Most "AI deployment" tutorials show you the happy path.

Here's what actually happened when I deployed a 120B parameter model on bare-metal Blackwell hardware — and the 3 mistakes that almost killed the project.

THE GOAL: Run nvidia/nemotron-3-super-120b on a single DGX Spark (128GB unified memory) for a production agentic stack.

THE FIRST WALL: Memory math.

120GB of FP8 weights + 8GB OS overhead = 128GB. That's 100% of the card with zero room left for the KV-cache (the model's working memory). The result wasn't a slow model — it was a full kernel panic that froze SSH access entirely.

Fix: switch to NVFP4 (4-bit). Same model, ~72GB footprint, 56GB of breathing room.

THE SECOND WALL: A symlink trap.

NVIDIA NIM and HuggingFace cache models using symlinks — shortcuts pointing to a separate "blobs" folder. I mounted only the snapshot folder into Docker. The shortcuts broke the moment the container started, because the blobs folder wasn't visible inside.

Fix: mount the parent hub directory, not just the snapshot.

THE THIRD WALL: Blackwell SM121 rejected standard containers outright with "illegal instruction" errors. Generic vLLM builds don't know this architecture exists yet.

Fix: vLLM v0.17.1-cu130 specifically, with the Marlin GEMM backend for the 4-bit kernels.

THE RESULT:

→ 16.4 tokens/sec sustained throughput → Stable 64k-token context window → Full OpenClaw agentic stack running on top (Discord, Telegram, tool-calling) → What should have taken 5 days took under 2 hours once I knew the fixes

I documented all 11 blockers in a 900-line runbook so the next person doesn't lose 5 days to the same mistakes.

If you're trying to deploy a large model on Blackwell hardware and hitting walls — happy to talk through what I learned. Drop a comment or DM.