Instructions to use HappyBulingling/MyModel-sft with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use HappyBulingling/MyModel-sft with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="HappyBulingling/MyModel-sft", trust_remote_code=True) messages = [ {"role": "user", "content": "Who are you?"}, ] pipe(messages)# pip install -U transformers accelerate # Load model directly from transformers import AutoModelForCausalLM model = AutoModelForCausalLM.from_pretrained("HappyBulingling/MyModel-sft", trust_remote_code=True, device_map="auto") - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use HappyBulingling/MyModel-sft with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "HappyBulingling/MyModel-sft" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "HappyBulingling/MyModel-sft", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker
docker model run hf.co/HappyBulingling/MyModel-sft
- SGLang
How to use HappyBulingling/MyModel-sft with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "HappyBulingling/MyModel-sft" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "HappyBulingling/MyModel-sft", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "HappyBulingling/MyModel-sft" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "HappyBulingling/MyModel-sft", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }' - Docker Model Runner
How to use HappyBulingling/MyModel-sft with Docker Model Runner:
docker model run hf.co/HappyBulingling/MyModel-sft
Upload MyModel checkpoint
Browse files- README.md +28 -1
- attention.py +124 -0
- block.py +65 -0
- chat_template.jinja +1 -0
- config.json +36 -0
- config.py +127 -0
- generation_config.json +7 -0
- model.py +287 -0
- model.safetensors +3 -0
- norm.py +19 -0
- rope.py +39 -0
- special_tokens_map.json +53 -0
- tokenizer.json +0 -0
- tokenizer_config.json +1064 -0
README.md
CHANGED
|
@@ -1,3 +1,30 @@
|
|
| 1 |
---
|
| 2 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
+
library_name: transformers
|
| 3 |
+
pipeline_tag: text-generation
|
| 4 |
+
tags:
|
| 5 |
+
- mymodel
|
| 6 |
+
- causal-lm
|
| 7 |
---
|
| 8 |
+
|
| 9 |
+
# MyModel-sft
|
| 10 |
+
|
| 11 |
+
MyModel 稠密文本模型(sft-4k),参数量 259M,训练 step 4,968,tokens 0.17B。
|
| 12 |
+
|
| 13 |
+
## 使用
|
| 14 |
+
|
| 15 |
+
```python
|
| 16 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 17 |
+
|
| 18 |
+
repo_id = "HappyBulingling/MyModel-sft"
|
| 19 |
+
tokenizer = AutoTokenizer.from_pretrained(repo_id)
|
| 20 |
+
model = AutoModelForCausalLM.from_pretrained(repo_id, trust_remote_code=True)
|
| 21 |
+
|
| 22 |
+
messages = [{"role": "user", "content": "你好"}]
|
| 23 |
+
inputs = tokenizer.apply_chat_template(
|
| 24 |
+
messages, add_generation_prompt=True, return_tensors="pt", enable_thinking=True
|
| 25 |
+
)
|
| 26 |
+
outputs = model.generate(
|
| 27 |
+
inputs, max_new_tokens=256, do_sample=True, temperature=0.8, top_p=0.9
|
| 28 |
+
)
|
| 29 |
+
print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True))
|
| 30 |
+
```
|
attention.py
ADDED
|
@@ -0,0 +1,124 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Grouped-query attention with q/k RMSNorm and SDPA."""
|
| 2 |
+
|
| 3 |
+
from __future__ import annotations
|
| 4 |
+
|
| 5 |
+
import torch
|
| 6 |
+
import torch.nn.functional as F
|
| 7 |
+
from torch import nn
|
| 8 |
+
|
| 9 |
+
from .config import MyModelConfig
|
| 10 |
+
from .norm import RMSNorm
|
| 11 |
+
from .rope import apply_rotary_pos_emb
|
| 12 |
+
|
| 13 |
+
|
| 14 |
+
def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
|
| 15 |
+
if n_rep == 1:
|
| 16 |
+
return hidden_states
|
| 17 |
+
batch, seq_len, num_kv_heads, head_dim = hidden_states.shape
|
| 18 |
+
hidden_states = hidden_states[:, :, :, None, :].expand(
|
| 19 |
+
batch, seq_len, num_kv_heads, n_rep, head_dim
|
| 20 |
+
)
|
| 21 |
+
return hidden_states.reshape(batch, seq_len, num_kv_heads * n_rep, head_dim)
|
| 22 |
+
|
| 23 |
+
|
| 24 |
+
def causal_mask(
|
| 25 |
+
q_len: int,
|
| 26 |
+
kv_len: int,
|
| 27 |
+
device: torch.device,
|
| 28 |
+
) -> torch.Tensor:
|
| 29 |
+
"""Boolean mask for a query block aligned to the end of the KV cache."""
|
| 30 |
+
rows = torch.arange(q_len, device=device).unsqueeze(1) + (kv_len - q_len)
|
| 31 |
+
cols = torch.arange(kv_len, device=device).unsqueeze(0)
|
| 32 |
+
return cols <= rows
|
| 33 |
+
|
| 34 |
+
|
| 35 |
+
class Attention(nn.Module):
|
| 36 |
+
def __init__(self, config: MyModelConfig):
|
| 37 |
+
super().__init__()
|
| 38 |
+
self.num_heads = config.num_attention_heads
|
| 39 |
+
self.num_kv_heads = config.num_key_value_heads
|
| 40 |
+
self.head_dim = config.head_dim
|
| 41 |
+
self.num_kv_groups = self.num_heads // self.num_kv_heads
|
| 42 |
+
self.dropout = config.dropout
|
| 43 |
+
|
| 44 |
+
self.q_proj = nn.Linear(
|
| 45 |
+
config.hidden_size,
|
| 46 |
+
self.num_heads * self.head_dim,
|
| 47 |
+
bias=config.attention_bias,
|
| 48 |
+
)
|
| 49 |
+
self.k_proj = nn.Linear(
|
| 50 |
+
config.hidden_size,
|
| 51 |
+
self.num_kv_heads * self.head_dim,
|
| 52 |
+
bias=config.attention_bias,
|
| 53 |
+
)
|
| 54 |
+
self.v_proj = nn.Linear(
|
| 55 |
+
config.hidden_size,
|
| 56 |
+
self.num_kv_heads * self.head_dim,
|
| 57 |
+
bias=config.attention_bias,
|
| 58 |
+
)
|
| 59 |
+
self.o_proj = nn.Linear(
|
| 60 |
+
self.num_heads * self.head_dim,
|
| 61 |
+
config.hidden_size,
|
| 62 |
+
bias=config.attention_bias,
|
| 63 |
+
)
|
| 64 |
+
self.q_norm = RMSNorm(self.head_dim, eps=config.rms_norm_eps) if config.qk_norm else nn.Identity()
|
| 65 |
+
self.k_norm = RMSNorm(self.head_dim, eps=config.rms_norm_eps) if config.qk_norm else nn.Identity()
|
| 66 |
+
|
| 67 |
+
def forward(
|
| 68 |
+
self,
|
| 69 |
+
hidden_states: torch.Tensor,
|
| 70 |
+
cos: torch.Tensor,
|
| 71 |
+
sin: torch.Tensor,
|
| 72 |
+
past_key_value: tuple[torch.Tensor, torch.Tensor] | None = None,
|
| 73 |
+
use_cache: bool = False,
|
| 74 |
+
attention_mask: torch.Tensor | None = None,
|
| 75 |
+
) -> tuple[torch.Tensor, tuple[torch.Tensor, torch.Tensor] | None]:
|
| 76 |
+
batch, seq_len, _ = hidden_states.shape
|
| 77 |
+
|
| 78 |
+
query = self.q_proj(hidden_states).view(batch, seq_len, self.num_heads, self.head_dim)
|
| 79 |
+
key = self.k_proj(hidden_states).view(batch, seq_len, self.num_kv_heads, self.head_dim)
|
| 80 |
+
value = self.v_proj(hidden_states).view(batch, seq_len, self.num_kv_heads, self.head_dim)
|
| 81 |
+
|
| 82 |
+
query = self.q_norm(query)
|
| 83 |
+
key = self.k_norm(key)
|
| 84 |
+
query, key = apply_rotary_pos_emb(query, key, cos, sin)
|
| 85 |
+
|
| 86 |
+
if past_key_value is not None:
|
| 87 |
+
key = torch.cat((past_key_value[0], key), dim=1)
|
| 88 |
+
value = torch.cat((past_key_value[1], value), dim=1)
|
| 89 |
+
|
| 90 |
+
present = (key, value) if use_cache else None
|
| 91 |
+
kv_len = key.shape[1]
|
| 92 |
+
|
| 93 |
+
key = repeat_kv(key, self.num_kv_groups).transpose(1, 2)
|
| 94 |
+
value = repeat_kv(value, self.num_kv_groups).transpose(1, 2)
|
| 95 |
+
query = query.transpose(1, 2)
|
| 96 |
+
|
| 97 |
+
dropout_p = self.dropout if self.training else 0.0
|
| 98 |
+
if attention_mask is None and past_key_value is None:
|
| 99 |
+
output = F.scaled_dot_product_attention(
|
| 100 |
+
query,
|
| 101 |
+
key,
|
| 102 |
+
value,
|
| 103 |
+
dropout_p=dropout_p,
|
| 104 |
+
is_causal=True,
|
| 105 |
+
)
|
| 106 |
+
else:
|
| 107 |
+
mask = causal_mask(seq_len, kv_len, hidden_states.device)
|
| 108 |
+
if attention_mask is not None:
|
| 109 |
+
if attention_mask.shape != (batch, kv_len):
|
| 110 |
+
raise ValueError(
|
| 111 |
+
f"attention_mask must have shape {(batch, kv_len)}, "
|
| 112 |
+
f"got {tuple(attention_mask.shape)}"
|
| 113 |
+
)
|
| 114 |
+
mask = mask.unsqueeze(0).unsqueeze(0) & attention_mask[:, None, None, :].bool()
|
| 115 |
+
output = F.scaled_dot_product_attention(
|
| 116 |
+
query,
|
| 117 |
+
key,
|
| 118 |
+
value,
|
| 119 |
+
attn_mask=mask,
|
| 120 |
+
dropout_p=dropout_p,
|
| 121 |
+
)
|
| 122 |
+
|
| 123 |
+
output = output.transpose(1, 2).reshape(batch, seq_len, -1)
|
| 124 |
+
return self.o_proj(output), present
|
block.py
ADDED
|
@@ -0,0 +1,65 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Transformer block: pre-norm GQA + SwiGLU with two residual paths."""
