-
Notifications
You must be signed in to change notification settings - Fork 5.5k
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
support ut of flash attention for more cases
- Loading branch information
Showing
4 changed files
with
261 additions
and
51 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,198 @@ | ||
| # Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved. | ||
| # | ||
| # Licensed under the Apache License, Version 2.0 (the "License"); | ||
| # you may not use this file except in compliance with the License. | ||
| # You may obtain a copy of the License at | ||
| # | ||
| # http://www.apache.org/licenses/LICENSE-2.0 | ||
| # | ||
| # Unless required by applicable law or agreed to in writing, software | ||
| # distributed under the License is distributed on an "AS IS" BASIS, | ||
| # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | ||
| # See the License for the specific language governing permissions and | ||
| # limitations under the License. | ||
|
|
||
| import random | ||
| import unittest | ||
|
|
||
| import numpy as np | ||
| from get_test_cover_info import ( | ||
| XPUOpTestWrapper, | ||
| create_test_class, | ||
| get_xpu_op_support_types, | ||
| ) | ||
| from op_test import convert_float_to_uint16 | ||
| from op_test_xpu import XPUOpTest | ||
|
|
||
| import paddle | ||
| import paddle.nn.functional as F | ||
| from paddle.base import core | ||
|
|
||
|
|
||
| def get_triangle_upper_mask(x): | ||
| mask = paddle.full_like(x, -1e4) | ||
| mask.stop_gradient = True | ||
| mask = paddle.triu(mask, diagonal=1) | ||
| mask.stop_gradient = True | ||
| return mask.astype(np.float32) | ||
|
|
||
|
|
||
| def attention_naive(q, k, v, bias, is_causal=True): | ||
| origin_dtype = q.dtype | ||
| assert k.dtype == origin_dtype | ||
| assert v.dtype == origin_dtype | ||
| if q.dtype != paddle.float32: | ||
| q = paddle.cast(q, "float32") | ||
| k = paddle.cast(k, "float32") | ||
| v = paddle.cast(v, "float32") | ||
| # real calculation | ||
| qt = paddle.transpose(q, [0, 2, 1, 3]) | ||
| kt = paddle.transpose(k, [0, 2, 1, 3]) | ||
| vt = paddle.transpose(v, [0, 2, 1, 3]) | ||
| scale = 1.0 / np.sqrt(q.shape[-1]) | ||
| s = paddle.matmul(qt, paddle.transpose(kt, [0, 1, 3, 2])) | ||
| s = paddle.scale(s, scale) | ||
| if bias is not None: | ||
| s = s + bias | ||
| if is_causal: | ||
| mask = get_triangle_upper_mask(s) | ||
| s = s + mask | ||
| softmax_lse = paddle.logsumexp(s, axis=3) | ||
| p = F.softmax(s) | ||
| o = paddle.matmul(p, vt) | ||
| o = paddle.cast(o, np.float32) | ||
| o = paddle.transpose(o, [0, 2, 1, 3]) | ||
| return o, softmax_lse | ||
|
|
||
|
|
||
| def is_flashattn_supported(): | ||
| xpu_version = core.get_xpu_device_version(0) | ||
| if xpu_version != core.XPUVersion.XPU3: | ||
| return False | ||
| xhpc_version = paddle.version.xpu_xhpc() | ||
| if xhpc_version == 'False': | ||
| return False | ||
| return True | ||
|
|
||
|
|
||
| class XPUTestFlashAttentionOp(XPUOpTestWrapper): | ||
| def __init__(self): | ||
| self.op_name = "flash_attn" | ||
| self.use_dynamic_create_class = False | ||
|
|
||
| @unittest.skipIf( | ||
| not is_flashattn_supported(), "only available on XPU3 with XHPC" | ||
| ) | ||
| class TestFlashAttentionOp(XPUOpTest): | ||
| def setUp(self): | ||
| self.op_type = "flash_attn" | ||
| self.tmp_seed = random.randint(1, 65536) | ||
| paddle.seed(self.tmp_seed) | ||
| self.init_dtype() | ||
