Hugging Face's logo

AlphaOxO
/
GPT-X2.5-125M-8bits-mlx

Text Generation
MLX
Safetensors
English
gptx2
custom_code
8-bit precision
Model card Files
xet
 Community
AlphaOxO commited on
Commit
a5263cd
·
verified ·
1 Parent(s): da6410f

Upload folder using huggingface_hub

Browse files
Files changed (12) hide show
  1. README.md +7 -0
  2. __init__.py +1 -0
  3. config.json +41 -0
  4. configuration_gptx2.py +40 -0
  5. export_to_hf.py +62 -0
  6. generation_config.json +9 -0
  7. model.safetensors +3 -0
  8. model.safetensors.index.json +702 -0
  9. modeling_gptx2.py +255 -0
  10. tokenizer.json +0 -0
  11. tokenizer_config.json +18 -0
  12. usage.py +22 -0
README.md ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ ---
2
+ language: en
3
+ library_name: mlx
4
+ pipeline_tag: text-generation
5
+ tags:
6
+ - mlx
7
+ ---
__init__.py ADDED
@@ -0,0 +1 @@
 
 
1
+ """GPT-X2.5 Hugging Face model package."""
config.json ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "GPTX2ForCausalLM"
4
+ ],
5
+ "attention_type": "grouped_query",
6
+ "auto_map": {
7
+ "AutoConfig": "configuration_gptx2.GPTX2Config",
8
+ "AutoModelForCausalLM": "modeling_gptx2.GPTX2ForCausalLM"
9
+ },
10
+ "bias": false,
11
+ "bos_token_id": 0,
12
+ "embedding_scale": false,
13
+ "eos_token_id": 0,
14
+ "head_dim": 64,
15
+ "hidden_act": "silu",
16
+ "hidden_size": 576,
17
+ "intermediate_size": 1728,
18
+ "max_position_embeddings": 8192,
19
+ "model_type": "gptx2",
20
+ "num_attention_heads": 9,
21
+ "num_hidden_layers": 30,
22
+ "num_key_value_heads": 3,
23
+ "pad_token_id": 1,
24
+ "qk_norm": false,
25
+ "quantization": {
26
+ "group_size": 64,
27
+ "bits": 8,
28
+ "mode": "affine"
29
+ },
30
+ "quantization_config": {
31
+ "group_size": 64,
32
+ "bits": 8,
33
+ "mode": "affine"
34
+ },
35
+ "rms_norm_eps": 1e-06,
36
+ "rope_theta": 100000.0,
37
+ "tie_word_embeddings": true,
38
+ "torch_dtype": "float32",
39
+ "vocab_size": 32770,
40
+ "xsa_projection": true
41
+ }
configuration_gptx2.py ADDED
@@ -0,0 +1,40 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """GPT-X2.5 model configuration."""
2
+
3
+ from transformers import PretrainedConfig
4
+
5
+
6
+ class GPTX2Config(PretrainedConfig):
7
+ model_type = "gptx2"
8
+
9
+ def __init__(
10
+ self,
11
+ vocab_size=32770,
12
+ hidden_size=576,
13
+ num_hidden_layers=30,
14
+ num_attention_heads=9,
15
+ num_key_value_heads=3,
16
+ head_dim=64,
17
+ intermediate_size=1728,
18
+ max_position_embeddings=8192,
19
+ rope_theta=100000.0,
20
+ rms_norm_eps=1e-6,
21
+ tie_word_embeddings=True,
22
+ xsa_projection=True,
23
+ qk_norm=False,
24
+ embedding_scale=False,
25
+ **kwargs,
26
+ ):
27
+ self.vocab_size = vocab_size
28
+ self.hidden_size = hidden_size
29
+ self.num_hidden_layers = num_hidden_layers
30
+ self.num_attention_heads = num_attention_heads
31
+ self.num_key_value_heads = num_key_value_heads
32
+ self.head_dim = head_dim
33
+ self.intermediate_size = intermediate_size
34
+ self.max_position_embeddings = max_position_embeddings
35
+ self.rope_theta = rope_theta
36
+ self.rms_norm_eps = rms_norm_eps
37
+ self.xsa_projection = xsa_projection
38
+ self.qk_norm = qk_norm
39
+ self.embedding_scale = embedding_scale
40
+ super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)
export_to_hf.py ADDED
@@ -0,0 +1,62 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Convert the final GPT-X2.5 training checkpoint to Hugging Face SafeTensors."""
2
+
3
+ import argparse
4
+ from pathlib import Path
5
+
6
+ import torch
7
+ from safetensors.torch import save_file
8
+
9
+
10
+ EXPECTED_TENSORS = 273
11
+ EXPECTED_UNIQUE_PARAMETERS = 135_032_256
12
+
13
+
14
+ def main():
15
+ parser = argparse.ArgumentParser()
16
+ parser.add_argument("checkpoint", type=Path, help="Training .pt checkpoint")
17
+ parser.add_argument("--output", type=Path, default=Path("model.safetensors"))
18
+ parser.add_argument(
19
+ "--dtype", choices=("bfloat16", "float16", "float32"), default="float32"
20
+ )
21
+ args = parser.parse_args()
22
+
23
+ checkpoint = torch.load(
24
+ args.checkpoint, map_location="cpu", weights_only=False, mmap=True
25
+ )
26
+ if "model" not in checkpoint:
27
+ raise KeyError("Checkpoint does not contain a 'model' state dictionary")
28
+ state = checkpoint["model"]
29
+ if len(state) != EXPECTED_TENSORS:
30
+ raise ValueError(f"Expected {EXPECTED_TENSORS} tensors, found {len(state)}")
31
+ if not torch.equal(state["transformer.wte.weight"], state["lm_head.weight"]):
32
+ raise ValueError("Tied embedding and LM-head tensors differ")
33
+
34
+ # Hugging Face restores lm_head.weight from the tied embedding. SafeTensors
35
+ # intentionally stores the shared tensor only once.
36
+ state = {key: value for key, value in state.items() if key != "lm_head.weight"}
37
+ parameter_count = sum(tensor.numel() for tensor in state.values())
38
+ if parameter_count != EXPECTED_UNIQUE_PARAMETERS:
39
+ raise ValueError(
40
+ f"Expected {EXPECTED_UNIQUE_PARAMETERS:,} unique parameters, "
41
+ f"found {parameter_count:,}"
42
+ )
43
+
44
+ dtype = getattr(torch, args.dtype)
45
+ converted = {
46
+ key: tensor.detach().to(dtype=dtype).contiguous() for key, tensor in state.items()
47
+ }
48
+ args.output.parent.mkdir(parents=True, exist_ok=True)
49
+ save_file(
50
+ converted,
51
+ str(args.output),
52
+ metadata={"format": "pt", "source_step": str(checkpoint.get("step", "unknown"))},
53
+ )
54
+ print(
55
+ f"Saved {args.output} ({parameter_count:,} parameters, {args.dtype}, "
56
+ f"step {checkpoint.get('step', 'unknown')})"
57
+ )
58
+
59
+
60
+ if __name__ == "__main__":
61
+ main()
62
+
generation_config.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token_id": 0,
3
+ "eos_token_id": 0,
4
+ "pad_token_id": 1,
5
+ "do_sample": true,
6
+ "temperature": 0.8,
7
+ "top_p": 0.95,
8
+ "top_k": 50
9
+ }
model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:85997330dbd2f8241a4cc113f5a9ba0d967897256277bfb32968a2595d66e9ea
3
+ size 152080932
model.safetensors.index.json ADDED
@@ -0,0 +1,702 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "metadata": {
3
+ "total_size": 152012304,
4
+ "total_parameters": 135032256
5
+ },
6
+ "weight_map": {
7
+ "model.h.0.attn.k_proj.biases": "model.safetensors",
8
+ "model.h.0.attn.k_proj.scales": "model.safetensors",
9
+ "model.h.0.attn.k_proj.weight": "model.safetensors",
10
+ "model.h.0.attn.o_proj.biases": "model.safetensors",
11
+ "model.h.0.attn.o_proj.scales": "model.safetensors",
12
+ "model.h.0.attn.o_proj.weight": "model.safetensors",
13
+ "model.h.0.attn.q_proj.biases": "model.safetensors",
14
+ "model.h.0.attn.q_proj.scales": "model.safetensors",
15
+ "model.h.0.attn.q_proj.weight": "model.safetensors",
16
+ "model.h.0.attn.v_proj.biases": "model.safetensors",
17
+ "model.h.0.attn.v_proj.scales": "model.safetensors",
18
+ "model.h.0.attn.v_proj.weight": "model.safetensors",
19
+ "model.h.0.ln_1.weight": "model.safetensors",
20
+ "model.h.0.ln_2.weight": "model.safetensors",
21
+ "model.h.0.mlp.w_down.biases": "model.safetensors",
22
+ "model.h.0.mlp.w_down.scales": "model.safetensors",
23
+ "model.h.0.mlp.w_down.weight": "model.safetensors",
24
+ "model.h.0.mlp.w_gate.biases": "model.safetensors",
25
