manbearian
commented
7 months ago Yes, no need to fix here... but we really need a way to address this... I wonder if we can modify the tutorials themselves with some conditional code to check if CPU is supported and test CPU back-end if it is?
ian Bearman Principal Software Engineering Manager AI Frameworks: Compilation and Optimization (Maia 100https://azure.microsoft.com/en-us/blog/microsoft-azure-delivers-purpose-built-cloud-infrastructure-in-the-era-of-ai/ and beyond!) Now hiring!https://aka.ms/ianb.jobs
BlackLivesMatter
/ Making your AI faster, cheaper, smarter! /
From: Nhat Nguyen @.> Sent: Wednesday, November 29, 2023 9:42 AM To: microsoft/triton-shared @.> Cc: Ian Bearman @.>; Review requested @.> Subject: Re: [microsoft/triton-shared] Add matmul kernel test for CPU backend (PR #67)
@nhat-nguyen commented on this pull request.
On python/examples/test_matmul.pyhttps://github.com/microsoft/triton-shared/pull/67#discussion_r1409656328:
yea this turns out harder than i expected:
- we need to modify the device type from cuda to cpu in the original code anyway
- our project is independent from the tutorial code which lives in the triton repo, so we can't directly access those files anyway
- Reply to this email directly, view it on GitHubhttps://github.com/microsoft/triton-shared/pull/67#discussion_r1409656328, or unsubscribehttps://github.com/notifications/unsubscribe-auth/AF7M3YVIE6NBG24LVJDKO2LYG5XXRAVCNFSM6AAAAAA77AOFAGVHI2DSMVQWIX3LMV43YUDVNRWFEZLROVSXG5CSMV3GSZLXHMYTONJVHA3DSNBQGM. You are receiving this because your review was requested.Message ID: @.**@.>>
Add the latest matmul triton tutorial as a test for the CPU backend. This exposed a bug in the TritonToLinalg pass where we need to initialize the init tensor with all zeroes first since allocated buffers aren't guaranteed to be zero-initialized.