neat-elevator-1 / config.json
quik-models's picture
Upload autoresearch experiment (val_bpb=2.415760)
7af3a9f verified
Raw
History Blame Contribute Delete
603 Bytes
{
"attention_backend": "sdpa",
"capacity_factor": 1.25,
"compute_dtype": "torch.bfloat16",
"expert_dim": 2048,
"img_size": 512,
"n_embd": 512,
"n_head": 4,
"n_kv_head": 4,
"n_layer": 8,
"num_experts": 8,
"num_experts_per_tok": 2,
"num_shared_experts": 1,
"patch_size": 14,
"projector_hidden": 1024,
"sequence_len": 2048,
"use_activation_checkpointing": true,
"use_expert_choice": false,
"use_moe": false,
"use_vision": true,
"vision_dim": 128,
"vision_heads": 4,
"vision_layers": 4,
"vocab_size": 16384,
"window_pattern": "SSSL"
}