|
| 2 |
+
|
| 3 |
+
from __future__ import annotations
|
| 4 |
+
|
| 5 |
+
import torch
|
| 6 |
+
from torch import nn
|
| 7 |
+
from transformers.activations import ACT2FN
|
| 8 |
+
|
| 9 |
+
from .attention import Attention
|
| 10 |
+
from .config import MyModelConfig
|
| 11 |
+
from .norm import RMSNorm
|
| 12 |
+
|
| 13 |
+
|
| 14 |
+
class FeedForward(nn.Module):
|
| 15 |
+
def __init__(self, config: MyModelConfig):
|
| 16 |
+
super().__init__()
|
| 17 |
+
self.gate_proj = nn.Linear(
|
| 18 |
+
config.hidden_size,
|
| 19 |
+
config.intermediate_size,
|
| 20 |
+
bias=config.mlp_bias,
|
| 21 |
+
)
|
| 22 |
+
self.up_proj = nn.Linear(
|
| 23 |
+
config.hidden_size,
|
| 24 |
+
config.intermediate_size,
|
| 25 |
+
bias=config.mlp_bias,
|
| 26 |
+
)
|
| 27 |
+
self.down_proj = nn.Linear(
|
| 28 |
+
config.intermediate_size,
|
| 29 |
+
config.hidden_size,
|
| 30 |
+
bias=config.mlp_bias,
|
| 31 |
+
)
|
| 32 |
+
self.act_fn = ACT2FN[config.hidden_act]
|
| 33 |
+
|
| 34 |
+
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
| 35 |
+
return self.down_proj(self.act_fn(self.gate_proj(hidden_states)) * self.up_proj(hidden_states))
|
| 36 |
+
|
| 37 |
+
|
| 38 |
+
class Block(nn.Module):
|
| 39 |
+
def __init__(self, config: MyModelConfig):
|
| 40 |
+
super().__init__()
|
| 41 |
+
self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 42 |
+
self.self_attn = Attention(config)
|
| 43 |
+
self.post_attention_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 44 |
+
self.mlp = FeedForward(config)
|
| 45 |
+
|
| 46 |
+
def forward(
|
| 47 |
+
self,
|
| 48 |
+
hidden_states: torch.Tensor,
|
| 49 |
+
cos: torch.Tensor,
|
| 50 |
+
sin: torch.Tensor,
|
| 51 |
+
past_key_value: tuple[torch.Tensor, torch.Tensor] | None = None,
|
| 52 |
+
use_cache: bool = False,
|
| 53 |
+
attention_mask: torch.Tensor | None = None,
|
| 54 |
+
) -> tuple[torch.Tensor, tuple[torch.Tensor, torch.Tensor] | None]:
|
| 55 |
+
attention_output, present = self.self_attn(
|
| 56 |
+
self.input_layernorm(hidden_states),
|
| 57 |
+
cos,
|
| 58 |
+
sin,
|
| 59 |
+
past_key_value,
|
| 60 |
+
use_cache,
|
| 61 |
+
attention_mask,
|
| 62 |
+
)
|
| 63 |
+
hidden_states = hidden_states + attention_output
|
| 64 |
+
hidden_states = hidden_states + self.mlp(self.post_attention_layernorm(hidden_states))
|
| 65 |
+
return hidden_states, present
|
chat_template.jinja
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n'}}{% if message['role'] == 'assistant' %}{% if message.get('reasoning_content') %}{{'<think>\n' + message['reasoning_content'] + '\n</think>\n\n' + message['content']}}{% else %}{{'<think>\n\n</think>\n\n' + message['content']}}{% endif %}{% else %}{{message['content']}}{% endif %}{{'<|im_end|>\n'}}{% endfor %}{% if add_generation_prompt %}{{'<|im_start|>assistant\n'}}{% if enable_thinking is defined and not enable_thinking %}{{'<think>\n\n</think>\n\n'}}{% else %}{{'<think>'}}{% endif %}{% endif %}
|
config.json
ADDED
|
@@ -0,0 +1,36 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"architectures": [
|
| 3 |
+
"MyModelForCausalLM"
|
| 4 |
+
],
|
| 5 |
+
"attention_bias": false,
|
| 6 |
+
"auto_map": {
|
| 7 |
+
"AutoConfig": "config.MyModelConfig",
|
| 8 |
+
"AutoModelForCausalLM": "model.MyModelForCausalLM"
|
| 9 |
+
},
|
| 10 |
+
"bos_token_id": 2,
|
| 11 |
+
"dropout": 0.0,
|
| 12 |
+
"dtype": "float32",
|
| 13 |
+
"eos_token_id": 0,
|
| 14 |
+
"ffn_type": "dense",
|
| 15 |
+
"head_dim": 64,
|
| 16 |
+
"hidden_act": "silu",
|
| 17 |
+
"hidden_size": 1024,
|
| 18 |
+
"im_end_token_id": 5,
|
| 19 |
+
"initializer_range": 0.02,
|
| 20 |
+
"intermediate_size": 2816,
|
| 21 |
+
"max_position_embeddings": 4096,
|
| 22 |
+
"mlp_bias": false,
|
| 23 |
+
"model_type": "mymodel",
|
| 24 |
+
"num_attention_heads": 16,
|
| 25 |
+
"num_hidden_layers": 20,
|
| 26 |
+
"num_key_value_heads": 4,
|
| 27 |
+
"pad_token_id": 1,
|
| 28 |
+
"parametrization": "standard",
|
| 29 |
+
"qk_norm": true,
|
| 30 |
+
"rms_norm_eps": 1e-06,
|
| 31 |
+
"rope_scaling": null,
|
| 32 |
+
"rope_theta": 1000000.0,
|
| 33 |
+
"transformers_version": "4.57.6",
|
| 34 |
+
"unk_token_id": 3,
|
| 35 |
+
"vocab_size": 32768
|
| 36 |
+
}
|
config.py
ADDED
|
@@ -0,0 +1,127 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Model configuration and the two presets used by this project."""
|
| 2 |
+
|
| 3 |
+
from __future__ import annotations
|
| 4 |
+
|
| 5 |
+
from transformers import PretrainedConfig
|
| 6 |
+
|
| 7 |
+
|
| 8 |
+
PRESETS = {
|
| 9 |
+
"tiny": {
|
| 10 |
+
"hidden_size": 256,
|
| 11 |
+
"num_hidden_layers": 4,
|
| 12 |
+
"num_attention_heads": 8,
|
| 13 |
+
"num_key_value_heads": 1,
|
| 14 |
+
"head_dim": 32,
|
| 15 |
+
"intermediate_size": 896,
|
| 16 |
+
"vocab_size": 32768,
|
| 17 |
+
},
|
| 18 |
+
"base": {
|
| 19 |
+
"hidden_size": 1024,
|
| 20 |
+
"num_hidden_layers": 20,
|
| 21 |
+
"num_attention_heads": 16,
|
| 22 |
+
"num_key_value_heads": 4,
|
| 23 |
+
"head_dim": 64,
|
| 24 |
+
"intermediate_size": 2816,
|
| 25 |
+
"vocab_size": 32768,
|
| 26 |
+
},
|
| 27 |
+
}
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
class MyModelConfig(PretrainedConfig):
|
| 31 |
+
"""Configuration for the 259M dense text backbone.
|
| 32 |
+
|
| 33 |
+
Defaults: 20 layers, d=1024, GQA 16:4, SwiGLU 2816, 32K vocabulary,
|
| 34 |
+
tied embeddings.
|
| 35 |
+
"""
|
| 36 |
+
|
| 37 |
+
model_type = "mymodel"
|
| 38 |
+
|
| 39 |
+
def __init__(
|
| 40 |
+
self,
|
| 41 |
+
hidden_size: int = 1024,
|
| 42 |
+
num_hidden_layers: int = 20,
|
| 43 |
+
num_attention_heads: int = 16,
|
| 44 |
+
num_key_value_heads: int = 4,
|
| 45 |
+
head_dim: int = 64,
|
| 46 |
+
intermediate_size: int = 2816,
|
| 47 |
+
vocab_size: int = 32768,
|
| 48 |
+
hidden_act: str = "silu",
|
| 49 |
+
rms_norm_eps: float = 1e-6,
|
| 50 |
+
qk_norm: bool = True,
|
| 51 |
+
rope_theta: float = 1_000_000.0,
|
| 52 |
+
max_position_embeddings: int = 4096,
|
| 53 |
+
rope_scaling: dict | None = None,
|
| 54 |
+
tie_word_embeddings: bool = True,
|
| 55 |
+
attention_bias: bool = False,
|
| 56 |
+
mlp_bias: bool = False,
|
| 57 |
+
dropout: float = 0.0,
|
| 58 |
+
ffn_type: str = "dense",
|
| 59 |
+
parametrization: str = "standard",
|
| 60 |
+
initializer_range: float = 0.02,
|
| 61 |
+
bos_token_id: int = 2,
|
| 62 |
+
eos_token_id: int = 0,
|
| 63 |
+
im_end_token_id: int = 5,
|
| 64 |
+
pad_token_id: int = 1,
|
| 65 |
+
unk_token_id: int = 3,
|
| 66 |
+
**kwargs,
|
| 67 |
+
):
|
| 68 |
+
super().__init__(
|
| 69 |
+
bos_token_id=bos_token_id,
|
| 70 |
+
eos_token_id=eos_token_id,
|
| 71 |
+
pad_token_id=pad_token_id,
|
| 72 |
+
unk_token_id=unk_token_id,
|
| 73 |
+
tie_word_embeddings=tie_word_embeddings,
|
| 74 |
+
**kwargs,
|
| 75 |
+
)
|
| 76 |
+
self.im_end_token_id = im_end_token_id
|
| 77 |
+
if hidden_size % num_attention_heads != 0:
|
| 78 |
+
raise ValueError("hidden_size must be divisible by num_attention_heads")
|
| 79 |
+
if num_attention_heads % num_key_value_heads != 0:
|
| 80 |
+
raise ValueError("num_attention_heads must be divisible by num_key_value_heads")
|
| 81 |
+
if head_dim * num_attention_heads != hidden_size:
|
| 82 |
+
raise ValueError("head_dim * num_attention_heads must equal hidden_size")
|
| 83 |
+
|
| 84 |
+
self.hidden_size = hidden_size
|
| 85 |
+
self.num_hidden_layers = num_hidden_layers
|
| 86 |
+
self.num_attention_heads = num_attention_heads
|
| 87 |
+
self.num_key_value_heads = num_key_value_heads
|
| 88 |
+
self.head_dim = head_dim
|
| 89 |
+
self.intermediate_size = intermediate_size
|
| 90 |
+
self.vocab_size = vocab_size
|
| 91 |
+
self.hidden_act = hidden_act
|
| 92 |
+
self.rms_norm_eps = rms_norm_eps
|
| 93 |
+
self.qk_norm = qk_norm
|
| 94 |
+
self.rope_theta = rope_theta
|
| 95 |
+
self.max_position_embeddings = max_position_embeddings
|
| 96 |
+
self.rope_scaling = rope_scaling
|
| 97 |
+
self.attention_bias = attention_bias
|
| 98 |
+
self.mlp_bias = mlp_bias
|
| 99 |
+
self.dropout = dropout
|
| 100 |
+
self.ffn_type = ffn_type
|
| 101 |
+
self.parametrization = parametrization
|
| 102 |
+
self.initializer_range = initializer_range
|
| 103 |
+
|
| 104 |
+
if self.ffn_type != "dense":
|
| 105 |
+
raise ValueError(f"unsupported ffn_type: {self.ffn_type}")
|
| 106 |
+
if self.parametrization != "standard":
|
| 107 |
+
raise ValueError(f"unsupported parametrization: {self.parametrization}")
|
| 108 |
+
|
| 109 |
+
@classmethod
|
| 110 |
+
def from_preset(cls, name: str, **overrides) -> "MyModelConfig":
|
| 111 |
+
if name not in PRESETS:
|
| 112 |
+
raise ValueError(f"unknown preset {name!r}; choose from {sorted(PRESETS)}")
|
| 113 |
+
values = dict(PRESETS[name])
|
| 114 |
+
values.update(overrides)
|
| 115 |
+
return cls(**values)
|
| 116 |
+
|
| 117 |
+
def estimate_parameters(self) -> int:
|
| 118 |
+
"""Formula-based parameter count, excluding q/k norm parameters."""
