vLLM lets compressed-tensors block-FP8 weights survive an engine restart VLLM merged CUDA IPC weight-cache restart support on November 15th, 2026 for serialized compressed-tensors models using 128-by-128 block FP8 linear weights, letting a restarted engine map processed GPU weights and scales held through CUDA inter-process communication instead of rebuilding that prepared storage. The merged path restores the metadata required by the validated CUTLASS block-FP8 configuration, and exact greedy output was checked on a B300. vLLM lets compressed-tensors block-FP8 weights survive an engine restart The merged change lets a restarted engine map processed CUTLASS block-FP8 weights through CUDA IPC, with exact greedy output checked on a B300. By RuntimeWire Staff https://runtimewire.com/author/runtimewire-staff · Published vLLM merged CUDA IPC weight-cache restart support on November 15th, 2026 for serialized compressed-tensors models that use 128-by-128 block FP8 linear weights. The change lets a restarted engine map processed GPU weights and scales held through CUDA inter-process communication instead of rebuilding that prepared storage from scratch. The merged path reuses processed weights and scales through CUDA IPC while restoring the metadata required by the validated CUTLASS block-FP8 configuration - explanatory diagram, not documentary evidence. RuntimeWire / AI-generated diagram.…