+ "model.h.0.mlp.w_gate.scales": "model.safetensors",
26
+ "model.h.0.mlp.w_gate.weight": "model.safetensors",
27
+ "model.h.0.mlp.w_up.biases": "model.safetensors",
28
+ "model.h.0.mlp.w_up.scales": "model.safetensors",
29
+ "model.h.0.mlp.w_up.weight": "model.safetensors",
30
+ "model.h.1.attn.k_proj.biases": "model.safetensors",
31
+ "model.h.1.attn.k_proj.scales": "model.safetensors",
32
+ "model.h.1.attn.k_proj.weight": "model.safetensors",
33
+ "model.h.1.attn.o_proj.biases": "model.safetensors",
34
+ "model.h.1.attn.o_proj.scales": "model.safetensors",
35
+ "model.h.1.attn.o_proj.weight": "model.safetensors",
36
+ "model.h.1.attn.q_proj.biases": "model.safetensors",
37
+ "model.h.1.attn.q_proj.scales": "model.safetensors",
38
+ "model.h.1.attn.q_proj.weight": "model.safetensors",
39
+ "model.h.1.attn.v_proj.biases": "model.safetensors",
40
+ "model.h.1.attn.v_proj.scales": "model.safetensors",
41
+ "model.h.1.attn.v_proj.weight": "model.safetensors",
42
+ "model.h.1.ln_1.weight": "model.safetensors",
43
+ "model.h.1.ln_2.weight": "model.safetensors",
44
+ "model.h.1.mlp.w_down.biases": "model.safetensors",
45
+ "model.h.1.mlp.w_down.scales": "model.safetensors",
46
+ "model.h.1.mlp.w_down.weight": "model.safetensors",
47
+ "model.h.1.mlp.w_gate.biases": "model.safetensors",
48
+ "model.h.1.mlp.w_gate.scales": "model.safetensors",
49
+ "model.h.1.mlp.w_gate.weight": "model.safetensors",
50
+ "model.h.1.mlp.w_up.biases": "model.safetensors",
51
+ "model.h.1.mlp.w_up.scales": "model.safetensors",
52
+ "model.h.1.mlp.w_up.weight": "model.safetensors",
53
+ "model.h.10.attn.k_proj.biases": "model.safetensors",
54
+ "model.h.10.attn.k_proj.scales": "model.safetensors",
55
+ "model.h.10.attn.k_proj.weight": "model.safetensors",
56
+ "model.h.10.attn.o_proj.biases": "model.safetensors",
57
+ "model.h.10.attn.o_proj.scales": "model.safetensors",
58
+ "model.h.10.attn.o_proj.weight": "model.safetensors",
59
+ "model.h.10.attn.q_proj.biases": "model.safetensors",
60
+ "model.h.10.attn.q_proj.scales": "model.safetensors",
61
+ "model.h.10.attn.q_proj.weight": "model.safetensors",
62
+ "model.h.10.attn.v_proj.biases": "model.safetensors",
63
+ "model.h.10.attn.v_proj.scales": "model.safetensors",
64
+ "model.h.10.attn.v_proj.weight": "model.safetensors",
65
+ "model.h.10.ln_1.weight": "model.safetensors",
66
+ "model.h.10.ln_2.weight": "model.safetensors",
67
+ "model.h.10.mlp.w_down.biases": "model.safetensors",
68
+ "model.h.10.mlp.w_down.scales": "model.safetensors",
69
+ "model.h.10.mlp.w_down.weight": "model.safetensors",
70
+ "model.h.10.mlp.w_gate.biases": "model.safetensors",
71
+ "model.h.10.mlp.w_gate.scales": "model.safetensors",
72
+ "model.h.10.mlp.w_gate.weight": "model.safetensors",
73
+ "model.h.10.mlp.w_up.biases": "model.safetensors",
74
+ "model.h.10.mlp.w_up.scales": "model.safetensors",
75
+ "model.h.10.mlp.w_up.weight": "model.safetensors",
76
+ "model.h.11.attn.k_proj.biases": "model.safetensors",
77
+ "model.h.11.attn.k_proj.scales": "model.safetensors",
78
+ "model.h.11.attn.k_proj.weight": "model.safetensors",
79
+ "model.h.11.attn.o_proj.biases": "model.safetensors",
80
+ "model.h.11.attn.o_proj.scales": "model.safetensors",
81
+ "model.h.11.attn.o_proj.weight": "model.safetensors",
82
+ "model.h.11.attn.q_proj.biases": "model.safetensors",
83
+ "model.h.11.attn.q_proj.scales": "model.safetensors",
84
+ "model.h.11.attn.q_proj.weight": "model.safetensors",
85
+ "model.h.11.attn.v_proj.biases": "model.safetensors",
86
+ "model.h.11.attn.v_proj.scales": "model.safetensors",
87
+ "model.h.11.attn.v_proj.weight": "model.safetensors",
88
+ "model.h.11.ln_1.weight": "model.safetensors",
89
+ "model.h.11.ln_2.weight": "model.safetensors",
90
+ "model.h.11.mlp.w_down.biases": "model.safetensors",
91
+ "model.h.11.mlp.w_down.scales": "model.safetensors",
92
+ "model.h.11.mlp.w_down.weight": "model.safetensors",
93
+ "model.h.11.mlp.w_gate.biases": "model.safetensors",
94
+ "model.h.11.mlp.w_gate.scales": "model.safetensors",
95
+ "model.h.11.mlp.w_gate.weight": "model.safetensors",
96
+ "model.h.11.mlp.w_up.biases": "model.safetensors",
97
+ "model.h.11.mlp.w_up.scales": "model.safetensors",
98
+ "model.h.11.mlp.w_up.weight": "model.safetensors",
99
+ "model.h.12.attn.k_proj.biases": "model.safetensors",
100
+ "model.h.12.attn.k_proj.scales": "model.safetensors",
101
+ "model.h.12.attn.k_proj.weight": "model.safetensors",
102
+ "model.h.12.attn.o_proj.biases": "model.safetensors",
103
+ "model.h.12.attn.o_proj.scales": "model.safetensors",
104
+ "model.h.12.attn.o_proj.weight": "model.safetensors",
105
+ "model.h.12.attn.q_proj.biases": "model.safetensors",
106
+ "model.h.12.attn.q_proj.scales": "model.safetensors",
107
+ "model.h.12.attn.q_proj.weight": "model.safetensors",
108
+ "model.h.12.attn.v_proj.biases": "model.safetensors",
109
+ "model.h.12.attn.v_proj.scales": "model.safetensors",
110
+ "model.h.12.attn.v_proj.weight": "model.safetensors",
111
+ "model.h.12.ln_1.weight": "model.safetensors",
112
+ "model.h.12.ln_2.weight": "model.safetensors",
113
+ "model.h.12.mlp.w_down.biases": "model.safetensors",
114
+ "model.h.12.mlp.w_down.scales": "model.safetensors",
115
+ "model.h.12.mlp.w_down.weight": "model.safetensors",
116
+ "model.h.12.mlp.w_gate.biases": "model.safetensors",
117
+ "model.h.12.mlp.w_gate.scales": "model.safetensors",
118
+ "model.h.12.mlp.w_gate.weight": "model.safetensors",
119
+ "model.h.12.mlp.w_up.biases": "model.safetensors",
120
+ "model.h.12.mlp.w_up.scales": "model.safetensors",
121
+ "model.h.12.mlp.w_up.weight": "model.safetensors",
122
+ "model.h.13.attn.k_proj.biases": "model.safetensors",
123
+ "model.h.13.attn.k_proj.scales": "model.safetensors",
124
+ "model.h.13.attn.k_proj.weight": "model.safetensors",
125
+ "model.h.13.attn.o_proj.biases": "model.safetensors",
126
+ "model.h.13.attn.o_proj.scales": "model.safetensors",
127
+ "model.h.13.attn.o_proj.weight": "model.safetensors",
128
+ "model.h.13.attn.q_proj.biases": "model.safetensors",
129
+ "model.h.13.attn.q_proj.scales": "model.safetensors",
130
+ "model.h.13.attn.q_proj.weight": "model.safetensors",
131
+ "model.h.13.attn.v_proj.biases": "model.safetensors",
132
+ "model.h.13.attn.v_proj.scales": "model.safetensors",
133
+ "model.h.13.attn.v_proj.weight": "model.safetensors",
134
+ "model.h.13.ln_1.weight": "model.safetensors",
135
+ "model.h.13.ln_2.weight": "model.safetensors",
136
+ "model.h.13.mlp.w_down.biases": "model.safetensors",
137
+ "model.h.13.mlp.w_down.scales": "model.safetensors",
138
+ "model.h.13.mlp.w_down.weight": "model.safetensors",
139
+ "model.h.13.mlp.w_gate.biases": "model.safetensors",
140
+ "model.h.13.mlp.w_gate.scales": "model.safetensors",
141
+ "model.h.13.mlp.w_gate.weight": "model.safetensors",
142
+ "model.h.13.mlp.w_up.biases": "model.safetensors",
143
+ "model.h.13.mlp.w_up.scales": "model.safetensors",
144
+ "model.h.13.mlp.w_up.weight": "model.safetensors",
145
+ "model.h.14.attn.k_proj.biases": "model.safetensors",
146
+ "model.h.14.attn.k_proj.scales": "model.safetensors",
147
+ "model.h.14.attn.k_proj.weight": "model.safetensors",
148
+ "model.h.14.attn.o_proj.biases": "model.safetensors",
149
