SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

[tests] Parameterized test_eager_matches_sdpa_inference (#36650)

Browse Source
This commit is contained in:
Joao Gante
2025-03-14 14:41:27 +00:00
committed by GitHub
parent 9215cc62d4
commit 42ebb6c23e
16 changed files with 285 additions and 1900 deletions
Download Patch File Download Diff File Expand all files Collapse all files

436
tests/test_modeling_common.py
View File

@@ -133,6 +133,23 @@ if is_deepspeed_available():
import deepspeed
# used in other test files e.g. when overwriting the test
TEST_EAGER_MATCHES_SDPA_INFERENCE_PARAMETERIZATION = [
(
# test name for the test runner
f"{dtype}_pad_{padding_side}{'' if use_attention_mask else '_no_attn_mask'}"
f"{'_output_attn' if output_attentions else ''}{'_sdpa_kernels' if enable_kernels else ''}",
# parameterization
*(dtype, padding_side, use_attention_mask, output_attentions, enable_kernels),
)
for dtype in ("fp16", "fp32", "bf16")
for padding_side in ("left", "right")
for use_attention_mask in (True, False)
for output_attentions in (True, False)
for enable_kernels in (True, False)
]
def _config_zero_init(config):
configs_no_init = copy.deepcopy(config)
for key in configs_no_init.__dict__.keys():
@@ -3543,31 +3560,39 @@ class ModelTesterMixin:
):
raise ValueError("The eager model should not have SDPA attention layers")
@parameterized.expand([("float16",), ("bfloat16",), ("float32",)])
@parameterized.expand(TEST_EAGER_MATCHES_SDPA_INFERENCE_PARAMETERIZATION)
@require_torch_sdpa
def test_eager_matches_sdpa_inference(self, torch_dtype: str):
def test_eager_matches_sdpa_inference(
self, name, torch_dtype, padding_side, use_attention_mask, output_attentions, enable_kernels
):
# TODO: we shouldn't need to do this skip, i.e. the test would be composable from the model tester. CLIP-like
# models have a custom mixin, which we detect to skip this test.
if not any(".ModelTesterMixin" in str(base) for base in self.__class__.__bases__):
self.skipTest(reason="CLIP-like models have a different `test_eager_matches_sdpa_inference`")
if not self.has_attentions:
self.skipTest(reason="Model architecture does not support attentions")
if not self.all_model_classes[0]._supports_sdpa:
self.skipTest(f"{self.all_model_classes[0].__name__} does not support SDPA")
if torch_dtype == "float16" and not is_torch_fp16_available_on_device(torch_device):
# convert shorthand name to torch.dtype
if torch_dtype == "fp16":
torch_dtype = torch.float16
elif torch_dtype == "bf16":
torch_dtype = torch.bfloat16
elif torch_dtype == "fp32":
torch_dtype = torch.float32
if not is_torch_fp16_available_on_device(torch_device) and torch_dtype == torch.float16:
self.skipTest(f"float16 not supported on {torch_device} (on the specific device currently used)")
if torch_dtype == "bfloat16" and not is_torch_bf16_available_on_device(torch_device):
if not is_torch_bf16_available_on_device(torch_device) and torch_dtype == torch.bfloat16:
self.skipTest(
f"bfloat16 not supported on {torch_device} (on the specific device currently used, e.g. Nvidia T4 GPU)"
)
# Not sure whether it's fine to put torch.XXX in a decorator if torch is not available so hacking it here instead.
if torch_dtype == "float16":
torch_dtype = torch.float16
elif torch_dtype == "bfloat16":
torch_dtype = torch.bfloat16
elif torch_dtype == "float32":
torch_dtype = torch.float32
# Dictionary of tolerances for eager <> sdpa tests. Key = (device, sdpa_kernels_enabled, dtype)
atols = {
("cpu", False, torch.float32): 1e-6,
("cpu", False, torch.float16): 5e-3,
@@ -3597,238 +3622,243 @@ class ModelTesterMixin:
("cuda", True, torch.float16): 5e-3,
}
def get_mean_reldiff(failcase, x, ref, atol, rtol):
return f"{failcase}: mean relative difference: {((x - ref).abs() / (ref.abs() + 1e-12)).mean():.3e}, torch atol = {atol}, torch rtol = {rtol}"
set_model_tester_for_less_flaky_test(self)
for model_class in self.all_model_classes:
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
set_config_for_less_flaky_test(config)
model = model_class(config)
# FIXME: we deactivate boolean mask for models using "use_mask_token" in their constructors.
# These models support masking only in the case `use_mask_token=True`. Otherwise they cannot consume an input mask.
# This means that the class needs to be instantiated much later, after `use_mask` is set, which means a significant refactor of the code.
