Skip to main content

​
RunPod GLM-5.3 Flash real-infrastructure evidence

Date: 2026-08-27
Verdict: not qualified
Run: glm53-runpod-20260827-r7 This record preserves a bounded paid negative result. It must not be cited as proof that InferCrane serves GLM-5.3 Flash. The raw, credential-free capture remains outside the repository at /tmp/infercrane-e2e-report/glm-acceptance/glm53-runpod-20260827-r7/ on the qualification host.

​
Exact attempted tuple

​
Measured boundary

  • Durable deploy submission started at 2026-08-27T00:41:41Z.
  • RunPod created Pod drgz7edz8pezmb at 2026-08-27T00:42:14.901Z.
  • Provider system logs showed the OCI transfer beginning at 00:42:33Z, extraction completing by approximately 00:45:28Z, and the container model snapshot fetch beginning at 00:45:32.188Z. The image path therefore consumed about 3m14s from Pod creation.
  • The pinned Hugging Face snapshot requested 72 files without an authenticated Hub token. It did not complete before the paid watchdog boundary. GPU and GPU-memory utilization were still zero at 01:01:29Z, consistent with model transfer rather than engine initialization.
  • The suite stopped at 01:49:55Z with exit code 124. The approximate paid exposure for this Pod was 68 minutes at the provider-observed $18.36/hour; provider billing remains authoritative.
Because the runtime never became ready, deploy-to-ready, TTFT, TPOT, request p95 at concurrency 1/8/32, streaming versus buffered behavior, model identity from /v1/models, and Release Guard behavior were not measured. They remain required before any support or performance claim.

​
Durable-operation and cleanup evidence

  • The foreground CLI was deliberately terminated after submission. Its captured exit was non-zero, while the persisted operation continued and retained the deterministic provider resource name.
  • The control plane was restarted during the unresolved deploy and resumed the durable operation.
  • Budget-stop cancellation was accepted for operation 1c4820f92983ffdf7e0a79e167e79bff.
  • At 2026-08-27T01:50:09Z, direct provider inventory returned zero Pods and zero Serverless endpoints. InferCrane deployment and orphan inventories were also empty.

​
Engineering conclusions

  1. The native Pod lifecycle, deterministic adoption, CLI-disconnect durability, process-restart durability, cancellation, and zero-resource teardown crossed a real provider boundary.
  2. An uncached 300+ GiB model transfer is not an acceptable default launch path. It is also not evidence that the runtime image or model is incompatible.
  3. The operation emitted repeated semantically identical checkpoints while waiting. The follow-up implementation deduplicates those events and exposes the persisted current step to the CLI.
  4. The general remediation is exact-model persistent cache/prewarm support, not a GLM-only recipe: native RunPod Pods can bind a verified model@commit network volume, standard Hub downloads use the persistent path, and filesystem-materialized profiles honor INFERCRANE_MODEL_DIR.
  5. A fresh paid qualification must start from that prepared volume, use an authenticated Hub path when transfer is required, and repeat the complete request/Release Guard/cleanup matrix.
The remediation described above was implemented after the paid run and therefore was not exercised by this evidence. Its contract tests belong to the later repository commit; real cache timing and a ready model tuple remain unqualified until a separate bounded run proves them.