CARVIEW |
Select Language
HTTP/2 200
date: Wed, 23 Jul 2025 18:30:14 GMT
content-type: text/html; charset=utf-8
cache-control: no-cache
content-security-policy: default-src 'none'; base-uri 'self'; child-src github.githubassets.com github.com/assets-cdn/worker/ github.com/assets/ gist.github.com/assets-cdn/worker/; connect-src 'self' uploads.github.com www.githubstatus.com collector.github.com raw.githubusercontent.com api.github.com github-cloud.s3.amazonaws.com github-production-repository-file-5c1aeb.s3.amazonaws.com github-production-upload-manifest-file-7fdce7.s3.amazonaws.com github-production-user-asset-6210df.s3.amazonaws.com *.rel.tunnels.api.visualstudio.com wss://*.rel.tunnels.api.visualstudio.com objects-origin.githubusercontent.com copilot-proxy.githubusercontent.com proxy.individual.githubcopilot.com proxy.business.githubcopilot.com proxy.enterprise.githubcopilot.com *.actions.githubusercontent.com wss://*.actions.githubusercontent.com productionresultssa0.blob.core.windows.net/ productionresultssa1.blob.core.windows.net/ productionresultssa2.blob.core.windows.net/ productionresultssa3.blob.core.windows.net/ productionresultssa4.blob.core.windows.net/ productionresultssa5.blob.core.windows.net/ productionresultssa6.blob.core.windows.net/ productionresultssa7.blob.core.windows.net/ productionresultssa8.blob.core.windows.net/ productionresultssa9.blob.core.windows.net/ productionresultssa10.blob.core.windows.net/ productionresultssa11.blob.core.windows.net/ productionresultssa12.blob.core.windows.net/ productionresultssa13.blob.core.windows.net/ productionresultssa14.blob.core.windows.net/ productionresultssa15.blob.core.windows.net/ productionresultssa16.blob.core.windows.net/ productionresultssa17.blob.core.windows.net/ productionresultssa18.blob.core.windows.net/ productionresultssa19.blob.core.windows.net/ github-production-repository-image-32fea6.s3.amazonaws.com github-production-release-asset-2e65be.s3.amazonaws.com insights.github.com wss://alive.github.com api.githubcopilot.com api.individual.githubcopilot.com api.business.githubcopilot.com api.enterprise.githubcopilot.com; font-src github.githubassets.com; form-action 'self' github.com gist.github.com copilot-workspace.githubnext.com objects-origin.githubusercontent.com; frame-ancestors 'none'; frame-src viewscreen.githubusercontent.com notebooks.githubusercontent.com; img-src 'self' data: blob: github.githubassets.com media.githubusercontent.com camo.githubusercontent.com identicons.github.com avatars.githubusercontent.com private-avatars.githubusercontent.com github-cloud.s3.amazonaws.com objects.githubusercontent.com release-assets.githubusercontent.com secured-user-images.githubusercontent.com/ user-images.githubusercontent.com/ private-user-images.githubusercontent.com opengraph.githubassets.com copilotprodattachments.blob.core.windows.net/github-production-copilot-attachments/ github-production-user-asset-6210df.s3.amazonaws.com customer-stories-feed.github.com spotlights-feed.github.com objects-origin.githubusercontent.com *.githubusercontent.com; manifest-src 'self'; media-src github.com user-images.githubusercontent.com/ secured-user-images.githubusercontent.com/ private-user-images.githubusercontent.com github-production-user-asset-6210df.s3.amazonaws.com gist.github.com; script-src github.githubassets.com; style-src 'unsafe-inline' github.githubassets.com; upgrade-insecure-requests; worker-src github.githubassets.com github.com/assets-cdn/worker/ github.com/assets/ gist.github.com/assets-cdn/worker/
referrer-policy: no-referrer-when-downgrade
