INNER CODE UNIT · Python

MXFP4_BYTES_PER_PARAM

FareedKhan-dev/kimi-k3-in-c · tools/budget.py:44

MXFP4_BYTES_PER_PARAM = 0.5 + 1.0 / 32.0


def classify(name: str) -> str:
    if ".block_sparse_moe.experts." in name:
        return "routed experts (STREAMABLE)"
    if ".block_sparse_moe.shared_experts." in name:
        return "shared experts (resident, dense)"
    if ".block_sparse_moe.routed_expert_" in name:
        return "MoE latent up/down/norm (resident)"
    if ".block_sparse_moe.gate" in name:
        return "MoE router (resident)"
    if ".self_attn." in name:
        return "attention (resident)"
    if "_res_norm" in name or "_res_proj" in name:
        return "attn-res aggregators (resident)"
    if "layernorm" in name or name.endswith(".norm.weight"):
        return "layer norms (resident)"

View source record →

📰 Research Paper
Loading…
⏳ Fetching content…