Add files using upload-large-folder tool
Browse files- README.md +40 -0
- chat_template.jinja +132 -0
- config.json +184 -0
- configuration_laguna.py +213 -0
- generation_config.json +12 -0
- model-00001-of-00014.safetensors +3 -0
- model-00002-of-00014.safetensors +3 -0
- model-00003-of-00014.safetensors +3 -0
- model-00004-of-00014.safetensors +3 -0
- model-00005-of-00014.safetensors +3 -0
- model-00006-of-00014.safetensors +3 -0
- model-00007-of-00014.safetensors +3 -0
- model-00008-of-00014.safetensors +3 -0
- model-00009-of-00014.safetensors +3 -0
- model-00010-of-00014.safetensors +3 -0
- model-00011-of-00014.safetensors +3 -0
- model-00012-of-00014.safetensors +3 -0
- model-00013-of-00014.safetensors +3 -0
- model-00014-of-00014.safetensors +3 -0
- model.safetensors.index.json +686 -0
- modeling_laguna.py +755 -0
- tokenizer.json +0 -0
- tokenizer_config.json +15 -0
README.md
ADDED
|
@@ -0,0 +1,40 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
library_name: mlx
|
| 3 |
+
inference: false
|
| 4 |
+
extra_gated_description: To learn more about how we process your personal data, please
|
| 5 |
+
read our <a href="https://poolside.ai/legal/privacy">Privacy Policy</a>.
|
| 6 |
+
tags:
|
| 7 |
+
- laguna-xs.2
|
| 8 |
+
- mlx
|
| 9 |
+
license: apache-2.0
|
| 10 |
+
pipeline_tag: text-generation
|
| 11 |
+
base_model: poolside/Laguna-XS.2
|
| 12 |
+
---
|
| 13 |
+
|
| 14 |
+
# mlx-community/Laguna-XS.2-bf16
|
| 15 |
+
|
| 16 |
+
This model [mlx-community/Laguna-XS.2-bf16](https://huggingface.co/mlx-community/Laguna-XS.2-bf16) was
|
| 17 |
+
converted to MLX format from [poolside/Laguna-XS.2](https://huggingface.co/poolside/Laguna-XS.2)
|
| 18 |
+
using mlx-lm version **0.31.3**.
|
| 19 |
+
|
| 20 |
+
## Use with mlx
|
| 21 |
+
|
| 22 |
+
```bash
|
| 23 |
+
pip install mlx-lm
|
| 24 |
+
```
|
| 25 |
+
|
| 26 |
+
```python
|
| 27 |
+
from mlx_lm import load, generate
|
| 28 |
+
|
| 29 |
+
model, tokenizer = load("mlx-community/Laguna-XS.2-bf16")
|
| 30 |
+
|
| 31 |
+
prompt = "hello"
|
| 32 |
+
|
| 33 |
+
if tokenizer.chat_template is not None:
|
| 34 |
+
messages = [{"role": "user", "content": prompt}]
|
| 35 |
+
prompt = tokenizer.apply_chat_template(
|
| 36 |
+
messages, add_generation_prompt=True, return_dict=False,
|
| 37 |
+
)
|
| 38 |
+
|
| 39 |
+
response = generate(model, tokenizer, prompt=prompt, verbose=True)
|
| 40 |
+
```
|
chat_template.jinja
ADDED
|
@@ -0,0 +1,132 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{#- Iteration on laguna_glm_thinking_v5/chat_template.jinja -#}
|
| 2 |
+
{#- Adds a default system message (used when no system message is provided in `messages`). -#}
|
| 3 |
+
{{- "〈|EOS|〉" -}}
|
| 4 |
+
{%- set enable_thinking = enable_thinking | default(false) -%}
|
| 5 |
+
{%- set render_assistant_messages_raw = render_assistant_messages_raw | default(false) -%}
|
| 6 |
+
{%- set add_generation_prompt = add_generation_prompt | default(false) -%}
|
| 7 |
+
|
| 8 |
+
{#- ───── header (system message) ───── -#}
|
| 9 |
+
{%- set system_message = "You are a helpful, conversationally-fluent assistant made by Poolside. You are here to be helpful to users through natural language conversations." -%}
|
| 10 |
+
{%- if messages and messages[0].role == "system" -%}
|
| 11 |
+
{%- set system_message = messages[0].content -%}
|
| 12 |
+
{%- endif -%}
|
| 13 |
+
|
| 14 |
+
{%- if (system_message and system_message.strip()) or tools -%}
|
| 15 |
+
{{- "<system>\n" -}}
|
| 16 |
+
|
| 17 |
+
{%- if system_message and system_message.strip() -%}
|
| 18 |
+
{{- "\n" -}}
|
| 19 |
+
{{- system_message.rstrip() -}}
|
| 20 |
+
{%- endif -%}
|
| 21 |
+
|
| 22 |
+
{%- if tools -%}
|
| 23 |
+
{{- "\n\n### Tools\n\n" -}}
|
| 24 |
+
{%- set ns = namespace(tool_string="You may call functions to assist with the user query.\n"
|
| 25 |
+
~ "All available function signatures are listed below:\n"
|
| 26 |
+
~ "<available_tools>\n") -%}
|
| 27 |
+
{%- for tool in tools -%}
|
| 28 |
+
{%- set ns.tool_string = ns.tool_string ~ (tool | tojson) ~ "\n" -%}
|
| 29 |
+
{%- endfor -%}
|
| 30 |
+
{%- if enable_thinking -%}
|
| 31 |
+
{%- set tool_string = ns.tool_string + "</available_tools>\n\n" ~
|
| 32 |
+
"Wrap your thinking in '<think>', '</think>' tags, followed by a function call. For each function call, return an unescaped XML-like object with function name and arguments within '<tool_call>' and '</tool_call>' tags, like here:\n" ~
|
| 33 |
+
"<think> your thoughts here </think>\n" ~
|
| 34 |
+
"<tool_call>function-name\n<arg_key>argument-key</arg_key>\n<arg_value>value-of-argument-key</arg_value>\n" ~
|
| 35 |
+
"</tool_call>" -%}
|
| 36 |
+
{%- else -%}
|
| 37 |
+
{%- set tool_string = ns.tool_string + "</available_tools>\n\n" ~
|
| 38 |
+
"For each function call, return an unescaped XML-like object " ~
|
| 39 |
+
"with function name and arguments within '<tool_call>' and '</tool_call>' tags, like here:\n" ~
|
| 40 |
+
"<tool_call>function-name\n<arg_key>argument-key</arg_key>\n<arg_value>value-of-argument-key</arg_value>\n" ~
|
| 41 |
+
"</tool_call>" -%}
|
| 42 |
+
{%- endif -%}
|
| 43 |
+
{{- tool_string -}}
|
| 44 |
+
{%- endif -%}
|
| 45 |
+
|
| 46 |
+
{{- "\n</system>\n" -}}
|
| 47 |
+
{%- endif -%}
|
| 48 |
+
|
| 49 |
+
{#- ───── main loop ───── -#}
|
| 50 |
+
{%- for message in messages -%}
|
| 51 |
+
{%- set content = message.content if message.content is string else "" -%}
|
| 52 |
+
{%- if message.role == "user" -%}
|
| 53 |
+
{{- "<user>\n" + content + "\n</user>\n" -}}
|
| 54 |
+
{%- elif message.role == "assistant" -%}
|
| 55 |
+
{%- generation -%}
|
| 56 |
+
{{- "<assistant>\n" -}}
|
| 57 |
+
{%- if render_assistant_messages_raw -%}
|
| 58 |
+
{#- Raw mode: prepend the generation prompt token, then dump content verbatim. -#}
|
| 59 |
+
{#- The generation prompt is <think> when enable_thinking, </think> otherwise. -#}
|
| 60 |
+
{#- Only prepend if content doesn't already start with it. -#}
|
| 61 |
+
{%- if enable_thinking -%}
|
| 62 |
+
{%- if not content.startswith('<think>') -%}
|
| 63 |
+
{{- '<think>' -}}
|
| 64 |
+
{%- endif -%}
|
| 65 |
+
{%- else -%}
|
| 66 |
+
{%- if not content.startswith('</think>') -%}
|
| 67 |
+
{{- '</think>' -}}
|
| 68 |
+
{%- endif -%}
|
| 69 |
+
{%- endif -%}
|
| 70 |
+
{{- content -}}
|
| 71 |
+
{#- Append closing tag if content doesn't already end with it. -#}
|
| 72 |
+
{%- if not content.endswith('</assistant>\n') and not content.endswith('</assistant>') -%}
|
| 73 |
+
{{- '\n</assistant>' -}}
|
| 74 |
+
{%- endif -%}
|
| 75 |
+
{{- "\n" -}}
|
| 76 |
+
{%- else -%}
|
| 77 |
+
{#- Extract reasoning content from message.reasoning (vLLM field name) or message.reasoning_content, or from <think> tags -#}
|
| 78 |
+
{%- set reasoning_content = '' %}
|
| 79 |
+
{%- if message.reasoning is string %}
|
| 80 |
+
{%- set reasoning_content = message.reasoning %}
|
| 81 |
+
{%- elif message.reasoning_content is string %}
|
| 82 |
+
{%- set reasoning_content = message.reasoning_content %}
|
| 83 |
+
{%- endif %}
|
| 84 |
+
{#- Always strip <think> tags from content if present to avoid duplication -#}
|
| 85 |
+
{%- if '</think>' in content %}
|
| 86 |
+
{%- if not reasoning_content %}
|
