OOOPS. Forgot to add model config
Browse files- model_config.yaml +58 -0
model_config.yaml
ADDED
|
@@ -0,0 +1,58 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
name: multiscreen-pure-9l256-tadam-4b
|
| 2 |
+
|
| 3 |
+
block_size: 1024
|
| 4 |
+
vocab_size: 4096
|
| 5 |
+
padded_vocab_size: 4096
|
| 6 |
+
|
| 7 |
+
n_layer: 9
|
| 8 |
+
n_embd: 256
|
| 9 |
+
n_head: 8
|
| 10 |
+
head_size: 32
|
| 11 |
+
n_query_groups: 1
|
| 12 |
+
|
| 13 |
+
intermediate_size: 1104
|
| 14 |
+
|
| 15 |
+
bias: false
|
| 16 |
+
attn_bias: false
|
| 17 |
+
|
| 18 |
+
norm_class_name: RMSNorm
|
| 19 |
+
norm_eps: 1.0e-05
|
| 20 |
+
parallel_residual: false
|
| 21 |
+
shared_attention_norm: false
|
| 22 |
+
post_attention_norm: false
|
| 23 |
+
post_mlp_norm: false
|
| 24 |
+
|
| 25 |
+
mlp_class_name: TileRoutedEulerGLUMLP
|
| 26 |
+
|
| 27 |
+
sparse_mlp_num_groups: 4
|
| 28 |
+
sparse_mlp_min_active_groups: 4
|
| 29 |
+
sparse_mlp_max_active_groups: 4
|
| 30 |
+
|
| 31 |
+
multiscreen_enabled: true
|
| 32 |
+
multiscreen_num_heads: 8
|
| 33 |
+
multiscreen_key_dim: 16
|
| 34 |
+
multiscreen_value_dim: 128
|
| 35 |
+
multiscreen_query_chunk_size: 128
|
| 36 |
+
multiscreen_window_threshold: 256
|
| 37 |
+
multiscreen_landmark_stride: 1
|
| 38 |
+
multiscreen_layer_interval: 0
|
| 39 |
+
multiscreen_mlp_enabled: false
|
| 40 |
+
multiscreen_triton_inference: true
|
| 41 |
+
|
| 42 |
+
sliding_window_size: 256
|
| 43 |
+
sliding_window_indices:
|
| 44 |
+
- 1
|
| 45 |
+
- 1
|
| 46 |
+
- 1
|
| 47 |
+
- 0
|
| 48 |
+
- 1
|
| 49 |
+
- 1
|
| 50 |
+
- 1
|
| 51 |
+
- 0
|
| 52 |
+
- 1
|
| 53 |
+
|
| 54 |
+
rotary_percentage: 1.0
|
| 55 |
+
rope_base: 10000
|
| 56 |
+
|
| 57 |
+
lm_head_bias: false
|
| 58 |
+
scale_embeddings: false
|