|
| 119 |
+
d = self.hidden_size
|
| 120 |
+
attn = 2 * d * (self.num_attention_heads * self.head_dim)
|
| 121 |
+
attn += 2 * d * (self.num_key_value_heads * self.head_dim)
|
| 122 |
+
ffn = 3 * d * self.intermediate_size
|
| 123 |
+
norms = self.num_hidden_layers * 2 * d + d
|
| 124 |
+
if self.qk_norm:
|
| 125 |
+
norms += self.num_hidden_layers * 2 * self.head_dim
|
| 126 |
+
embedding = self.vocab_size * d
|
| 127 |
+
return self.num_hidden_layers * (attn + ffn) + norms + embedding
|
generation_config.json
ADDED
|
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_from_model_config": true,
|
| 3 |
+
"bos_token_id": 2,
|
| 4 |
+
"eos_token_id": 0,
|
| 5 |
+
"pad_token_id": 1,
|
| 6 |
+
"transformers_version": "4.57.6"
|
| 7 |
+
}
|
model.py
ADDED
|
@@ -0,0 +1,287 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""MyModel backbone and causal language-model head."""
|
| 2 |
+
|
| 3 |
+
from __future__ import annotations
|
| 4 |
+
|
| 5 |
+
import math
|
| 6 |
+
from collections.abc import Sequence
|
| 7 |
+
from typing import Iterable
|
| 8 |
+
|
| 9 |
+
import torch
|
| 10 |
+
import torch.nn.functional as F
|
| 11 |
+
from torch import nn
|
| 12 |
+
from torch.utils.checkpoint import checkpoint
|
| 13 |
+
from transformers import GenerationMixin, PreTrainedModel
|
| 14 |
+
from transformers.modeling_outputs import CausalLMOutputWithPast
|
| 15 |
+
|
| 16 |
+
from .block import Block
|
| 17 |
+
from .config import MyModelConfig
|
| 18 |
+
from .norm import RMSNorm
|
| 19 |
+
from .rope import precompute_rope
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
def chunked_cross_entropy(
|
| 23 |
+
hidden_states: torch.Tensor,
|
| 24 |
+
labels: torch.Tensor,
|
| 25 |
+
lm_head: nn.Linear,
|
| 26 |
+
chunk_size: int,
|
| 27 |
+
ignore_index: int = -100,
|
| 28 |
+
) -> torch.Tensor:
|
| 29 |
+
"""Next-token cross entropy without materializing all logits at once."""
|
| 30 |
+
if hidden_states.shape[:2] != labels.shape:
|
| 31 |
+
raise ValueError(
|
| 32 |
+
f"hidden_states {tuple(hidden_states.shape[:2])} and labels "
|
| 33 |
+
f"{tuple(labels.shape)} must agree on batch and sequence length"
|
| 34 |
+
)
|
| 35 |
+
if chunk_size <= 0:
|
| 36 |
+
raise ValueError("chunk_size must be positive")
|
| 37 |
+
|
| 38 |
+
hidden = hidden_states[:, :-1, :]
|
| 39 |
+
target = labels[:, 1:]
|
| 40 |
+
total_loss = hidden_states.new_zeros(())
|
| 41 |
+
total_tokens = 0
|
| 42 |
+
for start in range(0, hidden.shape[1], chunk_size):
|
| 43 |
+
chunk_hidden = hidden[:, start : start + chunk_size, :]
|
| 44 |
+
chunk_target = target[:, start : start + chunk_size]
|
| 45 |
+
logits = lm_head(chunk_hidden)
|
| 46 |
+
total_loss = total_loss + F.cross_entropy(
|
| 47 |
+
logits.reshape(-1, logits.shape[-1]),
|
| 48 |
+
chunk_target.reshape(-1),
|
| 49 |
+
ignore_index=ignore_index,
|
| 50 |
+
reduction="sum",
|
| 51 |
+
)
|
| 52 |
+
total_tokens += int((chunk_target != ignore_index).sum())
|
| 53 |
+
return total_loss / max(total_tokens, 1)
|
| 54 |
+
|
| 55 |
+
|
| 56 |
+
class MyModelModel(nn.Module):
|
| 57 |
+
def __init__(self, config: MyModelConfig):
|
| 58 |
+
super().__init__()
|
| 59 |
+
self.config = config
|
| 60 |
+
self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
|
| 61 |
+
self.dropout = nn.Dropout(config.dropout)
|
| 62 |
+
self.layers = nn.ModuleList([Block(config) for _ in range(config.num_hidden_layers)])
|
| 63 |
+
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 64 |
+
|
| 65 |
+
cos, sin = precompute_rope(
|
| 66 |
+
config.head_dim,
|
| 67 |
+
config.max_position_embeddings,
|
| 68 |
+
config.rope_theta,
|
| 69 |
+
)
|
| 70 |
+
self.register_buffer("rope_cos", cos, persistent=False)
|
| 71 |
+
self.register_buffer("rope_sin", sin, persistent=False)
|
| 72 |
+
self.gradient_checkpointing = False
|
| 73 |
+
|
| 74 |
+
def gradient_checkpointing_enable(self) -> None:
|
| 75 |
+
self.gradient_checkpointing = True
|
| 76 |
+
|
| 77 |
+
def gradient_checkpointing_disable(self) -> None:
|
| 78 |
+
self.gradient_checkpointing = False
|
| 79 |
+
|
| 80 |
+
def forward(
|
| 81 |
+
self,
|
| 82 |
+
input_ids: torch.Tensor,
|
| 83 |
+
attention_mask: torch.Tensor | None = None,
|
| 84 |
+
past_key_values: Iterable[tuple[torch.Tensor, torch.Tensor] | None] | None = None,
|
| 85 |
+
use_cache: bool = False,
|
| 86 |
+
) -> tuple[torch.Tensor, tuple[tuple[torch.Tensor, torch.Tensor], ...] | None]:
|
| 87 |
+
batch, seq_len = input_ids.shape
|
| 88 |
+
if past_key_values is None:
|
| 89 |
+
past_key_values = [None] * len(self.layers)
|
| 90 |
+
else:
|
| 91 |
+
past_key_values = list(past_key_values)
|
| 92 |
+
if len(past_key_values) != len(self.layers):
|
| 93 |
+
raise ValueError(
|
| 94 |
+
f"expected {len(self.layers)} past key/value entries, "
|
| 95 |
+
f"got {len(past_key_values)}"
|
| 96 |
+
)
|
| 97 |
+
|
| 98 |
+
start_pos = 0 if past_key_values[0] is None else past_key_values[0][0].shape[1]
|
| 99 |
+
if start_pos + seq_len > self.config.max_position_embeddings:
|
| 100 |
+
raise ValueError(
|
| 101 |
+
f"sequence length {start_pos + seq_len} exceeds "
|
| 102 |
+
f"max_position_embeddings {self.config.max_position_embeddings}"
|
| 103 |
+
)
|
| 104 |
+
|
| 105 |
+
hidden_states = self.dropout(self.embed_tokens(input_ids))
|
| 106 |
+
cos = self.rope_cos[start_pos : start_pos + seq_len]
|
| 107 |
+
sin = self.rope_sin[start_pos : start_pos + seq_len]
|
| 108 |
+
|
| 109 |
+
presents: list[tuple[torch.Tensor, torch.Tensor] | None] = []
|
| 110 |
+
for layer, past_key_value in zip(self.layers, past_key_values):
|
| 111 |
+
if self.gradient_checkpointing and self.training:
|
| 112 |
+
hidden_states, _ = checkpoint(
|
| 113 |
+
layer,
|
| 114 |
+
hidden_states,
|
| 115 |
+
cos,
|
| 116 |
+
sin,
|
| 117 |
+
past_key_value,
|
| 118 |
+
False,
|
| 119 |
+
attention_mask,
|
| 120 |
+
use_reentrant=False,
|
| 121 |
+
)
|
| 122 |
+
presents.append(None)
|
| 123 |
+
else:
|
| 124 |
+
hidden_states, present = layer(
|
| 125 |
+
hidden_states,
|
| 126 |
+
cos,
|
| 127 |
+
sin,
|
| 128 |
+
past_key_value,
|
| 129 |
+
use_cache,
|
| 130 |
+
attention_mask,
|
| 131 |
+
)
|
| 132 |
+
presents.append(present)
|
| 133 |
+
|
| 134 |
+
hidden_states = self.norm(hidden_states)
|
| 135 |
+
if not use_cache:
|
| 136 |
+
return hidden_states, None
|
| 137 |
+
return hidden_states, tuple(presents)
|
| 138 |
+
|
| 139 |
+
|
| 140 |
+
class MyModelForCausalLM(PreTrainedModel, GenerationMixin):
|
| 141 |
+
config_class = MyModelConfig
|
| 142 |
+
_tied_weights_keys = ["lm_head.weight"]
|
| 143 |
+
|
| 144 |
+
def __init__(self, config: MyModelConfig):
|
| 145 |
+
super().__init__(config)
|
| 146 |
+
self.model = MyModelModel(config)
|
| 147 |
+
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
|
| 148 |
+
if config.tie_word_embeddings:
|
| 149 |
+
self.lm_head.weight = self.model.embed_tokens.weight
|
| 150 |
+
self.post_init()
|
| 151 |
+
self._scale_residual_projections()
|
| 152 |
+
|
| 153 |
+
def _init_weights(self, module: nn.Module) -> None:
|
| 154 |
+
if isinstance(module, nn.Linear):
|
| 155 |
+
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
|
| 156 |
+
if module.bias is not None:
|
| 157 |
+
module.bias.data.zero_()
|
| 158 |
+
elif isinstance(module, nn.Embedding):
|
| 159 |
+
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
|
| 160 |
+
|
| 161 |
+
def _scale_residual_projections(self) -> None:
|
| 162 |
+
scale = 1.0 / math.sqrt(2 * self.config.num_hidden_layers)
|
| 163 |
+
with torch.no_grad():
|
| 164 |
+
for layer in self.model.layers:
|
| 165 |
+
layer.self_attn.o_proj.weight.mul_(scale)
|
| 166 |
+
layer.mlp.down_proj.weight.mul_(scale)
|
| 167 |
+
|
| 168 |
+
def gradient_checkpointing_enable(self) -> None:
|
| 169 |
+
self.model.gradient_checkpointing_enable()
|
| 170 |
+
|
| 171 |
+
def gradient_checkpointing_disable(self) -> None:
|
| 172 |
+
self.model.gradient_checkpointing_disable()
