Same model, opposite curves: llama.cpp beats vLLM 7x at 90K context on DGX Spark

niacolhealth · reddit · 2026-09-10

A same-box, same-prompt comparison of Ling-3.0-flash on a 128GB DGX Spark shows the short-prompt ranking reverses with context depth (INT4/vLLM fork vs Q5KM/llama.cpp):

| Starting context | INT4 + vLLM | Q5KM + llama.cpp |

|---|---|---|

| Short prompt | 38.3 tok/s | 35.7 tok/s |

| 45K tokens | 7.9 tok/s | 33.6 tok/s |

| 90K tokens | 4.6 tok/s | 33.2 tok/s |

Setup: 262,144 max context, 103GB memory used; streaming measured with time-to-first-token separated from decode.

Caveats: creator-measured, not independently reproduced; two variables (runtime and quantization) change at once, so no pure vLLM-vs-llama.cpp effect is isolated, and the proposed CUDA-graph explanation is unproven. The repo's later 131,072-context default shouldn't be conflated with this table.

Practical takeaway: the deciding variable is context already present when generation starts — workflows carrying tens of thousands of tokens should favor llama.cpp here.

Original post →

More from Infra

Infra channel →