jgrusewski c731bef759 feat(infra): split training image into RL + supervised, rename ci-compile → ci-rl ...

RL models (DQN/PPO) only need basic CUDA runtime (~2GB), while supervised
models need cuDNN (~4GB). Splitting saves ~2GB pull time per RL job.
Rename the L4 GPU pool from ci-compile to ci-rl to reflect its actual use.
Add DaemonSet image pre-puller to cache training images on GPU nodes.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

2026-02-28 10:16:42 +01:00

..

Dockerfile.ci-builder

feat(infra): upgrade CI to L40S + mold linker for faster builds

2026-02-27 00:05:23 +01:00

Dockerfile.ci-builder-cpu

feat(ci): add CUDA-free CI builder image for service compilation

2026-02-28 00:16:28 +01:00

Dockerfile.infra-runner

fix(ci): deploy job uses infra-runner with scw + kubectl

2026-02-26 08:50:53 +01:00

Dockerfile.runtime

perf(ci): compile once with PVC sccache, package with Kaniko

2026-02-26 00:50:25 +01:00

Dockerfile.service

fix(ci): add AWS_REGION for sccache S3 in Dockerfiles

2026-02-25 21:39:39 +01:00

Dockerfile.training

feat(ml): add evaluate_supervised binary + PPO eval in CI

2026-02-26 21:31:45 +01:00

Dockerfile.training-rl

feat(infra): split training image into RL + supervised, rename ci-compile → ci-rl

2026-02-28 10:16:42 +01:00

Dockerfile.training-supervised

feat(infra): split training image into RL + supervised, rename ci-compile → ci-rl

2026-02-28 10:16:42 +01:00

Dockerfile.web-gateway

fix(ci): update Dockerfiles for restructured repo + migrate to SCW registry

2026-02-25 21:09:52 +01:00

Dockerfile.web-gateway-runtime

fix(docker): correct web-gateway binary path after CI compile split

2026-02-28 00:59:33 +01:00