wandering-thunder-99 / config.json
quik-models's picture
Upload autoresearch experiment (val_bpb=1.013770)
4ff6f71 verified
Raw
History Blame Contribute Delete
440 Bytes
{
"attention_backend": "sdpa",
"capacity_factor": 1.0,
"compute_dtype": "torch.bfloat16",
"expert_dim": 2048,
"n_embd": 512,
"n_head": 4,
"n_kv_head": 4,
"n_layer": 8,
"num_experts": 8,
"num_experts_per_tok": 2,
"num_shared_experts": 2,
"sequence_len": 2048,
"use_activation_checkpointing": true,
"use_expert_choice": true,
"use_moe": true,
"vocab_size": 16384,
"window_pattern": "SSSL"
}