# However masking there is not done at any layers that matters (i.e self-attention), therefore we can safely deactivate it.
deactivate_mask = "use_mask_token" in inspect.signature(model_class).parameters
is_encoder_decoder = model.config.is_encoder_decoder
# TODO: standardize the interfaces for musicgen models, see other todo in this test
if model.__class__.__name__ == "MusicgenMelodyForConditionalGeneration":
is_encoder_decoder = True
else:
is_encoder_decoder = model.config.is_encoder_decoder
with tempfile.TemporaryDirectory() as tmpdirname:
model.save_pretrained(tmpdirname)
model_from_pretrained_kwargs = {
"pretrained_model_name_or_path": tmpdirname,
"torch_dtype": torch_dtype,
}
if (
hasattr(config, "use_mask_token")
or "use_mask_token" in inspect.signature(model.__init__).parameters
):
model_from_pretrained_kwargs["use_mask_token"] = True
# TODO: remove this try/except, models should have a shared API
try:
model_sdpa = model_class.from_pretrained(
tmpdirname, torch_dtype=torch_dtype, attn_implementation="sdpa"
**model_from_pretrained_kwargs, attn_implementation="sdpa"
)
except ValueError:
model_sdpa = model_class.from_pretrained(tmpdirname, torch_dtype=torch_dtype)
model_sdpa = model_class.from_pretrained(**model_from_pretrained_kwargs)
model_sdpa = model_sdpa.eval().to(torch_device, dtype=torch_dtype)
model_eager = model_class.from_pretrained(
tmpdirname,
torch_dtype=torch_dtype,
attn_implementation="eager",
)
model_eager = model_class.from_pretrained(**model_from_pretrained_kwargs, attn_implementation="eager")
model_eager = model_eager.eval().to(torch_device, dtype=torch_dtype)
set_model_for_less_flaky_test(model_eager)
set_model_for_less_flaky_test(model_sdpa)
# We use these for loops instead of parameterized.expand just for the interest of avoiding loading/saving 16 times the model,
# but it would be nicer to have an efficient way to use parameterized.expand
fail_cases = []
for padding_side in ["left", "right"]:
for use_mask in [False, True]:
for output_attentions in [True, False]:
can_output_attn = "output_attentions" in inspect.signature(model_sdpa.forward).parameters
if not (self.has_attentions and can_output_attn) and output_attentions:
continue
# TODO: if we can also check with `batch_size=1` without being flaky?
for batch_size in [7]:
dummy_input = inputs_dict[model.main_input_name]
can_output_attn = "output_attentions" in inspect.signature(model_sdpa.forward).parameters
if not (self.has_attentions and can_output_attn) and output_attentions:
self.skipTest(reason="Model does not support output_attentions")
if dummy_input.dtype in [torch.float32, torch.bfloat16, torch.float16]:
dummy_input = dummy_input.to(torch_dtype)
# TODO: if we can also check with `batch_size=1` without being flaky?
for batch_size in [7]:
# musicgen decoder models; TODO: find better abstraction
if hasattr(self.model_tester, "num_codebooks") and not hasattr(model_eager, "text_encoder"):
input_data_batch_size = batch_size * self.model_tester.num_codebooks
else:
input_data_batch_size = batch_size
dummy_input = dummy_input[:batch_size]
if dummy_input.shape[0] != batch_size:
if dummy_input.dtype in [torch.float32, torch.bfloat16, torch.float16]:
extension = torch.rand(
batch_size - dummy_input.shape[0],
*dummy_input.shape[1:],
dtype=torch_dtype,
device=torch_device,
)
dummy_input = torch.cat((dummy_input, extension), dim=0).to(torch_device)
else:
extension = torch.randint(
high=5,
size=(batch_size - dummy_input.shape[0], *dummy_input.shape[1:]),
dtype=dummy_input.dtype,
device=torch_device,
)
dummy_input = torch.cat((dummy_input, extension), dim=0).to(torch_device)
dummy_input = inputs_dict[model.main_input_name]
if not use_mask:
dummy_attention_mask = None
else:
dummy_attention_mask = inputs_dict.get("attention_mask", None)
if dummy_attention_mask is None:
if is_encoder_decoder:
seqlen = inputs_dict.get("decoder_input_ids", dummy_input).shape[-1]
else:
seqlen = dummy_input.shape[-1]
dummy_attention_mask = (
torch.ones(batch_size, seqlen).to(torch.int64).to(torch_device)
)
if dummy_input.dtype in [torch.float32, torch.bfloat16, torch.float16]:
dummy_input = dummy_input.to(torch_dtype)
dummy_attention_mask = dummy_attention_mask[:batch_size]
if dummy_attention_mask.shape[0] != batch_size:
extension = torch.ones(
batch_size - dummy_attention_mask.shape[0],
*dummy_attention_mask.shape[1:],
dtype=dummy_attention_mask.dtype,
device=torch_device,
)
dummy_attention_mask = torch.cat((dummy_attention_mask, extension), dim=0)