|
| 173 |
+
|
| 174 |
+
def forward(
|
| 175 |
+
self,
|
| 176 |
+
input_ids: torch.Tensor,
|
| 177 |
+
attention_mask: torch.Tensor | None = None,
|
| 178 |
+
past_key_values: Iterable[tuple[torch.Tensor, torch.Tensor] | None] | None = None,
|
| 179 |
+
use_cache: bool = False,
|
| 180 |
+
labels: torch.Tensor | None = None,
|
| 181 |
+
chunk_size: int = 0,
|
| 182 |
+
logits_to_keep: int = 0,
|
| 183 |
+
**kwargs,
|
| 184 |
+
) -> CausalLMOutputWithPast:
|
| 185 |
+
hidden_states, presents = self.model(
|
| 186 |
+
input_ids,
|
| 187 |
+
attention_mask=attention_mask,
|
| 188 |
+
past_key_values=past_key_values,
|
| 189 |
+
use_cache=use_cache,
|
| 190 |
+
)
|
| 191 |
+
|
| 192 |
+
loss = None
|
| 193 |
+
logits = None
|
| 194 |
+
if labels is not None and chunk_size and chunk_size > 0:
|
| 195 |
+
loss = chunked_cross_entropy(hidden_states, labels, self.lm_head, chunk_size)
|
| 196 |
+
else:
|
| 197 |
+
logits = self.lm_head(hidden_states)
|
| 198 |
+
if labels is not None:
|
| 199 |
+
shift_logits = logits[:, :-1, :]
|
| 200 |
+
shift_labels = labels[:, 1:]
|
| 201 |
+
loss = F.cross_entropy(
|
| 202 |
+
shift_logits.reshape(-1, shift_logits.shape[-1]),
|
| 203 |
+
shift_labels.reshape(-1),
|
| 204 |
+
ignore_index=-100,
|
| 205 |
+
)
|
| 206 |
+
if logits_to_keep > 0:
|
| 207 |
+
logits = logits[:, -logits_to_keep:, :]
|
| 208 |
+
|
| 209 |
+
return CausalLMOutputWithPast(
|
| 210 |
+
loss=loss,
|
| 211 |
+
logits=logits,
|
| 212 |
+
past_key_values=presents,
|
| 213 |
+
hidden_states=hidden_states,
|
| 214 |
+
attentions=None,
|
| 215 |
+
)
|
| 216 |
+
|
| 217 |
+
@torch.inference_mode()
|
| 218 |
+
def generate(
|
| 219 |
+
self,
|
| 220 |
+
input_ids: torch.Tensor,
|
| 221 |
+
max_new_tokens: int = 64,
|
| 222 |
+
temperature: float = 1.0,
|
| 223 |
+
top_k: int = 0,
|
| 224 |
+
top_p: float = 1.0,
|
| 225 |
+
eos_token_id: int | Sequence[int] | None = None,
|
| 226 |
+
do_sample: bool = True,
|
| 227 |
+
use_cache: bool = True,
|
| 228 |
+
) -> torch.Tensor:
|
| 229 |
+
if temperature <= 0:
|
| 230 |
+
raise ValueError("temperature must be positive")
|
| 231 |
+
if eos_token_id is None:
|
| 232 |
+
stop_ids = [int(self.config.eos_token_id)]
|
| 233 |
+
im_end = getattr(self.config, "im_end_token_id", None)
|
| 234 |
+
if im_end is not None:
|
| 235 |
+
stop_ids.append(int(im_end))
|
| 236 |
+
elif isinstance(eos_token_id, int):
|
| 237 |
+
stop_ids = [eos_token_id]
|
| 238 |
+
else:
|
| 239 |
+
stop_ids = [int(token_id) for token_id in eos_token_id]
|
| 240 |
+
stop_ids = list(dict.fromkeys(stop_ids))
|
| 241 |
+
if not stop_ids:
|
| 242 |
+
raise ValueError("eos_token_id must contain at least one token id")
|
| 243 |
+
stop_tensor = torch.tensor(
|
| 244 |
+
stop_ids, dtype=torch.long, device=input_ids.device
|
| 245 |
+
)
|
| 246 |
+
fill_id = stop_ids[0]
|
| 247 |
+
|
| 248 |
+
past_key_values = None
|
| 249 |
+
finished = torch.zeros(input_ids.shape[0], dtype=torch.bool, device=input_ids.device)
|
| 250 |
+
for _ in range(max_new_tokens):
|
| 251 |
+
past_len = 0 if past_key_values is None else past_key_values[0][0].shape[1]
|
| 252 |
+
outputs = self(
|
| 253 |
+
input_ids[:, past_len:],
|
| 254 |
+
past_key_values=past_key_values,
|
| 255 |
+
use_cache=use_cache,
|
| 256 |
+
)
|
| 257 |
+
logits = outputs.logits[:, -1, :] / temperature
|
| 258 |
+
|
| 259 |
+
if top_k > 0:
|
| 260 |
+
threshold = torch.topk(logits, min(top_k, logits.shape[-1]), dim=-1).values[..., -1:]
|
| 261 |
+
logits = logits.masked_fill(logits < threshold, float("-inf"))
|
| 262 |
+
if top_p < 1.0:
|
| 263 |
+
sorted_logits, sorted_indices = torch.sort(logits, descending=True, dim=-1)
|
| 264 |
+
cumulative = torch.cumsum(torch.softmax(sorted_logits, dim=-1), dim=-1)
|
| 265 |
+
remove = cumulative > top_p
|
| 266 |
+
remove[..., 1:] = remove[..., :-1].clone()
|
| 267 |
+
remove[..., 0] = False
|
| 268 |
+
logits = logits.masked_fill(remove.scatter(1, sorted_indices, remove), float("-inf"))
|
| 269 |
+
|
| 270 |
+
if do_sample:
|
| 271 |
+
probs = torch.softmax(logits, dim=-1)
|
| 272 |
+
next_token = torch.multinomial(probs, num_samples=1)
|
| 273 |
+
else:
|
| 274 |
+
next_token = torch.argmax(logits, dim=-1, keepdim=True)
|
| 275 |
+
|
| 276 |
+
next_token = torch.where(
|
| 277 |
+
finished.unsqueeze(-1),
|
| 278 |
+
next_token.new_full((next_token.shape[0], 1), fill_id),
|
| 279 |
+
next_token,
|
| 280 |
+
)
|
| 281 |
+
input_ids = torch.cat((input_ids, next_token), dim=-1)
|
| 282 |
+
past_key_values = outputs.past_key_values if use_cache else None
|
| 283 |
+
|
| 284 |
+
finished |= (next_token == stop_tensor).any(dim=-1).reshape(-1)
|
| 285 |
+
if finished.all():
|
| 286 |
+
break
|
| 287 |
+
return input_ids
|
model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f7e91b5e476ad9d95071cad2a271ad79694e6e0e5fdaff1846f2575ce601589b
|
| 3 |
+
size 1036196104
|
norm.py
ADDED
|
@@ -0,0 +1,19 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""RMSNorm with fp32 statistics and parameter-free operations."""
|
| 2 |
+
|
| 3 |
+
from __future__ import annotations
|
| 4 |
+
|
| 5 |
+
import torch
|
| 6 |
+
from torch import nn
|
| 7 |
+
|
| 8 |
+
|
| 9 |
+
class RMSNorm(nn.Module):
|
| 10 |
+
def __init__(self, hidden_size: int, eps: float = 1e-6):
|
| 11 |
+
super().__init__()
|
| 12 |
+
self.weight = nn.Parameter(torch.ones(hidden_size))
|
| 13 |
+
self.eps = eps
|
| 14 |
+
|
| 15 |
+
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
| 16 |
+
dtype = hidden_states.dtype
|
| 17 |
+
x = hidden_states.float()
|
| 18 |
+
x = x * torch.rsqrt(x.pow(2).mean(dim=-1, keepdim=True) + self.eps)
|
| 19 |
+
return (self.weight.float() * x).to(dtype)
|
rope.py
ADDED
|
@@ -0,0 +1,39 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Rotary position embeddings."""
|
| 2 |
+
|
| 3 |
+
from __future__ import annotations
|
| 4 |
+
|
| 5 |
+
import torch
|
| 6 |
+
|
| 7 |
+
|
| 8 |
+
def precompute_rope(
|
| 9 |
+
head_dim: int,
|
| 10 |
+
max_position_embeddings: int,
|
| 11 |
+
rope_theta: float,
|
| 12 |
+
) -> tuple[torch.Tensor, torch.Tensor]:
|
| 13 |
+
if head_dim % 2 != 0:
|
| 14 |
+
raise ValueError("head_dim must be even for RoPE")
|
| 15 |
+
inv_freq = 1.0 / (
|
| 16 |
+
rope_theta ** (torch.arange(0, head_dim, 2, dtype=torch.float32) / head_dim)
|
| 17 |
+
)
|
| 18 |
+
positions = torch.arange(max_position_embeddings, dtype=torch.float32)
|
| 19 |
+
freqs = torch.outer(positions, inv_freq)
|
| 20 |
+
emb = torch.cat((freqs, freqs), dim=-1)
|
| 21 |
+
return emb.cos(), emb.sin()
|
| 22 |
+
|
| 23 |
+
|
| 24 |
+
def rotate_half(x: torch.Tensor) -> torch.Tensor:
|
| 25 |
+
half = x.shape[-1] // 2
|
| 26 |
+
return torch.cat((-x[..., half:], x[..., :half]), dim=-1)
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
def apply_rotary_pos_emb(
|
| 30 |
+
q: torch.Tensor,
|
| 31 |
+
k: torch.Tensor,
|
| 32 |
+
cos: torch.Tensor,
|
| 33 |
+
sin: torch.Tensor,
|
| 34 |
+
) -> tuple[torch.Tensor, torch.Tensor]:
|
| 35 |
+
cos = cos.to(q.device).unsqueeze(0).unsqueeze(2)
|
| 36 |
+
sin = sin.to(q.device).unsqueeze(0).unsqueeze(2)
|
| 37 |
+
q_embed = (q * cos) + (rotate_half(q) * sin)
|
| 38 |
+
k_embed = (k * cos) + (rotate_half(k) * sin)
|
| 39 |
+
return q_embed, k_embed
|
special_tokens_map.json
ADDED
|
@@ -0,0 +1,53 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"additional_special_tokens": [
|
| 3 |
+
"<|im_start|>",
|
| 4 |
+
"<|im_end|>",
|
| 5 |
+
"<|system|>",
|
| 6 |
+
"<|user|>",
|
| 7 |
+
"<|assistant|>",
|
| 8 |
+
"<|tool|>",
|
| 9 |
+
"<think>",
|
| 10 |
+
"</think>",
|
| 11 |
+
"<tool_call>",
|
| 12 |
+
"</tool_call>",
|
| 13 |
+
"<tool_response>",
|
| 14 |
+
"</tool_response>",
|
| 15 |
+
"<image_start>",
|
| 16 |
+
"<image_pad>",
|
| 17 |
+
"<image_end>",
|
| 18 |
+
"<audio_start>",
|
| 19 |
+
"<audio_pad>",
|
| 20 |
+
"<audio_end>",
|
| 21 |
+
"<speech_bos>",
|
| 22 |
+
"<speech_eos>",
|
| 23 |
+
"<speech_pad>"
