_without_padding = torch.compile( get_f(m_copy_without_padding, optim_without_padding) ) latency_without_padding = benchmarker.benchmark_gpu( lambda: opt_f_without_padding(*perf_args, **perf_kwargs) ) print( f"Latency with and without padding: {latency_with_padding:.3f} v.s. {latency_without_padding:.3f}" ) # profiling self.do_profiling( opt_f_with_padding, opt_f_without_padding, args=perf_args, kwargs=perf_kwargs, ) def test_nvidia_deeprecommender(self): """ Compared the perf with and without comprehensive padding. """ layer_sizes = [197951, 512, 512, 1024, 512, 512, 197951] x = torch.randn(4, layer_sizes[0]) class Model(nn.Module): def __init__(self) -> None: super().__init__() mod_list = [] for i in range(len(layer_sizes) - 1): mod_list.append(nn.Linear(layer_sizes[i], layer_sizes[i + 1])) mod_list.append(nn.SELU()) if i == 2: mod_list.append(nn.Dropout(0.8)) self.seq = nn.Sequential(*mod_list) def forward(self, x): return self.seq(x) m = Model() perf_inputs = torch.randn(256, layer_sizes[0]) self.run_acc_and_perf_test(m, x, perf_inputs) @unittest.skipIf(not DO_PERF_TEST or not HAS_TRANSFORMER, "Perf test not enabled") def test_longformer(self, bs=4): from transformers import AutoConfig, AutoModelForMaskedLM config = AutoConfig.from_pretrained("allenai/longformer-base-4096") model = AutoModelForMaskedLM.from_config(config) vocab_size = model.config.vocab_size seq_length = 1024 input_dict = gen_transformer_inputs(vocab_size, bs, seq_length) self.run_acc_and_perf_test(model, input_dict) @unittest.skipIf(not DO_PERF_TEST or not HAS_TRANSFORMER, "Perf test not enabled") def test_longformer_small_bs(self): """ The model exists in both HF and TB. In TB it uses a samller batch size. """ self.test_longformer(bs=2) @instantiate_parametrized_tests class PaddingTest(TestCaseBase): @unittest.skipIf(not DO_PERF_TEST, "Perf test not enabled") def test_mm_padding_perf(self): def naive_mm(a, b): return a @ b def _compute_padding(s, align): return (s + align - 1) // align * align - s @torch.compile def pad_mm(a, b, align=16): """ NOTE: this function only pad a single dimension which is good enough for testing. """ m_padding = _compute_padding(a.size(0), align) k_padding = _compute_padding(a.size(1), align) n_padding = _compute_padding(b.size(1), align) return pad_mm_pass.pad_mm(a, b, m_padding, k_padding, n_padding) for M, K, N, f in ( (8192, 768, 30523, naive_mm), (8192, 768, 30523, pad_mm), (8192, 768, 30528, naive_mm), (30523, 8192, 768, naive_mm), (30528, 8192, 768, naive_mm), ): a = torch.randn(M, K) b = torch.randn(K, N) ms = benchmarker.benchmark_gpu(lambda: f(a, b)) print(f"MxKxN {M}x{K}x{N} {f.__name__}: {ms:.3f}ms") @unittest.skipIf(not DO_PERF_TEST, "Perf test not enabled") def test_padmm(self): """ Latency between origional matmul and padded matmul: 2.717 v.s. 2.356 """ mat1_pad = torch.randn(8192, 30522, dtype=torch.float16) mat2_pad = torch.randn(30522, 768, dtype=torch.float16) def f(): return mat1_pad @ mat2_pad def pad_dim(x: Tensor, padded_length: int, dim: int) -> Tensor: pad = x.new_zeros(*x.shape[:dim], padded_length, *x.shape[dim + 1 :]) return torch.cat([x, pad], dim=dim) @torch.compile(fullgraph=True, options={"triton.cudagraphs": False}) def g(): mat1 = mat1_pad mat2 = mat2_pad mat1 = pad_dim(mat1, 6, 1) mat2 = pad_dim(mat2, 6, 0) return torch.ops.aten.mm(mat1, mat2) ori_time = benchmarker.benchmark_gpu(f) pad_time = benchmarker.benchmark_gpu(g) print( f"Latency between origional matmul and padded matmul: {ori_time:.3f} v.s. {pad_time:.3f}" ) self.do_profiling(f, g, "No MM Padding", "With mm padding") @unittest.skipIf(not DO_PERF_TEST, "Perf test not enabled") def test_matmul(self): """ Latency with good and bad shapes: 1.705 v.s. 2.625 """ x_good_shape = torch.randn(8192, 30528, dtype=torch.float16) weight_good_shape = torch.randn(30528, 768, dtype=torch.float16) out_good_shape = torch.randn(8192, 768, dtype=torch.float16) # Using stride (30522, 1) does not