From cecf7324dc95a2f9fe05c9ea45380a415ff7ea01 Mon Sep 17 00:00:00 2001 From: Hossein Firooz Date: Tue, 15 Sep 2026 14:31:19 +0300 Subject: [PATCH 1/4] gpu interactive limits I've added our policy on using interactive jobs with GPUs. --- triton/tut/interactive.rst | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/triton/tut/interactive.rst b/triton/tut/interactive.rst index 48e6bd695..b61673146 100644 --- a/triton/tut/interactive.rst +++ b/triton/tut/interactive.rst @@ -178,6 +178,12 @@ then sshes to there and connects to the shell:: slowly when sent across the general Internet. +Interactive shell with GPU nodes +------------------------------- + +For interactive jobs with gpu resources, it is advised to use `gpu-debug` partition. `gpu-debug` has a higher priority and your job will be scheduled faster. + +If you need a specific GPU type, you can submit your interactive job to other partitions too, but keep the time limit of the job to 4 hours. Also the number of GPUs requested in an interactive job should be kept to 1 GPUs. Checking your jobs ------------------ From 699d267779707de01a3d66790074cbecf25e93ce Mon Sep 17 00:00:00 2001 From: Richard Darst Date: Wed, 16 Sep 2026 16:04:38 +0300 Subject: [PATCH 2/4] Apply batched suggestions from code review Co-authored-by: Richard Darst --- triton/tut/interactive.rst | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/triton/tut/interactive.rst b/triton/tut/interactive.rst index b61673146..067daf9ee 100644 --- a/triton/tut/interactive.rst +++ b/triton/tut/interactive.rst @@ -179,9 +179,9 @@ then sshes to there and connects to the shell:: Interactive shell with GPU nodes -------------------------------- +-------------------------------- -For interactive jobs with gpu resources, it is advised to use `gpu-debug` partition. `gpu-debug` has a higher priority and your job will be scheduled faster. +For interactive jobs with gpu resources, it is advised to use ``gpu-debug`` partition. ``gpu-debug`` has a higher priority and your job will be scheduled faster. If you need a specific GPU type, you can submit your interactive job to other partitions too, but keep the time limit of the job to 4 hours. Also the number of GPUs requested in an interactive job should be kept to 1 GPUs. From 118512acd45fd9c2014c96015f1df0ed2ced73e6 Mon Sep 17 00:00:00 2001 From: Hossein Firooz Date: Wed, 16 Sep 2026 16:15:23 +0300 Subject: [PATCH 3/4] Update triton/tut/interactive.rst Co-authored-by: Richard Darst --- triton/tut/interactive.rst | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/triton/tut/interactive.rst b/triton/tut/interactive.rst index 067daf9ee..4641656be 100644 --- a/triton/tut/interactive.rst +++ b/triton/tut/interactive.rst @@ -181,7 +181,7 @@ then sshes to there and connects to the shell:: Interactive shell with GPU nodes -------------------------------- -For interactive jobs with gpu resources, it is advised to use ``gpu-debug`` partition. ``gpu-debug`` has a higher priority and your job will be scheduled faster. +For interactive jobs with GPU resources, it is advised to use ``gpu-debug`` partition. ``gpu-debug`` has a higher priority and your job will be scheduled faster. Batch GPUs are designed mainly for efficient batch computation and jobs which appear to be interactive and have low GPU usage may be cancelled with or without notice. If you need a specific GPU type, you can submit your interactive job to other partitions too, but keep the time limit of the job to 4 hours. Also the number of GPUs requested in an interactive job should be kept to 1 GPUs. From 3745d3ff0323c4d454704bb03a6bd9e19cdbb158 Mon Sep 17 00:00:00 2001 From: Hossein Firooz Date: Wed, 16 Sep 2026 16:15:35 +0300 Subject: [PATCH 4/4] Update triton/tut/interactive.rst Co-authored-by: Richard Darst --- triton/tut/interactive.rst | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/triton/tut/interactive.rst b/triton/tut/interactive.rst index 4641656be..821a56849 100644 --- a/triton/tut/interactive.rst +++ b/triton/tut/interactive.rst @@ -183,7 +183,7 @@ Interactive shell with GPU nodes For interactive jobs with GPU resources, it is advised to use ``gpu-debug`` partition. ``gpu-debug`` has a higher priority and your job will be scheduled faster. Batch GPUs are designed mainly for efficient batch computation and jobs which appear to be interactive and have low GPU usage may be cancelled with or without notice. -If you need a specific GPU type, you can submit your interactive job to other partitions too, but keep the time limit of the job to 4 hours. Also the number of GPUs requested in an interactive job should be kept to 1 GPUs. +If you need a specific GPU type, you can submit your interactive job to other partitions too, but keep the time limit of the job to 4 hours. Also the number of GPUs requested in an interactive job should be kept to 1 GPUs. Limited use of batch GPUs may be accepted if the reservation is kept to when you are actively at the computer using it, but frequently save because we can't guarantee your job won't be killed. Checking your jobs ------------------