dummy_attention_mask = dummy_attention_mask.to(torch_device)
dummy_input = dummy_input[:input_data_batch_size]
if dummy_input.shape[0] != input_data_batch_size:
if dummy_input.dtype in [torch.float32, torch.bfloat16, torch.float16]:
extension = torch.rand(
input_data_batch_size - dummy_input.shape[0],
*dummy_input.shape[1:],
dtype=torch_dtype,
device=torch_device,
)
dummy_input = torch.cat((dummy_input, extension), dim=0).to(torch_device)
else:
extension = torch.randint(
high=5,
size=(input_data_batch_size - dummy_input.shape[0], *dummy_input.shape[1:]),
dtype=dummy_input.dtype,
device=torch_device,
)
dummy_input = torch.cat((dummy_input, extension), dim=0).to(torch_device)
dummy_attention_mask[:] = 1
if padding_side == "left":
dummy_attention_mask[-1, :2] = 0
dummy_attention_mask[-1, 2:] = 1
elif padding_side == "right":
dummy_attention_mask[-1, -2:] = 0
dummy_attention_mask[-1, :-2] = 1
if not use_attention_mask:
dummy_attention_mask = None
else:
dummy_attention_mask = inputs_dict.get("attention_mask", None)
if dummy_attention_mask is None:
if is_encoder_decoder:
seqlen = inputs_dict.get("decoder_input_ids", dummy_input).shape[-1]
else:
seqlen = dummy_input.shape[-1]
dummy_attention_mask = torch.ones(batch_size, seqlen).to(torch.int64).to(torch_device)
for enable_kernels in [False, True]:
failcase = f"padding_side={padding_side}, use_mask={use_mask}, enable_kernels={enable_kernels}"
if is_encoder_decoder:
decoder_input_ids = inputs_dict.get("decoder_input_ids", dummy_input)[
:batch_size
]
if decoder_input_ids.shape[0] != batch_size:
extension = torch.ones(
batch_size - decoder_input_ids.shape[0],
*decoder_input_ids.shape[1:],
dtype=decoder_input_ids.dtype,
device=torch_device,
)
decoder_input_ids = torch.cat((decoder_input_ids, extension), dim=0)
decoder_input_ids = decoder_input_ids.to(torch_device)
dummy_attention_mask = dummy_attention_mask[:batch_size]
if dummy_attention_mask.shape[0] != batch_size:
extension = torch.ones(
batch_size - dummy_attention_mask.shape[0],
*dummy_attention_mask.shape[1:],
dtype=dummy_attention_mask.dtype,
device=torch_device,
)
dummy_attention_mask = torch.cat((dummy_attention_mask, extension), dim=0)
dummy_attention_mask = dummy_attention_mask.to(torch_device)
# TODO: never an `attention_mask` arg here?
processed_inputs = {
model.main_input_name: dummy_input,
"decoder_input_ids": decoder_input_ids,
"decoder_attention_mask": dummy_attention_mask,
"output_hidden_states": True,
}
else:
processed_inputs = {
model.main_input_name: dummy_input,
"output_hidden_states": True,
}
dummy_attention_mask[:] = 1
if padding_side == "left":
dummy_attention_mask[-1, :2] = 0
dummy_attention_mask[-1, 2:] = 1
elif padding_side == "right":
dummy_attention_mask[-1, -2:] = 0
dummy_attention_mask[-1, :-2] = 1
# Otherwise fails for e.g. WhisperEncoderModel
if "attention_mask" in inspect.signature(model_eager.forward).parameters:
processed_inputs["attention_mask"] = dummy_attention_mask
if is_encoder_decoder:
# musicgen encoder-decoder models; TODO: find better abstraction
if hasattr(self.model_tester, "num_codebooks"):
input_data_batch_size = batch_size * self.model_tester.num_codebooks
else:
input_data_batch_size = batch_size
if (
self.has_attentions
and "output_attentions" in inspect.signature(model_sdpa.forward).parameters
):
processed_inputs["output_attentions"] = output_attentions
if not deactivate_mask and (
"bool_masked_pos" in inspect.signature(model_eager.forward).parameters
):
dummy_mask = torch.ones((self.model_tester.num_masks,))
decoder_input_ids = inputs_dict.get("decoder_input_ids", dummy_input)[:input_data_batch_size]
if decoder_input_ids.shape[0] != input_data_batch_size:
extension = torch.ones(
input_data_batch_size - decoder_input_ids.shape[0],
*decoder_input_ids.shape[1:],
dtype=decoder_input_ids.dtype,
device=torch_device,
)
decoder_input_ids = torch.cat((decoder_input_ids, extension), dim=0)
decoder_input_ids = decoder_input_ids.to(torch_device)
# In case of additional token (like class) we define a custom `mask_length`
if hasattr(self.model_tester, "mask_length"):
mask_length = self.model_tester.mask_length - dummy_mask.size(0)
else:
mask_length = self.model_tester.seq_length - dummy_mask.size(0)
dummy_mask = torch.cat([dummy_mask, torch.zeros(mask_length)])
dummy_bool_masked_pos = dummy_mask.expand(batch_size, -1).bool()
processed_inputs["bool_masked_pos"] = dummy_bool_masked_pos.to(torch_device)
