-
DMI vs. vLLM Hidden State Extraction: 13x More Data with 15-17x Lower Overhead
On Qwen3-4B prefill, DMI captures much more internal data while keeping overhead far lower.
-
Beyond Hidden States: Comparing DMI with vLLM's New Extraction Feature
vLLM can now extract hidden states. DMI captures a broader internal-state trace across prefill and decode.
-
Doing Internal-State Research on vLLM? DMI Now Gives You the Tensors
DMI brings internal-state capture to the vLLM inference path for test-time research.