search

sustcsonglin / flash-linear-attention

Efficient implementations of state-of-the-art linear attention models in Pytorch and Triton
MIT License
1.33k stars 68 forks source link

Can you provide the configuration of the 340M parameter version models? #15

Closed helldog-star closed 6 months ago

helldog-star commented 6 months ago

Thanks！

yzhangcs commented 6 months ago

@helldog-star For GLA

{
  "attn_mode": "chunk",
  "bos_token_id": 1,
  "eos_token_id": 2,
  "expand_k": 0.5,
  "expand_v": 1,
  "fuse_cross_entropy": true,
  "fuse_norm": true,
  "hidden_act": "swish",
  "hidden_ratio": 4,
  "hidden_size": 1024,
  "initializer_range": 0.02,
  "intermediate_size": null,
  "max_position_embeddings": 2048,
  "model_type": "gla",
  "num_heads": 4,
  "num_hidden_layers": 24,
  "rms_norm_eps": 1e-06,
  "tie_word_embeddings": true,
  "transformers_version": "4.40.0",
  "use_cache": true,
  "use_gk": true,
  "use_gv": false,
  "vocab_size": 32000
}

and for RetNet

{
  "attn_mode": "chunk",
  "bos_token_id": 1,
  "eos_token_id": 2,
  "expand_k": 1,
  "expand_v": 2,
  "fuse_cross_entropy": true,
  "fuse_norm": true,
  "hidden_act": "swish",
  "hidden_ratio": 2,
  "hidden_size": 1024,
  "max_position_embeddings": 2048,
  "model_type": "retnet",
  "num_heads": 4,
  "num_hidden_layers": 24,
  "rms_norm_eps": 1e-06,
  "tie_word_embeddings": true,
  "transformers_version": "4.40.0",
  "use_cache": true,
  "vocab_size": 32000
}