vLLM-MoetCopyright 2026 the vLLM-Moet authors This product includes software developed for serving DeepSeek-V4-Flash on a singleSM120 GPU: 2-bit MoE expert quantization, an FP4 "delta" tier, and hand-written SASSkernels (the `cubit` toolchain). Built on and incorporating: - vLLM (https://github.com/vllm-project/vllm), Apache License 2.0,  Copyright contributors to the vLLM project. The base is the official tag v0.24.0;  `patch/vllm-moet-v0.24.0.patch` is applied on top of it. DeepSeek-V4 model support,  sparse-MLA / MTP / indexer and SM120 enablement ship in upstream vLLM and were authored  by vLLM contributors — not by this project. This project's own contribution is the  2-bit MoE + FP4 delta cache + confidence gate + the cubit kernels, plus small SM120  fixes to the upstream base (carried in the patch). - DeepGEMM (https://github.com/deepseek-ai/DeepGEMM), MIT License, Copyright (c) 2025  DeepSeek. Used at runtime on SM120 — FP8 dense/einsum GEMMs, the indexer's MQA logits,  hc_prenorm, and the stock DEEPGEMM_MXFP4 MoE path the MTP drafter runs on — via the  nv-dev pin `a6b593d2`, whose SM120 backend was authored by the NVIDIA DevTech APAC team  (deepseek-ai/DeepGEMM PR #324). The prebuilt wheel  `patch/deep_gemm-2.5.0+a6b593d-cp312-cp312-linux_x86_64.whl` is redistributed here under  that MIT license and copyright. - The DeepSeek-V4-Flash model weights are NOT included or redistributed. Obtain the  official checkpoint separately; the 2-bit planes are built from it at load time. cubit (the SASS assembler and the hand-written kernels under `kernels/`) and the 2-bitMoE + FP4-delta quantization (`patch/vllm-new-files/`) are the original contribution ofthis project.