Has anyone had any luck using tensorRT for this comp? I see that kaggle now uses 3.12 instead of the working examples from 3.11. Seems we cannot convert to tensorRT on 3.12 but we can use the engine if the files are created already. So I have been trying to do python 3.11 in a virtual environment and load needed packages on vast.ai to convert models and then bring those all back into kaggle for inference but I have been having a miserable time. I’m certain many in the community could benefit from a greater understanding of this if someone has managed to figure it out
", "messages": [ { "id": "3400594", "postDate": "02/01/2026 21:23:49", "content": "Has anyone had any luck using tensorRT for this comp? I see that kaggle now uses 3.12 instead of the working examples from 3.11. Seems we cannot convert to tensorRT on 3.12 but we can use the engine if the files are created already. So I have been trying to do python 3.11 in a virtual environment and load needed packages on vast.ai to convert models and then bring those all back into kaggle for inference but I have been having a miserable time. I’m certain many in the community could benefit from a greater understanding of this if someone has managed to figure it out
", "rawMarkdown": "Has anyone had any luck using tensorRT for this comp? I see that kaggle now uses 3.12 instead of the working examples from 3.11. Seems we cannot convert to tensorRT on 3.12 but we can use the engine if the files are created already. So I have been trying to do python 3.11 in a virtual environment and load needed packages on vast.ai to convert models and then bring those all back into kaggle for inference but I have been having a miserable time. I’m certain many in the community could benefit from a greater understanding of this if someone has managed to figure it out", "votes": null }, { "id": "3401623", "postDate": "02/04/2026 06:36:14", "content": "I haven't tried it on Kaggle yet, but in my local environment TensorRT seems to provide a 2‑3× speedup—very impressive
", "rawMarkdown": "I haven't tried it on Kaggle yet, but in my local environment TensorRT seems to provide a 2‑3× speedup—very impressive", "votes": null }, { "id": "3403728", "postDate": "02/09/2026 06:23:33", "content": "if you do it locally and then It is easy to convert, I guess.
", "rawMarkdown": "if you do it locally and then It is easy to convert, I guess.", "votes": null }, { "id": "3404323", "postDate": "02/10/2026 10:07:03", "content": "you can fork a notebook from previous date - the metric demo uses 3.11 and then just make sure you keep it pinned to the original environment.
", "rawMarkdown": "you can fork a notebook from previous date - the metric demo uses 3.11 and then just make sure you keep it pinned to the original environment.", "votes": null }, { "id": "3404761", "postDate": "02/11/2026 07:04:23", "content": "i hope the kaggle product team (or some kaggler) can make a Kaggle Agent Copilot to write code like this (e.g. compile for tensorRT, multithreads for 2x T4 GPU, using TPU …).
\nPersonally, chatgppt helps me to translate training code to DDP and torch compile which saves me a lot of work.
", "rawMarkdown": "i hope the kaggle product team (or some kaggler) can make a Kaggle Agent Copilot to write code like this (e.g. compile for tensorRT, multithreads for 2x T4 GPU, using TPU ...).\n\nPersonally, chatgppt helps me to translate training code to DDP and torch compile which saves me a lot of work.", "votes": null }, { "id": "3404764", "postDate": "02/11/2026 07:14:43", "content": "Codex and ClaudeCode helped me a lot with DDP and training optimization. Before this, I was a complete novice… I can't even imagine how people participated in competitions and coding before the AI era.
", "rawMarkdown": "Codex and ClaudeCode helped me a lot with DDP and training optimization. Before this, I was a complete novice... I can't even imagine how people participated in competitions and coding before the AI era.", "votes": null }, { "id": "3404774", "postDate": "02/11/2026 08:01:34", "content": "To be honest, I hope kaggle can announce an official code pilot develop competition, asking kaggler to develop kaggle product, that would be very cool.
", "rawMarkdown": "To be honest, I hope kaggle can announce an official code pilot develop competition, asking kaggler to develop kaggle product, that would be very cool.", "votes": null }, { "id": "3404818", "postDate": "02/11/2026 10:17:38", "content": "Actually At that time most comps were tabular,and cv and NLP comp were really simple
", "rawMarkdown": "Actually At that time most comps were tabular,and cv and NLP comp were really simple", "votes": null }, { "id": "3404824", "postDate": "02/11/2026 10:22:54", "content": "How about one start kaggle agent framework. then people start to submit or contribute their code document md as dataset. The person that make the agent win past kaggle competition wins and the trained kaggle agent can be opened source.