make a difference here. x_bad_shape = rand_strided( (8192, 30522), (30528, 1), device="cuda", dtype=torch.float16 ) weight_bad_shape = torch.randn(30522, 768, dtype=torch.float16) out_bad_shape = torch.randn(8192, 768, dtype=torch.float16) def f(x, weight, out): torch.mm(x, weight, out=out) return out f1 = torch.compile( functools.partial(f, x_good_shape, weight_good_shape, out_good_shape) ) f2 = torch.compile( functools.partial(f, x_bad_shape, weight_bad_shape, out_bad_shape) ) latency_good_shape = benchmarker.benchmark_gpu(f1) latency_bad_shape = benchmarker.benchmark_gpu(f2) print( f"Latency with good and bad shapes: {latency_good_shape:.3f} v.s. {latency_bad_shape:.3f}" ) self.do_profiling(f1, f2) @serialTest() def test_nobias_LinearAndSoftmax_codegen(self): self.test_LinearAndSoftmax_codegen(bias=False) def test_LinearAndSoftmax_codegen(self, bias=True): m_bad_shape = LinearAndSoftmax(vocab_size=30523, bias=bias) inputs_bad_shape = m_bad_shape.get_example_inputs() m_bad_shape_opt = torch.compile(copy.deepcopy(m_bad_shape)) _, wrapper_codes = run_and_get_code( forward_and_backward_pass, m_bad_shape_opt, inputs_bad_shape ) forward_and_backward_pass(m_bad_shape, inputs_bad_shape) self.assertEqual( m_bad_shape.linear.weight.grad, m_bad_shape_opt.linear.weight.grad ) self.assertTrue(len(wrapper_codes) == 2) # one for forward and oen for backward forward_wrapper = wrapper_codes[0] # make sure the load for softmax is aligned self.assertTrue( "tl.load(in_ptr0 + (r1 + (30528*x0))" in forward_wrapper, f"forward_wrapper: {forward_wrapper}", ) if DO_PERF_TEST: latency = benchmarker.benchmark_gpu( lambda: forward_and_backward_pass(m_bad_shape_opt, inputs_bad_shape) ) print(f"latency: {latency:.3f}ms") @config.patch(pattern_matcher=False) def test_attention(self): batch_size, seq_len, num_heads, hidden_size = 1, 4, 1, 16 inv_scale = (num_heads / hidden_size) ** 0.5 class Attention(nn.Module): def __init__(self) -> None: super().__init__() self.query = nn.Linear(hidden_size, hidden_size) self.key = nn.Linear(hidden_size, hidden_size) self.value = nn.Linear(hidden_size, hidden_size) @staticmethod def reshape(x): return x.view(batch_size, seq_len, num_heads, -1).permute(0, 2, 1, 3) @staticmethod def cancel_reshape(x): return x.permute(0, 2, 1, 3).view(batch_size, seq_len, hidden_size) def forward(self, x): query, key, value = self.query(x), self.key(x), self.value(x) weights = ( torch.matmul( self.reshape(query), self.reshape(key).permute(0, 1, 3, 2) ) * inv_scale ).softmax(dim=-1) return self.cancel_reshape(torch.matmul(weights, self.reshape(value))) attn = Attention() x = torch.randn(batch_size, seq_len, hidden_size) self.common_numeric_check(attn, x) def test_view(self): def f(x): return x.view(3, 3, 3) x = torch.randn(3, 9) self.common_numeric_check(f, x) def test_pad_strides(self): """ Note that dim0's stride is also padded even though its previous value is already multiple of 16. The reason is we padded dim1's stride. We have to correspondingly increase the stride for dim0. """ sizes = [2, 16, 2047] in_strides = [2047 * 16, 2047, 1] out_strides = list(ir.Layout._pad_strides(in_strides, sizes, torch.float32)) expected_strides = [2048 * 16, 2048, 1] self.assertEqual( expected_strides, out_strides, f"{expected_strides} v.s. {out_strides}" ) def test_pad_strides_skip(self): """ The padding is skipped to avoid too much memory overhead. """ sizes = [2, 32, 127] in_strides = [4064, 127, 1] out_strides = list(ir.Layout._pad_strides(in_strides, sizes, torch.float32)) expected_strides = [4064, 127, 1] self.assertEqual( expected_strides, out_strides, f"{expected_strides} v.s. {out_strides}" ) def test_pad_3d_tensor(self): """ Constructing this test case guided by the fact that we don't pad placeholder or user visible output's strides. Add a matmul in the beginning and end so we can pad strides for intermediate tensors. """ def f(x, y): x = torch.matmul(x, y) x = x + 1 return