[Inference]add blha_get_max_len op & modify block_multihead_attention op

[ming1753]

PR Category

Inference

PR Types

New features

Description

Pcard-71500

新增算子：blha_get_max_len，输入为seq_lens_encoder、seq_lens_decoder、bsz，输出为max_enc_len_this_time、max_dec_len_this_time

使用示例：

import paddle
paddle.device.set_device('gpu')
seq_lens_encoder = paddle.cast(paddle.randn(shape=[10]), dtype=paddle.int32)
seq_lens_decoder = paddle.cast(paddle.randn(shape=[10]), dtype=paddle.int32)
bsz = 10
batch_size = paddle.ones(shape=[bsz])
max_enc_len_this_time, max_dec_len_this_time = paddle.incubate.nn.functional.blha_get_max_len(seq_lens_encoder, seq_lens_decoder, batch_size)

修改block_multihead_attention，新增两个可选参数max_enc_len_this_time和max_dec_len_this_time，在传入时不在kernel内部计算

[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.

[vivienfanghuagood]

应该in_dynamic_or_pir_mode

[ming1753]

已修改

[carryyu]

谢谢，已收到！

[jzhang533]

hi, running sample code in paddle.incubate.nn.functional.block_multihead_attention raised NotImplementedError in PR-CI-Static-Check.

btw: change to api no longer requires my approval since #64212 , but please feel free to assign me if you think my review helps.

[ming1753]

Thanks for review. PR-CI-Static-Check raised NotImplementedError in because there is no implementation of FlashAttention on it, not PR's fault.

[jzhang533]

I think we have two options:

• surrounding the sample codes with doctest: +SKIP("reason"), doctest: -SKIP directives to instruct doctest skip running the sample codes.

• upgrade the env of PR-CI-Static-Check.