| 87 |
+
{%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
|
| 88 |
+
{%- endif %}
|
| 89 |
+
{%- set content = content.split('</think>')[-1].lstrip('\n') %}
|
| 90 |
+
{%- endif %}
|
| 91 |
+
{#- Display reasoning content for all messages -#}
|
| 92 |
+
{%- if reasoning_content -%}
|
| 93 |
+
{{- '<think>\n' + reasoning_content.strip() + '\n</think>\n' -}}
|
| 94 |
+
{%- else -%}
|
| 95 |
+
{{- '</think>\n' -}}
|
| 96 |
+
{%- endif -%}
|
| 97 |
+
{#- Display main content -#}
|
| 98 |
+
{%- if content.strip() -%}
|
| 99 |
+
{{- content.strip() ~ "\n" -}}
|
| 100 |
+
{%- endif -%}
|
| 101 |
+
{%- if message.tool_calls -%}
|
| 102 |
+
{%- for tool_call in message.tool_calls -%}
|
| 103 |
+
{%- set function_data = tool_call.function -%}
|
| 104 |
+
{{- '<tool_call>' + function_data.name }}
|
| 105 |
+
{% set _args = function_data.arguments %}
|
| 106 |
+
{%- for k, v in _args.items() -%}
|
| 107 |
+
{{- "<arg_key>" ~ k ~ "</arg_key>\n" -}}
|
| 108 |
+
{{- "<arg_value>"}}{{ v | tojson(ensure_ascii=False) if v is not string else v }}{{ "</arg_value>\n" -}}
|
| 109 |
+
{%- endfor -%}
|
| 110 |
+
{{- "</tool_call>\n" -}}
|
| 111 |
+
{%- endfor -%}
|
| 112 |
+
{%- endif -%}
|
| 113 |
+
{{- "</assistant>\n" -}}
|
| 114 |
+
{%- endif -%}
|
| 115 |
+
{%- endgeneration -%}
|
| 116 |
+
{%- elif message.role == "tool" -%}
|
| 117 |
+
{{- "<tool_response>\n" + content + "\n</tool_response>\n" -}}
|
| 118 |
+
{%- elif message.role == "system" and loop.index0 != 0 -%}
|
| 119 |
+
{#- Render additional system messages (skip the first one which is handled separately in the header) -#}
|
| 120 |
+
{{- "<system>\n" + content + "\n</system>\n" -}}
|
| 121 |
+
{%- endif -%}
|
| 122 |
+
{%- endfor -%}
|
| 123 |
+
{#- ───── generation prompt ───── -#}
|
| 124 |
+
{%- if add_generation_prompt -%}
|
| 125 |
+
{{- "<assistant>\n" -}}
|
| 126 |
+
{#- ───── Include reasoning mode directive ───── -#}
|
| 127 |
+
{%- if not enable_thinking %}
|
| 128 |
+
{{- '</think>' -}}
|
| 129 |
+
{%- else %}
|
| 130 |
+
{{- '<think>' -}}
|
| 131 |
+
{%- endif %}
|
| 132 |
+
{%- endif -%}
|
config.json
ADDED
|
@@ -0,0 +1,184 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"architectures": [
|
| 3 |
+
"LagunaForCausalLM"
|
| 4 |
+
],
|
| 5 |
+
"attention_bias": false,
|
| 6 |
+
"attention_dropout": 0.0,
|
| 7 |
+
"auto_map": {
|
| 8 |
+
"AutoConfig": "configuration_laguna.LagunaConfig",
|
| 9 |
+
"AutoModelForCausalLM": "modeling_laguna.LagunaForCausalLM"
|
| 10 |
+
},
|
| 11 |
+
"bos_token_id": 2,
|
| 12 |
+
"eos_token_id": [
|
| 13 |
+
2,
|
| 14 |
+
24
|
| 15 |
+
],
|
| 16 |
+
"gating": true,
|
| 17 |
+
"head_dim": 128,
|
| 18 |
+
"hidden_size": 2048,
|
| 19 |
+
"intermediate_size": 8192,
|
| 20 |
+
"layer_types": [
|
| 21 |
+
"full_attention",
|
| 22 |
+
"sliding_attention",
|
| 23 |
+
"sliding_attention",
|
| 24 |
+
"sliding_attention",
|
| 25 |
+
"full_attention",
|
| 26 |
+
"sliding_attention",
|
| 27 |
+
"sliding_attention",
|
| 28 |
+
"sliding_attention",
|
| 29 |
+
"full_attention",
|
| 30 |
+
"sliding_attention",
|
| 31 |
+
"sliding_attention",
|
| 32 |
+
"sliding_attention",
|
| 33 |
+
"full_attention",
|
| 34 |
+
"sliding_attention",
|
| 35 |
+
"sliding_attention",
|
| 36 |
+
"sliding_attention",
|
| 37 |
+
"full_attention",
|
| 38 |
+
"sliding_attention",
|
| 39 |
+
"sliding_attention",
|
| 40 |
+
"sliding_attention",
|
| 41 |
+
"full_attention",
|
| 42 |
+
"sliding_attention",
|
| 43 |
+
"sliding_attention",
|
| 44 |
+
"sliding_attention",
|
| 45 |
+
"full_attention",
|
| 46 |
+
"sliding_attention",
|
| 47 |
+
"sliding_attention",
|
| 48 |
+
"sliding_attention",
|
| 49 |
+
"full_attention",
|
| 50 |
+
"sliding_attention",
|
| 51 |
+
"sliding_attention",
|
| 52 |
+
"sliding_attention",
|
| 53 |
+
"full_attention",
|
| 54 |
+
"sliding_attention",
|
| 55 |
+
"sliding_attention",
|
| 56 |
+
"sliding_attention",
|
| 57 |
+
"full_attention",
|
| 58 |
+
"sliding_attention",
|
| 59 |
+
"sliding_attention",
|
| 60 |
+
"sliding_attention"
|
| 61 |
+
],
|
| 62 |
+
"max_position_embeddings": 131072,
|
| 63 |
+
"mlp_layer_types": [
|
| 64 |
+
"dense",
|
| 65 |
+
"sparse",
|
| 66 |
+
"sparse",
|
| 67 |
+
"sparse",
|
| 68 |
+
"sparse",
|
| 69 |
+
"sparse",
|
| 70 |
+
"sparse",
|
| 71 |
+
"sparse",
|
| 72 |
+
"sparse",
|
| 73 |
+
"sparse",
|
| 74 |
+
"sparse",
|
| 75 |
+
"sparse",
|
| 76 |
+
"sparse",
|
| 77 |
+
"sparse",
|
| 78 |
+
"sparse",
|
| 79 |
+
"sparse",
|
| 80 |
+
"sparse",
|
| 81 |
+
"sparse",
|
| 82 |
+
"sparse",
|
| 83 |
+
"sparse",
|
| 84 |
+
"sparse",
|
| 85 |
+
"sparse",
|
| 86 |
+
"sparse",
|
| 87 |
+
"sparse",
|
| 88 |
+
"sparse",
|
| 89 |
+
"sparse",
|
| 90 |
+
"sparse",
|
| 91 |
+
"sparse",
|
| 92 |
+
"sparse",
|
| 93 |
+
"sparse",
|
| 94 |
+
"sparse",
|
| 95 |
+
"sparse",
|
| 96 |
+
"sparse",
|
| 97 |
+
"sparse",
|
| 98 |
+
"sparse",
|
| 99 |
+
"sparse",
|
| 100 |
+
"sparse",
|
| 101 |
+
"sparse",
|
| 102 |
+
"sparse",
|
| 103 |
+
"sparse"
|
| 104 |
+
],
|
| 105 |
+
"model_type": "laguna",
|
| 106 |
+
"moe_apply_router_weight_on_input": false,
|
| 107 |
+
"moe_intermediate_size": 512,
|
| 108 |
+
"moe_routed_scaling_factor": 2.5,
|
| 109 |
+
"num_attention_heads": 48,
|
| 110 |
+
"num_attention_heads_per_layer": [
|
| 111 |
+
48,
|
| 112 |
+
64,
|
| 113 |
+
64,
|
| 114 |
+
64,
|
| 115 |
+
48,
|
| 116 |
+
64,
|
| 117 |
+
64,
|
| 118 |
+
64,
|
| 119 |
+
48,
|
| 120 |
+
64,
|
| 121 |
+
64,
|
| 122 |
+
64,
|
| 123 |
+
48,
|
| 124 |
+
64,
|
| 125 |
+
64,
|
| 126 |
+
64,
|
| 127 |
+
48,
|
| 128 |
+
64,
|
| 129 |
+
64,
|
| 130 |
+
64,
|
| 131 |
+
48,
|
| 132 |
+
64,
|
| 133 |
+
64,
|
| 134 |
+
64,
|
| 135 |
+
48,
|
| 136 |
+
64,
|
| 137 |
+
64,
|
| 138 |
+
64,
|
| 139 |
+
48,
|
| 140 |
+
64,
|
| 141 |
+
64,
|
| 142 |
+
64,
|
| 143 |
+
48,
|
| 144 |
+
64,
|
| 145 |
+
64,
|
| 146 |
+
64,
|
| 147 |
+
48,
|
| 148 |
+
64,
|
| 149 |
+
64,
|
| 150 |
+
64
|
| 151 |
+
],
|
| 152 |
+
"num_experts": 256,
|
| 153 |
+
"num_experts_per_tok": 8,
|
| 154 |
+
"num_hidden_layers": 40,
|
| 155 |
+
"num_key_value_heads": 8,
|
| 156 |
+
"pad_token_id": 9,
|
| 157 |
+
"partial_rotary_factor": 0.5,
|
| 158 |
+
"rms_norm_eps": 1e-06,
|
| 159 |
+
"rope_parameters": {
|
| 160 |
+
"full_attention": {
|
| 161 |
+
"rope_theta": 500000.0,
|
| 162 |
+
"rope_type": "yarn",
|
| 163 |
+
"factor": 32.0,
|
| 164 |
+
"original_max_position_embeddings": 4096,
|
| 165 |
+
"beta_slow": 1.0,
|
| 166 |
+
"beta_fast": 64.0,
|
| 167 |
+
"attention_factor": 1.0,
|
| 168 |
+
"partial_rotary_factor": 0.5
|
| 169 |
+
},
|
| 170 |
+
"sliding_attention": {
|
| 171 |
+
"rope_type": "default",
|
| 172 |
+
"rope_theta": 10000.0,
|
| 173 |
+
"partial_rotary_factor": 1.0
|
| 174 |
+
},
|
| 175 |
+
"original_max_position_embeddings": 4096
|
| 176 |
+
},
|
| 177 |
+
"router_aux_loss_coef": 0.0,
|
| 178 |
+
"shared_expert_intermediate_size": 512,
|
| 179 |
+
"sliding_window": 512,
|
| 180 |
+
"tie_word_embeddings": false,
|
| 181 |
+
"torch_dtype": "bfloat16",
|
| 182 |
+
"use_cache": true,
|
| 183 |
+
"vocab_size": 100352
|
| 184 |
+
}
|
configuration_laguna.py
ADDED
|
@@ -0,0 +1,213 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Copyright 2026 Poolside and the HuggingFace Inc. team. All rights reserved.
