The defining engineering story of PoCSmith is completing a real 7B fine-tune on an RTX 4050 laptop GPU with 6GB of VRAM. The run finished in 3 hours 17 minutes across 885 steps (295 per epoch) at ~12 seconds per step, with evaluation every 100 steps, holding VRAM at 5.9/6.0GB (96%) and the GPU around 73°C, with no OOM crashes and roughly 228MB of headroom.

That result rests on a stack of deliberate memory optimizations: 4-bit NF4 quantization (13GB→4GB), gradient checkpointing (recompute instead of store activations), a paged 8-bit optimizer (offload state to RAM), bf16 precision (native on RTX 40-series), batch size 1 with gradient accumulation 4, and disabled packing for predictable per-step memory. The documented challenges were toolchain and bookkeeping rather than modeling: a TRL API migration from TrainingArguments to SFTConfig, a Hugging Face cache-directory fix (copying files to the correct commit-hash folder), and accepted eval-time VRAM spikes that the 228MB headroom absorbed. The lesson the project demonstrates is that specializing a 7B model is now an evening’s work on consumer hardware, provided the memory budget is characterized rather than guessed.