+ "model.h.14.attn.o_proj.scales": "model.safetensors",
150
+ "model.h.14.attn.o_proj.weight": "model.safetensors",
151
+ "model.h.14.attn.q_proj.biases": "model.safetensors",
152
+ "model.h.14.attn.q_proj.scales": "model.safetensors",
153
+ "model.h.14.attn.q_proj.weight": "model.safetensors",
154
+ "model.h.14.attn.v_proj.biases": "model.safetensors",
155
+ "model.h.14.attn.v_proj.scales": "model.safetensors",
156
+ "model.h.14.attn.v_proj.weight": "model.safetensors",
157
+ "model.h.14.ln_1.weight": "model.safetensors",
158
+ "model.h.14.ln_2.weight": "model.safetensors",
159
+ "model.h.14.mlp.w_down.biases": "model.safetensors",
160
+ "model.h.14.mlp.w_down.scales": "model.safetensors",
161
+ "model.h.14.mlp.w_down.weight": "model.safetensors",
162
+ "model.h.14.mlp.w_gate.biases": "model.safetensors",
163
+ "model.h.14.mlp.w_gate.scales": "model.safetensors",
164
+ "model.h.14.mlp.w_gate.weight": "model.safetensors",
165
+ "model.h.14.mlp.w_up.biases": "model.safetensors",
166
+ "model.h.14.mlp.w_up.scales": "model.safetensors",
167
+ "model.h.14.mlp.w_up.weight": "model.safetensors",
168
+ "model.h.15.attn.k_proj.biases": "model.safetensors",
169
+ "model.h.15.attn.k_proj.scales": "model.safetensors",
170
+ "model.h.15.attn.k_proj.weight": "model.safetensors",
171
+ "model.h.15.attn.o_proj.biases": "model.safetensors",
172
+ "model.h.15.attn.o_proj.scales": "model.safetensors",
173
+ "model.h.15.attn.o_proj.weight": "model.safetensors",
174
+ "model.h.15.attn.q_proj.biases": "model.safetensors",
175
+ "model.h.15.attn.q_proj.scales": "model.safetensors",
176
+ "model.h.15.attn.q_proj.weight": "model.safetensors",
177
+ "model.h.15.attn.v_proj.biases": "model.safetensors",
178
+ "model.h.15.attn.v_proj.scales": "model.safetensors",
179
+ "model.h.15.attn.v_proj.weight": "model.safetensors",
180
+ "model.h.15.ln_1.weight": "model.safetensors",
181
+ "model.h.15.ln_2.weight": "model.safetensors",
182
+ "model.h.15.mlp.w_down.biases": "model.safetensors",
183
+ "model.h.15.mlp.w_down.scales": "model.safetensors",
184
+ "model.h.15.mlp.w_down.weight": "model.safetensors",
185
+ "model.h.15.mlp.w_gate.biases": "model.safetensors",
186
+ "model.h.15.mlp.w_gate.scales": "model.safetensors",
187
+ "model.h.15.mlp.w_gate.weight": "model.safetensors",
188
+ "model.h.15.mlp.w_up.biases": "model.safetensors",
189
+ "model.h.15.mlp.w_up.scales": "model.safetensors",
190
+ "model.h.15.mlp.w_up.weight": "model.safetensors",
191
+ "model.h.16.attn.k_proj.biases": "model.safetensors",
192
+ "model.h.16.attn.k_proj.scales": "model.safetensors",
193
+ "model.h.16.attn.k_proj.weight": "model.safetensors",
194
+ "model.h.16.attn.o_proj.biases": "model.safetensors",
195
+ "model.h.16.attn.o_proj.scales": "model.safetensors",
196
+ "model.h.16.attn.o_proj.weight": "model.safetensors",
197
+ "model.h.16.attn.q_proj.biases": "model.safetensors",
198
+ "model.h.16.attn.q_proj.scales": "model.safetensors",
199
+ "model.h.16.attn.q_proj.weight": "model.safetensors",
200
+ "model.h.16.attn.v_proj.biases": "model.safetensors",
201
+ "model.h.16.attn.v_proj.scales": "model.safetensors",
202
+ "model.h.16.attn.v_proj.weight": "model.safetensors",
203
+ "model.h.16.ln_1.weight": "model.safetensors",
204
+ "model.h.16.ln_2.weight": "model.safetensors",
205
+ "model.h.16.mlp.w_down.biases": "model.safetensors",
206
+ "model.h.16.mlp.w_down.scales": "model.safetensors",
207
+ "model.h.16.mlp.w_down.weight": "model.safetensors",
208
+ "model.h.16.mlp.w_gate.biases": "model.safetensors",
209
+ "model.h.16.mlp.w_gate.scales": "model.safetensors",
210
+ "model.h.16.mlp.w_gate.weight": "model.safetensors",
211
+ "model.h.16.mlp.w_up.biases": "model.safetensors",
212
+ "model.h.16.mlp.w_up.scales": "model.safetensors",
213
+ "model.h.16.mlp.w_up.weight": "model.safetensors",
214
+ "model.h.17.attn.k_proj.biases": "model.safetensors",
215
+ "model.h.17.attn.k_proj.scales": "model.safetensors",
216
+ "model.h.17.attn.k_proj.weight": "model.safetensors",
217
+ "model.h.17.attn.o_proj.biases": "model.safetensors",
218
+ "model.h.17.attn.o_proj.scales": "model.safetensors",
219
+ "model.h.17.attn.o_proj.weight": "model.safetensors",
220
+ "model.h.17.attn.q_proj.biases": "model.safetensors",
221
+ "model.h.17.attn.q_proj.scales": "model.safetensors",
222
+ "model.h.17.attn.q_proj.weight": "model.safetensors",
223
+ "model.h.17.attn.v_proj.biases": "model.safetensors",
224
+ "model.h.17.attn.v_proj.scales": "model.safetensors",
225
+ "model.h.17.attn.v_proj.weight": "model.safetensors",
226
+ "model.h.17.ln_1.weight": "model.safetensors",
227
+ "model.h.17.ln_2.weight": "model.safetensors",
228
+ "model.h.17.mlp.w_down.biases": "model.safetensors",
229
+ "model.h.17.mlp.w_down.scales": "model.safetensors",
230
+ "model.h.17.mlp.w_down.weight": "model.safetensors",
231
+ "model.h.17.mlp.w_gate.biases": "model.safetensors",
232
+ "model.h.17.mlp.w_gate.scales": "model.safetensors",
233
+ "model.h.17.mlp.w_gate.weight": "model.safetensors",
234
+ "model.h.17.mlp.w_up.biases": "model.safetensors",
235
+ "model.h.17.mlp.w_up.scales": "model.safetensors",
236
+ "model.h.17.mlp.w_up.weight": "model.safetensors",
237
+ "model.h.18.attn.k_proj.biases": "model.safetensors",
238
+ "model.h.18.attn.k_proj.scales": "model.safetensors",
239
+ "model.h.18.attn.k_proj.weight": "model.safetensors",
240
+ "model.h.18.attn.o_proj.biases": "model.safetensors",
241
+ "model.h.18.attn.o_proj.scales": "model.safetensors",
242
+ "model.h.18.attn.o_proj.weight": "model.safetensors",
243
+ "model.h.18.attn.q_proj.biases": "model.safetensors",
244
+ "model.h.18.attn.q_proj.scales": "model.safetensors",
245
+ "model.h.18.attn.q_proj.weight": "model.safetensors",
246
+ "model.h.18.attn.v_proj.biases": "model.safetensors",
247
+ "model.h.18.attn.v_proj.scales": "model.safetensors",
248
+ "model.h.18.attn.v_proj.weight": "model.safetensors",
249
+ "model.h.18.ln_1.weight": "model.safetensors",
250
+ "model.h.18.ln_2.weight": "model.safetensors",
251
+ "model.h.18.mlp.w_down.biases": "model.safetensors",
252
+ "model.h.18.mlp.w_down.scales": "model.safetensors",
253
+ "model.h.18.mlp.w_down.weight": "model.safetensors",
254
+ "model.h.18.mlp.w_gate.biases": "model.safetensors",
255
+ "model.h.18.mlp.w_gate.scales": "model.safetensors",
256
+ "model.h.18.mlp.w_gate.weight": "model.safetensors",
257
+ "model.h.18.mlp.w_up.biases": "model.safetensors",
258
+ "model.h.18.mlp.w_up.scales": "model.safetensors",
259
+ "model.h.18.mlp.w_up.weight": "model.safetensors",
260
+ "model.h.19.attn.k_proj.biases": "model.safetensors",
261
+ "model.h.19.attn.k_proj.scales": "model.safetensors",
262
+ "model.h.19.attn.k_proj.weight": "model.safetensors",
263
+ "model.h.19.attn.o_proj.biases": "model.safetensors",
264
+ "model.h.19.attn.o_proj.scales": "model.safetensors",
265
+ "model.h.19.attn.o_proj.weight": "model.safetensors",
266
+ "model.h.19.attn.q_proj.biases": "model.safetensors",
267
+ "model.h.19.attn.q_proj.scales": "model.safetensors",
268
+ "model.h.19.attn.q_proj.weight": "model.safetensors",
269
+ "model.h.19.attn.v_proj.biases": "model.safetensors",
270
+ "model.h.19.attn.v_proj.scales": "model.safetensors",
271
