transformer_lens.model_bridge.supported_architectures.seed_oss module¶
Seed-OSS architecture adapter.
ByteDance’s Seed-OSS (SeedOssForCausalLM) is a Llama-layout decoder —
RMSNorm + RoPE + GQA + gated MLP under identical module paths — with
config-gated attention/MLP biases (handled by the shared weight machinery)
and no BOS prepending.
- class transformer_lens.model_bridge.supported_architectures.seed_oss.SeedOssArchitectureAdapter(cfg: Any)¶
Bases:
LlamaArchitectureAdapterArchitecture adapter for SeedOssForCausalLM models.