LLM Ops & Infrastructure.

Operationalizing and scaling large language model deployments. Covers 12 sub-skills.

• LLM Ops & Infrastructure · 1 of 3

  Fine-Tuning Workflows

  A team needs to fine-tune a 13B model on a single 24GB GPU. Which approach is the standard fit?

  AQLoRA with 4-bit base weights and LoRA adaptersBAdapter-fusion across multiple full-rank checkpointsCPipeline parallel training split across 4 GPU shardsDFull-parameter FT in fp16 with gradient checkpoint

• LLM Ops & Infrastructure · 2 of 3

  Fine-Tuning Workflows

  A team has 50k noisy SFT rows scraped from logs. What's the leverage move before launching a training run?

  AFilter and de-duplicate to a smaller high-quality subsetBSwitch optimizer to a noise-tolerant variant for the first epochCIncrease the global batch size to dilute the bad-row signalDTrain longer and rely on validation loss to mask the noise

• LLM Ops & Infrastructure · 3 of 3

  Fine-Tuning Workflows

  A team running multi-tenant adapter serving on vLLM sees latency spike every time a new adapter is requested for the first time. What is the primary contributor?

  AKV cache invalidation triggered by a base-model parameter refreshBCold-load of adapter weights and graph compilation on first requestCTokenizer reset between tenants forcing a full vocab table rebuildDContinuous batching disabled for adapter-routed requests by default