LightOnOCR-2 memory spike sparks GLM-OCR caution

Hot take: this looks more like expected multimodal inference behavior than a setup mistake, especially once generation starts and caches explode.

• –Post-prompt spikes are common in OCR VLMs because image tokens, long outputs, and KV cache can dominate memory more than raw model weights.
• –LightOnOCR-2’s own guidance emphasizes rendering constraints and vLLM runtime flags, which suggests serving/config choices strongly affect peak memory.
• –GLM-OCR is marketed as a compact 0.9B model with efficiency-focused decoding, but it is still in the same document-VLM class and can spike on large pages or long outputs.
• –On 16-18GB unified-memory laptops, reliable local runs usually require stricter page batching, lower pixel budgets, and tighter token limits.