|
| 2 |
+
#
|
| 3 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 4 |
+
# you may not use this file except in compliance with the License.
|
| 5 |
+
# You may obtain a copy of the License at
|
| 6 |
+
#
|
| 7 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 8 |
+
#
|
| 9 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 10 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 11 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 12 |
+
# See the License for the specific language governing permissions and
|
| 13 |
+
# limitations under the License.
|
| 14 |
+
from typing import Any, Literal
|
| 15 |
+
|
| 16 |
+
from huggingface_hub.dataclasses import strict
|
| 17 |
+
|
| 18 |
+
from transformers.configuration_utils import PreTrainedConfig
|
| 19 |
+
from transformers.modeling_rope_utils import RopeParameters
|
| 20 |
+
from transformers.utils import auto_docstring
|
| 21 |
+
|
| 22 |
+
|
| 23 |
+
@auto_docstring(checkpoint="poolside/laguna-XS.2")
|
| 24 |
+
@strict
|
| 25 |
+
class LagunaConfig(PreTrainedConfig):
|
| 26 |
+
r"""
|
| 27 |
+
partial_rotary_factor (`float`, *optional*):
|
| 28 |
+
Fraction of ``head_dim`` to rotate. Folded into each ``rope_parameters[layer_type]``
|
| 29 |
+
entry by ``__post_init__``.
|
| 30 |
+
num_attention_heads_per_layer (`list[int]`, *optional*):
|
| 31 |
+
Per-layer override for ``num_attention_heads``. Length must equal ``num_hidden_layers``.
|
| 32 |
+
mlp_layer_types (`list[str]`, *optional*):
|
| 33 |
+
Per-layer MLP type — ``"dense"`` or ``"sparse"``. Length must equal
|
| 34 |
+
``num_hidden_layers``. Defaults to first layer dense, rest sparse.
|
| 35 |
+
moe_routed_scaling_factor (`float`, *optional*, defaults to 1.0):
|
| 36 |
+
Scalar applied to routed-expert output before combining with the shared-expert output.
|
| 37 |
+
moe_apply_router_weight_on_input (`bool`, *optional*, defaults to `False`):
|
| 38 |
+
Whether to apply router weights to the MoE input rather than the output. Not supported
|
| 39 |
+
in transformers yet; ``True`` will raise a ``NotImplementedError`` for now.
|
| 40 |
+
moe_router_logit_softcapping (`float`, *optional*, defaults to 0.0):
|
| 41 |
+
Scaling factor when applying tanh softcapping on the logits of the MoE router logits.
|
| 42 |
+
|
| 43 |
+
Example:
|
| 44 |
+
|
| 45 |
+
```python
|
| 46 |
+
>>> from transformers import LagunaModel, LagunaConfig
|
| 47 |
+
|
| 48 |
+
>>> configuration = LagunaConfig()
|
| 49 |
+
>>> model = LagunaModel(configuration)
|
| 50 |
+
>>> configuration = model.config
|
| 51 |
+
```
|
| 52 |
+
"""
|
| 53 |
+
|
| 54 |
+
model_type = "laguna"
|
| 55 |
+
keys_to_ignore_at_inference = ["past_key_values"]
|
| 56 |
+
base_model_tp_plan = {
|
| 57 |
+
"layers.*.self_attn.q_proj": "colwise",
|
| 58 |
+
"layers.*.self_attn.k_proj": "colwise",
|
| 59 |
+
"layers.*.self_attn.v_proj": "colwise",
|
| 60 |
+
"layers.*.self_attn.g_proj": "colwise",
|
| 61 |
+
"layers.*.self_attn.o_proj": "rowwise",
|
| 62 |
+
"layers.*.self_attn.q_norm": "replicated_with_grad_allreduce",
|
| 63 |
+
"layers.*.self_attn.k_norm": "replicated_with_grad_allreduce",
|
| 64 |
+
"layers.*.mlp.gate_proj": "colwise",
|
| 65 |
+
"layers.*.mlp.up_proj": "colwise",
|
| 66 |
+
"layers.*.mlp.down_proj": "rowwise",
|
| 67 |
+
"layers.*.mlp.experts.gate_up_proj": "packed_colwise",
|
| 68 |
+
"layers.*.mlp.experts.down_proj": "rowwise",
|
| 69 |
+
"layers.*.mlp.experts": "moe_tp_experts",
|
| 70 |
+
"layers.*.mlp.shared_experts.gate_proj": "colwise",
|
| 71 |
+
"layers.*.mlp.shared_experts.up_proj": "colwise",
|
| 72 |
+
"layers.*.mlp.shared_experts.down_proj": "rowwise",
|
| 73 |
+
}
|
| 74 |
+
base_model_pp_plan = {
|
| 75 |
+
"embed_tokens": (["input_ids"], ["inputs_embeds"]),
|
| 76 |
+
"layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
|
| 77 |
+
"norm": (["hidden_states"], ["hidden_states"]),
|
| 78 |
+
}
|
| 79 |
+
|
| 80 |
+
# Qwen2Moe-inherited defaults we want to override for Laguna's typical shape.
|
| 81 |
+
vocab_size: int = 100352
|
| 82 |
+
hidden_size: int = 2048
|
| 83 |
+
intermediate_size: int = 8192
|
| 84 |
+
num_hidden_layers: int = 40
|
| 85 |
+
num_attention_heads: int = 48
|
| 86 |
+
num_key_value_heads: int = 8
|
| 87 |
+
hidden_act: str = "silu"
|
| 88 |
+
max_position_embeddings: int = 131072
|
| 89 |
+
initializer_range: float = 0.02
|
| 90 |
+
rms_norm_eps: float = 1e-6
|
| 91 |
+
use_cache: bool = True
|
| 92 |
+
tie_word_embeddings: bool = False
|
| 93 |
+
rope_parameters: RopeParameters | dict | None = None
|
| 94 |
+
sliding_window: int | None = None
|
| 95 |
+
attention_dropout: float | int = 0.0
|
| 96 |
+
moe_intermediate_size: int = 512
|
| 97 |
+
shared_expert_intermediate_size: int = 512
|
| 98 |
+
num_experts_per_tok: int = 8
|
| 99 |
+
num_experts: int = 256
|
| 100 |
+
output_router_logits: bool = False
|
| 101 |
+
router_aux_loss_coef: float = 0.001
|
| 102 |
+
layer_types: list[str] | None = None
|
| 103 |
+
pad_token_id: int | None = None
|
| 104 |
+
bos_token_id: int | None = None
|
| 105 |
+
eos_token_id: int | list[int] | None = None
|
| 106 |
+
|
| 107 |
+
# Laguna-specific attention
|
| 108 |
+
head_dim: int = 128
|
| 109 |
+
attention_bias: bool = False
|
| 110 |
+
partial_rotary_factor: float | None = None
|
| 111 |
+
num_attention_heads_per_layer: list[int] | None = None
|
| 112 |
+
# Laguna-specific MoE
|
| 113 |
+
mlp_layer_types: list[str] | None = None
|
| 114 |
+
moe_routed_scaling_factor: float = 1.0
|
| 115 |
+
moe_apply_router_weight_on_input: bool = False
|
| 116 |
+
moe_router_logit_softcapping: float = 0.0
|
| 117 |
+
|
| 118 |
+
def __post_init__(self, **kwargs):
|
| 119 |
+
if self.layer_types is None:
|
| 120 |
+
self.layer_types = ["full_attention"] * self.num_hidden_layers
|
| 121 |
+
if self.mlp_layer_types is None:
|
| 122 |
+
self.mlp_layer_types = ["dense"] + ["sparse"] * (self.num_hidden_layers - 1)
|
| 123 |
+
if self.num_attention_heads_per_layer is None:
|
| 124 |
+
self.num_attention_heads_per_layer = [self.num_attention_heads] * self.num_hidden_layers
|
| 125 |
+
|
| 126 |
+
default_rope_params: dict[Literal["full_attention", "sliding_attention"], dict[str, Any]] = {
|
| 127 |
+
"full_attention": {"rope_type": "default", "rope_theta": 500000.0},
|
| 128 |
+
"sliding_attention": {"rope_type": "default", "rope_theta": 10000.0},
|
| 129 |
+
}
|
| 130 |
+
if self.rope_parameters is None:
|
| 131 |
+
self.rope_parameters = default_rope_params
|
| 132 |
+
|
| 133 |
+
self._normalize_rope_parameters()
|
| 134 |
+
# Skip ``Qwen2MoeConfig.__post_init__`` — it references ``mlp_only_layers`` /
|
| 135 |
+
# ``use_sliding_window`` / ``max_window_layers`` which Laguna drops above.
|
| 136 |
+
super().__post_init__(**kwargs)
|
| 137 |
+
|
| 138 |
+
def _normalize_rope_parameters(self):
|
| 139 |
+
"""Coerce ``rope_parameters`` to the nested ``{layer_type: {...}}`` shape.
