Pull requests / #912
#912 setup: show the GPU's PCIe link in step 1, warn on fewer lanes than the card has
closed · @1872183316 · 0 comments · View on GitHub
BenchmarksSetup & installAMD / HIPNVIDIA / CUDAModels & quants
Description
Step 1 (and `--check`) says the GPU's PCIe link, and warns when the card runs on fewer lanes than it has. On an X99 board with an RTX 4060 Ti (`--check`, this branch): ``` === Step 1: checking your PC === [ok] GPU: NVIDIA GeForce RTX 4060 Ti, 16.0 GB VRAM, compute capability 8.9, driver 580.178.04 [ok] RAM: 94 GB [ok] CPU: Intel(R) Xeon(R) CPU E5-2673 v3 @ 2.40GHz (AVX2) [ok] PCIe: 3.0 x8 (the card supports 4.0, the board 3.0) - up to ~8 GB/s to the GPU: long prompts are read slower than on a PCIe 4.0 x16 PC (their experts are copied over it); the engine measures the link at start for the output speed ``` and with a card on fewer lanes than it has: ``` [!] the GPU runs on 4 of its 16 PCIe lanes: is it in the right slot (the one wired x16), fully seated, and not sharing lanes with an M.2 drive? (the BIOS can say) ``` Why: the prompt path streams experts over the link, so its rate decides prompt speed (on this PC 714 tok/s for a 3,454-token prompt on Q2_0, against 2,650 at 32K on the RTX 5070), and a card in the wrong slot or sharing lanes is a common and silent loss on older boards. How: `nvidia-smi --query-gpu=pcie.link.gen.max,pcie.link.gen.gpumax,pcie.link.gen.hostmax,pcie.link.width.current, pcie.link.width.max`. The current generation is not asked: an idle card drops to a lower one (this one reported gen 1 at idle). NVIDIA only (AMD gets no line); nothing changes when nvidia-smi does not answer. Tests: `tools/test_setup_pcie.py` (4: parsing, an unparsable answer, a board-limited slow link, a fast link with no extra text, a card on 4 of 16 lanes). All `tools/test_setup_*.py` pass. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
Related on strata.com
Editorial links to help you install, pick models, or read release notes — not part of the upstream thread.