# TODO: never an `attention_mask` arg here?
processed_inputs = {
model.main_input_name: dummy_input,
"decoder_input_ids": decoder_input_ids,
"decoder_attention_mask": dummy_attention_mask,
"output_hidden_states": True,
}
else:
processed_inputs = {
model.main_input_name: dummy_input,
"output_hidden_states": True,
}
if "noise" in inspect.signature(model_eager.forward).parameters:
np.random.seed(2)
num_patches = int(
(self.model_tester.image_size // self.model_tester.patch_size) ** 2
)
noise = np.random.uniform(size=(batch_size, num_patches))
processed_inputs["noise"] = torch.from_numpy(noise)
# Otherwise fails for e.g. WhisperEncoderModel
if "attention_mask" in inspect.signature(model_eager.forward).parameters:
processed_inputs["attention_mask"] = dummy_attention_mask
# TODO: test gradients as well (& for FA2 as well!)
with torch.no_grad():
with sdpa_kernel(
enable_flash=enable_kernels,
enable_math=True,
enable_mem_efficient=enable_kernels,
):
prepared_inputs = self._prepare_for_class(processed_inputs, model_class)
outputs_eager = model_eager(**prepared_inputs)
outputs_sdpa = model_sdpa(**prepared_inputs)
if (
self.has_attentions
and "output_attentions" in inspect.signature(model_sdpa.forward).parameters
):
processed_inputs["output_attentions"] = output_attentions
if "bool_masked_pos" in inspect.signature(model_eager.forward).parameters:
dummy_mask = torch.ones((self.model_tester.num_masks,))
if hasattr(outputs_eager, "vision_hidden_states"):
logits_eager = outputs_eager.vision_hidden_states[-1]
logits_sdpa = outputs_sdpa.vision_hidden_states[-1]
else:
logits_eager = (
outputs_eager.hidden_states[-1]
if not is_encoder_decoder
else outputs_eager.decoder_hidden_states[-1]
)
logits_sdpa = (
outputs_sdpa.hidden_states[-1]
if not is_encoder_decoder
else outputs_sdpa.decoder_hidden_states[-1]
)
# In case of additional token (like class) we define a custom `mask_length`
if hasattr(self.model_tester, "mask_length"):
mask_length = self.model_tester.mask_length - dummy_mask.size(0)
else:
mask_length = self.model_tester.seq_length - dummy_mask.size(0)
dummy_mask = torch.cat([dummy_mask, torch.zeros(mask_length)])
dummy_bool_masked_pos = dummy_mask.expand(batch_size, -1).bool()
processed_inputs["bool_masked_pos"] = dummy_bool_masked_pos.to(torch_device)
if torch_device in ["cpu", "cuda"]:
atol = atols[torch_device, enable_kernels, torch_dtype]
rtol = rtols[torch_device, enable_kernels, torch_dtype]
elif torch_device == "xpu":
# As of PyTorch 2.5 XPU backend supports only torch.nn.attention.SDPBackend.MATH
# which is implemented on PyTorch level using aten operators and is
# device agnostic with respect to implementation of each aten operator.
atol = atols["cuda", False, torch_dtype]
rtol = rtols["cuda", False, torch_dtype]
else:
atol = 1e-7
rtol = 1e-4
if "noise" in inspect.signature(model_eager.forward).parameters:
np.random.seed(2)
num_patches = int((self.model_tester.image_size // self.model_tester.patch_size) ** 2)
noise = np.random.uniform(size=(batch_size, num_patches))
processed_inputs["noise"] = torch.from_numpy(noise)
