llama 3.3 70b q4 generates valid kubernetes yaml but wrong resource limits past 24k context
ok so tested at 24.2k yesterday generating k8s deployment yaml and the syntax is perfect but resource limits are just wrong like it generates `limits: {memory: "512Mi", cpu: "500m"}` when the actual requirement in the prompt was 2gb memory and 2 cpu cores. the yaml structure is flawless, indentation perfect, but the actual values are plausible-but-wrong tested at 17k context on same prompts and values were correct. seems like another case of syntax preservation but semantic drift past ~20k
we hit same thing at 25.1k with k8s deployments - memory limits were completely wrong, like 256Mi when actual requirement is 4Gi. yaml syntax was perfect tho
tested llama 3.3 70b q4_k_m at 25.3k yesterday with k8s statefulset yaml and hit exact same thing - cpu requests were set to 100m when schema specifies 500m, memory limits completely wrong. yaml validated fine but deployment failed
can you share the exact yaml structure that triggered this? we're running llama 3.3 70b q4 for k8s manifest generation in a ci pipeline and if resource limits are drifting past 24k we need to know now before it hits prod
yeah share the yaml, we're using llama 3.3 70b q4 in ci for k8s manifests and if limits are drifting past 24k that's a production incident waiting to happen
yeah share yaml, we're using q4 for k8s in ci and if limits drift past 24k that breaks everything
hit this yesterday on llama 3.3 70b q4_k_m at 24.7k context generating k8s yaml for a postgres statefulset. memory limits were set to 512Mi when our actual schema specifies 2Gi, and cpu requests were completely inverted (requested more than limit). yaml syntax was perfect though, passed yamllint with zero errors.
yeah we hit this too.... the model knows the schema structure but past ~24k it just starts guessing plausible values instead of following the actual spec. cost us like 6 hours of debugging in prod last week