| self.set_attrs() | ||
| self.set_shape() | ||
| self.init_data() | ||
| self.tolerance = 5e-4 | ||
|
|
||
| def init_dtype(self): | ||
| self.dtype = self.in_type | ||
|
|
||
| def set_shape(self): | ||
| # [b, l, h, d] | ||
| self.q_shape = [1, 128, 2, 32] | ||
| self.k_shape = [1, 128, 2, 32] | ||
| self.v_shape = [1, 128, 2, 32] | ||
| # [b, h, l, l] | ||
| self.bias_shape = [1, 2, 128, 128] | ||
|
|
||
| def set_attrs(self): | ||
| self.is_causal = True | ||
| self.with_bias = False | ||
|
|
||
| def init_data(self): | ||
| q = np.random.random(self.q_shape) | ||
| k = np.random.random(self.k_shape) | ||
| v = np.random.random(self.v_shape) | ||
| q_ = paddle.to_tensor(q, stop_gradient=False) | ||
| k_ = paddle.to_tensor(k, stop_gradient=False) | ||
| v_ = paddle.to_tensor(v, stop_gradient=False) | ||
| # fixed the seed & offset to pass the check of seed_offset | ||
| fixed_seed_offset = paddle.to_tensor( | ||
| np.array([self.tmp_seed, 0]).astype(np.int32) | ||
| ) | ||
| self.inputs = { | ||
| "q": convert_float_to_uint16(q) | ||
| if self.dtype == np.uint16 | ||
| else q.astype(self.dtype), | ||
| "k": convert_float_to_uint16(k) | ||
| if self.dtype == np.uint16 | ||
| else k.astype(self.dtype), | ||
| "v": convert_float_to_uint16(v) | ||
| if self.dtype == np.uint16 | ||
| else v.astype(self.dtype), | ||
| "fixed_seed_offset": fixed_seed_offset, | ||
| } | ||
| bias_ = None | ||
| if self.with_bias: | ||
| bias = np.random.random(self.bias_shape).astype(np.float32) | ||
| self.inputs["attn_mask"] = bias | ||
| bias_ = paddle.to_tensor(bias, stop_gradient=True) | ||
|
|
||
| out, softmax_lse = attention_naive( | ||
| q_, k_, v_, bias=bias_, is_causal=self.is_causal | ||
| ) | ||
| out.backward() | ||
| self.dq = q_.grad.numpy() | ||
| self.dk = k_.grad.numpy() | ||
| self.dv = v_.grad.numpy() | ||
| self.dout = paddle.ones_like(out, dtype=self.dtype) | ||
| self.attrs = { | ||
| 'dropout': 0.0, | ||
| 'causal': self.is_causal, | ||
| 'return_softmax': False, | ||
| 'rng_name': '', | ||
| } | ||
| softmax_lse = softmax_lse.numpy() | ||
| self.outputs = { | ||
| "out": convert_float_to_uint16(out) | ||
| if self.dtype == np.uint16 | ||
| else out.astype(self.dtype), | ||
| "softmax": np.array([]), # not used | ||
| "softmax_lse": softmax_lse, | ||
| "seed_offset": fixed_seed_offset, | ||
| } | ||
|
|
||
| def test_check_output(self): | ||
| self.check_output_with_place( | ||
| paddle.XPUPlace(0), atol=self.tolerance, rtol=self.tolerance | ||
| ) | ||
|
|
||
| def test_check_grad(self): | ||
| self.check_grad( | ||
| ['q', 'k', 'v'], | ||
| 'out', | ||
| user_defined_grads=[self.dq, self.dk, self.dv], | ||
| user_defined_grad_outputs=self.dout, | ||
| numeric_grad_delta=self.tolerance, | ||
| max_relative_error=self.tolerance, | ||
| ) | ||
|
|
||
| # class TestFlashAttentionOp2_with_bias(TestFlashAttentionOp): | ||
| # def set_attrs(self): | ||
| # self.is_causal = True | ||
| # self.with_bias = True | ||
|
|
||
| # class TestFlashAttentionOp3_uncausal_with_bias(TestFlashAttentionOp): //WIP | ||
|
|
||
| # def set_attrs(self): | ||
| # self.is_causal = False | ||
| # self.with_bias = True | ||
|
|
||
|
|
||
| support_types = get_xpu_op_support_types("flash_attn") | ||
| for stype in support_types: | ||
| create_test_class(globals(), XPUTestFlashAttentionOp, stype) | ||
|
|
||
| if __name__ == '__main__': | ||
| paddle.disable_static() | ||
| unittest.main() |