torch.matmul(x, y) x = torch.randn(2, 16, 2047) y = torch.randn(2047, 2047) self.common_numeric_check(f, x, y, tol=1e-2) self.assertTrue(metrics.num_comprehensive_padding > 0) def test_conv(self): """ Padding the input for convolution may cause extra copy kernel being called. Check this example trace: https://gist.github.com/shunting314/ce45398f7d51a63ce05fc8d411faddb3 """ x_shape = (1, 128, 640, 959) x1 = torch.randn(*x_shape) padded_stride = ir.Layout._pad_strides(x1.stride(), x1.shape, torch.float32) x2 = rand_strided(x_shape, padded_stride, device="cuda") x2.copy_(x1) weight = torch.randn(64, 128, 3, 3) def fun(x, weight): return torch.convolution( x, weight, stride=(1, 1), padding=(1, 1), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1, bias=None, ) ref = fun(x1, weight) act = fun(x2, weight) self.check_close(ref, act) if DO_PERF_TEST: latency_with_padding = benchmarker.benchmark_gpu(lambda: fun(x2, weight)) latency_without_padding = benchmarker.benchmark_gpu(lambda: fun(x1, weight)) print( f"Latency with and without padding: {latency_with_padding:.3f} v.s. {latency_without_padding:.3f}" ) self.do_profiling(lambda: fun(x2, weight), lambda: fun(x1, weight)) @unittest.skipIf(not DO_PERF_TEST, "Perf test not enabled") def test_cat(self): """ Compare the perf between aten cat and compiled cat. Latency between eager and compiled: 1.596 v.s. 0.601 Eager cat can be 2.66x slower than inductor kernel. """ x = torch.randn(8192, 30522, dtype=torch.float16) def f(x): pad = x.new_zeros(x.size(0), 6) return torch.cat([x, pad], dim=1) # disable cudagraphs since cudagraphs need copy the input which # distort the latency a lot! (double the latency here for compiled # version) with config.patch("triton.cudagraphs", False): opt_f = torch.compile(f) opt_f(x) eager_time = benchmarker.benchmark_gpu(lambda: f(x)) opt_time = benchmarker.benchmark_gpu(lambda: opt_f(x)) print( f"Latency between eager and compiled: {eager_time:.3f} v.s. {opt_time:.3f}" ) self.do_profiling(lambda: f(x), lambda: opt_f(x), "Eager Cat", "Compiled Cat") def test_pad_channels_last(self): t = torch.randn(2, 3, 5, 1025) in_strides = t.stride() out_strides = ir.Layout._pad_strides(in_strides, t.shape, torch.float32) self.assertTrue(in_strides != out_strides) t = t.to(memory_format=torch.channels_last) in_strides = t.stride() out_strides = ir.Layout._pad_strides(in_strides, t.shape, torch.float32) self.assertTrue(in_strides == out_strides) @parametrize("alignment_bytes", (32, 128)) @parametrize("shape", [(21, 19), (3, 5, 71)]) @parametrize("dtype", (torch.float16, torch.float32)) def test_pad_outputs( self, dtype: torch.dtype, shape: Tuple[int], alignment_bytes: int ): """ Tests padding output tensors to a specific alignment. This is enabled by a config flag. """ func = torch.add inputs = tuple(torch.randn(*shape, dtype=dtype) for input_idx in range(2)) # Compile and run with config.patch( { "comprehensive_padding": True, "padding_alignment_bytes": alignment_bytes, "padding_stride_threshold": 0, "pad_outputs": True, } ): compiled_func = torch.compile(func) compiled_out = compiled_func(*inputs) # Check numerics eager_out = func(*inputs) self.check_close(eager_out, compiled_out) # Compute the expected padding element_size = torch.tensor([], dtype=dtype).element_size() self.assertGreater(alignment_bytes, element_size) self.assertEqual(alignment_bytes % element_size, 0) alignment_elements = alignment_bytes // element_size contiguous_stride = inputs[0].stride() expected_stride = [1] for dim in reversed(shape[1:]): slice_size = dim * expected_stride[0] new_stride = alignment_elements * ceildiv(slice_size, alignment_elements) expected_stride.insert(0, new_stride) expected_stride = tuple(expected_stride) self.assertNotEqual(expected_stride, contiguous_stride) # Check strides self.assertFalse(compiled_out.is_contiguous()) self.assertEqual(compiled_out.stride(), expected_stride) if __name__ == "__main__": if HAS_CUDA: run_tests()