优化to_tensor函数中的bf16转换 #73050

Qin-sx · 2025-06-02T15:56:29Z

PR Category

User Experience

PR Types

Improvements

Description

resolve #72484

需求：

保持原numpy类型不变，直接转为一个paddle.Tensor

返回paddle.Tensor，astype为bfloat16

测试代码

import numpy as np
import paddle
import time

x = np.random.randn(100000).astype(np.float32)
tensor_bfloat16 = paddle.to_tensor(x, dtype=paddle.bfloat16)


num_runs = 1000
total_time = 0

for _ in range(num_runs):
    start_time = time.time()
    tensor_bfloat16 = paddle.to_tensor(x, dtype=paddle.bfloat16)
    paddle.device.synchronize()
    end_time = time.time()
    total_time += (end_time - start_time)

avg_time = total_time / num_runs

print(f"avg time ({num_runs} runs): {avg_time * 1000:.4f} ms")

原本方式平均耗时为 40.4780 ms，修改后平均耗时为 0.2655 ms。

modified: python/paddle/tensor/creation.py modified: test/dygraph_to_static/test_to_tensor.py

paddle-bot · 2025-06-02T15:56:34Z

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

codecov-commenter · 2025-06-02T18:00:32Z

Codecov Report

All modified and coverable lines are covered by tests ✅

Please upload report for BASE (develop@425c14d). Learn more about missing BASE report.

Additional details and impacted files

@@             Coverage Diff             @@
##             develop    #73050   +/-   ##
===========================================
  Coverage           ?   100.00%           
===========================================
  Files              ?         1           
  Lines              ?        10           
  Branches           ?         0           
===========================================
  Hits               ?        10           
  Misses             ?         0           
  Partials           ?         0

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

SigureMo · 2025-06-03T02:18:50Z

test/dygraph_to_static/test_to_tensor.py

这段代码并没有修改静态图分支，不需要添加动转静单测，单测添加到 test/legacy_test/test_eager_tensor.py

只是加了一个普通的函数跑不到，且添加 unittest.skipIf 的方式也是错的

tensor = core.eager.Tensor( value=data, place=place, persistable=False, zero_copy=False, name=None, stop_gradient=stop_gradient, ) # tensor = tensor.astype(dtype) tensor = paddle.cast(tensor, dtype) return tensor

您好，可否帮忙看一下，在使用转换后，x.grad会变为None

x = paddle.to_tensor( 1e6, dtype=paddle.bfloat16, stop_gradient=False) print("x:", x) y = x * x y.backward() print("x.grad:", x.grad)

修改后结果

x: Tensor(shape=[], dtype=bfloat16, place=Place(cpu), stop_gradient=False, -9.9942e+05) x.grad: None

原本结果

x: Tensor(shape=[], dtype=bfloat16, place=Place(cpu), stop_gradient=False, -9.9942e+05) x.grad: Tensor(shape=[], dtype=bfloat16, place=Place(cpu), stop_gradient=False, -1.9988e+06)

modified: python/paddle/tensor/creation.py modified: test/dygraph_to_static/test_to_tensor.py

zhwesky2010 · 2025-06-04T09:40:19Z

python/paddle/tensor/creation.py

+            tensor.stop_gradient = stop_gradient
+            return tensor
+        else:
+            data = _handle_np_dtype(data, dtype)


可以把_handle_np_dtype逻辑挪出来，移除掉原来不正确的bf16分支代码

请问这个是指将_handle_np_dtype函数删除，然后把代码放在else:中吗？因为_handle_np_dtype函数在前面也被调用过。

请问这个是指将_handle_np_dtype函数删除，然后把代码放在else:中吗？因为_handle_np_dtype函数在前面也被调用过。

这个逻辑很简单，可以都挪出来，直接处理掉原来错误的bf16分支

这个逻辑很简单，可以都挪出来，直接处理掉原来错误的bf16分支

嗯，收到，已修改

zhwesky2010 · 2025-06-04T09:48:57Z

python/paddle/tensor/creation.py

+                name=None,
+                stop_gradient=stop_gradient,
+            )
+            tensor = tensor.detach().astype(dtype)


这样写呢：

tensor = core.eager.Tensor( value=data, place=place, persistable=False, zero_copy=False, name=None, stop_gradient=True, ) tensor = tensor.astype('bfloat16') tensor.stop_gradient = stop_gradient return tensor

这样好像不行，上一个版本是类似的做法。我个人理解是只能数据进行转换，不能带着gradient一起转换。如果不用detach，会导致grad为None。

tensor = core.eager.Tensor( value=data, place=place, persistable=False, zero_copy=False, name=None, stop_gradient=stop_gradient, ) # tensor = tensor.astype(dtype) tensor = paddle.cast(tensor, dtype) return tensor

您好，可否帮忙看一下，在使用转换后，x.grad会变为None

x = paddle.to_tensor( 1e6, dtype=paddle.bfloat16, stop_gradient=False) print("x:", x) y = x * x y.backward() print("x.grad:", x.grad)

修改后结果

x: Tensor(shape=[], dtype=bfloat16, place=Place(cpu), stop_gradient=False, -9.9942e+05) x.grad: None

原本结果

x: Tensor(shape=[], dtype=bfloat16, place=Place(cpu), stop_gradient=False, -9.9942e+05) x.grad: Tensor(shape=[], dtype=bfloat16, place=Place(cpu), stop_gradient=False, -1.9988e+06)

这样好像不行，上一个版本是类似的做法。我个人理解是只能数据进行转换，不能带着gradient一起转换。如果不用detach，会导致grad为None。

tensor = core.eager.Tensor( value=data, place=place, persistable=False, zero_copy=False, name=None, stop_gradient=stop_gradient, ) # tensor = tensor.astype(dtype) tensor = paddle.cast(tensor, dtype) return tensor