|
| 140 |
+
|
| 141 |
+
Accepts an already-nested dict as-is, or a flat dict that gets broadcast to every
|
| 142 |
+
layer type. A top-level ``partial_rotary_factor`` is folded into each sub-dict as
|
| 143 |
+
a default.
|
| 144 |
+
"""
|
| 145 |
+
layer_types = set(self.layer_types)
|
| 146 |
+
rope_params = self.rope_parameters or {}
|
| 147 |
+
is_nested = isinstance(rope_params, dict) and any(k in layer_types for k in rope_params)
|
| 148 |
+
if is_nested:
|
| 149 |
+
nested = {lt: dict(rope_params.get(lt, {})) for lt in layer_types}
|
| 150 |
+
else:
|
| 151 |
+
nested = {lt: dict(rope_params) for lt in layer_types}
|
| 152 |
+
|
| 153 |
+
if self.partial_rotary_factor is not None:
|
| 154 |
+
for params in nested.values():
|
| 155 |
+
params.setdefault("partial_rotary_factor", self.partial_rotary_factor)
|
| 156 |
+
|
| 157 |
+
for params in nested.values():
|
| 158 |
+
params.setdefault("rope_type", "default")
|
| 159 |
+
|
| 160 |
+
self.rope_parameters = nested
|
| 161 |
+
# Null the top-level field now that its value lives in each sub-dict — otherwise
|
| 162 |
+
# ``standardize_rope_params`` would overwrite per-type values with the global one.
|
| 163 |
+
self.partial_rotary_factor = None
|
| 164 |
+
|
| 165 |
+
def convert_rope_params_to_dict(self, **kwargs):
|
| 166 |
+
# No need to handle BC for new models, because they have no old-format `rope_scaling`
|
| 167 |
+
return kwargs
|
| 168 |
+
|
| 169 |
+
def _validate_yarn_rope_parameters(self, rope_parameters: dict, ignore_keys=None):
|
| 170 |
+
"""Override: parent reads ``self.rope_parameters["original_max_position_embeddings"]``
|
| 171 |
+
for its post-hoc factor sanity-check, which works for flat rope configs but raises
|
| 172 |
+
``KeyError`` when ``self.rope_parameters`` is the Laguna/Gemma3-style per-layer-type
|
| 173 |
+
map (its keys are layer types like ``"full_attention"``). Fix locally by reading
|
| 174 |
+
from the per-call ``rope_parameters`` dict that ``validate_rope`` already passes in.
|
| 175 |
+
"""
|
| 176 |
+
# Delegate to parent for the shared checks by temporarily swapping in a flat
|
| 177 |
+
# ``self.rope_parameters`` that has the key the parent expects. Cheapest way to
|
| 178 |
+
# share the parent's logic without reimplementing it here.
|
| 179 |
+
flat = getattr(self, "rope_parameters", None)
|
| 180 |
+
self.rope_parameters = rope_parameters
|
| 181 |
+
try:
|
| 182 |
+
super()._validate_yarn_rope_parameters(rope_parameters, ignore_keys=ignore_keys)
|
| 183 |
+
finally:
|
| 184 |
+
self.rope_parameters = flat
|
| 185 |
+
|
| 186 |
+
def validate_architecture(self):
|
| 187 |
+
"""Part of ``@strict``-powered validation."""
|
| 188 |
+
if self.moe_apply_router_weight_on_input:
|
| 189 |
+
raise NotImplementedError(
|
| 190 |
+
"moe_apply_router_weight_on_input=True is not yet supported in the "
|
| 191 |
+
"transformers implementation of Laguna."
|
| 192 |
+
)
|
| 193 |
+
if (
|
| 194 |
+
self.num_attention_heads_per_layer is not None
|
| 195 |
+
and len(self.num_attention_heads_per_layer) != self.num_hidden_layers
|
| 196 |
+
):
|
| 197 |
+
raise ValueError(
|
| 198 |
+
f"num_attention_heads_per_layer length ({len(self.num_attention_heads_per_layer)}) "
|
| 199 |
+
f"must equal num_hidden_layers ({self.num_hidden_layers})."
|
| 200 |
+
)
|
| 201 |
+
if len(self.layer_types) != self.num_hidden_layers:
|
| 202 |
+
raise ValueError(
|
| 203 |
+
f"layer_types length ({len(self.layer_types)}) "
|
| 204 |
+
f"must equal num_hidden_layers ({self.num_hidden_layers})."
|
| 205 |
+
)
|
| 206 |
+
if len(self.mlp_layer_types) != self.num_hidden_layers:
|
| 207 |
+
raise ValueError(
|
| 208 |
+
f"mlp_layer_types length ({len(self.mlp_layer_types)}) "
|
| 209 |
+
f"must equal num_hidden_layers ({self.num_hidden_layers})."
|
| 210 |
+
)
|
| 211 |
+
|
| 212 |
+
|
| 213 |
+
__all__ = ["LagunaConfig"]
|
generation_config.json
ADDED
|
@@ -0,0 +1,12 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"bos_token_id": 2,
|
| 3 |
+
"do_sample": true,
|
| 4 |
+
"eos_token_id": [
|
| 5 |
+
2,
|
| 6 |
+
24
|
| 7 |
+
],
|
| 8 |
+
"max_new_tokens": 2048,
|
| 9 |
+
"pad_token_id": 9,
|
| 10 |
+
"temperature": 0.7,
|
| 11 |
+
"top_p": 0.9
|
| 12 |
+
}
|
model-00001-of-00014.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f9fe4500c6fa1674cc0cb4af6ee39c771fe3c4f5a50a04cd4bd484be57e70ac7
|
| 3 |
+
size 5108631760
|
model-00002-of-00014.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:bec7bc2b1f85933f19e24fc36b63ffd08d9f8491c9223938ce54155bcbef2807
|
| 3 |
+
size 5064328168
|
model-00003-of-00014.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:1c7c2a3a2bff6128e3142c784b734137753fcdb229ba5bf5fea1cef616f56609
|
| 3 |
+
size 5064328152
|
model-00004-of-00014.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:7ac36bbdb16cc0f3aacc1f18d31ab2b357be892d309b24d26799c7b5eb92d158
|
| 3 |
+
size 5064328201
|
model-00005-of-00014.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:0808b34a8cc8335826c82f39ca584766ecb99c1d327f75c2ef3c8f35de1d498d
|
| 3 |
+
size 5081170971
|
model-00006-of-00014.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:e62695f20640e3c7004b62f9cfe5c9aaf2ab25ecce08b8ff5be6d4e75f585162
|
| 3 |
+
size 5064328219
|
model-00007-of-00014.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:fd3e10d357f725365bd09705ce4af92c11008b0adb92681702601fb2fea2d372
|
| 3 |
+
size 5064328167
|
model-00008-of-00014.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:e802c51a04c31410a0fb69a495cfcfaa9b2c134ddbca888f9de295d5463f6a74
|
| 3 |
+
size 5064328179
|
model-00009-of-00014.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:aeaa943ff2cbb4c08e3106eb143430e212d8290273bdd8b2c631d42f15778e12
|
| 3 |
+
size 5081170967
|
model-00010-of-00014.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ab0a5e67a8c1b0c6234ad83008615dcac627288feeba4329f2cb58f989c2af35
|
| 3 |
+
size 5064328173
|
model-00011-of-00014.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d25feaac0fa88400e53c6c84db94018813d0343b6f0ae4dc415a096b4c4fb32b
|
| 3 |
+
size 5064328207
|
model-00012-of-00014.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:4e7d3eddfae6108bd949fb9044b0853e341fff9290e18dcbd5dfe1a8ebb5a394
|
| 3 |
+
size 5064328203
|
model-00013-of-00014.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:5c04663e58796c20df7aa858bc96e749658a74be28e1c33b1fb63ebb230f26f3
|
| 3 |
+
size 5081170929
|
model-00014-of-00014.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:030a59809f2e98c4429b87ffeae3758011e54e2a194810c8f0dd77d34e397e93
|
| 3 |
+
size 954217395
|
model.safetensors.index.json
ADDED
|
@@ -0,0 +1,686 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"metadata": {
|
| 3 |
+
"total_size": 66885234176,
|
| 4 |
+
"total_parameters": 33442607104
|
| 5 |
+
},
|
| 6 |
+
"weight_map": {
|
| 7 |
+
"lm_head.weight": "model-00014-of-00014.safetensors",
|
| 8 |
+
"model.embed_tokens.weight": "model-00001-of-00014.safetensors",
|
| 9 |
+
"model.layers.0.input_layernorm.weight": "model-00001-of-00014.safetensors",
|
| 10 |
+
"model.layers.0.mlp.down_proj.weight": "model-00001-of-00014.safetensors",
|
| 11 |
+
"model.layers.0.mlp.gate_proj.weight": "model-00001-of-00014.safetensors",
|
| 12 |
+
"model.layers.0.mlp.up_proj.weight": "model-00001-of-00014.safetensors",
|
| 13 |
+
"model.layers.0.post_attention_layernorm.weight": "model-00001-of-00014.safetensors",
|
| 14 |
+