+ "model.h.19.attn.v_proj.weight": "model.safetensors",
272
+ "model.h.19.ln_1.weight": "model.safetensors",
273
+ "model.h.19.ln_2.weight": "model.safetensors",
274
+ "model.h.19.mlp.w_down.biases": "model.safetensors",
275
+ "model.h.19.mlp.w_down.scales": "model.safetensors",
276
+ "model.h.19.mlp.w_down.weight": "model.safetensors",
277
+ "model.h.19.mlp.w_gate.biases": "model.safetensors",
278
+ "model.h.19.mlp.w_gate.scales": "model.safetensors",
279
+ "model.h.19.mlp.w_gate.weight": "model.safetensors",
280
+ "model.h.19.mlp.w_up.biases": "model.safetensors",
281
+ "model.h.19.mlp.w_up.scales": "model.safetensors",
282
+ "model.h.19.mlp.w_up.weight": "model.safetensors",
283
+ "model.h.2.attn.k_proj.biases": "model.safetensors",
284
+ "model.h.2.attn.k_proj.scales": "model.safetensors",
285
+ "model.h.2.attn.k_proj.weight": "model.safetensors",
286
+ "model.h.2.attn.o_proj.biases": "model.safetensors",
287
+ "model.h.2.attn.o_proj.scales": "model.safetensors",
288
+ "model.h.2.attn.o_proj.weight": "model.safetensors",
289
+ "model.h.2.attn.q_proj.biases": "model.safetensors",
290
+ "model.h.2.attn.q_proj.scales": "model.safetensors",
291
+ "model.h.2.attn.q_proj.weight": "model.safetensors",
292
+ "model.h.2.attn.v_proj.biases": "model.safetensors",
293
+ "model.h.2.attn.v_proj.scales": "model.safetensors",
294
+ "model.h.2.attn.v_proj.weight": "model.safetensors",
295
+ "model.h.2.ln_1.weight": "model.safetensors",
296
+ "model.h.2.ln_2.weight": "model.safetensors",
297
+ "model.h.2.mlp.w_down.biases": "model.safetensors",
298
+ "model.h.2.mlp.w_down.scales": "model.safetensors",
299
+ "model.h.2.mlp.w_down.weight": "model.safetensors",
300
+ "model.h.2.mlp.w_gate.biases": "model.safetensors",
301
+ "model.h.2.mlp.w_gate.scales": "model.safetensors",
302
+ "model.h.2.mlp.w_gate.weight": "model.safetensors",
303
+ "model.h.2.mlp.w_up.biases": "model.safetensors",
304
+ "model.h.2.mlp.w_up.scales": "model.safetensors",
305
+ "model.h.2.mlp.w_up.weight": "model.safetensors",
306
+ "model.h.20.attn.k_proj.biases": "model.safetensors",
307
+ "model.h.20.attn.k_proj.scales": "model.safetensors",
308
+ "model.h.20.attn.k_proj.weight": "model.safetensors",
309
+ "model.h.20.attn.o_proj.biases": "model.safetensors",
310
+ "model.h.20.attn.o_proj.scales": "model.safetensors",
311
+ "model.h.20.attn.o_proj.weight": "model.safetensors",
312
+ "model.h.20.attn.q_proj.biases": "model.safetensors",
313
+ "model.h.20.attn.q_proj.scales": "model.safetensors",
314
+ "model.h.20.attn.q_proj.weight": "model.safetensors",
315
+ "model.h.20.attn.v_proj.biases": "model.safetensors",
316
+ "model.h.20.attn.v_proj.scales": "model.safetensors",
317
+ "model.h.20.attn.v_proj.weight": "model.safetensors",
318
+ "model.h.20.ln_1.weight": "model.safetensors",
319
+ "model.h.20.ln_2.weight": "model.safetensors",
320
+ "model.h.20.mlp.w_down.biases": "model.safetensors",
321
+ "model.h.20.mlp.w_down.scales": "model.safetensors",
322
+ "model.h.20.mlp.w_down.weight": "model.safetensors",
323
+ "model.h.20.mlp.w_gate.biases": "model.safetensors",
324
+ "model.h.20.mlp.w_gate.scales": "model.safetensors",
325
+ "model.h.20.mlp.w_gate.weight": "model.safetensors",
326
+ "model.h.20.mlp.w_up.biases": "model.safetensors",
327
+ "model.h.20.mlp.w_up.scales": "model.safetensors",
328
+ "model.h.20.mlp.w_up.weight": "model.safetensors",
329
+ "model.h.21.attn.k_proj.biases": "model.safetensors",
330
+ "model.h.21.attn.k_proj.scales": "model.safetensors",
331
+ "model.h.21.attn.k_proj.weight": "model.safetensors",
332
+ "model.h.21.attn.o_proj.biases": "model.safetensors",
333
+ "model.h.21.attn.o_proj.scales": "model.safetensors",
334
+ "model.h.21.attn.o_proj.weight": "model.safetensors",
335
+ "model.h.21.attn.q_proj.biases": "model.safetensors",
336
+ "model.h.21.attn.q_proj.scales": "model.safetensors",
337
+ "model.h.21.attn.q_proj.weight": "model.safetensors",
338
+ "model.h.21.attn.v_proj.biases": "model.safetensors",
339
+ "model.h.21.attn.v_proj.scales": "model.safetensors",
340
+ "model.h.21.attn.v_proj.weight": "model.safetensors",
341
+ "model.h.21.ln_1.weight": "model.safetensors",
342
+ "model.h.21.ln_2.weight": "model.safetensors",
343
+ "model.h.21.mlp.w_down.biases": "model.safetensors",
344
+ "model.h.21.mlp.w_down.scales": "model.safetensors",
345
+ "model.h.21.mlp.w_down.weight": "model.safetensors",
346
+ "model.h.21.mlp.w_gate.biases": "model.safetensors",
347
+ "model.h.21.mlp.w_gate.scales": "model.safetensors",
348
+ "model.h.21.mlp.w_gate.weight": "model.safetensors",
349
+ "model.h.21.mlp.w_up.biases": "model.safetensors",
350
+ "model.h.21.mlp.w_up.scales": "model.safetensors",
351
+ "model.h.21.mlp.w_up.weight": "model.safetensors",
352
+ "model.h.22.attn.k_proj.biases": "model.safetensors",
353
+ "model.h.22.attn.k_proj.scales": "model.safetensors",
354
+ "model.h.22.attn.k_proj.weight": "model.safetensors",
355
+ "model.h.22.attn.o_proj.biases": "model.safetensors",
356
+ "model.h.22.attn.o_proj.scales": "model.safetensors",
357
+ "model.h.22.attn.o_proj.weight": "model.safetensors",
358
+ "model.h.22.attn.q_proj.biases": "model.safetensors",
359
+ "model.h.22.attn.q_proj.scales": "model.safetensors",
360
+ "model.h.22.attn.q_proj.weight": "model.safetensors",
361
+ "model.h.22.attn.v_proj.biases": "model.safetensors",
362
+ "model.h.22.attn.v_proj.scales": "model.safetensors",
363
+ "model.h.22.attn.v_proj.weight": "model.safetensors",
364
+ "model.h.22.ln_1.weight": "model.safetensors",
365
+ "model.h.22.ln_2.weight": "model.safetensors",
366
+ "model.h.22.mlp.w_down.biases": "model.safetensors",
367
+ "model.h.22.mlp.w_down.scales": "model.safetensors",
368
+ "model.h.22.mlp.w_down.weight": "model.safetensors",
369
+ "model.h.22.mlp.w_gate.biases": "model.safetensors",
370
+ "model.h.22.mlp.w_gate.scales": "model.safetensors",
371
+ "model.h.22.mlp.w_gate.weight": "model.safetensors",
372
+ "model.h.22.mlp.w_up.biases": "model.safetensors",
373
+ "model.h.22.mlp.w_up.scales": "model.safetensors",
374
+ "model.h.22.mlp.w_up.weight": "model.safetensors",
375
+ "model.h.23.attn.k_proj.biases": "model.safetensors",
376
+ "model.h.23.attn.k_proj.scales": "model.safetensors",
377
+ "model.h.23.attn.k_proj.weight": "model.safetensors",
378
+ "model.h.23.attn.o_proj.biases": "model.safetensors",
379
+ "model.h.23.attn.o_proj.scales": "model.safetensors",
380
+ "model.h.23.attn.o_proj.weight": "model.safetensors",
381
+ "model.h.23.attn.q_proj.biases": "model.safetensors",
382
+ "model.h.23.attn.q_proj.scales": "model.safetensors",
383
+ "model.h.23.attn.q_proj.weight": "model.safetensors",
384
+ "model.h.23.attn.v_proj.biases": "model.safetensors",
385
+ "model.h.23.attn.v_proj.scales": "model.safetensors",
386
+ "model.h.23.attn.v_proj.weight": "model.safetensors",
387
+ "model.h.23.ln_1.weight": "model.safetensors",
388
+ "model.h.23.ln_2.weight": "model.safetensors",
389
+ "model.h.23.mlp.w_down.biases": "model.safetensors",
390
+ "model.h.23.mlp.w_down.scales": "model.safetensors",
391
+ "model.h.23.mlp.w_down.weight": "model.safetensors",
392
+ "model.h.23.mlp.w_gate.biases": "model.safetensors",
393
+ "model.h.23.mlp.w_gate.scales": "model.safetensors",
394
