DeepSeek-V4-Flash-0731 on 8x RTX 3090 (SM86) with vLLM — verified FP8 serving, benchmarks, build guide, and reproducible release.
-
Updated
Aug 9, 2026 - Python
DeepSeek-V4-Flash-0731 on 8x RTX 3090 (SM86) with vLLM — verified FP8 serving, benchmarks, build guide, and reproducible release.
Empirical LLM inference profiling on NVIDIA RTX A400 (Ampere SM86, 4GB): roofline ridge point, SDPA kernels, batching, CUDA Graphs, and WDDM PCIe paging telemetry.
Native NVIDIA DLSS Frame Generation on GeForce RTX 30 in ELDEN RING - DLSS-G on Ampere (sm_86), with no FSR fallback. Combines ERSS-FG with dlssg_for_sm86, adds a settings profile and a verifier that proves from the logs which frame generation actually ran.
To associate your repository with the sm86 topic, visit your repo's landing page and select "manage topics."