[PHI] Support int64 for paddle.kthvalue param k

[cangtianhuang]

PR Category

Operator Mechanism

PR Types

Bug fixes

Description

Modification:

1. Change the parameter k signature from int to int64_t. Preposition PR:

• [PHI] Fix paddle.kthvalue for big tensor #73745

2. Fix paddle/fluid/framework/infershape_utils.cc 、 paddle/fluid/framework/operator.cc to support int to int64_t , referring to the handling of vector<int> : https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/infershape_utils.cc#L825

Correctness Test:

x_np = np.random.random(4294967295) - 0.5 k = 4200000000 x = paddle.to_tensor(x_np, dtype="float32") result = paddle.kthvalue(x, k) x = torch.tensor(x_np, dtype=torch.float32) result = torch.kthvalue(x, k)

Now paddle.kthvalue supports int64_t for the parameter k, and computes faster than torch. Paddle return the first index of the k-th value, whereas torch return it randomly.

Compatibility Test:

To test whether the models of old/new IR can be loaded before and after modification (using paddle daily build to save the models)

import paddle pir_path = "models/kthvalue_pir" legacy_ir_path = "models/kthvalue_legacy_ir" class KthValueModel(paddle.nn.Layer): def __init__(self): super().__init__() def forward(self, x): values, indices = paddle.kthvalue(x, 2) return values, indices model = KthValueModel() # paddle.jit.save( # model, # pir_path, # input_spec=[ # paddle.static.InputSpec(shape=[None, None], dtype='float32', name='x'), # ], # ) # with paddle.pir_utils.DygraphOldIrGuard(): # paddle.jit.save( # model, # legacy_ir_path, # input_spec=[ # paddle.static.InputSpec( # shape=[None, None], dtype='float32', name='x' # ), # ], # ) x = paddle.randn([4, 6], dtype='float32') load_model_pir = paddle.jit.load(pir_path) out = load_model_pir(x) print(out) with paddle.pir_utils.DygraphOldIrGuard(): load_model_legacy_ir = paddle.jit.load(legacy_ir_path) out_legacy_ir = load_model_legacy_ir(x) print(out_legacy_ir) 

new IR report error:

It can be addressed for Release version by changing pir patch version:

• 【save/load】change pir patch version #73051

old IR pass:

Pcard-85711

[paddle-bot]

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.

[wanghuancoder]

LGTM 这个PR再找黄济懿、陈志扬帮忙看看

[huangjiyi]

另外 Static-Check 的 approval 问题需要解决一下，旧 IR 的 Op 定义是不希望改变的，同样参考 #72639 里在 static_ops.yaml, static_backward.yaml, ops_exclude.yaml 对 lp_pool 的改动

[cangtianhuang]

另外 Static-Check 的 approval 问题需要解决一下，旧 IR 的 Op 定义是不希望改变的，同样参考 #72639 里在 static_ops.yaml, static_backward.yaml, ops_exclude.yaml 对 lp_pool 的改动

done

[huangjiyi]

LGTM

[codecov-commenter]

Codecov Report

Attention: Patch coverage is 82.60870% with 4 lines in your changes missing coverage. Please review.

Please upload report for BASE (develop@b8b4c24). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
paddle/fluid/framework/infershape_utils.cc	83.33%	2 Missing ⚠️
paddle/fluid/framework/operator.cc	81.81%	2 Missing ⚠️

❌ Your patch status has failed because the patch coverage (82.60%) is below the target coverage (90.00%). You can increase the patch coverage or adjust the target coverage.

Additional details and impacted files

@@ Coverage Diff @@ ## develop #74002 +/- ## ========================================== Coverage ? 82.60% ========================================== Files ? 2 Lines ? 23 Branches ? 0 ========================================== Hits ? 19 Misses ? 4 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.

[wanghuancoder]

LGTM

[huangjiyi]

LGTM

[cangtianhuang]

/re-run all-failed