", "rawMarkdown": "How about one start kaggle agent framework. then people start to submit or contribute their code document md as dataset. The person that make the agent win past kaggle competition wins and the trained kaggle agent can be opened source.", "votes": null }, { "id": "3407610", "postDate": "02/18/2026 18:22:54", "content": "Yep this ended up being the direction we went. There are ways to rent a T4 to build engine files and then use them on the current kaggle inference but its a bit more painful
", "rawMarkdown": "Yep this ended up being the direction we went. There are ways to rent a T4 to build engine files and then use them on the current kaggle inference but its a bit more painful", "votes": null }, { "id": "3410550", "postDate": "02/23/2026 08:59:21", "content": "@hengck23 That’s a really cool idea. I’d love to see Kaggle introduce multiple “Agent Teams.” Imagine competing not only against other people, but also against AI agents, it would add a whole new dimension to the game.
\nThat said, if the agents completely dominate the leaderboard, the competition could start to feel meaningless for human participants. I suspect this would be most likely in tabular competitions, where strong automated pipelines can already achieve near state-of-the-art performance.
\nFor more complex challenges like Vesuvius or the Stanford RNA competitions, the difficulty and domain-specific nuances probably still require significant human intuition, creativity, and iterative experimentation. Those types of problems might remain more resilient to full automation, at least for now.
\nAll I know is that some agent is absolutely going to try a brute-force approach and end up burning through Google’s data centers in the process.
", "rawMarkdown": "hengck23 That’s a really cool idea. I’d love to see Kaggle introduce multiple “Agent Teams.” Imagine competing not only against other people, but also against AI agents, it would add a whole new dimension to the game.\n\nThat said, if the agents completely dominate the leaderboard, the competition could start to feel meaningless for human participants. I suspect this would be most likely in tabular competitions, where strong automated pipelines can already achieve near state-of-the-art performance.\n\nFor more complex challenges like Vesuvius or the Stanford RNA competitions, the difficulty and domain-specific nuances probably still require significant human intuition, creativity, and iterative experimentation. Those types of problems might remain more resilient to full automation, at least for now.\n\nAll I know is that some agent is absolutely going to try a brute-force approach and end up burning through Google’s data centers in the process.", "votes": null }, { "id": "3410588", "postDate": "02/23/2026 10:24:58", "content": "“if the agents completely dominate the leaderboard, the competition could start to feel meaningless for human participants. ”
\nInstead of human training model, it is human training agent
\nKaggle agent (skill) training competition, especially ml or datascience skill
", "rawMarkdown": "“if the agents completely dominate the leaderboard, the competition could start to feel meaningless for human participants. ” \n\nInstead of human training model, it is human training agent\n\nKaggle agent (skill) training competition, especially ml or datascience skill", "votes": null } ], "comments": [ { "id": 3401623, "author_name": "wzyfromhust", "author_url": "", "post_date": "02/04/2026 06:36:14", "content": "I haven't tried it on Kaggle yet, but in my local environment TensorRT seems to provide a 2‑3× speedup—very impressive
", "votes": null, "replies": [] }, { "id": 3403728, "author_name": "dragonzhang", "author_url": "", "post_date": "02/09/2026 06:23:33", "content": "if you do it locally and then It is easy to convert, I guess.
", "votes": null, "replies": [] }, { "id": 3404323, "author_name": "mccocoful", "author_url": "", "post_date": "02/10/2026 10:07:03", "content": "you can fork a notebook from previous date - the metric demo uses 3.11 and then just make sure you keep it pinned to the original environment.
", "votes": null, "replies": [ { "id": 3407610, "author_name": "cody11null", "author_url": "", "post_date": "02/18/2026 18:22:54", "content": "Yep this ended up being the direction we went. There are ways to rent a T4 to build engine files and then use them on the current kaggle inference but its a bit more painful
", "votes": null, "replies": [] } ] }, { "id": 3404761, "author_name": "hengck23", "author_url": "", "post_date": "02/11/2026 07:04:23", "content": "i hope the kaggle product team (or some kaggler) can make a Kaggle Agent Copilot to write code like this (e.g. compile for tensorRT, multithreads for 2x T4 GPU, using TPU …).
\nPersonally, chatgppt helps me to translate training code to DDP and torch compile which saves me a lot of work.
", "votes": null, "replies": [ { "id": 3404764, "author_name": "wzyfromhust", "author_url": "", "post_date": "02/11/2026 07:14:43", "content": "Codex and ClaudeCode helped me a lot with DDP and training optimization. Before this, I was a complete novice… I can't even imagine how people participated in competitions and coding before the AI era.
", "votes": null, "replies": [ { "id": 3404818, "author_name": "tahaalshatiri", "author_url": "", "post_date": "02/11/2026 10:17:38", "content": "Actually At that time most comps were tabular,and cv and NLP comp were really simple
", "votes": null, "replies": [] } ] }, { "id": 3404774, "author_name": "tom99763", "author_url": "", "post_date": "02/11/2026 08:01:34", "content": "To be honest, I hope kaggle can announce an official code pilot develop competition, asking kaggler to develop kaggle product, that would be very cool.
