Radar ·
vLLM cuts memory by loading Qwen PLE rows selectively
A report published on November 1, 2026 says vLLM now reads only the required rows of Qwen’s 51B PLE table from the original safetensors file. On a DGX Spark GB10, loaded memory fell from 99.09 to 72.09 GiB, with decode within about 3.5% at c=1.
A report published on November 1, 2026 says vLLM now reads only the required rows of Qwen’s 51B PLE table from the original safetensors file. On the cited DGX Spark GB10, loaded memory fell from 99.09 to 72.09 GiB, while the decode speed difference stayed within approximately 3.5% at c=1. According to the report, this may make it easier to serve large models with PLE on memory-constrained hardware.
To confirm the figures and understand the test conditions, consult the original publication and check the configuration, memory metric, and meaning of c=1. If you use AI to study or apply this result, avoid including personal data or internal documents, and verify conclusions against documentation and your own tests.
Original source: vLLM lê linhas PLE do Qwen dos safetensors para reduzir memória