Vadim21221 commited on
Commit
970d7a4
·
verified ·
1 Parent(s): fce1367

Upload 6 files

Browse files
pile_mul_fractal_2_8/config.json ADDED
@@ -0,0 +1,51 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model_name": "google/gemma-2-2b",
3
+ "layer": 12,
4
+ "hook_point": "resid_post",
5
+ "act_size": 2304,
6
+ "sae_type": "mul_fractal_topk",
7
+ "dict_size": 65536,
8
+ "aux_penalty": 0.03125,
9
+ "input_unit_norm": true,
10
+ "batch_norm_on_queries": false,
11
+ "affine_batch_norm": false,
12
+ "linear_heads": 0,
13
+ "topk2": 50,
14
+ "topk1": 50,
15
+ "topk2_warmup_steps_fraction": 0.0,
16
+ "start_topk2": 50,
17
+ "topk1_warmup_steps_fraction": 0.0,
18
+ "start_topk1": 50,
19
+ "topk2_aux": 512,
20
+ "cartesian_op": "mul",
21
+ "router_depth": 2,
22
+ "router_tree_width": null,
23
+ "num_mkeys": 2,
24
+ "num_nkeys": 8,
25
+ "num_heads": 4096,
26
+ "n_batches_to_dead": 10,
27
+ "lr": 0.0008,
28
+ "bandwidth": 0.001,
29
+ "l1_coeff": 0.0018,
30
+ "num_tokens": 799634235,
31
+ "seq_len": 1024,
32
+ "model_batch_size": 64,
33
+ "num_batches_in_buffer": 5,
34
+ "max_grad_norm": 1.0,
35
+ "batch_size": 8192,
36
+ "weight_decay": 0.0,
37
+ "warmup_fraction": 0.1,
38
+ "scheduler_type": "cosine_with_min_lr",
39
+ "device": "cuda",
40
+ "dtype": "torch.float32",
41
+ "sae_dtype": "torch.float32",
42
+ "dataset_path": "cerebras/SlimPajama-627B",
43
+ "wandb_project": "turbo-llama-lens",
44
+ "enable_wandb": true,
45
+ "sae_name": "sae",
46
+ "seed": 42,
47
+ "performance_log_steps": 100,
48
+ "save_checkpoint_steps": 15000000,
49
+ "wandb_run_suffix": "exp80_bench",
50
+ "sweep_pair": "{'dict_size': 65536, 'num_heads': 4096, 'num_mkeys': 2, 'num_nkeys': 8, 'num_tokens': 799634235, 'sae_type': 'mul_fractal_topk', 'start_topk1': 50, 'start_topk2': 50, 'topk1': 50}"
51
+ }
pile_mul_fractal_2_8/sae.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f5c54175b1ab9e3a9e4c40601edd77d228932af7933c15d6dd5d421fdd1ad426
3
+ size 981904728
pile_mul_fractal_4_4/config.json ADDED
@@ -0,0 +1,51 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model_name": "google/gemma-2-2b",
3
+ "layer": 12,
4
+ "hook_point": "resid_post",
5
+ "act_size": 2304,
6
+ "sae_type": "mul_fractal_topk",
7
+ "dict_size": 65536,
8
+ "aux_penalty": 0.03125,
9
+ "input_unit_norm": true,
10
+ "batch_norm_on_queries": false,
11
+ "affine_batch_norm": false,
12
+ "linear_heads": 0,
13
+ "topk2": 50,
14
+ "topk1": 50,
15
+ "topk2_warmup_steps_fraction": 0.0,
16
+ "start_topk2": 50,
17
+ "topk1_warmup_steps_fraction": 0.0,
18
+ "start_topk1": 50,
19
+ "topk2_aux": 512,
20
+ "cartesian_op": "mul",
21
+ "router_depth": 2,
22
+ "router_tree_width": null,
23
+ "num_mkeys": 4,
24
+ "num_nkeys": 4,
25
+ "num_heads": 4096,
26
+ "n_batches_to_dead": 10,
27
+ "lr": 0.0008,
28
+ "bandwidth": 0.001,
29
