Scouter

RL

GitHub · hecate0821.github.io

RL

RL

GQA
MLA
DSA
GDN / KDA

Ring

Ulysses

W [N,K]KN—
X [M,K]K—M
dY [M,N]—NM

TP
DP
EP
CP / DCP
PP
FSDP

    
        

    TP4 × 16TP8 × 8DEP32 × 2DEP64
    8763 / 10.5927 / 8.6998 / 8.01,100 / 7.3
    321,857 / 17.22,645 / 12.13,428 / 9.33,723 / 8.6
    1285,153 / 24.85,713 / 22.48,567 / 14.99,007 / 14.2

    MTP
    EAGLE-3
    DFlash / DFlash2
    DSpark

    —Miles tools/convert_hf_to_{mxfp8,nvfp4}.py
    Miles --fp8-recipe mxfp8 / --fp4-recipe nvfp4 + NVTE_*
    —

    
        

    WxAy
    MX / NV
    RHT
    SR
    PTQ / QAT

    sgl-project/sglang
    sgl-project/sgl-cookbook
    vllm-project/vllm
    lightseekorg/tokenspeed
    DeepInfra
    DeepInferdeepseek-ai/DeepInfer
    pytorch/torchtitan
    fla-org/flash-linear-attention
    TRT-LLMNVIDIA/TensorRT-LLM

    SemiAnalysis InferenceX