AI control plane to schedule, run, debug, observe, and automate AI/ML workloads, agents, and sandboxes on your infra. Community: https://teapawt.org/t.co/z7RVr8R4Tr
Joined December 2016
- Tweets353
- Following576
- Followers1.1K
- Likes156
Our plans for persistent sandbox workspaces, suspend/resume, activity-aware idle suspension, and ready pools on Kubernetes, and where AX and Substrate differ and can be used within Polyaxon:
polyaxon.com/blog/polyaxon-a…
Quantization, pruning, or distillation: which path actually helps your model?
Start from one checkpoint. A Polyaxon DAG schedules the GPU jobs and keeps each candidate's evidence together. Measure speed only after a compatible export.
polyaxon.com/blog/compare-mo…
A Jev-like decision classifier needs more than a one-letter output.
Our new tutorial trains a Qwen LoRA adapter in a Polyaxon GPU job, logs its lineage, and checks held-out routing cases before serving it with vLLM.
polyaxon.com/blog/train-a-de…
polyaxon sandbox shell creates a new PTY each time. To resume an existing live terminal from an application, keep its ID and use the SDK's pty.attach().
Part 3 of our remote development series: polyaxon.com/blog/work-inter…
Connect your terminal or IDE to a Polyaxon workbench over SSH, and forward an application port with -L. Attach to tmux explicitly when you want to resume a shell.
Part 2 of our remote development series: polyaxon.com/blog/connect-yo…
An ordinary ops shell starts fresh.
With "plugins.tmux" sessions return to the same shell, from CLI or UI, variables, and running commands while the container stays alive.
polyaxon.com/blog/reconnect-…
Polyaxon query language now supports nested queries.
You can shortlist runs with low loss OR high accuracy:
status:succeeded AND kind:job AND (metrics.loss:<=0.08 OR metrics.accuracy:>=0.95)
Combine conditions across fields in the Polyaxon UI and CLI: polyaxon.com/blog/combine-ru…
Three learning rates × two LoRA ranks = six candidates.
Keep the evaluator fixed, compare quality against GPU-hours, and repeat finalists before the final test. A practical tuning workflow with Polyaxon:
polyaxon.com/blog/tune-llm-h…
Polyaxon accepts now shorter format for params. Instead of writing:
batch_size: {value: 32}
users can past:
batch_size: 32
The specification also infers the optional flag based on input value. Less YAML for your next training run.
polyaxon.com/blog/write-shor…
Uploading multiple files and folders with Polyaxon CLI is now possible!
polyaxon.com/blog/upload-mul…
Design isolated Kubernetes execution for AI agents with a threat model, stronger runtimes, scoped identity, controlled egress, bounded storage, and auditable cleanup.
polyaxon.com/blog/sandbox-ai…
Polyaxon's new sandbox client:
from polyaxon.client import SandboxClient
sandbox = SandboxClient(project="demo")
sandbox.ping()
result = sandbox.process.exec(["sh", "-lc", "echo ok"])
print(result.stdout)
print(result.stderr)
print(result.exit_code)
Fs ops:
sandbox.fs.write_text("/tmp/hello.txt", "hello\n")
sandbox.fs.read_text("/tmp/hello.txt")
sandbox.fs.ls("/tmp")
sandbox.fs.upload_file("./local.txt", "/tmp/f.txt")
sandbox.fs.download_file("/tmp/f.txt", "./copy.txt")
Two new CLI commands are coming in beta in the next Polyaxon version:
`polyaxon sandbox` to launch isolated environments
`polyaxon ssh` to attach to workloads with governed remote access
Built for faster debugging loops across AI/ML workloads and agents.