+ "model.h.23.mlp.w_gate.weight": "model.safetensors",
395
+ "model.h.23.mlp.w_up.biases": "model.safetensors",
396
+ "model.h.23.mlp.w_up.scales": "model.safetensors",
397
+ "model.h.23.mlp.w_up.weight": "model.safetensors",
398
+ "model.h.24.attn.k_proj.biases": "model.safetensors",
399
+ "model.h.24.attn.k_proj.scales": "model.safetensors",
400
+ "model.h.24.attn.k_proj.weight": "model.safetensors",
401
+ "model.h.24.attn.o_proj.biases": "model.safetensors",
402
+ "model.h.24.attn.o_proj.scales": "model.safetensors",
403
+ "model.h.24.attn.o_proj.weight": "model.safetensors",
404
+ "model.h.24.attn.q_proj.biases": "model.safetensors",
405
+ "model.h.24.attn.q_proj.scales": "model.safetensors",
406
+ "model.h.24.attn.q_proj.weight": "model.safetensors",
407
+ "model.h.24.attn.v_proj.biases": "model.safetensors",
408
+ "model.h.24.attn.v_proj.scales": "model.safetensors",
409
+ "model.h.24.attn.v_proj.weight": "model.safetensors",
410
+ "model.h.24.ln_1.weight": "model.safetensors",
411
+ "model.h.24.ln_2.weight": "model.safetensors",
412
+ "model.h.24.mlp.w_down.biases": "model.safetensors",
413
+ "model.h.24.mlp.w_down.scales": "model.safetensors",
414
+ "model.h.24.mlp.w_down.weight": "model.safetensors",
415
+ "model.h.24.mlp.w_gate.biases": "model.safetensors",
416
+ "model.h.24.mlp.w_gate.scales": "model.safetensors",
417
+ "model.h.24.mlp.w_gate.weight": "model.safetensors",
418
+ "model.h.24.mlp.w_up.biases": "model.safetensors",
419
+ "model.h.24.mlp.w_up.scales": "model.safetensors",
420
+ "model.h.24.mlp.w_up.weight": "model.safetensors",
421
+ "model.h.25.attn.k_proj.biases": "model.safetensors",
422
+ "model.h.25.attn.k_proj.scales": "model.safetensors",
423
+ "model.h.25.attn.k_proj.weight": "model.safetensors",
424
+ "model.h.25.attn.o_proj.biases": "model.safetensors",
425
+ "model.h.25.attn.o_proj.scales": "model.safetensors",
426
+ "model.h.25.attn.o_proj.weight": "model.safetensors",
427
+ "model.h.25.attn.q_proj.biases": "model.safetensors",
428
+ "model.h.25.attn.q_proj.scales": "model.safetensors",
429
+ "model.h.25.attn.q_proj.weight": "model.safetensors",
430
+ "model.h.25.attn.v_proj.biases": "model.safetensors",
431
+ "model.h.25.attn.v_proj.scales": "model.safetensors",
432
+ "model.h.25.attn.v_proj.weight": "model.safetensors",
433
+ "model.h.25.ln_1.weight": "model.safetensors",
434
+ "model.h.25.ln_2.weight": "model.safetensors",
435
+ "model.h.25.mlp.w_down.biases": "model.safetensors",
436
+ "model.h.25.mlp.w_down.scales": "model.safetensors",
437
+ "model.h.25.mlp.w_down.weight": "model.safetensors",
438
+ "model.h.25.mlp.w_gate.biases": "model.safetensors",
439
+ "model.h.25.mlp.w_gate.scales": "model.safetensors",
440
+ "model.h.25.mlp.w_gate.weight": "model.safetensors",
441
+ "model.h.25.mlp.w_up.biases": "model.safetensors",
442
+ "model.h.25.mlp.w_up.scales": "model.safetensors",
443
+ "model.h.25.mlp.w_up.weight": "model.safetensors",
444
+ "model.h.26.attn.k_proj.biases": "model.safetensors",
445
+ "model.h.26.attn.k_proj.scales": "model.safetensors",
446
+ "model.h.26.attn.k_proj.weight": "model.safetensors",
447
+ "model.h.26.attn.o_proj.biases": "model.safetensors",
448
+ "model.h.26.attn.o_proj.scales": "model.safetensors",
449
+ "model.h.26.attn.o_proj.weight": "model.safetensors",
450
+ "model.h.26.attn.q_proj.biases": "model.safetensors",
451
+ "model.h.26.attn.q_proj.scales": "model.safetensors",
452
+ "model.h.26.attn.q_proj.weight": "model.safetensors",
453
+ "model.h.26.attn.v_proj.biases": "model.safetensors",
454
+ "model.h.26.attn.v_proj.scales": "model.safetensors",
455
+ "model.h.26.attn.v_proj.weight": "model.safetensors",
456
+ "model.h.26.ln_1.weight": "model.safetensors",
457
+ "model.h.26.ln_2.weight": "model.safetensors",
458
+ "model.h.26.mlp.w_down.biases": "model.safetensors",
459
+ "model.h.26.mlp.w_down.scales": "model.safetensors",
460
+ "model.h.26.mlp.w_down.weight": "model.safetensors",
461
+ "model.h.26.mlp.w_gate.biases": "model.safetensors",
462
+ "model.h.26.mlp.w_gate.scales": "model.safetensors",
463
+ "model.h.26.mlp.w_gate.weight": "model.safetensors",
464
+ "model.h.26.mlp.w_up.biases": "model.safetensors",
465
+ "model.h.26.mlp.w_up.scales": "model.safetensors",
466
+ "model.h.26.mlp.w_up.weight": "model.safetensors",
467
+ "model.h.27.attn.k_proj.biases": "model.safetensors",
468
+ "model.h.27.attn.k_proj.scales": "model.safetensors",
469
+ "model.h.27.attn.k_proj.weight": "model.safetensors",
470
+ "model.h.27.attn.o_proj.biases": "model.safetensors",
471
+ "model.h.27.attn.o_proj.scales": "model.safetensors",
472
+ "model.h.27.attn.o_proj.weight": "model.safetensors",
473
+ "model.h.27.attn.q_proj.biases": "model.safetensors",
474
+ "model.h.27.attn.q_proj.scales": "model.safetensors",
475
+ "model.h.27.attn.q_proj.weight": "model.safetensors",
476
+ "model.h.27.attn.v_proj.biases": "model.safetensors",
477
+ "model.h.27.attn.v_proj.scales": "model.safetensors",
478
+ "model.h.27.attn.v_proj.weight": "model.safetensors",
479
+ "model.h.27.ln_1.weight": "model.safetensors",
480
+ "model.h.27.ln_2.weight": "model.safetensors",
481
+ "model.h.27.mlp.w_down.biases": "model.safetensors",
482
+ "model.h.27.mlp.w_down.scales": "model.safetensors",
483
+ "model.h.27.mlp.w_down.weight": "model.safetensors",
484
+ "model.h.27.mlp.w_gate.biases": "model.safetensors",
485
+ "model.h.27.mlp.w_gate.scales": "model.safetensors",
486
+ "model.h.27.mlp.w_gate.weight": "model.safetensors",
487
+ "model.h.27.mlp.w_up.biases": "model.safetensors",
488
+ "model.h.27.mlp.w_up.scales": "model.safetensors",
489
+ "model.h.27.mlp.w_up.weight": "model.safetensors",
490
+ "model.h.28.attn.k_proj.biases": "model.safetensors",
491
+ "model.h.28.attn.k_proj.scales": "model.safetensors",
492
+ "model.h.28.attn.k_proj.weight": "model.safetensors",
493
+ "model.h.28.attn.o_proj.biases": "model.safetensors",
494
+ "model.h.28.attn.o_proj.scales": "model.safetensors",
495
+ "model.h.28.attn.o_proj.weight": "model.safetensors",
496
+ "model.h.28.attn.q_proj.biases": "model.safetensors",
497
+ "model.h.28.attn.q_proj.scales": "model.safetensors",
498
+ "model.h.28.attn.q_proj.weight": "model.safetensors",
499
+ "model.h.28.attn.v_proj.biases": "model.safetensors",
500
+ "model.h.28.attn.v_proj.scales": "model.safetensors",
501
+ "model.h.28.attn.v_proj.weight": "model.safetensors",
502
+ "model.h.28.ln_1.weight": "model.safetensors",
503
+ "model.h.28.ln_2.weight": "model.safetensors",
504
+ "model.h.28.mlp.w_down.biases": "model.safetensors",
505
+ "model.h.28.mlp.w_down.scales": "model.safetensors",
506
+ "model.h.28.mlp.w_down.weight": "model.safetensors",
507
+ "model.h.28.mlp.w_gate.biases": "model.safetensors",
508
+ "model.h.28.mlp.w_gate.scales": "model.safetensors",
509
+ "model.h.28.mlp.w_gate.weight": "model.safetensors",
510
+ "model.h.28.mlp.w_up.biases": "model.safetensors",
511
+ "model.h.28.mlp.w_up.scales": "model.safetensors",
512
+ "model.h.28.mlp.w_up.weight": "model.safetensors",
513
+ "model.h.29.attn.k_proj.biases": "model.safetensors",
514
+ "model.h.29.attn.k_proj.scales": "model.safetensors",
515
+ "model.h.29.attn.k_proj.weight": "model.safetensors",
516
+ "model.h.29.attn.o_proj.biases": "model.safetensors",