", "votes": null, "replies": [ { "id": 3404824, "author_name": "hengck23", "author_url": "", "post_date": "02/11/2026 10:22:54", "content": "How about one start kaggle agent framework. then people start to submit or contribute their code document md as dataset. The person that make the agent win past kaggle competition wins and the trained kaggle agent can be opened source.
", "votes": null, "replies": [ { "id": 3410550, "author_name": "tom99763", "author_url": "", "post_date": "02/23/2026 08:59:21", "content": "@hengck23 That’s a really cool idea. I’d love to see Kaggle introduce multiple “Agent Teams.” Imagine competing not only against other people, but also against AI agents, it would add a whole new dimension to the game.
\nThat said, if the agents completely dominate the leaderboard, the competition could start to feel meaningless for human participants. I suspect this would be most likely in tabular competitions, where strong automated pipelines can already achieve near state-of-the-art performance.
\nFor more complex challenges like Vesuvius or the Stanford RNA competitions, the difficulty and domain-specific nuances probably still require significant human intuition, creativity, and iterative experimentation. Those types of problems might remain more resilient to full automation, at least for now.
\nAll I know is that some agent is absolutely going to try a brute-force approach and end up burning through Google’s data centers in the process.
", "votes": null, "replies": [ { "id": 3410588, "author_name": "hengck23", "author_url": "", "post_date": "02/23/2026 10:24:58", "content": "“if the agents completely dominate the leaderboard, the competition could start to feel meaningless for human participants. ”
\nInstead of human training model, it is human training agent
\nKaggle agent (skill) training competition, especially ml or datascience skill
", "votes": null, "replies": [] } ] } ] } ] } ] } ], "raw_markdown_by_id": { "3400594": "Has anyone had any luck using tensorRT for this comp? I see that kaggle now uses 3.12 instead of the working examples from 3.11. Seems we cannot convert to tensorRT on 3.12 but we can use the engine if the files are created already. So I have been trying to do python 3.11 in a virtual environment and load needed packages on vast.ai to convert models and then bring those all back into kaggle for inference but I have been having a miserable time. I’m certain many in the community could benefit from a greater understanding of this if someone has managed to figure it out", "3401623": "I haven't tried it on Kaggle yet, but in my local environment TensorRT seems to provide a 2‑3× speedup—very impressive", "3403728": "if you do it locally and then It is easy to convert, I guess.", "3404323": "you can fork a notebook from previous date - the metric demo uses 3.11 and then just make sure you keep it pinned to the original environment.", "3404761": "i hope the kaggle product team (or some kaggler) can make a Kaggle Agent Copilot to write code like this (e.g. compile for tensorRT, multithreads for 2x T4 GPU, using TPU ...).\n\nPersonally, chatgppt helps me to translate training code to DDP and torch compile which saves me a lot of work.", "3404764": "Codex and ClaudeCode helped me a lot with DDP and training optimization. Before this, I was a complete novice... I can't even imagine how people participated in competitions and coding before the AI era.", "3404774": "To be honest, I hope kaggle can announce an official code pilot develop competition, asking kaggler to develop kaggle product, that would be very cool.", "3404818": "Actually At that time most comps were tabular,and cv and NLP comp were really simple", "3404824": "How about one start kaggle agent framework. then people start to submit or contribute their code document md as dataset. The person that make the agent win past kaggle competition wins and the trained kaggle agent can be opened source.", "3407610": "Yep this ended up being the direction we went. There are ways to rent a T4 to build engine files and then use them on the current kaggle inference but its a bit more painful", "3410550": "hengck23 That’s a really cool idea. I’d love to see Kaggle introduce multiple “Agent Teams.” Imagine competing not only against other people, but also against AI agents, it would add a whole new dimension to the game.\n\nThat said, if the agents completely dominate the leaderboard, the competition could start to feel meaningless for human participants. I suspect this would be most likely in tabular competitions, where strong automated pipelines can already achieve near state-of-the-art performance.\n\nFor more complex challenges like Vesuvius or the Stanford RNA competitions, the difficulty and domain-specific nuances probably still require significant human intuition, creativity, and iterative experimentation. Those types of problems might remain more resilient to full automation, at least for now.\n\nAll I know is that some agent is absolutely going to try a brute-force approach and end up burning through Google’s data centers in the process.", "3410588": "“if the agents completely dominate the leaderboard, the competition could start to feel meaningless for human participants. ” \n\nInstead of human training model, it is human training agent\n\nKaggle agent (skill) training competition, especially ml or datascience skill" }, "source": "meta" }