|
| 24 |
+
],
|
| 25 |
+
"bos_token": {
|
| 26 |
+
"content": "<|bos|>",
|
| 27 |
+
"lstrip": false,
|
| 28 |
+
"normalized": false,
|
| 29 |
+
"rstrip": false,
|
| 30 |
+
"single_word": false
|
| 31 |
+
},
|
| 32 |
+
"eos_token": {
|
| 33 |
+
"content": "<|eos|>",
|
| 34 |
+
"lstrip": false,
|
| 35 |
+
"normalized": false,
|
| 36 |
+
"rstrip": false,
|
| 37 |
+
"single_word": false
|
| 38 |
+
},
|
| 39 |
+
"pad_token": {
|
| 40 |
+
"content": "<|pad|>",
|
| 41 |
+
"lstrip": false,
|
| 42 |
+
"normalized": false,
|
| 43 |
+
"rstrip": false,
|
| 44 |
+
"single_word": false
|
| 45 |
+
},
|
| 46 |
+
"unk_token": {
|
| 47 |
+
"content": "<|unk|>",
|
| 48 |
+
"lstrip": false,
|
| 49 |
+
"normalized": false,
|
| 50 |
+
"rstrip": false,
|
| 51 |
+
"single_word": false
|
| 52 |
+
}
|
| 53 |
+
}
|
tokenizer.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
tokenizer_config.json
ADDED
|
@@ -0,0 +1,1064 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"add_bos_token": false,
|
| 3 |
+
"add_eos_token": false,
|
| 4 |
+
"add_prefix_space": false,
|
| 5 |
+
"added_tokens_decoder": {
|
| 6 |
+
"0": {
|
| 7 |
+
"content": "<|eos|>",
|
| 8 |
+
"lstrip": false,
|
| 9 |
+
"normalized": false,
|
| 10 |
+
"rstrip": false,
|
| 11 |
+
"single_word": false,
|
| 12 |
+
"special": true
|
| 13 |
+
},
|
| 14 |
+
"1": {
|
| 15 |
+
"content": "<|pad|>",
|
| 16 |
+
"lstrip": false,
|
| 17 |
+
"normalized": false,
|
| 18 |
+
"rstrip": false,
|
| 19 |
+
"single_word": false,
|
| 20 |
+
"special": true
|
| 21 |
+
},
|
| 22 |
+
"2": {
|
| 23 |
+
"content": "<|bos|>",
|
| 24 |
+
"lstrip": false,
|
| 25 |
+
"normalized": false,
|
| 26 |
+
"rstrip": false,
|
| 27 |
+
"single_word": false,
|
| 28 |
+
"special": true
|
| 29 |
+
},
|
| 30 |
+
"3": {
|
| 31 |
+
"content": "<|unk|>",
|
| 32 |
+
"lstrip": false,
|
| 33 |
+
"normalized": false,
|
| 34 |
+
"rstrip": false,
|
| 35 |
+
"single_word": false,
|
| 36 |
+
"special": true
|
| 37 |
+
},
|
| 38 |
+
"4": {
|
| 39 |
+
"content": "<|im_start|>",
|
| 40 |
+
"lstrip": false,
|
| 41 |
+
"normalized": false,
|
| 42 |
+
"rstrip": false,
|
| 43 |
+
"single_word": false,
|
| 44 |
+
"special": true
|
| 45 |
+
},
|
| 46 |
+
"5": {
|
| 47 |
+
"content": "<|im_end|>",
|
| 48 |
+
"lstrip": false,
|
| 49 |
+
"normalized": false,
|
| 50 |
+
"rstrip": false,
|
| 51 |
+
"single_word": false,
|
| 52 |
+
"special": true
|
| 53 |
+
},
|
| 54 |
+
"6": {
|
| 55 |
+
"content": "<|system|>",
|
| 56 |
+
"lstrip": false,
|
| 57 |
+
"normalized": false,
|
| 58 |
+
"rstrip": false,
|
| 59 |
+
"single_word": false,
|
| 60 |
+
"special": true
|
| 61 |
+
},
|
| 62 |
+
"7": {
|
| 63 |
+
"content": "<|user|>",
|
| 64 |
+
"lstrip": false,
|
| 65 |
+
"normalized": false,
|
| 66 |
+
"rstrip": false,
|
| 67 |
+
"single_word": false,
|
| 68 |
+
"special": true
|
| 69 |
+
},
|
| 70 |
+
"8": {
|
| 71 |
+
"content": "<|assistant|>",
|
| 72 |
+
"lstrip": false,
|
| 73 |
+
"normalized": false,
|
| 74 |
+
"rstrip": false,
|
| 75 |
+
"single_word": false,
|
| 76 |
+
"special": true
|
| 77 |
+
},
|
| 78 |
+
"9": {
|
| 79 |
+
"content": "<|tool|>",
|
| 80 |
+
"lstrip": false,
|
| 81 |
+
"normalized": false,
|
| 82 |
+
"rstrip": false,
|
| 83 |
+
"single_word": false,
|
| 84 |
+
"special": true
|
| 85 |
+
},
|
| 86 |
+
"10": {
|
| 87 |
+
"content": "<think>",
|
| 88 |
+
"lstrip": false,
|
| 89 |
+
"normalized": false,
|
| 90 |
+
"rstrip": false,
|
| 91 |
+
"single_word": false,
|
| 92 |
+
"special": true
|
| 93 |
+
},
|
| 94 |
+
"11": {
|
| 95 |
+
"content": "</think>",
|
| 96 |
+
"lstrip": false,
|
| 97 |
+
"normalized": false,
|
| 98 |
+
"rstrip": false,
|
| 99 |
+
"single_word": false,
|
| 100 |
+
"special": true
|
| 101 |
+
},
|
| 102 |
+
"12": {
|
| 103 |
+
"content": "<tool_call>",
|
| 104 |
+
"lstrip": false,
|
| 105 |
+
"normalized": false,
|
| 106 |
+
"rstrip": false,
|
| 107 |
+
"single_word": false,
|
| 108 |
+
"special": true
|
| 109 |
+
},
|
| 110 |
+
"13": {
|
| 111 |
+
"content": "</tool_call>",
|
| 112 |
+
"lstrip": false,
|
| 113 |
+
"normalized": false,
|
| 114 |
+
"rstrip": false,
|
| 115 |
+
"single_word": false,
|
| 116 |
+
"special": true
|
| 117 |
+
},
|
| 118 |
+
"14": {
|
| 119 |
+
"content": "<tool_response>",
|
| 120 |
+
"lstrip": false,
|
| 121 |
+
"normalized": false,
|
| 122 |
+
"rstrip": false,
|
| 123 |
+
"single_word": false,
|
| 124 |
+
"special": true
|
| 125 |
+
},
|
| 126 |
+
"15": {
|
| 127 |
+
"content": "</tool_response>",
|
| 128 |
+
"lstrip": false,
|
| 129 |
+
"normalized": false,
|
| 130 |
+
"rstrip": false,
|
| 131 |
+
"single_word": false,
|
| 132 |
+
"special": true
|
| 133 |
+
},
|
| 134 |
+
"16": {
|
| 135 |
+
"content": "<image_start>",
|
| 136 |
+
"lstrip": false,
|
| 137 |
+
"normalized": false,
|
| 138 |
+
"rstrip": false,
|
| 139 |
+
"single_word": false,
|
| 140 |
+
"special": true
|
| 141 |
+
},
|
| 142 |
+
"17": {
|
| 143 |
+
"content": "<image_pad>",
|
| 144 |
+
"lstrip": false,
|
| 145 |
+
"normalized": false,
|
| 146 |
+
"rstrip": false,
|
| 147 |
+
"single_word": false,
|
| 148 |
+
"special": true
|
| 149 |
+
},
|
| 150 |
+
"18": {
|
| 151 |
+
"content": "<image_end>",
|
| 152 |
+
"lstrip": false,
|
| 153 |
+
"normalized": false,
|
| 154 |
+
"rstrip": false,
|
| 155 |
+
"single_word": false,
|
| 156 |
+
"special": true
|
| 157 |
+
},
|
| 158 |
+
"19": {
|
| 159 |
+
"content": "<audio_start>",
|
| 160 |
+
"lstrip": false,
|
| 161 |
+
"normalized": false,
|
| 162 |
+
"rstrip": false,
|
| 163 |
+
"single_word": false,
|
| 164 |
+
"special": true
|
| 165 |
+
},
|
| 166 |
+
"20": {
|
| 167 |
+
"content": "<audio_pad>",
|
| 168 |
+
"lstrip": false,
|
| 169 |
+
"normalized": false,
|
| 170 |
+
"rstrip": false,
|
| 171 |
+
"single_word": false,
|
| 172 |
+
"special": true
|
| 173 |
+
},
|
| 174 |
+
"21": {
|
| 175 |
+
"content": "<audio_end>",
|
| 176 |
+
"lstrip": false,
|
| 177 |
+
"normalized": false,
|
| 178 |
+
"rstrip": false,
|
| 179 |
+
"single_word": false,
|
| 180 |
+
"special": true
|
| 181 |
+
},
|
| 182 |
+
"22": {
|
| 183 |
+
"content": "<speech_bos>",
|
| 184 |
+
"lstrip": false,
|
| 185 |
+
"normalized": false,
|
| 186 |
+
"rstrip": false,
|
| 187 |
+
"single_word": false,
|
| 188 |
+
"special": true
|
| 189 |
+
},
|
| 190 |
+
"23": {
|
| 191 |
+
"content": "<speech_eos>",
|
| 192 |
+
"lstrip": false,
|
| 193 |
+
"normalized": false,
|
| 194 |
+
"rstrip": false,
|
| 195 |
+
"single_word": false,
|
| 196 |
+
"special": true
|
| 197 |
+
},
|
| 198 |
+
"24": {
|
| 199 |
+
"content": "<speech_pad>",
|
| 200 |
+
"lstrip": false,
|
| 201 |
+
"normalized": false,
|
| 202 |
+
"rstrip": false,
|
| 203 |
+
"single_word": false,
|
| 204 |
+
"special": true
|
| 205 |
+
},
|
| 206 |
+
"25": {
|
| 207 |
+
"content": "<|reserved_1|>",
|
| 208 |
+
"lstrip": false,
|
| 209 |
+
"normalized": false,
|
| 210 |
+
"rstrip": false,
|
| 211 |
+
"single_word": false,
|
| 212 |
+
"special": true
|
| 213 |
+
},
|
| 214 |
+
"26": {
|
| 215 |
+
"content": "<|reserved_2|>",
|
| 216 |
+
"lstrip": false,
|
| 217 |
+
"normalized": false,
|
| 218 |
+
"rstrip": false,
|
| 219 |
+
"single_word": false,
|
| 220 |
+
"special": true
|
| 221 |
+
},
|
| 222 |
+
"27": {
|
| 223 |
+
"content": "<|reserved_3|>",
|
| 224 |
+
"lstrip": false,
|
| 225 |
+
"normalized": false,
|
| 226 |
+
"rstrip": false,
|
| 227 |
+
"single_word": false,
|
| 228 |
+
"special": true
|
| 229 |
+
},
|
| 230 |
+
"28": {
|
| 231 |
+
"content": "<|reserved_4|>",
|
| 232 |
+
"lstrip": false,
|
| 233 |
+
"normalized": false,
|
| 234 |
+
"rstrip": false,
|
| 235 |
+
"single_word": false,
|
| 236 |
+
"special": true
|
| 237 |
+
},
|
| 238 |
+
"29": {
|
| 239 |
+
"content": "<|reserved_5|>",
|
| 240 |
+
"lstrip": false,
|
| 241 |
+
"normalized": false,
|
| 242 |
+
"rstrip": false,
|
| 243 |
+
"single_word": false,
|
| 244 |
+
"special": true
|
| 245 |
+
},
|
| 246 |
+
"30": {
|
| 247 |
+
"content": "<|reserved_6|>",
|
| 248 |
+
"lstrip": false,
|
| 249 |