517
+ "model.h.29.attn.o_proj.scales": "model.safetensors",
518
+ "model.h.29.attn.o_proj.weight": "model.safetensors",
519
+ "model.h.29.attn.q_proj.biases": "model.safetensors",
520
+ "model.h.29.attn.q_proj.scales": "model.safetensors",
521
+ "model.h.29.attn.q_proj.weight": "model.safetensors",
522
+ "model.h.29.attn.v_proj.biases": "model.safetensors",
523
+ "model.h.29.attn.v_proj.scales": "model.safetensors",
524
+ "model.h.29.attn.v_proj.weight": "model.safetensors",
525
+ "model.h.29.ln_1.weight": "model.safetensors",
526
+ "model.h.29.ln_2.weight": "model.safetensors",
527
+ "model.h.29.mlp.w_down.biases": "model.safetensors",
528
+ "model.h.29.mlp.w_down.scales": "model.safetensors",
529
+ "model.h.29.mlp.w_down.weight": "model.safetensors",
530
+ "model.h.29.mlp.w_gate.biases": "model.safetensors",
531
+ "model.h.29.mlp.w_gate.scales": "model.safetensors",
532
+ "model.h.29.mlp.w_gate.weight": "model.safetensors",
533
+ "model.h.29.mlp.w_up.biases": "model.safetensors",
534
+ "model.h.29.mlp.w_up.scales": "model.safetensors",
535
+ "model.h.29.mlp.w_up.weight": "model.safetensors",
536
+ "model.h.3.attn.k_proj.biases": "model.safetensors",
537
+ "model.h.3.attn.k_proj.scales": "model.safetensors",
538
+ "model.h.3.attn.k_proj.weight": "model.safetensors",
539
+ "model.h.3.attn.o_proj.biases": "model.safetensors",
540
+ "model.h.3.attn.o_proj.scales": "model.safetensors",
541
+ "model.h.3.attn.o_proj.weight": "model.safetensors",
542
+ "model.h.3.attn.q_proj.biases": "model.safetensors",
543
+ "model.h.3.attn.q_proj.scales": "model.safetensors",
544
+ "model.h.3.attn.q_proj.weight": "model.safetensors",
545
+ "model.h.3.attn.v_proj.biases": "model.safetensors",
546
+ "model.h.3.attn.v_proj.scales": "model.safetensors",
547
+ "model.h.3.attn.v_proj.weight": "model.safetensors",
548
+ "model.h.3.ln_1.weight": "model.safetensors",
549
+ "model.h.3.ln_2.weight": "model.safetensors",
550
+ "model.h.3.mlp.w_down.biases": "model.safetensors",
551
+ "model.h.3.mlp.w_down.scales": "model.safetensors",
552
+ "model.h.3.mlp.w_down.weight": "model.safetensors",
553
+ "model.h.3.mlp.w_gate.biases": "model.safetensors",
554
+ "model.h.3.mlp.w_gate.scales": "model.safetensors",
555
+ "model.h.3.mlp.w_gate.weight": "model.safetensors",
556
+ "model.h.3.mlp.w_up.biases": "model.safetensors",
557
+ "model.h.3.mlp.w_up.scales": "model.safetensors",
558
+ "model.h.3.mlp.w_up.weight": "model.safetensors",
559
+ "model.h.4.attn.k_proj.biases": "model.safetensors",
560
+ "model.h.4.attn.k_proj.scales": "model.safetensors",
561
+ "model.h.4.attn.k_proj.weight": "model.safetensors",
562
+ "model.h.4.attn.o_proj.biases": "model.safetensors",
563
+ "model.h.4.attn.o_proj.scales": "model.safetensors",
564
+ "model.h.4.attn.o_proj.weight": "model.safetensors",
565
+ "model.h.4.attn.q_proj.biases": "model.safetensors",
566
+ "model.h.4.attn.q_proj.scales": "model.safetensors",
567
+ "model.h.4.attn.q_proj.weight": "model.safetensors",
568
+ "model.h.4.attn.v_proj.biases": "model.safetensors",
569
+ "model.h.4.attn.v_proj.scales": "model.safetensors",
570
+ "model.h.4.attn.v_proj.weight": "model.safetensors",
571
+ "model.h.4.ln_1.weight": "model.safetensors",
572
+ "model.h.4.ln_2.weight": "model.safetensors",
573
+ "model.h.4.mlp.w_down.biases": "model.safetensors",
574
+ "model.h.4.mlp.w_down.scales": "model.safetensors",
575
+ "model.h.4.mlp.w_down.weight": "model.safetensors",
576
+ "model.h.4.mlp.w_gate.biases": "model.safetensors",
577
+ "model.h.4.mlp.w_gate.scales": "model.safetensors",
578
+ "model.h.4.mlp.w_gate.weight": "model.safetensors",
579
+ "model.h.4.mlp.w_up.biases": "model.safetensors",
580
+ "model.h.4.mlp.w_up.scales": "model.safetensors",
581
+ "model.h.4.mlp.w_up.weight": "model.safetensors",
582
+ "model.h.5.attn.k_proj.biases": "model.safetensors",
583
+ "model.h.5.attn.k_proj.scales": "model.safetensors",
584
+ "model.h.5.attn.k_proj.weight": "model.safetensors",
585
+ "model.h.5.attn.o_proj.biases": "model.safetensors",
586
+ "model.h.5.attn.o_proj.scales": "model.safetensors",
587
+ "model.h.5.attn.o_proj.weight": "model.safetensors",
588
+ "model.h.5.attn.q_proj.biases": "model.safetensors",
589
+ "model.h.5.attn.q_proj.scales": "model.safetensors",
590
+ "model.h.5.attn.q_proj.weight": "model.safetensors",
591
+ "model.h.5.attn.v_proj.biases": "model.safetensors",
592
+ "model.h.5.attn.v_proj.scales": "model.safetensors",
593
+ "model.h.5.attn.v_proj.weight": "model.safetensors",
594
+ "model.h.5.ln_1.weight": "model.safetensors",
595
+ "model.h.5.ln_2.weight": "model.safetensors",
596
+ "model.h.5.mlp.w_down.biases": "model.safetensors",
597
+ "model.h.5.mlp.w_down.scales": "model.safetensors",
598
+ "model.h.5.mlp.w_down.weight": "model.safetensors",
599
+ "model.h.5.mlp.w_gate.biases": "model.safetensors",
600
+ "model.h.5.mlp.w_gate.scales": "model.safetensors",
601
+ "model.h.5.mlp.w_gate.weight": "model.safetensors",
602
+ "model.h.5.mlp.w_up.biases": "model.safetensors",
603
+ "model.h.5.mlp.w_up.scales": "model.safetensors",
604
+ "model.h.5.mlp.w_up.weight": "model.safetensors",
605
+ "model.h.6.attn.k_proj.biases": "model.safetensors",
606
+ "model.h.6.attn.k_proj.scales": "model.safetensors",
607
+ "model.h.6.attn.k_proj.weight": "model.safetensors",
608
+ "model.h.6.attn.o_proj.biases": "model.safetensors",
609
+ "model.h.6.attn.o_proj.scales": "model.safetensors",
610
+ "model.h.6.attn.o_proj.weight": "model.safetensors",
611
+ "model.h.6.attn.q_proj.biases": "model.safetensors",
612
+ "model.h.6.attn.q_proj.scales": "model.safetensors",
613
+ "model.h.6.attn.q_proj.weight": "model.safetensors",
614
+ "model.h.6.attn.v_proj.biases": "model.safetensors",
615
+ "model.h.6.attn.v_proj.scales": "model.safetensors",
616
+ "model.h.6.attn.v_proj.weight": "model.safetensors",
617
+ "model.h.6.ln_1.weight": "model.safetensors",
618
+ "model.h.6.ln_2.weight": "model.safetensors",
619
+ "model.h.6.mlp.w_down.biases": "model.safetensors",
620
+ "model.h.6.mlp.w_down.scales": "model.safetensors",
621
+ "model.h.6.mlp.w_down.weight": "model.safetensors",
622
+ "model.h.6.mlp.w_gate.biases": "model.safetensors",
623
+ "model.h.6.mlp.w_gate.scales": "model.safetensors",
624
+ "model.h.6.mlp.w_gate.weight": "model.safetensors",
625
+ "model.h.6.mlp.w_up.biases": "model.safetensors",
626
+ "model.h.6.mlp.w_up.scales": "model.safetensors",
627
+ "model.h.6.mlp.w_up.weight": "model.safetensors",
628
+ "model.h.7.attn.k_proj.biases": "model.safetensors",
629
+ "model.h.7.attn.k_proj.scales": "model.safetensors",
630
+ "model.h.7.attn.k_proj.weight": "model.safetensors",
631
+ "model.h.7.attn.o_proj.biases": "model.safetensors",
632
+ "model.h.7.attn.o_proj.scales": "model.safetensors",
633
+ "model.h.7.attn.o_proj.weight": "model.safetensors",
634
+ "model.h.7.attn.q_proj.biases": "model.safetensors",
635
+ "model.h.7.attn.q_proj.scales": "model.safetensors",
636
+ "model.h.7.attn.q_proj.weight": "model.safetensors",
637
+ "model.h.7.attn.v_proj.biases": "model.safetensors",
638
+ "model.h.7.attn.v_proj.scales": "model.safetensors",
639
+ "model.h.7.attn.v_proj.weight": "model.safetensors",
640
+ "model.h.7.ln_1.weight": "model.safetensors",
