Samba-7B

No samba-7b config.json is reachable on the HF hub under sambanovasystems (the org now hosts unrelated repos). The arXiv:2406.07522 paper documents the sliding-window attention design (window 2048, RoPE base 10000) but does not publish the head counts, so the KV-geometry fields are left null. This is a source-retrieval gap: the paper is cited and the model exists, the config does not. 3.8B parameters, 3.2T training tokens per the Samba paper (arXiv:2406.07522). Layer-wise interleaving of Mamba with SwiGLU and sliding-window attention (SWA). SWA window size w=2048 chosen because 'FlashAttention 2 has the same training speed as Mamba's selective parallel scan at the sequence length of 2048'. Reports 3.73x higher throughput vs Transformers with GQA for 128K prompts and 3.64x speedup generating 64K tokens. The exact Mamba:SWA layer ratio and KV head geometry are not stated in the abstract and the config.json was not accessible, so kv_cache_bytes_per_token is null. The throughput claims are the paper's own, measured on the authors' hardware. CONTEXT_SCALING_METHOD (none-verified): both Samba mirrors return HTTP 401 to an unauthenticated fetch; config unreadable.