"model.layers.0.self_attn.g_proj.weight": "model-00001-of-00014.safetensors",
|
| 15 |
+
"model.layers.0.self_attn.k_norm.weight": "model-00001-of-00014.safetensors",
|
| 16 |
+
"model.layers.0.self_attn.k_proj.weight": "model-00001-of-00014.safetensors",
|
| 17 |
+
"model.layers.0.self_attn.o_proj.weight": "model-00001-of-00014.safetensors",
|
| 18 |
+
"model.layers.0.self_attn.q_norm.weight": "model-00001-of-00014.safetensors",
|
| 19 |
+
"model.layers.0.self_attn.q_proj.weight": "model-00001-of-00014.safetensors",
|
| 20 |
+
"model.layers.0.self_attn.v_proj.weight": "model-00001-of-00014.safetensors",
|
| 21 |
+
"model.layers.1.input_layernorm.weight": "model-00001-of-00014.safetensors",
|
| 22 |
+
"model.layers.1.mlp.gate.e_score_correction_bias": "model-00001-of-00014.safetensors",
|
| 23 |
+
"model.layers.1.mlp.gate.proj.weight": "model-00001-of-00014.safetensors",
|
| 24 |
+
"model.layers.1.mlp.shared_expert.down_proj.weight": "model-00001-of-00014.safetensors",
|
| 25 |
+
"model.layers.1.mlp.shared_expert.gate_proj.weight": "model-00001-of-00014.safetensors",
|
| 26 |
+
"model.layers.1.mlp.shared_expert.up_proj.weight": "model-00001-of-00014.safetensors",
|
| 27 |
+
"model.layers.1.mlp.switch_mlp.down_proj.weight": "model-00001-of-00014.safetensors",
|
| 28 |
+
"model.layers.1.mlp.switch_mlp.gate_proj.weight": "model-00001-of-00014.safetensors",
|
| 29 |
+
"model.layers.1.mlp.switch_mlp.up_proj.weight": "model-00001-of-00014.safetensors",
|
| 30 |
+
"model.layers.1.post_attention_layernorm.weight": "model-00001-of-00014.safetensors",
|
| 31 |
+
"model.layers.1.self_attn.g_proj.weight": "model-00001-of-00014.safetensors",
|
| 32 |
+
"model.layers.1.self_attn.k_norm.weight": "model-00001-of-00014.safetensors",
|
| 33 |
+
"model.layers.1.self_attn.k_proj.weight": "model-00001-of-00014.safetensors",
|
| 34 |
+
"model.layers.1.self_attn.o_proj.weight": "model-00001-of-00014.safetensors",
|
| 35 |
+
"model.layers.1.self_attn.q_norm.weight": "model-00001-of-00014.safetensors",
|
| 36 |
+
"model.layers.1.self_attn.q_proj.weight": "model-00001-of-00014.safetensors",
|
| 37 |
+
"model.layers.1.self_attn.v_proj.weight": "model-00001-of-00014.safetensors",
|
| 38 |
+
"model.layers.10.input_layernorm.weight": "model-00004-of-00014.safetensors",
|
| 39 |
+
"model.layers.10.mlp.gate.e_score_correction_bias": "model-00004-of-00014.safetensors",
|
| 40 |
+
"model.layers.10.mlp.gate.proj.weight": "model-00004-of-00014.safetensors",
|
| 41 |
+
"model.layers.10.mlp.shared_expert.down_proj.weight": "model-00004-of-00014.safetensors",
|
| 42 |
+
"model.layers.10.mlp.shared_expert.gate_proj.weight": "model-00004-of-00014.safetensors",
|
| 43 |
+
"model.layers.10.mlp.shared_expert.up_proj.weight": "model-00004-of-00014.safetensors",
|
| 44 |
+
"model.layers.10.mlp.switch_mlp.down_proj.weight": "model-00004-of-00014.safetensors",
|
| 45 |
+
"model.layers.10.mlp.switch_mlp.gate_proj.weight": "model-00004-of-00014.safetensors",
|
| 46 |
+
"model.layers.10.mlp.switch_mlp.up_proj.weight": "model-00004-of-00014.safetensors",
|
| 47 |
+
"model.layers.10.post_attention_layernorm.weight": "model-00004-of-00014.safetensors",
|
| 48 |
+
"model.layers.10.self_attn.g_proj.weight": "model-00004-of-00014.safetensors",
|
| 49 |
+
"model.layers.10.self_attn.k_norm.weight": "model-00004-of-00014.safetensors",
|
| 50 |
+
"model.layers.10.self_attn.k_proj.weight": "model-00004-of-00014.safetensors",
|
| 51 |
+
"model.layers.10.self_attn.o_proj.weight": "model-00004-of-00014.safetensors",
|
| 52 |
+
"model.layers.10.self_attn.q_norm.weight": "model-00004-of-00014.safetensors",
|
| 53 |
+
"model.layers.10.self_attn.q_proj.weight": "model-00004-of-00014.safetensors",
|
| 54 |
+
"model.layers.10.self_attn.v_proj.weight": "model-00004-of-00014.safetensors",
|
| 55 |
+
"model.layers.11.input_layernorm.weight": "model-00004-of-00014.safetensors",
|
| 56 |
+
"model.layers.11.mlp.gate.e_score_correction_bias": "model-00004-of-00014.safetensors",
|
| 57 |
+
"model.layers.11.mlp.gate.proj.weight": "model-00004-of-00014.safetensors",
|
| 58 |
+
"model.layers.11.mlp.shared_expert.down_proj.weight": "model-00004-of-00014.safetensors",
|
| 59 |
+
"model.layers.11.mlp.shared_expert.gate_proj.weight": "model-00004-of-00014.safetensors",
|
| 60 |
+
"model.layers.11.mlp.shared_expert.up_proj.weight": "model-00004-of-00014.safetensors",
|
| 61 |
+
"model.layers.11.mlp.switch_mlp.down_proj.weight": "model-00004-of-00014.safetensors",
|
| 62 |
+
"model.layers.11.mlp.switch_mlp.gate_proj.weight": "model-00004-of-00014.safetensors",
|
| 63 |
+
"model.layers.11.mlp.switch_mlp.up_proj.weight": "model-00004-of-00014.safetensors",
|
| 64 |
+
"model.layers.11.post_attention_layernorm.weight": "model-00004-of-00014.safetensors",
|
| 65 |
+
"model.layers.11.self_attn.g_proj.weight": "model-00004-of-00014.safetensors",
|
| 66 |
+
"model.layers.11.self_attn.k_norm.weight": "model-00004-of-00014.safetensors",
|
| 67 |
+
"model.layers.11.self_attn.k_proj.weight": "model-00004-of-00014.safetensors",
|
| 68 |
+
"model.layers.11.self_attn.o_proj.weight": "model-00004-of-00014.safetensors",
|
| 69 |
+
"model.layers.11.self_attn.q_norm.weight": "model-00004-of-00014.safetensors",
|
| 70 |
+
"model.layers.11.self_attn.q_proj.weight": "model-00004-of-00014.safetensors",
|
| 71 |
+
"model.layers.11.self_attn.v_proj.weight": "model-00004-of-00014.safetensors",
|
| 72 |
+
"model.layers.12.input_layernorm.weight": "model-00005-of-00014.safetensors",
|
| 73 |
+
"model.layers.12.mlp.gate.e_score_correction_bias": "model-00004-of-00014.safetensors",
|
| 74 |
+
"model.layers.12.mlp.gate.proj.weight": "model-00004-of-00014.safetensors",
|
| 75 |
+
"model.layers.12.mlp.shared_expert.down_proj.weight": "model-00005-of-00014.safetensors",
|
| 76 |
+
"model.layers.12.mlp.shared_expert.gate_proj.weight": "model-00005-of-00014.safetensors",
|
| 77 |
+
"model.layers.12.mlp.shared_expert.up_proj.weight": "model-00005-of-00014.safetensors",
|
| 78 |
+
"model.layers.12.mlp.switch_mlp.down_proj.weight": "model-00005-of-00014.safetensors",
|
| 79 |
+
"model.layers.12.mlp.switch_mlp.gate_proj.weight": "model-00004-of-00014.safetensors",
|
| 80 |
+
"model.layers.12.mlp.switch_mlp.up_proj.weight": "model-00004-of-00014.safetensors",
|
| 81 |
+
"model.layers.12.post_attention_layernorm.weight": "model-00005-of-00014.safetensors",
|
| 82 |
+
"model.layers.12.self_attn.g_proj.weight": "model-00004-of-00014.safetensors",
|
| 83 |
+
"model.layers.12.self_attn.k_norm.weight": "model-00004-of-00014.safetensors",
|
| 84 |
+
"model.layers.12.self_attn.k_proj.weight": "model-00004-of-00014.safetensors",
|
| 85 |
+
"model.layers.12.self_attn.o_proj.weight": "model-00004-of-00014.safetensors",
|
| 86 |
+
"model.layers.12.self_attn.q_norm.weight": "model-00004-of-00014.safetensors",
|
| 87 |
+
"model.layers.12.self_attn.q_proj.weight": "model-00004-of-00014.safetensors",
|
| 88 |
+
"model.layers.12.self_attn.v_proj.weight": "model-00004-of-00014.safetensors",
|
| 89 |
+
"model.layers.13.input_layernorm.weight": "model-00005-of-00014.safetensors",
|
| 90 |
+
"model.layers.13.mlp.gate.e_score_correction_bias": "model-00005-of-00014.safetensors",
|
| 91 |
+
"model.layers.13.mlp.gate.proj.weight": "model-00005-of-00014.safetensors",
|
| 92 |
+
"model.layers.13.mlp.shared_expert.down_proj.weight": "model-00005-of-00014.safetensors",
|
| 93 |
+
"model.layers.13.mlp.shared_expert.gate_proj.weight": "model-00005-of-00014.safetensors",