641
+ "model.h.7.ln_2.weight": "model.safetensors",
642
+ "model.h.7.mlp.w_down.biases": "model.safetensors",
643
+ "model.h.7.mlp.w_down.scales": "model.safetensors",
644
+ "model.h.7.mlp.w_down.weight": "model.safetensors",
645
+ "model.h.7.mlp.w_gate.biases": "model.safetensors",
646
+ "model.h.7.mlp.w_gate.scales": "model.safetensors",
647
+ "model.h.7.mlp.w_gate.weight": "model.safetensors",
648
+ "model.h.7.mlp.w_up.biases": "model.safetensors",
649
+ "model.h.7.mlp.w_up.scales": "model.safetensors",
650
+ "model.h.7.mlp.w_up.weight": "model.safetensors",
651
+ "model.h.8.attn.k_proj.biases": "model.safetensors",
652
+ "model.h.8.attn.k_proj.scales": "model.safetensors",
653
+ "model.h.8.attn.k_proj.weight": "model.safetensors",
654
+ "model.h.8.attn.o_proj.biases": "model.safetensors",
655
+ "model.h.8.attn.o_proj.scales": "model.safetensors",
656
+ "model.h.8.attn.o_proj.weight": "model.safetensors",
657
+ "model.h.8.attn.q_proj.biases": "model.safetensors",
658
+ "model.h.8.attn.q_proj.scales": "model.safetensors",
659
+ "model.h.8.attn.q_proj.weight": "model.safetensors",
660
+ "model.h.8.attn.v_proj.biases": "model.safetensors",
661
+ "model.h.8.attn.v_proj.scales": "model.safetensors",
662
+ "model.h.8.attn.v_proj.weight": "model.safetensors",
663
+ "model.h.8.ln_1.weight": "model.safetensors",
664
+ "model.h.8.ln_2.weight": "model.safetensors",
665
+ "model.h.8.mlp.w_down.biases": "model.safetensors",
666
+ "model.h.8.mlp.w_down.scales": "model.safetensors",
667
+ "model.h.8.mlp.w_down.weight": "model.safetensors",
668
+ "model.h.8.mlp.w_gate.biases": "model.safetensors",
669
+ "model.h.8.mlp.w_gate.scales": "model.safetensors",
670
+ "model.h.8.mlp.w_gate.weight": "model.safetensors",
671
+ "model.h.8.mlp.w_up.biases": "model.safetensors",
672
+ "model.h.8.mlp.w_up.scales": "model.safetensors",
673
+ "model.h.8.mlp.w_up.weight": "model.safetensors",
674
+ "model.h.9.attn.k_proj.biases": "model.safetensors",
675
+ "model.h.9.attn.k_proj.scales": "model.safetensors",
676
+ "model.h.9.attn.k_proj.weight": "model.safetensors",
677
+ "model.h.9.attn.o_proj.biases": "model.safetensors",
678
+ "model.h.9.attn.o_proj.scales": "model.safetensors",
679
+ "model.h.9.attn.o_proj.weight": "model.safetensors",
680
+ "model.h.9.attn.q_proj.biases": "model.safetensors",
681
+ "model.h.9.attn.q_proj.scales": "model.safetensors",
682
+ "model.h.9.attn.q_proj.weight": "model.safetensors",
683
+ "model.h.9.attn.v_proj.biases": "model.safetensors",
684
+ "model.h.9.attn.v_proj.scales": "model.safetensors",
685
+ "model.h.9.attn.v_proj.weight": "model.safetensors",
686
+ "model.h.9.ln_1.weight": "model.safetensors",
687
+ "model.h.9.ln_2.weight": "model.safetensors",
688
+ "model.h.9.mlp.w_down.biases": "model.safetensors",
689
+ "model.h.9.mlp.w_down.scales": "model.safetensors",
690
+ "model.h.9.mlp.w_down.weight": "model.safetensors",
691
+ "model.h.9.mlp.w_gate.biases": "model.safetensors",
692
+ "model.h.9.mlp.w_gate.scales": "model.safetensors",
693
+ "model.h.9.mlp.w_gate.weight": "model.safetensors",
694
+ "model.h.9.mlp.w_up.biases": "model.safetensors",
695
+ "model.h.9.mlp.w_up.scales": "model.safetensors",
696
+ "model.h.9.mlp.w_up.weight": "model.safetensors",
697
+ "model.ln_f.weight": "model.safetensors",
698
+ "model.wte.biases": "model.safetensors",
699
+ "model.wte.scales": "model.safetensors",
700
+ "model.wte.weight": "model.safetensors"
701
+ }
702
+ }
modeling_gptx2.py ADDED
@@ -0,0 +1,255 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """GPT-X2.5 causal language model for Hugging Face Transformers."""
2
+
3
+ from typing import Optional
4
+
5
+ import torch
6
+ import torch.nn as nn
7
+ from torch.nn import functional as F
8
+ from transformers import PreTrainedModel
9
+ from transformers.cache_utils import DynamicCache
10
+ from transformers.generation.utils import GenerationMixin
11
+ from transformers.modeling_outputs import CausalLMOutputWithPast
12
+
13
+ from .configuration_gptx2 import GPTX2Config
14
+
15
+
16
+ class RMSNorm(nn.Module):
17
+ def __init__(self, dim, eps=1e-6):
18
+ super().__init__()
19
+ self.eps = eps
20
+ self.weight = nn.Parameter(torch.ones(dim))
21
+
22
+ def forward(self, x):
23
+ rms = torch.rsqrt(x.float().pow(2).mean(-1, keepdim=True) + self.eps)
24
+ return (x.float() * rms).type_as(x) * self.weight
25
+
26
+
27
+ def precompute_rope_cos_sin(head_dim, seq_len, theta=100000.0, device=None):
28
+ inv_freq = 1.0 / (
29
+ theta ** (torch.arange(0, head_dim, 2, dtype=torch.float32, device=device) / head_dim)
30
+ )
31
+ positions = torch.arange(seq_len, dtype=torch.float32, device=device)
32
+ freqs = torch.outer(positions, inv_freq)
33
+ return freqs.cos(), freqs.sin()
34
+
35
+
36
+ def apply_rotary_emb(q, k, rope_cos, rope_sin):
37
+ cos = rope_cos.unsqueeze(0).unsqueeze(0)
38
+ sin = rope_sin.unsqueeze(0).unsqueeze(0)
39
+
40
+ q_float = q.float().reshape(*q.shape[:-1], -1, 2)
41
+ k_float = k.float().reshape(*k.shape[:-1], -1, 2)
42
+ q_even, q_odd = q_float.unbind(-1)
43
+ k_even, k_odd = k_float.unbind(-1)
44
+
45
+ q_out = torch.stack(
46
+ (q_even * cos - q_odd * sin, q_even * sin + q_odd * cos), dim=-1
47
+ ).flatten(-2)
48
+ k_out = torch.stack(
49
+ (k_even * cos - k_odd * sin, k_even * sin + k_odd * cos), dim=-1
50
+ ).flatten(-2)
51
+ return q_out.type_as(q), k_out.type_as(k)
52
+
53
+
54
+ class GPTX2Attention(nn.Module):
55
+ def __init__(self, config, layer_idx):
56
+ super().__init__()
57
+ self.layer_idx = layer_idx
58
+ self.n_head = config.num_attention_heads
59
+ self.n_kv_heads = config.num_key_value_heads
60
+ self.head_dim = config.head_dim
61
+ self.n_rep = self.n_head // self.n_kv_heads
62
+ self.xsa_projection = config.xsa_projection
63
+
64
+ self.q_proj = nn.Linear(config.hidden_size, self.n_head * self.head_dim, bias=False)
65
+ self.k_proj = nn.Linear(config.hidden_size, self.n_kv_heads * self.head_dim, bias=False)
66
+ self.v_proj = nn.Linear(config.hidden_size, self.n_kv_heads * self.head_dim, bias=False)
67
+ self.o_proj = nn.Linear(self.n_head * self.head_dim, config.hidden_size, bias=False)
68
+ self.o_proj.NANOGPT_SCALE_INIT = 1
69
+
70
+ def forward(self, x, rope_cos, rope_sin, past_key_value=None, attention_mask=None):
71
+ batch_size, query_length, _ = x.size()
72
+ q = self.q_proj(x).view(batch_size, query_length, self.n_head, self.head_dim).transpose(1, 2)
73
+ k = self.k_proj(x).view(batch_size, query_length, self.n_kv_heads, self.head_dim).transpose(1, 2)
74
+ v = self.v_proj(x).view(batch_size, query_length, self.n_kv_heads, self.head_dim).transpose(1, 2)
75
+
76
+ q, k = apply_rotary_emb(q, k, rope_cos, rope_sin)
77
+ current_v = v
78
+ if past_key_value is not None:
79
+ k, v = past_key_value.update(k, v, self.layer_idx)
80
+
81
+ key_length = k.size(2)
82
+ k_repeated = k.repeat_interleave(self.n_rep, dim=1)
83
+ v_repeated = v.repeat_interleave(self.n_rep, dim=1)
84
+
85
+ past_length = key_length - query_length
86
+ is_causal = query_length > 1 and past_length == 0
87
+ attn_mask = None
88
+ # An explicit causal mask is required for cached multi-token chunks and
89
+ # whenever a padding mask is present (the latter disables is_causal).
