gte-large on GeForce RTX 2080 Ti

Can I run gte-large on a GeForce RTX 2080 Ti?

Yes. The best fit is Q8_0 at 65,536 context with f16 KV — 7.11 GiB of 10.23 GiB usable, leaving 3.12 GiB headroom. Expect roughly 68 tokens/sec (modeled, ±12.9%).

From the file· weights summed from filesFrom the file· KV computed per layerPredicted· speed and compute buffer
KV cache dtypef16q8_0q4_0— 11 GB card, 10.23 GiB usable after overhead

Every quantization at every context

total memory required; green fits, red does not
QuantWeights●4K◐8K◐16K◐32K◐64K◐128K◐
Q8_00.33 GiB1.51.92.64.17.113.1
Q6_K0.26 GiB1.41.82.54.07.013.0
Q5_K_M0.23 GiB1.41.82.54.07.013.0
Q5_00.22 GiB1.41.72.54.07.013.0
Q5_K_S0.22 GiB1.41.72.54.07.013.0
Q4_K_M0.20 GiB1.41.72.54.07.013.0
Q4_K_S0.19 GiB1.31.72.54.07.013.0
Q4_00.19 GiB1.31.72.54.07.013.0
Q3_K_L0.18 GiB1.31.72.54.07.013.0
Q3_K_M0.17 GiB1.31.72.43.96.912.9
Q3_K_S0.15 GiB1.31.72.43.96.912.9
Q2_K0.13 GiB1.31.72.43.96.912.9
●From the file◐Predictedwhat these mean

Figures are GiB of total memory: weights plus KV cache plus compute buffer and backend overhead. Weights and KV are near-exact; the overhead term is modeled. Hover any cell for the breakdown.

Run it

The best-fitting configuration above, as a command:

llama-cli -hf thenlper/gte-large \
  --ctx-size 65536 \
  -ngl auto

Recent llama.cpp defaults to --fit on with -ngl auto, so it will size the offload for you. The question worth your attention is not how many layers to offload but what context and quantization you are willing to live with — which is what the grid above is for.

Why other calculators disagree

A parameters × bits ÷ 8 estimate ignores two things that dominate at long context. First, the weights themselves are not the nominal rate — quantizations are mixtures, so the real file is consistently larger than the label implies. Second, the KV cache grows linearly with context and, past about 32K, becomes larger than the weights for many models.