|
| 94 |
+
"model.layers.13.mlp.shared_expert.up_proj.weight": "model-00005-of-00014.safetensors",
|
| 95 |
+
"model.layers.13.mlp.switch_mlp.down_proj.weight": "model-00005-of-00014.safetensors",
|
| 96 |
+
"model.layers.13.mlp.switch_mlp.gate_proj.weight": "model-00005-of-00014.safetensors",
|
| 97 |
+
"model.layers.13.mlp.switch_mlp.up_proj.weight": "model-00005-of-00014.safetensors",
|
| 98 |
+
"model.layers.13.post_attention_layernorm.weight": "model-00005-of-00014.safetensors",
|
| 99 |
+
"model.layers.13.self_attn.g_proj.weight": "model-00005-of-00014.safetensors",
|
| 100 |
+
"model.layers.13.self_attn.k_norm.weight": "model-00005-of-00014.safetensors",
|
| 101 |
+
"model.layers.13.self_attn.k_proj.weight": "model-00005-of-00014.safetensors",
|
| 102 |
+
"model.layers.13.self_attn.o_proj.weight": "model-00005-of-00014.safetensors",
|
| 103 |
+
"model.layers.13.self_attn.q_norm.weight": "model-00005-of-00014.safetensors",
|
| 104 |
+
"model.layers.13.self_attn.q_proj.weight": "model-00005-of-00014.safetensors",
|
| 105 |
+
"model.layers.13.self_attn.v_proj.weight": "model-00005-of-00014.safetensors",
|
| 106 |
+
"model.layers.14.input_layernorm.weight": "model-00005-of-00014.safetensors",
|
| 107 |
+
"model.layers.14.mlp.gate.e_score_correction_bias": "model-00005-of-00014.safetensors",
|
| 108 |
+
"model.layers.14.mlp.gate.proj.weight": "model-00005-of-00014.safetensors",
|
| 109 |
+
"model.layers.14.mlp.shared_expert.down_proj.weight": "model-00005-of-00014.safetensors",
|
| 110 |
+
"model.layers.14.mlp.shared_expert.gate_proj.weight": "model-00005-of-00014.safetensors",
|
| 111 |
+
"model.layers.14.mlp.shared_expert.up_proj.weight": "model-00005-of-00014.safetensors",
|
| 112 |
+
"model.layers.14.mlp.switch_mlp.down_proj.weight": "model-00005-of-00014.safetensors",
|
| 113 |
+
"model.layers.14.mlp.switch_mlp.gate_proj.weight": "model-00005-of-00014.safetensors",
|
| 114 |
+
"model.layers.14.mlp.switch_mlp.up_proj.weight": "model-00005-of-00014.safetensors",
|
| 115 |
+
"model.layers.14.post_attention_layernorm.weight": "model-00005-of-00014.safetensors",
|
| 116 |
+
"model.layers.14.self_attn.g_proj.weight": "model-00005-of-00014.safetensors",
|
| 117 |
+
"model.layers.14.self_attn.k_norm.weight": "model-00005-of-00014.safetensors",
|
| 118 |
+
"model.layers.14.self_attn.k_proj.weight": "model-00005-of-00014.safetensors",
|
| 119 |
+
"model.layers.14.self_attn.o_proj.weight": "model-00005-of-00014.safetensors",
|
| 120 |
+
"model.layers.14.self_attn.q_norm.weight": "model-00005-of-00014.safetensors",
|
| 121 |
+
"model.layers.14.self_attn.q_proj.weight": "model-00005-of-00014.safetensors",
|
| 122 |
+
"model.layers.14.self_attn.v_proj.weight": "model-00005-of-00014.safetensors",
|
| 123 |
+
"model.layers.15.input_layernorm.weight": "model-00006-of-00014.safetensors",
|
| 124 |
+
"model.layers.15.mlp.gate.e_score_correction_bias": "model-00005-of-00014.safetensors",
|
| 125 |
+
"model.layers.15.mlp.gate.proj.weight": "model-00005-of-00014.safetensors",
|
| 126 |
+
"model.layers.15.mlp.shared_expert.down_proj.weight": "model-00006-of-00014.safetensors",
|
| 127 |
+
"model.layers.15.mlp.shared_expert.gate_proj.weight": "model-00006-of-00014.safetensors",
|
| 128 |
+
"model.layers.15.mlp.shared_expert.up_proj.weight": "model-00006-of-00014.safetensors",
|
| 129 |
+
"model.layers.15.mlp.switch_mlp.down_proj.weight": "model-00006-of-00014.safetensors",
|
| 130 |
+
"model.layers.15.mlp.switch_mlp.gate_proj.weight": "model-00005-of-00014.safetensors",
|
| 131 |
+
"model.layers.15.mlp.switch_mlp.up_proj.weight": "model-00005-of-00014.safetensors",
|
| 132 |
+
"model.layers.15.post_attention_layernorm.weight": "model-00006-of-00014.safetensors",
|
| 133 |
+
"model.layers.15.self_attn.g_proj.weight": "model-00005-of-00014.safetensors",
|
| 134 |
+
"model.layers.15.self_attn.k_norm.weight": "model-00005-of-00014.safetensors",
|
| 135 |
+
"model.layers.15.self_attn.k_proj.weight": "model-00005-of-00014.safetensors",
|
| 136 |
+
"model.layers.15.self_attn.o_proj.weight": "model-00005-of-00014.safetensors",
|
| 137 |
+
"model.layers.15.self_attn.q_norm.weight": "model-00005-of-00014.safetensors",
|
| 138 |
+
"model.layers.15.self_attn.q_proj.weight": "model-00005-of-00014.safetensors",
|
| 139 |
+
"model.layers.15.self_attn.v_proj.weight": "model-00005-of-00014.safetensors",
|
| 140 |
+
"model.layers.16.input_layernorm.weight": "model-00006-of-00014.safetensors",
|
| 141 |
+
"model.layers.16.mlp.gate.e_score_correction_bias": "model-00006-of-00014.safetensors",
|
| 142 |
+
"model.layers.16.mlp.gate.proj.weight": "model-00006-of-00014.safetensors",
|
| 143 |
+
"model.layers.16.mlp.shared_expert.down_proj.weight": "model-00006-of-00014.safetensors",
|
| 144 |
+
"model.layers.16.mlp.shared_expert.gate_proj.weight": "model-00006-of-00014.safetensors",
|
| 145 |
+
"model.layers.16.mlp.shared_expert.up_proj.weight": "model-00006-of-00014.safetensors",
|
| 146 |
+
"model.layers.16.mlp.switch_mlp.down_proj.weight": "model-00006-of-00014.safetensors",
|
| 147 |
+
"model.layers.16.mlp.switch_mlp.gate_proj.weight": "model-00006-of-00014.safetensors",
|
| 148 |
+
"model.layers.16.mlp.switch_mlp.up_proj.weight": "model-00006-of-00014.safetensors",
|
| 149 |
+
"model.layers.16.post_attention_layernorm.weight": "model-00006-of-00014.safetensors",
|
| 150 |
+
"model.layers.16.self_attn.g_proj.weight": "model-00006-of-00014.safetensors",
|
| 151 |
+
"model.layers.16.self_attn.k_norm.weight": "model-00006-of-00014.safetensors",
|
| 152 |
+
"model.layers.16.self_attn.k_proj.weight": "model-00006-of-00014.safetensors",
|
| 153 |
+
"model.layers.16.self_attn.o_proj.weight": "model-00006-of-00014.safetensors",
|
| 154 |
+
"model.layers.16.self_attn.q_norm.weight": "model-00006-of-00014.safetensors",
|
| 155 |
+
"model.layers.16.self_attn.q_proj.weight": "model-00006-of-00014.safetensors",
|
| 156 |
+
"model.layers.16.self_attn.v_proj.weight": "model-00006-of-00014.safetensors",
|
| 157 |
+
"model.layers.17.input_layernorm.weight": "model-00006-of-00014.safetensors",
|
| 158 |
+
"model.layers.17.mlp.gate.e_score_correction_bias": "model-00006-of-00014.safetensors",
|
| 159 |
+
"model.layers.17.mlp.gate.proj.weight": "model-00006-of-00014.safetensors",
|
| 160 |
+
"model.layers.17.mlp.shared_expert.down_proj.weight": "model-00006-of-00014.safetensors",
|
| 161 |
+
"model.layers.17.mlp.shared_expert.gate_proj.weight": "model-00006-of-00014.safetensors",
|
| 162 |
+
"model.layers.17.mlp.shared_expert.up_proj.weight": "model-00006-of-00014.safetensors",
|
| 163 |
+
"model.layers.17.mlp.switch_mlp.down_proj.weight": "model-00006-of-00014.safetensors",
|
| 164 |
+
"model.layers.17.mlp.switch_mlp.gate_proj.weight": "model-00006-of-00014.safetensors",
|
| 165 |
+
"model.layers.17.mlp.switch_mlp.up_proj.weight": "model-00006-of-00014.safetensors",
|
| 166 |
+
"model.layers.17.post_attention_layernorm.weight": "model-00006-of-00014.safetensors",
|
| 167 |
+
"model.layers.17.self_attn.g_proj.weight": "model-00006-of-00014.safetensors",
|
| 168 |
+
"model.layers.17.self_attn.k_norm.weight": "model-00006-of-00014.safetensors",
|
| 169 |
+
"model.layers.17.self_attn.k_proj.weight": "model-00006-of-00014.safetensors",
|
| 170 |
+
"model.layers.17.self_attn.o_proj.weight": "model-00006-of-00014.safetensors",
|
| 171 |
+
"model.layers.17.self_attn.q_norm.weight": "model-00006-of-00014.safetensors",