+
"normalized": false,
|
| 250 |
+
"rstrip": false,
|
| 251 |
+
"single_word": false,
|
| 252 |
+
"special": true
|
| 253 |
+
},
|
| 254 |
+
"31": {
|
| 255 |
+
"content": "<|reserved_7|>",
|
| 256 |
+
"lstrip": false,
|
| 257 |
+
"normalized": false,
|
| 258 |
+
"rstrip": false,
|
| 259 |
+
"single_word": false,
|
| 260 |
+
"special": true
|
| 261 |
+
},
|
| 262 |
+
"32": {
|
| 263 |
+
"content": "<|reserved_8|>",
|
| 264 |
+
"lstrip": false,
|
| 265 |
+
"normalized": false,
|
| 266 |
+
"rstrip": false,
|
| 267 |
+
"single_word": false,
|
| 268 |
+
"special": true
|
| 269 |
+
},
|
| 270 |
+
"33": {
|
| 271 |
+
"content": "<|reserved_9|>",
|
| 272 |
+
"lstrip": false,
|
| 273 |
+
"normalized": false,
|
| 274 |
+
"rstrip": false,
|
| 275 |
+
"single_word": false,
|
| 276 |
+
"special": true
|
| 277 |
+
},
|
| 278 |
+
"34": {
|
| 279 |
+
"content": "<|reserved_10|>",
|
| 280 |
+
"lstrip": false,
|
| 281 |
+
"normalized": false,
|
| 282 |
+
"rstrip": false,
|
| 283 |
+
"single_word": false,
|
| 284 |
+
"special": true
|
| 285 |
+
},
|
| 286 |
+
"35": {
|
| 287 |
+
"content": "<|reserved_11|>",
|
| 288 |
+
"lstrip": false,
|
| 289 |
+
"normalized": false,
|
| 290 |
+
"rstrip": false,
|
| 291 |
+
"single_word": false,
|
| 292 |
+
"special": true
|
| 293 |
+
},
|
| 294 |
+
"36": {
|
| 295 |
+
"content": "<|reserved_12|>",
|
| 296 |
+
"lstrip": false,
|
| 297 |
+
"normalized": false,
|
| 298 |
+
"rstrip": false,
|
| 299 |
+
"single_word": false,
|
| 300 |
+
"special": true
|
| 301 |
+
},
|
| 302 |
+
"37": {
|
| 303 |
+
"content": "<|reserved_13|>",
|
| 304 |
+
"lstrip": false,
|
| 305 |
+
"normalized": false,
|
| 306 |
+
"rstrip": false,
|
| 307 |
+
"single_word": false,
|
| 308 |
+
"special": true
|
| 309 |
+
},
|
| 310 |
+
"38": {
|
| 311 |
+
"content": "<|reserved_14|>",
|
| 312 |
+
"lstrip": false,
|
| 313 |
+
"normalized": false,
|
| 314 |
+
"rstrip": false,
|
| 315 |
+
"single_word": false,
|
| 316 |
+
"special": true
|
| 317 |
+
},
|
| 318 |
+
"39": {
|
| 319 |
+
"content": "<|reserved_15|>",
|
| 320 |
+
"lstrip": false,
|
| 321 |
+
"normalized": false,
|
| 322 |
+
"rstrip": false,
|
| 323 |
+
"single_word": false,
|
| 324 |
+
"special": true
|
| 325 |
+
},
|
| 326 |
+
"40": {
|
| 327 |
+
"content": "<|reserved_16|>",
|
| 328 |
+
"lstrip": false,
|
| 329 |
+
"normalized": false,
|
| 330 |
+
"rstrip": false,
|
| 331 |
+
"single_word": false,
|
| 332 |
+
"special": true
|
| 333 |
+
},
|
| 334 |
+
"41": {
|
| 335 |
+
"content": "<|reserved_17|>",
|
| 336 |
+
"lstrip": false,
|
| 337 |
+
"normalized": false,
|
| 338 |
+
"rstrip": false,
|
| 339 |
+
"single_word": false,
|
| 340 |
+
"special": true
|
| 341 |
+
},
|
| 342 |
+
"42": {
|
| 343 |
+
"content": "<|reserved_18|>",
|
| 344 |
+
"lstrip": false,
|
| 345 |
+
"normalized": false,
|
| 346 |
+
"rstrip": false,
|
| 347 |
+
"single_word": false,
|
| 348 |
+
"special": true
|
| 349 |
+
},
|
| 350 |
+
"43": {
|
| 351 |
+
"content": "<|reserved_19|>",
|
| 352 |
+
"lstrip": false,
|
| 353 |
+
"normalized": false,
|
| 354 |
+
"rstrip": false,
|
| 355 |
+
"single_word": false,
|
| 356 |
+
"special": true
|
| 357 |
+
},
|
| 358 |
+
"44": {
|
| 359 |
+
"content": "<|reserved_20|>",
|
| 360 |
+
"lstrip": false,
|
| 361 |
+
"normalized": false,
|
| 362 |
+
"rstrip": false,
|
| 363 |
+
"single_word": false,
|
| 364 |
+
"special": true
|
| 365 |
+
},
|
| 366 |
+
"45": {
|
| 367 |
+
"content": "<|reserved_21|>",
|
| 368 |
+
"lstrip": false,
|
| 369 |
+
"normalized": false,
|
| 370 |
+
"rstrip": false,
|
| 371 |
+
"single_word": false,
|
| 372 |
+
"special": true
|
| 373 |
+
},
|
| 374 |
+
"46": {
|
| 375 |
+
"content": "<|reserved_22|>",
|
| 376 |
+
"lstrip": false,
|
| 377 |
+
"normalized": false,
|
| 378 |
+
"rstrip": false,
|
| 379 |
+
"single_word": false,
|
| 380 |
+
"special": true
|
| 381 |
+
},
|
| 382 |
+
"47": {
|
| 383 |
+
"content": "<|reserved_23|>",
|
| 384 |
+
"lstrip": false,
|
| 385 |
+
"normalized": false,
|
| 386 |
+
"rstrip": false,
|
| 387 |
+
"single_word": false,
|
| 388 |
+
"special": true
|
| 389 |
+
},
|
| 390 |
+
"48": {
|
| 391 |
+
"content": "<|reserved_24|>",
|
| 392 |
+
"lstrip": false,
|
| 393 |
+
"normalized": false,
|
| 394 |
+
"rstrip": false,
|
| 395 |
+
"single_word": false,
|
| 396 |
+
"special": true
|
| 397 |
+
},
|
| 398 |
+
"49": {
|
| 399 |
+
"content": "<|reserved_25|>",
|
| 400 |
+
"lstrip": false,
|
| 401 |
+
"normalized": false,
|
| 402 |
+
"rstrip": false,
|
| 403 |
+
"single_word": false,
|
| 404 |
+
"special": true
|
| 405 |
+
},
|
| 406 |
+
"50": {
|
| 407 |
+
"content": "<|reserved_26|>",
|
| 408 |
+
"lstrip": false,
|
| 409 |
+
"normalized": false,
|
| 410 |
+
"rstrip": false,
|
| 411 |
+
"single_word": false,
|
| 412 |
+
"special": true
|
| 413 |
+
},
|
| 414 |
+
"51": {
|
| 415 |
+
"content": "<|reserved_27|>",
|
| 416 |
+
"lstrip": false,
|
| 417 |
+
"normalized": false,
|
| 418 |
+
"rstrip": false,
|
| 419 |
+
"single_word": false,
|
| 420 |
+
"special": true
|
| 421 |
+
},
|
| 422 |
+
"52": {
|
| 423 |
+
"content": "<|reserved_28|>",
|
| 424 |
+
"lstrip": false,
|
| 425 |
+
"normalized": false,
|
| 426 |
+
"rstrip": false,
|
| 427 |
+
"single_word": false,
|
| 428 |
+
"special": true
|
| 429 |
+
},
|
| 430 |
+
"53": {
|
| 431 |
+
"content": "<|reserved_29|>",
|
| 432 |
+
"lstrip": false,
|
| 433 |
+
"normalized": false,
|
| 434 |
+
"rstrip": false,
|
| 435 |
+
"single_word": false,
|
| 436 |
+
"special": true
|
| 437 |
+
},
|
| 438 |
+
"54": {
|
| 439 |
+
"content": "<|reserved_30|>",
|
| 440 |
+
"lstrip": false,
|
| 441 |
+
"normalized": false,
|
| 442 |
+
"rstrip": false,
|
| 443 |
+
"single_word": false,
|
| 444 |
+
"special": true
|
| 445 |
+
},
|
| 446 |
+
"55": {
|
| 447 |
+
"content": "<|reserved_31|>",
|
| 448 |
+
"lstrip": false,
|
| 449 |
+
"normalized": false,
|
| 450 |
+
"rstrip": false,
|
| 451 |
+
"single_word": false,
|
| 452 |
+
"special": true
|
| 453 |
+
},
|
| 454 |
+
"56": {
|
| 455 |
+
"content": "<|reserved_32|>",
|
| 456 |
+
"lstrip": false,
|
| 457 |
+
"normalized": false,
|
| 458 |
+
"rstrip": false,
|
| 459 |
+
"single_word": false,
|
| 460 |
+
"special": true
|
| 461 |
+
},
|
| 462 |
+
"57": {
|
| 463 |
+
"content": "<|reserved_33|>",
|
| 464 |
+
"lstrip": false,
|
| 465 |
+
"normalized": false,
|
| 466 |
+
"rstrip": false,
|
| 467 |
+
"single_word": false,
|
| 468 |
+
"special": true
|
| 469 |
+
},
|
| 470 |
+
"58": {
|
| 471 |
+
"content": "<|reserved_34|>",
|
| 472 |
+
"lstrip": false,
|
| 473 |
+
"normalized": false,
|
| 474 |
+
"rstrip": false,
|
| 475 |
+
"single_word": false,
|
| 476 |
+
"special": true
|
| 477 |
+
},
|
| 478 |
+
"59": {
|
| 479 |
+
"content": "<|reserved_35|>",
|
| 480 |
+
"lstrip": false,
|
| 481 |
+
"normalized": false,
|
| 482 |
+
"rstrip": false,
|
| 483 |
+
"single_word": false,
|
| 484 |
+
"special": true
|
| 485 |
+
},
|
| 486 |
+
"60": {
|
| 487 |
+
"content": "<|reserved_36|>",
|
| 488 |
+
"lstrip": false,
|
| 489 |
+
"normalized": false,
|
| 490 |
+
"rstrip": false,
|
| 491 |
+
"single_word": false,
|
| 492 |
+
"special": true
|
| 493 |
+
},
|
| 494 |
+
"61": {
|
| 495 |
+
"content": "<|reserved_37|>",
|
| 496 |
+
"lstrip": false,
|
| 497 |
+
"normalized": false,
|
| 498 |
+
"rstrip": false,
|
| 499 |
+
"single_word": false,
|
| 500 |
+
"special": true
|
| 501 |
+
},
|
| 502 |
+
"62": {
|
| 503 |
+
"content": "<|reserved_38|>",
|
| 504 |
+
"lstrip": false,
|
| 505 |
+
"normalized": false,
|
| 506 |
+
"rstrip": false,
|
| 507 |
+
"single_word": false,
|
| 508 |
+
"special": true
|
| 509 |
+
},
|
| 510 |
+
"63": {
|
| 511 |
+
"content": "<|reserved_39|>",
|
| 512 |
+
"lstrip": false,
|
| 513 |
+
"normalized": false,
|
| 514 |
+
"rstrip": false,
|
| 515 |
+
"single_word": false,
|
| 516 |
+
"special": true
|
| 517 |
+
},
|
| 518 |
+
"64": {
|
| 519 |
+
"content": "<|reserved_40|>",
|
| 520 |
+
"lstrip": false,
|
| 521 |
+
"normalized": false,
|
| 522 |
+
"rstrip": false,