+ "l1_coeff": 0.0018,
30
+ "num_tokens": 999238222,
31
+ "seq_len": 1024,
32
+ "model_batch_size": 64,
33
+ "num_batches_in_buffer": 5,
34
+ "max_grad_norm": 1.0,
35
+ "batch_size": 8192,
36
+ "weight_decay": 0.0,
37
+ "warmup_fraction": 0.1,
38
+ "scheduler_type": "cosine_with_min_lr",
39
+ "device": "cuda",
40
+ "dtype": "torch.float32",
41
+ "sae_dtype": "torch.float32",
42
+ "dataset_path": "cerebras/SlimPajama-627B",
43
+ "wandb_project": "turbo-llama-lens",
44
+ "enable_wandb": true,
45
+ "sae_name": "sae",
46
+ "seed": 42,
47
+ "performance_log_steps": 100,
48
+ "save_checkpoint_steps": 15000000,
49
+ "wandb_run_suffix": "exp80_bench",
50
+ "sweep_pair": "{'dict_size': 65536, 'num_heads': 4096, 'num_mkeys': 4, 'num_nkeys': 4, 'num_tokens': 999238222, 'sae_type': 'mul_fractal_topk', 'start_topk1': 50, 'start_topk2': 50, 'topk1': 50}"
51
+ }
pile_mul_fractal_4_4/sae.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c1e3d3e6b52962698843f4d715c6d74de506b0fca85b1162650e0029f81c37e7
3
+ size 906374488
pile_topk_l0_50_65k/config.json ADDED
@@ -0,0 +1,51 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model_name": "google/gemma-2-2b",
3
+ "layer": 12,
4
+ "hook_point": "resid_post",
5
+ "act_size": 2304,
6
+ "sae_type": "topk",
7
+ "dict_size": 65536,
8
+ "aux_penalty": 0.03125,
9
+ "input_unit_norm": true,
10
+ "batch_norm_on_queries": false,
11
+ "affine_batch_norm": false,
12
+ "linear_heads": 0,
13
+ "topk2": 50,
14
+ "topk1": 50,
15
+ "topk2_warmup_steps_fraction": 0.0,
16
+ "start_topk2": 50,
17
+ "topk1_warmup_steps_fraction": 0.0,
18
+ "start_topk1": 50,
19
+ "topk2_aux": 512,
20
+ "cartesian_op": "mul",
21
+ "router_depth": 2,
22
+ "router_tree_width": null,
23
+ "num_mkeys": null,
24
+ "num_nkeys": null,
25
+ "num_heads": -1,
26
+ "n_batches_to_dead": 10,
27
+ "lr": 0.0008,
28
+ "bandwidth": 0.001,
29
+ "l1_coeff": 0.0018,
30
+ "num_tokens": 500000000,
31
+ "seq_len": 1024,
32
+ "model_batch_size": 64,
33
+ "num_batches_in_buffer": 5,
34
+ "max_grad_norm": 1.0,
35
+ "batch_size": 8192,
36
+ "weight_decay": 0.0,
37
+ "warmup_fraction": 0.1,
38
+ "scheduler_type": "cosine_with_min_lr",
39
+ "device": "cuda",
40
+ "dtype": "torch.float32",
41
+ "sae_dtype": "torch.float32",
42
+ "dataset_path": "cerebras/SlimPajama-627B",
43
+ "wandb_project": "turbo-llama-lens",
44
+ "enable_wandb": true,
45
+ "sae_name": "sae",
46
+ "seed": 42,
47
+ "performance_log_steps": 100,
48
+ "save_checkpoint_steps": 15000000,
49
+ "wandb_run_suffix": "exp80_bench",
50
+ "sweep_pair": "{'dict_size': 65536, 'num_tokens': 500000000, 'sae_type': 'topk', 'start_topk1': 50, 'start_topk2': 50, 'topk1': 50}"
51
+ }
pile_topk_l0_50_65k/sae.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3a23fac935d5096f0b297a19fc1f885596589d33f7e50c651138aa54360f7464
3
+ size 1208495512