|
| 172 |
+
"model.layers.17.self_attn.q_proj.weight": "model-00006-of-00014.safetensors",
|
| 173 |
+
"model.layers.17.self_attn.v_proj.weight": "model-00006-of-00014.safetensors",
|
| 174 |
+
"model.layers.18.input_layernorm.weight": "model-00007-of-00014.safetensors",
|
| 175 |
+
"model.layers.18.mlp.gate.e_score_correction_bias": "model-00006-of-00014.safetensors",
|
| 176 |
+
"model.layers.18.mlp.gate.proj.weight": "model-00006-of-00014.safetensors",
|
| 177 |
+
"model.layers.18.mlp.shared_expert.down_proj.weight": "model-00007-of-00014.safetensors",
|
| 178 |
+
"model.layers.18.mlp.shared_expert.gate_proj.weight": "model-00007-of-00014.safetensors",
|
| 179 |
+
"model.layers.18.mlp.shared_expert.up_proj.weight": "model-00007-of-00014.safetensors",
|
| 180 |
+
"model.layers.18.mlp.switch_mlp.down_proj.weight": "model-00007-of-00014.safetensors",
|
| 181 |
+
"model.layers.18.mlp.switch_mlp.gate_proj.weight": "model-00006-of-00014.safetensors",
|
| 182 |
+
"model.layers.18.mlp.switch_mlp.up_proj.weight": "model-00006-of-00014.safetensors",
|
| 183 |
+
"model.layers.18.post_attention_layernorm.weight": "model-00007-of-00014.safetensors",
|
| 184 |
+
"model.layers.18.self_attn.g_proj.weight": "model-00006-of-00014.safetensors",
|
| 185 |
+
"model.layers.18.self_attn.k_norm.weight": "model-00006-of-00014.safetensors",
|
| 186 |
+
"model.layers.18.self_attn.k_proj.weight": "model-00006-of-00014.safetensors",
|
| 187 |
+
"model.layers.18.self_attn.o_proj.weight": "model-00006-of-00014.safetensors",
|
| 188 |
+
"model.layers.18.self_attn.q_norm.weight": "model-00006-of-00014.safetensors",
|
| 189 |
+
"model.layers.18.self_attn.q_proj.weight": "model-00006-of-00014.safetensors",
|
| 190 |
+
"model.layers.18.self_attn.v_proj.weight": "model-00006-of-00014.safetensors",
|
| 191 |
+
"model.layers.19.input_layernorm.weight": "model-00007-of-00014.safetensors",
|
| 192 |
+
"model.layers.19.mlp.gate.e_score_correction_bias": "model-00007-of-00014.safetensors",
|
| 193 |
+
"model.layers.19.mlp.gate.proj.weight": "model-00007-of-00014.safetensors",
|
| 194 |
+
"model.layers.19.mlp.shared_expert.down_proj.weight": "model-00007-of-00014.safetensors",
|
| 195 |
+
"model.layers.19.mlp.shared_expert.gate_proj.weight": "model-00007-of-00014.safetensors",
|
| 196 |
+
"model.layers.19.mlp.shared_expert.up_proj.weight": "model-00007-of-00014.safetensors",
|
| 197 |
+
"model.layers.19.mlp.switch_mlp.down_proj.weight": "model-00007-of-00014.safetensors",
|
| 198 |
+
"model.layers.19.mlp.switch_mlp.gate_proj.weight": "model-00007-of-00014.safetensors",
|
| 199 |
+
"model.layers.19.mlp.switch_mlp.up_proj.weight": "model-00007-of-00014.safetensors",
|
| 200 |
+
"model.layers.19.post_attention_layernorm.weight": "model-00007-of-00014.safetensors",
|
| 201 |
+
"model.layers.19.self_attn.g_proj.weight": "model-00007-of-00014.safetensors",
|
| 202 |
+
"model.layers.19.self_attn.k_norm.weight": "model-00007-of-00014.safetensors",
|
| 203 |
+
"model.layers.19.self_attn.k_proj.weight": "model-00007-of-00014.safetensors",
|
| 204 |
+
"model.layers.19.self_attn.o_proj.weight": "model-00007-of-00014.safetensors",
|
| 205 |
+
"model.layers.19.self_attn.q_norm.weight": "model-00007-of-00014.safetensors",
|
| 206 |
+
"model.layers.19.self_attn.q_proj.weight": "model-00007-of-00014.safetensors",
|
| 207 |
+
"model.layers.19.self_attn.v_proj.weight": "model-00007-of-00014.safetensors",
|
| 208 |
+
"model.layers.2.input_layernorm.weight": "model-00001-of-00014.safetensors",
|
| 209 |
+
"model.layers.2.mlp.gate.e_score_correction_bias": "model-00001-of-00014.safetensors",
|
| 210 |
+
"model.layers.2.mlp.gate.proj.weight": "model-00001-of-00014.safetensors",
|
| 211 |
+
"model.layers.2.mlp.shared_expert.down_proj.weight": "model-00001-of-00014.safetensors",
|
| 212 |
+
"model.layers.2.mlp.shared_expert.gate_proj.weight": "model-00001-of-00014.safetensors",
|
| 213 |
+
"model.layers.2.mlp.shared_expert.up_proj.weight": "model-00001-of-00014.safetensors",
|
| 214 |
+
"model.layers.2.mlp.switch_mlp.down_proj.weight": "model-00001-of-00014.safetensors",
|
| 215 |
+
"model.layers.2.mlp.switch_mlp.gate_proj.weight": "model-00001-of-00014.safetensors",
|
| 216 |
+
"model.layers.2.mlp.switch_mlp.up_proj.weight": "model-00001-of-00014.safetensors",
|
| 217 |
+
"model.layers.2.post_attention_layernorm.weight": "model-00001-of-00014.safetensors",
|
| 218 |
+
"model.layers.2.self_attn.g_proj.weight": "model-00001-of-00014.safetensors",
|
| 219 |
+
"model.layers.2.self_attn.k_norm.weight": "model-00001-of-00014.safetensors",
|
| 220 |
+
"model.layers.2.self_attn.k_proj.weight": "model-00001-of-00014.safetensors",
|
| 221 |
+
"model.layers.2.self_attn.o_proj.weight": "model-00001-of-00014.safetensors",
|
| 222 |
+
"model.layers.2.self_attn.q_norm.weight": "model-00001-of-00014.safetensors",
|
| 223 |
+
"model.layers.2.self_attn.q_proj.weight": "model-00001-of-00014.safetensors",
|
| 224 |
+
"model.layers.2.self_attn.v_proj.weight": "model-00001-of-00014.safetensors",
|
| 225 |
+
"model.layers.20.input_layernorm.weight": "model-00007-of-00014.safetensors",
|
| 226 |
+
"model.layers.20.mlp.gate.e_score_correction_bias": "model-00007-of-00014.safetensors",
|
| 227 |
+
"model.layers.20.mlp.gate.proj.weight": "model-00007-of-00014.safetensors",
|
| 228 |
+
"model.layers.20.mlp.shared_expert.down_proj.weight": "model-00007-of-00014.safetensors",
|
| 229 |
+
"model.layers.20.mlp.shared_expert.gate_proj.weight": "model-00007-of-00014.safetensors",
|
| 230 |
+
"model.layers.20.mlp.shared_expert.up_proj.weight": "model-00007-of-00014.safetensors",
|
| 231 |
+
"model.layers.20.mlp.switch_mlp.down_proj.weight": "model-00007-of-00014.safetensors",
|
| 232 |
+
"model.layers.20.mlp.switch_mlp.gate_proj.weight": "model-00007-of-00014.safetensors",
|
| 233 |
+
"model.layers.20.mlp.switch_mlp.up_proj.weight": "model-00007-of-00014.safetensors",
|
| 234 |
+
"model.layers.20.post_attention_layernorm.weight": "model-00007-of-00014.safetensors",
|
| 235 |
+
"model.layers.20.self_attn.g_proj.weight": "model-00007-of-00014.safetensors",
|
| 236 |
+
"model.layers.20.self_attn.k_norm.weight": "model-00007-of-00014.safetensors",
|
| 237 |
+
"model.layers.20.self_attn.k_proj.weight": "model-00007-of-00014.safetensors",
|
| 238 |
+
"model.layers.20.self_attn.o_proj.weight": "model-00007-of-00014.safetensors",
|
| 239 |
+
"model.layers.20.self_attn.q_norm.weight": "model-00007-of-00014.safetensors",
|
| 240 |
+
"model.layers.20.self_attn.q_proj.weight": "model-00007-of-00014.safetensors",
|
| 241 |
+
"model.layers.20.self_attn.v_proj.weight": "model-00007-of-00014.safetensors",
|
| 242 |
+
"model.layers.21.input_layernorm.weight": "model-00008-of-00014.safetensors",
|
| 243 |
+
"model.layers.21.mlp.gate.e_score_correction_bias": "model-00007-of-00014.safetensors",
|
| 244 |
+
"model.layers.21.mlp.gate.proj.weight": "model-00007-of-00014.safetensors",
|
| 245 |
+
"model.layers.21.mlp.shared_expert.down_proj.weight": "model-00008-of-00014.safetensors",
|
| 246 |
+
"model.layers.21.mlp.shared_expert.gate_proj.weight": "model-00008-of-00014.safetensors",
|
| 247 |
+
"model.layers.21.mlp.shared_expert.up_proj.weight": "model-00008-of-00014.safetensors",
|
| 248 |
+
"model.layers.21.mlp.switch_mlp.down_proj.weight": "model-00008-of-00014.safetensors",
|
| 249 |
+
"model.layers.21.mlp.switch_mlp.gate_proj.weight": "model-00007-of-00014.safetensors",