90
+ if query_length > 1 and (past_length > 0 or attention_mask is not None):
91
+ causal = torch.ones(
92
+ query_length, key_length, dtype=torch.bool, device=x.device
93
+ ).tril(diagonal=past_length)
94
+ attn_mask = causal[None, None, :, :]
95
+ if attention_mask is not None:
96
+ key_padding = attention_mask[:, None, None, :key_length].to(torch.bool)
97
+ attn_mask = key_padding if attn_mask is None else (key_padding & attn_mask)
98
+ is_causal = False
99
+
100
+ y = F.scaled_dot_product_attention(
101
+ q, k_repeated, v_repeated, attn_mask=attn_mask, is_causal=is_causal
102
+ )
103
+
104
+ # XSA: remove each query-head output's component parallel to its
105
+ # corresponding current-token KV value. This exactly matches training.
106
+ if self.xsa_projection:
107
+ y = y.view(batch_size, self.n_kv_heads, self.n_rep, query_length, self.head_dim)
108
+ v_grouped = current_v.unsqueeze(2)
109
+ denominator = v_grouped.pow(2).sum(dim=-1, keepdim=True).clamp_min(1e-6)
110
+ y = y - ((y * v_grouped).sum(dim=-1, keepdim=True) / denominator) * v_grouped
111
+ y = y.view(batch_size, self.n_head, query_length, self.head_dim)
112
+
113
+ y = y.transpose(1, 2).contiguous().view(
114
+ batch_size, query_length, self.n_head * self.head_dim
115
+ )
116
+ return self.o_proj(y)
117
+
118
+
119
+ class GPTX2SwiGLUMLP(nn.Module):
120
+ def __init__(self, config):
121
+ super().__init__()
122
+ self.w_gate = nn.Linear(config.hidden_size, config.intermediate_size, bias=False)
123
+ self.w_up = nn.Linear(config.hidden_size, config.intermediate_size, bias=False)
124
+ self.w_down = nn.Linear(config.intermediate_size, config.hidden_size, bias=False)
125
+ self.w_down.NANOGPT_SCALE_INIT = 1
126
+
127
+ def forward(self, x):
128
+ return self.w_down(F.silu(self.w_gate(x)) * self.w_up(x))
129
+
130
+
131
+ class GPTX2Block(nn.Module):
132
+ def __init__(self, config, layer_idx):
133
+ super().__init__()
134
+ self.ln_1 = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
135
+ self.attn = GPTX2Attention(config, layer_idx)
136
+ self.ln_2 = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
137
+ self.mlp = GPTX2SwiGLUMLP(config)
138
+
139
+ def forward(self, x, rope_cos, rope_sin, past_key_value=None, attention_mask=None):
140
+ x = x + self.attn(
141
+ self.ln_1(x), rope_cos, rope_sin, past_key_value, attention_mask
142
+ )
143
+ return x + self.mlp(self.ln_2(x))
144
+
145
+
146
+ class GPTX2PreTrainedModel(PreTrainedModel):
147
+ config_class = GPTX2Config
148
+ base_model_prefix = "transformer"
149
+ supports_gradient_checkpointing = False
150
+
151
+ def _init_weights(self, module):
152
+ std = 0.02
153
+ if hasattr(module, "NANOGPT_SCALE_INIT"):
154
+ std *= 2 * self.config.num_hidden_layers**-0.5
155
+ if isinstance(module, nn.Linear):
156
+ nn.init.normal_(module.weight, mean=0.0, std=std)
157
+ elif isinstance(module, nn.Embedding):
158
+ nn.init.normal_(module.weight, mean=0.0, std=0.02)
159
+
160
+
161
+ class GPTX2ForCausalLM(GPTX2PreTrainedModel, GenerationMixin):
162
+ _tied_weights_keys = {"lm_head.weight": "transformer.wte.weight"}
163
+
164
+ def __init__(self, config):
165
+ super().__init__(config)
166
+ self.transformer = nn.ModuleDict(
167
+ {
168
+ "wte": nn.Embedding(config.vocab_size, config.hidden_size),
169
+ "h": nn.ModuleList(
170
+ [GPTX2Block(config, i) for i in range(config.num_hidden_layers)]
171
+ ),
172
+ "ln_f": RMSNorm(config.hidden_size, eps=config.rms_norm_eps),
173
+ }
174
+ )
175
+ self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
176
+ if config.tie_word_embeddings:
177
+ self.lm_head.weight = self.transformer["wte"].weight
178
+ self._rope_cache = None
179
+ self.post_init()
180
+
181
+ def get_input_embeddings(self):
182
+ return self.transformer["wte"]
183
+
184
+ def set_input_embeddings(self, value):
185
+ self.transformer["wte"] = value
186
+
187
+ def get_output_embeddings(self):
188
+ return self.lm_head
189
+
190
+ def set_output_embeddings(self, value):
191
+ self.lm_head = value
192
+
193
+ def prepare_inputs_for_generation(
194
+ self, input_ids, past_key_values=None, attention_mask=None, **kwargs
195
+ ):
196
+ if past_key_values is not None and past_key_values.get_seq_length() > 0:
197
+ input_ids = input_ids[:, -1:]
198
+ return {
199
+ "input_ids": input_ids,
200
+ "attention_mask": attention_mask,
201
+ "past_key_values": past_key_values,
202
+ "use_cache": True,
203
+ }
204
+
205
+ def _get_rope(self, seq_len, device):
206
+ cache = self._rope_cache
207
+ if cache is None or cache[0].device != device or cache[0].size(0) < seq_len:
208
+ cache = precompute_rope_cos_sin(
209
+ self.config.head_dim, seq_len, self.config.rope_theta, device=device
210
+ )
211
+ self._rope_cache = cache
212
+ return cache[0][:seq_len], cache[1][:seq_len]
213
+
214
+ def forward(
215
+ self,
216
+ input_ids,
217
+ attention_mask=None,
218
+ labels=None,
219
+ past_key_values: Optional[DynamicCache] = None,
220
+ use_cache=False,
221
+ **kwargs,
222
+ ):
223
+ _, query_length = input_ids.size()
224
+ if use_cache and past_key_values is None:
225
+ past_key_values = DynamicCache()
226
+ past_length = past_key_values.get_seq_length() if past_key_values is not None else 0
227
+ if past_length + query_length > self.config.max_position_embeddings:
228
+ raise ValueError(
229
+ f"Sequence length {past_length + query_length} exceeds "
230
+ f"max_position_embeddings={self.config.max_position_embeddings}"
231
+ )
232
+
233
+ x = self.transformer["wte"](input_ids)
234
+ rope_cos, rope_sin = self._get_rope(past_length + query_length, input_ids.device)
235
+ rope_cos = rope_cos[past_length:]
236
+ rope_sin = rope_sin[past_length:]
237
+
238
+ cache = past_key_values if use_cache else None
239
+ for block in self.transformer["h"]:
240
+ x = block(x, rope_cos, rope_sin, cache, attention_mask)
241
+ logits = self.lm_head(self.transformer["ln_f"](x))
242
+
243
+ loss = None
244
+ if labels is not None:
245
+ shift_logits = logits[..., :-1, :].contiguous()
246
+ shift_labels = labels[..., 1:].contiguous()
247
+ loss = F.cross_entropy(
248
+ shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)
249
+ )
250
+
251
+ return CausalLMOutputWithPast(
252
+ loss=loss,
253
+ logits=logits,
254
+ past_key_values=past_key_values if use_cache else None,
255
+ )
tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
tokenizer_config.json ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_prefix_space": false,
3
+ "backend": "tokenizers",
4
+ "bos_token": "<|endoftext|>",
5
+ "clean_up_tokenization_spaces": false,
6
+ "eos_token": "<|endoftext|>",
7
+ "extra_special_tokens": [
8
+ "<|im_start|>",
9
+ "<|im_end|>"
10
+ ],
11
+ "is_local": true,
12
+ "local_files_only": false,
13
+ "model_max_length": 8192,
14
+ "pad_token": "<|padding|>",
15
+ "tokenizer_class": "TokenizersBackend",
16
+ "unk_token": "<|endoftext|>",
17
+ "vocab_size": 32768
18
+ }
usage.py ADDED
@@ -0,0 +1,22 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from transformers import AutoModelForCausalLM, AutoTokenizer
2
+ import torch
3
+
4
+
5
+ model_name = "/Users/alpha/model/GPT-X2.5-135M"
6
+ tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
7
+ model = AutoModelForCausalLM.from_pretrained(
8
+ model_name,
9
+ trust_remote_code=True,
10
+ dtype=torch.float32
11
+ )
12
+
13
+ prompt = "The main is"
14
+ inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
15
+ with torch.inference_mode():
16
+ output = model.generate(
17
+ **inputs,
18
+ max_new_tokens=120,
19
+ temperature=0.8
20
+ )
21
+
22
+ print(tokenizer.decode(output[0], skip_special_tokens=True))