server-timing: pull_request_layout-fragment;desc="pull_request_layout fragment";dur=410.908066,conversation_content-fragment;desc="conversation_content fragment";dur=657.701037,conversation_sidebar-fragment;desc="conversation_sidebar fragment";dur=394.634537,nginx;desc="NGINX";dur=1.297807,glb;desc="GLB";dur=100.906869
strict-transport-security: max-age=31536000; includeSubdomains; preload
vary: X-PJAX, X-PJAX-Container, Turbo-Visit, Turbo-Frame, X-Requested-With,Accept-Encoding, Accept, X-Requested-With
x-content-type-options: nosniff
x-frame-options: deny
x-voltron-version: fd8fbbc
x-xss-protection: 0
server: github.com
content-encoding: gzip
accept-ranges: bytes
set-cookie: _gh_sess=dA4pYVRluUtLb%2BfW6J0q1%2Bad7esD8GVhGS2MTxHZyZk8ZG3lNfih%2FjAqKq1ETGp9v7EoLVYamJhEtISi2x3OZ7mjNWF1bGieFhP69k5EPPDhVM9FMjfBvrSpzwXbFa%2FZ84V5ylrpKgV8dk7cJ4nxRv1hNkhNWE6FuNwagabh%2FhhcvNyqioed4dLqVErawinREydDJpf9xnnmPGA0WefqsCqrf%2BrIlKdvU9aegFreJsDmNxoWjH6nCBhRzc7ZfZuUmgJYGEv8gQeuskbyZfml2g%3D%3D--2Kly1%2B%2F3BlYGFYYG--DPwH5p6BtwOcs2DxQYwRpQ%3D%3D; Path=/; HttpOnly; Secure; SameSite=Lax
set-cookie: _octo=GH1.1.700729161.1753295413; Path=/; Domain=github.com; Expires=Thu, 23 Jul 2026 18:30:13 GMT; Secure; SameSite=Lax
set-cookie: logged_in=no; Path=/; Domain=github.com; Expires=Thu, 23 Jul 2026 18:30:13 GMT; HttpOnly; Secure; SameSite=Lax
x-github-request-id: BD40:221257:7646:9157:68812A35
[PHI] fix paddle.incubate.softmax_mask_fuse for big tensor by DanielSun11 · Pull Request #73096 · PaddlePaddle/Paddle · GitHub
Skip to content
Navigation Menu
{{ message }}
-
Notifications
You must be signed in to change notification settings - Fork 5.8k
[PHI] fix paddle.incubate.softmax_mask_fuse for big tensor #73096
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Merged
lshpku
merged 5 commits into
PaddlePaddle:develop
from
DanielSun11:fix_fused_softmax_mask
Jun 12, 2025
Merged
[PHI] fix paddle.incubate.softmax_mask_fuse for big tensor #73096
lshpku
merged 5 commits into
PaddlePaddle:develop
from
DanielSun11:fix_fused_softmax_mask
Jun 12, 2025
Conversation
This file contains hidden or 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
你的PR提交成功,感谢你对开源项目的贡献! |
/re-run all-failed |
/re-run all-failed |
lshpku
approved these changes
Jun 12, 2025
wanghuancoder
approved these changes
Jun 12, 2025
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
LGTM
shanjiang7
pushed a commit
to shanjiang7/Paddle
that referenced
this pull request
Jun 12, 2025
…dle#73096) * fix fused_softmax_mask and its grad for big tensor * fix forward kernel block config * using v1 and v2 kernel to keep the performance * fix data type for windows
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
You can’t perform that action at this time.
PR Category
Execute Infrastructure
PR Types
Bug fixes
Description
修复paddle.incubate.softmax_mask_fuse使其能支持big tensor。
对softmax_mask_fuse的前向和反向的kernel都进行了修改;
存在问题1: 访存越界
成因:
前向和反向kernel中访问全局显存时使用的是int类型的offset,导致访问超过2^31以上的offset时出现溢出,导致出现非法地址。
修复方案:
将关键的索引部分替换为int64类型
存在问题2:launch kernel时的非法配置
成因:
launch kernel时 gird的各个维度的配置为
dim3 blocks(query_seq_len / batches_per_block, attn_heads, batches);
由于gridDim.y/z的最大值为65535,因此当attn_heads或者batches超过此值时会报cuda error 9
修复方案:
计算总共需要的block数,然后只使用gridDim中的x维度(最大为2^32 - 1),在cuda kernel中重新计算mask Tensor对应的内存索引。
Grid中block的配置修复前后的对比
CUDA Kernel中重新计算每个block的任务分配
当前支持的Tensor的最大numel为 batches * attn_heads * query_seq_len * key_seq_len = (2^32 - 1 )*8192 ,batches * attn_heads * query_seq_len最大值为(2^32 - 1 ),以float16 为例,此时需要的显存约为65536GB,可以应对绝大多数场景。
性能测试
通过nsight system和nsight compute进行深入的profiling发现。修改后的cuda kernel 每个thread所需的register由30涨到了37,进而导致warp的occupancy 下降,从而影响了性能。但是修复后的cuda kernel更加通用可以应对更大规模的Tensor。为了保证性能和支持Big Tensor,对Tensor中数据规模较小时,仍采用原有的cuda kernel(SoftmaxMaskFuseV1GPUKernel)以保证性能。当Tensor中数据规模较大时采用修复后的SoftmaxMaskFuseV2GPUKernel。
正确性测试
pcard-67164