Skip to content

seed_oss

seed_oss

Seed-OSS on vLLM (vllm.model_executor.models.seed_oss).

Llama's names and Llama's fused block, with vLLM's own classes: the block takes and returns the stream as (hidden_states, residual). On a checkpoint that ties its embeddings vLLM's lm_head is the embedding module itself, which still unembeds through the logits processor.

RENAME module-attribute

RENAME = {'model.embed_tokens': 'embed_tokens', 'model.layers': 'layers', 'model.norm': 'norm'}

ENVOYS module-attribute

ENVOYS = {SeedOssDecoderLayer: Layer, SeedOssAttention: Attention, SeedOssMLP: Mlp}

Layer

Layer(*args: Any, **kwargs: Any)

Bases: FusedLayer

The decoder block: forward(positions, hidden_states, residual) -> (hidden_states, residual).