|
| 523 |
+
"single_word": false,
|
| 524 |
+
"special": true
|
| 525 |
+
},
|
| 526 |
+
"65": {
|
| 527 |
+
"content": "<|reserved_41|>",
|
| 528 |
+
"lstrip": false,
|
| 529 |
+
"normalized": false,
|
| 530 |
+
"rstrip": false,
|
| 531 |
+
"single_word": false,
|
| 532 |
+
"special": true
|
| 533 |
+
},
|
| 534 |
+
"66": {
|
| 535 |
+
"content": "<|reserved_42|>",
|
| 536 |
+
"lstrip": false,
|
| 537 |
+
"normalized": false,
|
| 538 |
+
"rstrip": false,
|
| 539 |
+
"single_word": false,
|
| 540 |
+
"special": true
|
| 541 |
+
},
|
| 542 |
+
"67": {
|
| 543 |
+
"content": "<|reserved_43|>",
|
| 544 |
+
"lstrip": false,
|
| 545 |
+
"normalized": false,
|
| 546 |
+
"rstrip": false,
|
| 547 |
+
"single_word": false,
|
| 548 |
+
"special": true
|
| 549 |
+
},
|
| 550 |
+
"68": {
|
| 551 |
+
"content": "<|reserved_44|>",
|
| 552 |
+
"lstrip": false,
|
| 553 |
+
"normalized": false,
|
| 554 |
+
"rstrip": false,
|
| 555 |
+
"single_word": false,
|
| 556 |
+
"special": true
|
| 557 |
+
},
|
| 558 |
+
"69": {
|
| 559 |
+
"content": "<|reserved_45|>",
|
| 560 |
+
"lstrip": false,
|
| 561 |
+
"normalized": false,
|
| 562 |
+
"rstrip": false,
|
| 563 |
+
"single_word": false,
|
| 564 |
+
"special": true
|
| 565 |
+
},
|
| 566 |
+
"70": {
|
| 567 |
+
"content": "<|reserved_46|>",
|
| 568 |
+
"lstrip": false,
|
| 569 |
+
"normalized": false,
|
| 570 |
+
"rstrip": false,
|
| 571 |
+
"single_word": false,
|
| 572 |
+
"special": true
|
| 573 |
+
},
|
| 574 |
+
"71": {
|
| 575 |
+
"content": "<|reserved_47|>",
|
| 576 |
+
"lstrip": false,
|
| 577 |
+
"normalized": false,
|
| 578 |
+
"rstrip": false,
|
| 579 |
+
"single_word": false,
|
| 580 |
+
"special": true
|
| 581 |
+
},
|
| 582 |
+
"72": {
|
| 583 |
+
"content": "<|reserved_48|>",
|
| 584 |
+
"lstrip": false,
|
| 585 |
+
"normalized": false,
|
| 586 |
+
"rstrip": false,
|
| 587 |
+
"single_word": false,
|
| 588 |
+
"special": true
|
| 589 |
+
},
|
| 590 |
+
"73": {
|
| 591 |
+
"content": "<|reserved_49|>",
|
| 592 |
+
"lstrip": false,
|
| 593 |
+
"normalized": false,
|
| 594 |
+
"rstrip": false,
|
| 595 |
+
"single_word": false,
|
| 596 |
+
"special": true
|
| 597 |
+
},
|
| 598 |
+
"74": {
|
| 599 |
+
"content": "<|reserved_50|>",
|
| 600 |
+
"lstrip": false,
|
| 601 |
+
"normalized": false,
|
| 602 |
+
"rstrip": false,
|
| 603 |
+
"single_word": false,
|
| 604 |
+
"special": true
|
| 605 |
+
},
|
| 606 |
+
"75": {
|
| 607 |
+
"content": "<|reserved_51|>",
|
| 608 |
+
"lstrip": false,
|
| 609 |
+
"normalized": false,
|
| 610 |
+
"rstrip": false,
|
| 611 |
+
"single_word": false,
|
| 612 |
+
"special": true
|
| 613 |
+
},
|
| 614 |
+
"76": {
|
| 615 |
+
"content": "<|reserved_52|>",
|
| 616 |
+
"lstrip": false,
|
| 617 |
+
"normalized": false,
|
| 618 |
+
"rstrip": false,
|
| 619 |
+
"single_word": false,
|
| 620 |
+
"special": true
|
| 621 |
+
},
|
| 622 |
+
"77": {
|
| 623 |
+
"content": "<|reserved_53|>",
|
| 624 |
+
"lstrip": false,
|
| 625 |
+
"normalized": false,
|
| 626 |
+
"rstrip": false,
|
| 627 |
+
"single_word": false,
|
| 628 |
+
"special": true
|
| 629 |
+
},
|
| 630 |
+
"78": {
|
| 631 |
+
"content": "<|reserved_54|>",
|
| 632 |
+
"lstrip": false,
|
| 633 |
+
"normalized": false,
|
| 634 |
+
"rstrip": false,
|
| 635 |
+
"single_word": false,
|
| 636 |
+
"special": true
|
| 637 |
+
},
|
| 638 |
+
"79": {
|
| 639 |
+
"content": "<|reserved_55|>",
|
| 640 |
+
"lstrip": false,
|
| 641 |
+
"normalized": false,
|
| 642 |
+
"rstrip": false,
|
| 643 |
+
"single_word": false,
|
| 644 |
+
"special": true
|
| 645 |
+
},
|
| 646 |
+
"80": {
|
| 647 |
+
"content": "<|reserved_56|>",
|
| 648 |
+
"lstrip": false,
|
| 649 |
+
"normalized": false,
|
| 650 |
+
"rstrip": false,
|
| 651 |
+
"single_word": false,
|
| 652 |
+
"special": true
|
| 653 |
+
},
|
| 654 |
+
"81": {
|
| 655 |
+
"content": "<|reserved_57|>",
|
| 656 |
+
"lstrip": false,
|
| 657 |
+
"normalized": false,
|
| 658 |
+
"rstrip": false,
|
| 659 |
+
"single_word": false,
|
| 660 |
+
"special": true
|
| 661 |
+
},
|
| 662 |
+
"82": {
|
| 663 |
+
"content": "<|reserved_58|>",
|
| 664 |
+
"lstrip": false,
|
| 665 |
+
"normalized": false,
|
| 666 |
+
"rstrip": false,
|
| 667 |
+
"single_word": false,
|
| 668 |
+
"special": true
|
| 669 |
+
},
|
| 670 |
+
"83": {
|
| 671 |
+
"content": "<|reserved_59|>",
|
| 672 |
+
"lstrip": false,
|
| 673 |
+
"normalized": false,
|
| 674 |
+
"rstrip": false,
|
| 675 |
+
"single_word": false,
|
| 676 |
+
"special": true
|
| 677 |
+
},
|
| 678 |
+
"84": {
|
| 679 |
+
"content": "<|reserved_60|>",
|
| 680 |
+
"lstrip": false,
|
| 681 |
+
"normalized": false,
|
| 682 |
+
"rstrip": false,
|
| 683 |
+
"single_word": false,
|
| 684 |
+
"special": true
|
| 685 |
+
},
|
| 686 |
+
"85": {
|
| 687 |
+
"content": "<|reserved_61|>",
|
| 688 |
+
"lstrip": false,
|
| 689 |
+
"normalized": false,
|
| 690 |
+
"rstrip": false,
|
| 691 |
+
"single_word": false,
|
| 692 |
+
"special": true
|
| 693 |
+
},
|
| 694 |
+
"86": {
|
| 695 |
+
"content": "<|reserved_62|>",
|
| 696 |
+
"lstrip": false,
|
| 697 |
+
"normalized": false,
|
| 698 |
+
"rstrip": false,
|
| 699 |
+
"single_word": false,
|
| 700 |
+
"special": true
|
| 701 |
+
},
|
| 702 |
+
"87": {
|
| 703 |
+
"content": "<|reserved_63|>",
|
| 704 |
+
"lstrip": false,
|
| 705 |
+
"normalized": false,
|
| 706 |
+
"rstrip": false,
|
| 707 |
+
"single_word": false,
|
| 708 |
+
"special": true
|
| 709 |
+
},
|
| 710 |
+
"88": {
|
| 711 |
+
"content": "<|reserved_64|>",
|
| 712 |
+
"lstrip": false,
|
| 713 |
+
"normalized": false,
|
| 714 |
+
"rstrip": false,
|
| 715 |
+
"single_word": false,
|
| 716 |
+
"special": true
|
| 717 |
+
},
|
| 718 |
+
"89": {
|
| 719 |
+
"content": "<|reserved_65|>",
|
| 720 |
+
"lstrip": false,
|
| 721 |
+
"normalized": false,
|
| 722 |
+
"rstrip": false,
|
| 723 |
+
"single_word": false,
|
| 724 |
+
"special": true
|
| 725 |
+
},
|
| 726 |
+
"90": {
|
| 727 |
+
"content": "<|reserved_66|>",
|
| 728 |
+
"lstrip": false,
|
| 729 |
+
"normalized": false,
|
| 730 |
+
"rstrip": false,
|
| 731 |
+
"single_word": false,
|
| 732 |
+
"special": true
|
| 733 |
+
},
|
| 734 |
+
"91": {
|
| 735 |
+
"content": "<|reserved_67|>",
|
| 736 |
+
"lstrip": false,
|
| 737 |
+
"normalized": false,
|
| 738 |
+
"rstrip": false,
|
| 739 |
+
"single_word": false,
|
| 740 |
+
"special": true
|
| 741 |
+
},
|
| 742 |
+
"92": {
|
| 743 |
+
"content": "<|reserved_68|>",
|
| 744 |
+
"lstrip": false,
|
| 745 |
+
"normalized": false,
|
| 746 |
+
"rstrip": false,
|
| 747 |
+
"single_word": false,
|
| 748 |
+
"special": true
|
| 749 |
+
},
|
| 750 |
+
"93": {
|
| 751 |
+
"content": "<|reserved_69|>",
|
| 752 |
+
"lstrip": false,
|
| 753 |
+
"normalized": false,
|
| 754 |
+
"rstrip": false,
|
| 755 |
+
"single_word": false,
|
| 756 |
+
"special": true
|
| 757 |
+
},
|
| 758 |
+
"94": {
|
| 759 |
+
"content": "<|reserved_70|>",
|
| 760 |
+
"lstrip": false,
|
| 761 |
+
"normalized": false,
|
| 762 |
+
"rstrip": false,
|
| 763 |
+
"single_word": false,
|
| 764 |
+
"special": true
|
| 765 |
+
},
|
| 766 |
+
"95": {
|
| 767 |
+
"content": "<|reserved_71|>",
|
| 768 |
+
"lstrip": false,
|
| 769 |
+
"normalized": false,
|
| 770 |
+
"rstrip": false,
|
| 771 |
+
"single_word": false,
|
| 772 |
+
"special": true
|
| 773 |
+
},
|
| 774 |
+
"96": {
|
| 775 |
+
"content": "<|reserved_72|>",
|
| 776 |
+
"lstrip": false,
|
| 777 |
+
"normalized": false,
|
| 778 |
+
"rstrip": false,
|
| 779 |
+
"single_word": false,
|
| 780 |
+
"special": true
|
| 781 |
+
},
|
| 782 |
+
"97": {
|
| 783 |
+
"content": "<|reserved_73|>",
|
| 784 |
+
"lstrip": false,
|
| 785 |
+
"normalized": false,
|
| 786 |
+
"rstrip": false,
|
| 787 |
+
"single_word": false,
|
| 788 |
+
"special": true
|
| 789 |
+
},
|
| 790 |
+
"98": {
|
| 791 |
+
"content": "<|reserved_74|>",
|
| 792 |
+
"lstrip": false,
|
| 793 |
+
"normalized": false,
|
| 794 |
+
"rstrip": false,
|
| 795 |
+
"single_word": false,
|
| 796 |