您好，可否帮忙看一下，在使用转换后，x.grad会变为None

x = paddle.to_tensor( 1e6, dtype=paddle.bfloat16, stop_gradient=False) print("x:", x) y = x * x y.backward() print("x.grad:", x.grad)

修改后结果

x: Tensor(shape=[], dtype=bfloat16, place=Place(cpu), stop_gradient=False, -9.9942e+05) x.grad: None

原本结果

x: Tensor(shape=[], dtype=bfloat16, place=Place(cpu), stop_gradient=False, -9.9942e+05) x.grad: Tensor(shape=[], dtype=bfloat16, place=Place(cpu), stop_gradient=False, -1.9988e+06)

这样写是可以的，这是因为之前的写法，没有设置stop_gradient=True，使得第一个tensor与cast后的tensor建立了反向关系，梯度传导到第一个tensor了，而cast后的tensor的梯度就被清理掉了。

目前把第一个tensor的stop_gradient设置为True，就避免了这个问题。

嗯，收到，我测试了一下，确实没问题，感谢

modified: python/paddle/tensor/creation.py

zhwesky2010 · 2025-06-09T04:10:24Z

python/paddle/tensor/creation.py

            # Windows default type is 'int32', while Linux/Mac is 'int64'. Unify they.
            if data.dtype in ['int32']:
                data = data.astype("int64")

    if dtype:
-        data = _handle_np_dtype(data, dtype)
+        if (


这里不用搞这么多判断吧，代码注意逻辑清晰，可读性强。

这里要么全统一用 convert_dtype(dtype) 判断，要么统一用dtype判断，不用在这里反复冗余判断。ndarray也没必要判断吧

嗯，收到，已按照下面的方式修改

zhwesky2010 · 2025-06-09T04:13:57Z

python/paddle/tensor/creation.py

+            tensor.stop_gradient = stop_gradient
+            return tensor
+        else:
+            if convert_dtype(dtype) != convert_dtype(data.dtype):


if dtype and convert_dtype(dtype) != convert_dtype(data.dtype): if convert_dtype(dtype) == 'uint16': ... else: data = data.astype(convert_dtype(dtype))

这样可以吗，这里的分支显得又多又乱

嗯，收到，已修改，感谢

zhwesky2010 · 2025-06-09T04:20:05Z

python/paddle/tensor/creation.py

@@ -757,13 +742,35 @@ def _handle_np_dtype(
                        if default_type in ['float16', 'float32']
                        else 'complex128'
                    )
-                data = _handle_np_dtype(data, default_type)
+                if convert_dtype(default_type) != convert_dtype(data.dtype):


这里统一写成：

if convert_dtype(default_type) != convert_dtype(data.dtype): dtype = default_type

然后交到下面的逻辑里去处理，代码更简洁

嗯，收到，已修改，感谢

zhwesky2010

代码注意下可读性方面

modified: python/paddle/tensor/creation.py

zhwesky2010

LGTM

* optimized bf16 convert in to_tensor modified: python/paddle/tensor/creation.py modified: test/dygraph_to_static/test_to_tensor.py * modified for grad modified: python/paddle/tensor/creation.py modified: test/dygraph_to_static/test_to_tensor.py * changed core.eager.Tensor para modified: python/paddle/tensor/creation.py * deleted _handle_np_dtype modified: python/paddle/tensor/creation.py * updated conditions modified: python/paddle/tensor/creation.py

optimized bf16 convert in to_tensor

76c54e0

modified: python/paddle/tensor/creation.py modified: test/dygraph_to_static/test_to_tensor.py

Qin-sx requested review from SigureMo, zrr1999 and gouzil as code owners June 2, 2025 15:56

paddle-bot bot added the contributor External developers label Jun 2, 2025

SigureMo reviewed Jun 3, 2025

View reviewed changes

modified for grad

c4bb5c9

modified: python/paddle/tensor/creation.py modified: test/dygraph_to_static/test_to_tensor.py

zhwesky2010 reviewed Jun 4, 2025

View reviewed changes

Qin-sx added 2 commits June 7, 2025 00:02

changed core.eager.Tensor para

f09a899

modified: python/paddle/tensor/creation.py

deleted _handle_np_dtype

bcb2b12

modified: python/paddle/tensor/creation.py

zhwesky2010 reviewed Jun 9, 2025

View reviewed changes

updated conditions

5e44308

modified: python/paddle/tensor/creation.py

zhwesky2010 approved these changes Jun 10, 2025

View reviewed changes

zhwesky2010 merged commit d2d3b1b into PaddlePaddle:develop Jun 12, 2025
50 checks passed

优化to_tensor函数中的bf16转换 #73050

优化to_tensor函数中的bf16转换 #73050

Uh oh!

Conversation

Qin-sx commented Jun 2, 2025 • edited by zhwesky2010 Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

PR Category

PR Types

Description

Uh oh!

paddle-bot bot commented Jun 2, 2025

Uh oh!

codecov-commenter commented Jun 2, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

zhwesky2010 Jun 9, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

zhwesky2010 Jun 9, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

zhwesky2010 Jun 9, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

zhwesky2010 left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

zhwesky2010 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Qin-sx commented Jun 2, 2025 •

edited by zhwesky2010

Loading

codecov-commenter commented Jun 2, 2025 •

edited

Loading

zhwesky2010 Jun 9, 2025 •

edited

Loading

zhwesky2010 Jun 9, 2025 •

edited

Loading

zhwesky2010 Jun 9, 2025 •

edited

Loading

zhwesky2010 left a comment •

edited

Loading