Skip to content

PyPI v4.0.17 - ClearML

Latest

Choose a tag to compare

@matanliba-clearml matanliba-clearml released this 10 Nov 22:40
· 1 commit to master since this release

v4.0.17 Release Notes

New Features and Bug Fixes

  • Add SNUG, an opt-in LLM usage metering module (agent.snug.enabled, default off)
    • Meter Anthropic, OpenAI and Gemini calls made by tasks, with exact token counts (estimates when no usage is returned)
    • Report prompt-cache read/write tokens, per-model usage, tool calls and per-chat scalars (tokens, latency, requests) to the task's SCALARS tab
    • Report usage events to the AI usage dashboard (enterprise server required)
    • Support HTTP/2 and gzip/zstd-compressed traffic
    • Add app-mode for metering desktop and CLI AI apps through a local forward proxy (Claude Desktop, Codex, OpenCode, Kiro CLI, Hermes, and generic CLI apps via wrap_bins / wrap_kind)
    • Add configurable host whitelist, header injection and optional call-history capture (credentials redacted)
  • Report container CPU/memory usage via cgroups when running inside a container, instead of host machine metrics
  • Fail the task with a visible reason when --use-owner-token cannot impersonate the task owner (enterprise server required)
  • Mask secret-looking env vars (KEY/SECRET/TOKEN/PASS/CREDENTIAL) in printed docker commands and configuration
  • Add agent.vcs_cache.lock_timeout_sec, agent.vcs_cache.temp_cache_on_lock_timeout, agent.vcs_cache.cleanup_orphans and agent.vcs_cache.orphan_min_age_sec configuration options; on vcs cache lock timeout, clone directly into the task folder instead of leaving duplicate cache folders
  • Add agent.crash_backoff / CLEARML_AGENT_CRASH_BACKOFF for exponential backoff of the daemon crash loop
  • Fix --dynamic-gpus over-pulling tasks onto the same GPU, reporting fractional slots as full GPUs, crashing when no available_gpus property exists, and getting stuck after a transient server error (enterprise server required)
  • Fix explicitly empty --entrypoint "" being stripped by agent.forbidden_docker_args
  • Fix conda update_envs regression
  • Fix slow startup when the fileserver is disabled (skip external IP lookup when API host is set)
  • Fix compiled bootstrap Python provisioning in air-gapped environments and when CLEARML_AGENT_SKIP_PYTHON_ENV_INSTALL is set but the image has no Python
  • Document CLEARML_BOOTSTRAP_UV_EXEC (#270)
  • Use ssl_minimum_version with urllib3 v2+ (removes deprecation warning)
  • Bump urllib3>=2.8.0 (Python 3.10+), virtualenv>=21.14.13 (Python 3.9+) and pyjwt>=2.13.0 for security fixes
  • Kubernetes glue
    • Report running tasks as pending again when their pod is replaced
    • Re-enqueue tasks rejected by a namespace ResourceQuota instead of failing them
    • Abort tasks whose running pod was deleted, and clean up pods of tasks stopped before they started running (enterprise server required)
    • Add opt-in node capacity check before launching a task (CLEARML_K8S_GLUE_CHECK_NODE_RESOURCES) (enterprise server required)
    • Add per-task extra containers (sidecars) (enterprise server required)
    • Tear down multi-node tasks whose pods cannot all be scheduled together (enterprise server required)
    • Add clearml-task-id pod label
    • Redact auth tokens and secrets from kubectl output shown in the task log
    • Fix GPU fractions not reported when the pod has CPU/memory limits (enterprise server required)
    • Fix crash on non-ASCII init scripts
    • Fix !!str YAML tag leaking into template values