+
"special": true
|
| 797 |
+
},
|
| 798 |
+
"99": {
|
| 799 |
+
"content": "<|reserved_75|>",
|
| 800 |
+
"lstrip": false,
|
| 801 |
+
"normalized": false,
|
| 802 |
+
"rstrip": false,
|
| 803 |
+
"single_word": false,
|
| 804 |
+
"special": true
|
| 805 |
+
},
|
| 806 |
+
"100": {
|
| 807 |
+
"content": "<|reserved_76|>",
|
| 808 |
+
"lstrip": false,
|
| 809 |
+
"normalized": false,
|
| 810 |
+
"rstrip": false,
|
| 811 |
+
"single_word": false,
|
| 812 |
+
"special": true
|
| 813 |
+
},
|
| 814 |
+
"101": {
|
| 815 |
+
"content": "<|reserved_77|>",
|
| 816 |
+
"lstrip": false,
|
| 817 |
+
"normalized": false,
|
| 818 |
+
"rstrip": false,
|
| 819 |
+
"single_word": false,
|
| 820 |
+
"special": true
|
| 821 |
+
},
|
| 822 |
+
"102": {
|
| 823 |
+
"content": "<|reserved_78|>",
|
| 824 |
+
"lstrip": false,
|
| 825 |
+
"normalized": false,
|
| 826 |
+
"rstrip": false,
|
| 827 |
+
"single_word": false,
|
| 828 |
+
"special": true
|
| 829 |
+
},
|
| 830 |
+
"103": {
|
| 831 |
+
"content": "<|reserved_79|>",
|
| 832 |
+
"lstrip": false,
|
| 833 |
+
"normalized": false,
|
| 834 |
+
"rstrip": false,
|
| 835 |
+
"single_word": false,
|
| 836 |
+
"special": true
|
| 837 |
+
},
|
| 838 |
+
"104": {
|
| 839 |
+
"content": "<|reserved_80|>",
|
| 840 |
+
"lstrip": false,
|
| 841 |
+
"normalized": false,
|
| 842 |
+
"rstrip": false,
|
| 843 |
+
"single_word": false,
|
| 844 |
+
"special": true
|
| 845 |
+
},
|
| 846 |
+
"105": {
|
| 847 |
+
"content": "<|reserved_81|>",
|
| 848 |
+
"lstrip": false,
|
| 849 |
+
"normalized": false,
|
| 850 |
+
"rstrip": false,
|
| 851 |
+
"single_word": false,
|
| 852 |
+
"special": true
|
| 853 |
+
},
|
| 854 |
+
"106": {
|
| 855 |
+
"content": "<|reserved_82|>",
|
| 856 |
+
"lstrip": false,
|
| 857 |
+
"normalized": false,
|
| 858 |
+
"rstrip": false,
|
| 859 |
+
"single_word": false,
|
| 860 |
+
"special": true
|
| 861 |
+
},
|
| 862 |
+
"107": {
|
| 863 |
+
"content": "<|reserved_83|>",
|
| 864 |
+
"lstrip": false,
|
| 865 |
+
"normalized": false,
|
| 866 |
+
"rstrip": false,
|
| 867 |
+
"single_word": false,
|
| 868 |
+
"special": true
|
| 869 |
+
},
|
| 870 |
+
"108": {
|
| 871 |
+
"content": "<|reserved_84|>",
|
| 872 |
+
"lstrip": false,
|
| 873 |
+
"normalized": false,
|
| 874 |
+
"rstrip": false,
|
| 875 |
+
"single_word": false,
|
| 876 |
+
"special": true
|
| 877 |
+
},
|
| 878 |
+
"109": {
|
| 879 |
+
"content": "<|reserved_85|>",
|
| 880 |
+
"lstrip": false,
|
| 881 |
+
"normalized": false,
|
| 882 |
+
"rstrip": false,
|
| 883 |
+
"single_word": false,
|
| 884 |
+
"special": true
|
| 885 |
+
},
|
| 886 |
+
"110": {
|
| 887 |
+
"content": "<|reserved_86|>",
|
| 888 |
+
"lstrip": false,
|
| 889 |
+
"normalized": false,
|
| 890 |
+
"rstrip": false,
|
| 891 |
+
"single_word": false,
|
| 892 |
+
"special": true
|
| 893 |
+
},
|
| 894 |
+
"111": {
|
| 895 |
+
"content": "<|reserved_87|>",
|
| 896 |
+
"lstrip": false,
|
| 897 |
+
"normalized": false,
|
| 898 |
+
"rstrip": false,
|
| 899 |
+
"single_word": false,
|
| 900 |
+
"special": true
|
| 901 |
+
},
|
| 902 |
+
"112": {
|
| 903 |
+
"content": "<|reserved_88|>",
|
| 904 |
+
"lstrip": false,
|
| 905 |
+
"normalized": false,
|
| 906 |
+
"rstrip": false,
|
| 907 |
+
"single_word": false,
|
| 908 |
+
"special": true
|
| 909 |
+
},
|
| 910 |
+
"113": {
|
| 911 |
+
"content": "<|reserved_89|>",
|
| 912 |
+
"lstrip": false,
|
| 913 |
+
"normalized": false,
|
| 914 |
+
"rstrip": false,
|
| 915 |
+
"single_word": false,
|
| 916 |
+
"special": true
|
| 917 |
+
},
|
| 918 |
+
"114": {
|
| 919 |
+
"content": "<|reserved_90|>",
|
| 920 |
+
"lstrip": false,
|
| 921 |
+
"normalized": false,
|
| 922 |
+
"rstrip": false,
|
| 923 |
+
"single_word": false,
|
| 924 |
+
"special": true
|
| 925 |
+
},
|
| 926 |
+
"115": {
|
| 927 |
+
"content": "<|reserved_91|>",
|
| 928 |
+
"lstrip": false,
|
| 929 |
+
"normalized": false,
|
| 930 |
+
"rstrip": false,
|
| 931 |
+
"single_word": false,
|
| 932 |
+
"special": true
|
| 933 |
+
},
|
| 934 |
+
"116": {
|
| 935 |
+
"content": "<|reserved_92|>",
|
| 936 |
+
"lstrip": false,
|
| 937 |
+
"normalized": false,
|
| 938 |
+
"rstrip": false,
|
| 939 |
+
"single_word": false,
|
| 940 |
+
"special": true
|
| 941 |
+
},
|
| 942 |
+
"117": {
|
| 943 |
+
"content": "<|reserved_93|>",
|
| 944 |
+
"lstrip": false,
|
| 945 |
+
"normalized": false,
|
| 946 |
+
"rstrip": false,
|
| 947 |
+
"single_word": false,
|
| 948 |
+
"special": true
|
| 949 |
+
},
|
| 950 |
+
"118": {
|
| 951 |
+
"content": "<|reserved_94|>",
|
| 952 |
+
"lstrip": false,
|
| 953 |
+
"normalized": false,
|
| 954 |
+
"rstrip": false,
|
| 955 |
+
"single_word": false,
|
| 956 |
+
"special": true
|
| 957 |
+
},
|
| 958 |
+
"119": {
|
| 959 |
+
"content": "<|reserved_95|>",
|
| 960 |
+
"lstrip": false,
|
| 961 |
+
"normalized": false,
|
| 962 |
+
"rstrip": false,
|
| 963 |
+
"single_word": false,
|
| 964 |
+
"special": true
|
| 965 |
+
},
|
| 966 |
+
"120": {
|
| 967 |
+
"content": "<|reserved_96|>",
|
| 968 |
+
"lstrip": false,
|
| 969 |
+
"normalized": false,
|
| 970 |
+
"rstrip": false,
|
| 971 |
+
"single_word": false,
|
| 972 |
+
"special": true
|
| 973 |
+
},
|
| 974 |
+
"121": {
|
| 975 |
+
"content": "<|reserved_97|>",
|
| 976 |
+
"lstrip": false,
|
| 977 |
+
"normalized": false,
|
| 978 |
+
"rstrip": false,
|
| 979 |
+
"single_word": false,
|
| 980 |
+
"special": true
|
| 981 |
+
},
|
| 982 |
+
"122": {
|
| 983 |
+
"content": "<|reserved_98|>",
|
| 984 |
+
"lstrip": false,
|
| 985 |
+
"normalized": false,
|
| 986 |
+
"rstrip": false,
|
| 987 |
+
"single_word": false,
|
| 988 |
+
"special": true
|
| 989 |
+
},
|
| 990 |
+
"123": {
|
| 991 |
+
"content": "<|reserved_99|>",
|
| 992 |
+
"lstrip": false,
|
| 993 |
+
"normalized": false,
|
| 994 |
+
"rstrip": false,
|
| 995 |
+
"single_word": false,
|
| 996 |
+
"special": true
|
| 997 |
+
},
|
| 998 |
+
"124": {
|
| 999 |
+
"content": "<|reserved_100|>",
|
| 1000 |
+
"lstrip": false,
|
| 1001 |
+
"normalized": false,
|
| 1002 |
+
"rstrip": false,
|
| 1003 |
+
"single_word": false,
|
| 1004 |
+
"special": true
|
| 1005 |
+
},
|
| 1006 |
+
"125": {
|
| 1007 |
+
"content": "<|reserved_101|>",
|
| 1008 |
+
"lstrip": false,
|
| 1009 |
+
"normalized": false,
|
| 1010 |
+
"rstrip": false,
|
| 1011 |
+
"single_word": false,
|
| 1012 |
+
"special": true
|
| 1013 |
+
},
|
| 1014 |
+
"126": {
|
| 1015 |
+
"content": "<|reserved_102|>",
|
| 1016 |
+
"lstrip": false,
|
| 1017 |
+
"normalized": false,
|
| 1018 |
+
"rstrip": false,
|
| 1019 |
+
"single_word": false,
|
| 1020 |
+
"special": true
|
| 1021 |
+
},
|
| 1022 |
+
"127": {
|
| 1023 |
+
"content": "<|reserved_103|>",
|
| 1024 |
+
"lstrip": false,
|
| 1025 |
+
"normalized": false,
|
| 1026 |
+
"rstrip": false,
|
| 1027 |
+
"single_word": false,
|
| 1028 |
+
"special": true
|
| 1029 |
+
}
|
| 1030 |
+
},
|
| 1031 |
+
"additional_special_tokens": [
|
| 1032 |
+
"<|im_start|>",
|
| 1033 |
+
"<|im_end|>",
|
| 1034 |
+
"<|system|>",
|
| 1035 |
+
"<|user|>",
|
| 1036 |
+
"<|assistant|>",
|
| 1037 |
+
"<|tool|>",
|
| 1038 |
+
"<think>",
|
| 1039 |
+
"</think>",
|
| 1040 |
+
"<tool_call>",
|
| 1041 |
+
"</tool_call>",
|
| 1042 |
+
"<tool_response>",
|
| 1043 |
+
"</tool_response>",
|
| 1044 |
+
"<image_start>",
|
| 1045 |
+
"<image_pad>",
|
| 1046 |
+
"<image_end>",
|
| 1047 |
+
"<audio_start>",
|
| 1048 |
+
"<audio_pad>",
|
| 1049 |
+
"<audio_end>",
|
| 1050 |
+
"<speech_bos>",
|
| 1051 |
+
"<speech_eos>",
|
| 1052 |
+
"<speech_pad>"
|
| 1053 |
+
],
|
| 1054 |
+
"bos_token": "<|bos|>",
|
| 1055 |
+
"clean_up_tokenization_spaces": false,
|
| 1056 |
+
"eos_token": "<|eos|>",
|
| 1057 |
+
"extra_special_tokens": {},
|
| 1058 |
+
"legacy": true,
|
| 1059 |
+
"model_max_length": 4096,
|
| 1060 |
+
"pad_token": "<|pad|>",
|
| 1061 |
+
"spaces_between_special_tokens": false,
|
| 1062 |
+
"tokenizer_class": "PreTrainedTokenizerFast",
|
| 1063 |
+
"unk_token": "<|unk|>"
|
| 1064 |
+
}
|