# Masked tokens output slightly deviates - we don't mind that.
if use_mask:
_logits_sdpa = torch.zeros_like(input=logits_sdpa)
_logits_eager = torch.zeros_like(input=logits_eager)
# TODO: test gradients as well (& for FA2 as well!)
with torch.no_grad():
with sdpa_kernel(
enable_flash=enable_kernels,
enable_math=True,
enable_mem_efficient=enable_kernels,
):
prepared_inputs = self._prepare_for_class(processed_inputs, model_class)
outputs_eager = model_eager(**prepared_inputs)
outputs_sdpa = model_sdpa(**prepared_inputs)
_logits_sdpa[:-1] = logits_sdpa[:-1]
_logits_eager[:-1] = logits_eager[:-1]
# TODO: rename logits -> hidden_states
if hasattr(outputs_eager, "vision_hidden_states"):
logits_eager = outputs_eager.vision_hidden_states[-1]
logits_sdpa = outputs_sdpa.vision_hidden_states[-1]
elif hasattr(outputs_eager, "audio_values"):
logits_eager = outputs_eager.audio_values
logits_sdpa = outputs_sdpa.audio_values
else:
logits_eager = (
outputs_eager.decoder_hidden_states[-1]
if hasattr(outputs_eager, "decoder_hidden_states")
else outputs_eager.hidden_states[-1]
)
logits_sdpa = (
outputs_sdpa.decoder_hidden_states[-1]
if hasattr(outputs_sdpa, "decoder_hidden_states")
else outputs_sdpa.hidden_states[-1]
)
if padding_side == "left":
_logits_sdpa[-1:, 2:] = logits_sdpa[-1:, 2:]
_logits_eager[-1:, 2:] = logits_eager[-1:, 2:]
if torch_device in ["cpu", "cuda"]:
atol = atols[torch_device, enable_kernels, torch_dtype]
rtol = rtols[torch_device, enable_kernels, torch_dtype]
elif torch_device == "xpu":
# As of PyTorch 2.5 XPU backend supports only torch.nn.attention.SDPBackend.MATH
# which is implemented on PyTorch level using aten operators and is
# device agnostic with respect to implementation of each aten operator.
atol = atols["cuda", False, torch_dtype]
rtol = rtols["cuda", False, torch_dtype]
else:
atol = 1e-7
rtol = 1e-4
elif padding_side == "right":
_logits_sdpa[-1:, 2:] = logits_sdpa[-1:, :-2]
_logits_eager[-1:, 2:] = logits_eager[-1:, :-2]
# Masked tokens output slightly deviates - we don't mind that.
if use_attention_mask:
_logits_sdpa = torch.zeros_like(input=logits_sdpa)
_logits_eager = torch.zeros_like(input=logits_eager)
logits_sdpa = _logits_sdpa
logits_eager = _logits_eager
_logits_sdpa[:-1] = logits_sdpa[:-1]
_logits_eager[:-1] = logits_eager[:-1]
results = [
torch.allclose(_logits_sdpa, _logits_eager, atol=atol, rtol=rtol)
for (_logits_sdpa, _logits_eager) in zip(logits_sdpa, logits_eager)
]
# If 80% batch elements have matched results, it's fine
if np.mean(results) < 0.8:
fail_cases.append(
get_mean_reldiff(failcase, logits_sdpa, logits_eager, atol, rtol)
)
if padding_side == "left":
_logits_sdpa[-1:, 2:] = logits_sdpa[-1:, 2:]
_logits_eager[-1:, 2:] = logits_eager[-1:, 2:]
self.assertTrue(len(fail_cases) == 0, "\n".join(fail_cases))
elif padding_side == "right":
_logits_sdpa[-1:, 2:] = logits_sdpa[-1:, :-2]
_logits_eager[-1:, 2:] = logits_eager[-1:, :-2]
logits_sdpa = _logits_sdpa
logits_eager = _logits_eager
results = [
torch.allclose(_logits_sdpa, _logits_eager, atol=atol, rtol=rtol)
for (_logits_sdpa, _logits_eager) in zip(logits_sdpa, logits_eager)
]
# If 80% batch elements have matched results, it's fine
if np.mean(results) < 0.8:
mean_relative_diff = ((logits_sdpa - logits_eager).abs() / (logits_eager.abs() + 1e-12)).mean()
raise ValueError(
f"mean relative difference: {mean_relative_diff:.3e}, torch atol = {atol}, torch rtol = "
f"{rtol}"
)
@require_torch_sdpa
@require_torch_gpu