|
| 250 |
+
"model.layers.21.mlp.switch_mlp.up_proj.weight": "model-00007-of-00014.safetensors",
|
| 251 |
+
"model.layers.21.post_attention_layernorm.weight": "model-00008-of-00014.safetensors",
|
| 252 |
+
"model.layers.21.self_attn.g_proj.weight": "model-00007-of-00014.safetensors",
|
| 253 |
+
"model.layers.21.self_attn.k_norm.weight": "model-00007-of-00014.safetensors",
|
| 254 |
+
"model.layers.21.self_attn.k_proj.weight": "model-00007-of-00014.safetensors",
|
| 255 |
+
"model.layers.21.self_attn.o_proj.weight": "model-00007-of-00014.safetensors",
|
| 256 |
+
"model.layers.21.self_attn.q_norm.weight": "model-00007-of-00014.safetensors",
|
| 257 |
+
"model.layers.21.self_attn.q_proj.weight": "model-00007-of-00014.safetensors",
|
| 258 |
+
"model.layers.21.self_attn.v_proj.weight": "model-00007-of-00014.safetensors",
|
| 259 |
+
"model.layers.22.input_layernorm.weight": "model-00008-of-00014.safetensors",
|
| 260 |
+
"model.layers.22.mlp.gate.e_score_correction_bias": "model-00008-of-00014.safetensors",
|
| 261 |
+
"model.layers.22.mlp.gate.proj.weight": "model-00008-of-00014.safetensors",
|
| 262 |
+
"model.layers.22.mlp.shared_expert.down_proj.weight": "model-00008-of-00014.safetensors",
|
| 263 |
+
"model.layers.22.mlp.shared_expert.gate_proj.weight": "model-00008-of-00014.safetensors",
|
| 264 |
+
"model.layers.22.mlp.shared_expert.up_proj.weight": "model-00008-of-00014.safetensors",
|
| 265 |
+
"model.layers.22.mlp.switch_mlp.down_proj.weight": "model-00008-of-00014.safetensors",
|
| 266 |
+
"model.layers.22.mlp.switch_mlp.gate_proj.weight": "model-00008-of-00014.safetensors",
|
| 267 |
+
"model.layers.22.mlp.switch_mlp.up_proj.weight": "model-00008-of-00014.safetensors",
|
| 268 |
+
"model.layers.22.post_attention_layernorm.weight": "model-00008-of-00014.safetensors",
|
| 269 |
+
"model.layers.22.self_attn.g_proj.weight": "model-00008-of-00014.safetensors",
|
| 270 |
+
"model.layers.22.self_attn.k_norm.weight": "model-00008-of-00014.safetensors",
|
| 271 |
+
"model.layers.22.self_attn.k_proj.weight": "model-00008-of-00014.safetensors",
|
| 272 |
+
"model.layers.22.self_attn.o_proj.weight": "model-00008-of-00014.safetensors",
|
| 273 |
+
"model.layers.22.self_attn.q_norm.weight": "model-00008-of-00014.safetensors",
|
| 274 |
+
"model.layers.22.self_attn.q_proj.weight": "model-00008-of-00014.safetensors",
|
| 275 |
+
"model.layers.22.self_attn.v_proj.weight": "model-00008-of-00014.safetensors",
|
| 276 |
+
"model.layers.23.input_layernorm.weight": "model-00008-of-00014.safetensors",
|
| 277 |
+
"model.layers.23.mlp.gate.e_score_correction_bias": "model-00008-of-00014.safetensors",
|
| 278 |
+
"model.layers.23.mlp.gate.proj.weight": "model-00008-of-00014.safetensors",
|
| 279 |
+
"model.layers.23.mlp.shared_expert.down_proj.weight": "model-00008-of-00014.safetensors",
|
| 280 |
+
"model.layers.23.mlp.shared_expert.gate_proj.weight": "model-00008-of-00014.safetensors",
|
| 281 |
+
"model.layers.23.mlp.shared_expert.up_proj.weight": "model-00008-of-00014.safetensors",
|
| 282 |
+
"model.layers.23.mlp.switch_mlp.down_proj.weight": "model-00008-of-00014.safetensors",
|
| 283 |
+
"model.layers.23.mlp.switch_mlp.gate_proj.weight": "model-00008-of-00014.safetensors",
|
| 284 |
+
"model.layers.23.mlp.switch_mlp.up_proj.weight": "model-00008-of-00014.safetensors",
|
| 285 |
+
"model.layers.23.post_attention_layernorm.weight": "model-00008-of-00014.safetensors",
|
| 286 |
+
"model.layers.23.self_attn.g_proj.weight": "model-00008-of-00014.safetensors",
|
| 287 |
+
"model.layers.23.self_attn.k_norm.weight": "model-00008-of-00014.safetensors",
|
| 288 |
+
"model.layers.23.self_attn.k_proj.weight": "model-00008-of-00014.safetensors",
|
| 289 |
+
"model.layers.23.self_attn.o_proj.weight": "model-00008-of-00014.safetensors",
|
| 290 |
+
"model.layers.23.self_attn.q_norm.weight": "model-00008-of-00014.safetensors",
|
| 291 |
+
"model.layers.23.self_attn.q_proj.weight": "model-00008-of-00014.safetensors",
|
| 292 |
+
"model.layers.23.self_attn.v_proj.weight": "model-00008-of-00014.safetensors",
|
| 293 |
+
"model.layers.24.input_layernorm.weight": "model-00009-of-00014.safetensors",
|
| 294 |
+
"model.layers.24.mlp.gate.e_score_correction_bias": "model-00008-of-00014.safetensors",
|
| 295 |
+
"model.layers.24.mlp.gate.proj.weight": "model-00008-of-00014.safetensors",
|
| 296 |
+
"model.layers.24.mlp.shared_expert.down_proj.weight": "model-00009-of-00014.safetensors",
|
| 297 |
+
"model.layers.24.mlp.shared_expert.gate_proj.weight": "model-00009-of-00014.safetensors",
|
| 298 |
+
"model.layers.24.mlp.shared_expert.up_proj.weight": "model-00009-of-00014.safetensors",
|
| 299 |
+
"model.layers.24.mlp.switch_mlp.down_proj.weight": "model-00009-of-00014.safetensors",
|
| 300 |
+
"model.layers.24.mlp.switch_mlp.gate_proj.weight": "model-00008-of-00014.safetensors",
|
| 301 |
+
"model.layers.24.mlp.switch_mlp.up_proj.weight": "model-00008-of-00014.safetensors",
|
| 302 |
+
"model.layers.24.post_attention_layernorm.weight": "model-00009-of-00014.safetensors",
|
| 303 |
+
"model.layers.24.self_attn.g_proj.weight": "model-00008-of-00014.safetensors",
|
| 304 |
+
"model.layers.24.self_attn.k_norm.weight": "model-00008-of-00014.safetensors",
|
| 305 |
+
"model.layers.24.self_attn.k_proj.weight": "model-00008-of-00014.safetensors",
|
| 306 |
+
"model.layers.24.self_attn.o_proj.weight": "model-00008-of-00014.safetensors",
|
| 307 |
+
"model.layers.24.self_attn.q_norm.weight": "model-00008-of-00014.safetensors",
|
| 308 |
+
"model.layers.24.self_attn.q_proj.weight": "model-00008-of-00014.safetensors",
|
| 309 |
+
"model.layers.24.self_attn.v_proj.weight": "model-00008-of-00014.safetensors",
|
| 310 |
+
"model.layers.25.input_layernorm.weight": "model-00009-of-00014.safetensors",
|
| 311 |
+
"model.layers.25.mlp.gate.e_score_correction_bias": "model-00009-of-00014.safetensors",
|
| 312 |
+
"model.layers.25.mlp.gate.proj.weight": "model-00009-of-00014.safetensors",
|
| 313 |
+
"model.layers.25.mlp.shared_expert.down_proj.weight": "model-00009-of-00014.safetensors",
|
| 314 |
+
"model.layers.25.mlp.shared_expert.gate_proj.weight": "model-00009-of-00014.safetensors",
|
| 315 |
+
"model.layers.25.mlp.shared_expert.up_proj.weight": "model-00009-of-00014.safetensors",
|
| 316 |
+
"model.layers.25.mlp.switch_mlp.down_proj.weight": "model-00009-of-00014.safetensors",
|
| 317 |
+
"model.layers.25.mlp.switch_mlp.gate_proj.weight": "model-00009-of-00014.safetensors",
|
| 318 |
+
"model.layers.25.mlp.switch_mlp.up_proj.weight": "model-00009-of-00014.safetensors",
|
| 319 |
+
"model.layers.25.post_attention_layernorm.weight": "model-00009-of-00014.safetensors",
|
| 320 |
+
"model.layers.25.self_attn.g_proj.weight": "model-00009-of-00014.safetensors",
|
| 321 |
+
"model.layers.25.self_attn.k_norm.weight": "model-00009-of-00014.safetensors",
|
| 322 |
+
"model.layers.25.self_attn.k_proj.weight": "model-00009-of-00014.safetensors",
|
| 323 |
+
"model.layers.25.self_attn.o_proj.weight": "model-00009-of-00014.safetensors",
|
| 324 |
+
"model.layers.25.self_attn.q_norm.weight": "model-00009-of-00014.safetensors",
|
| 325 |
+
"model.layers.25.self_attn.q_proj.weight": "model-00009-of-00014.safetensors",
|
| 326 |
+
"model.layers.25.self_attn.v_proj.weight": "model-00009-of-00014.safetensors",
|
| 327 |
+
"model.layers.26.input_layernorm.weight": "model-00009-of-00014.safetensors",
|
| 328 |
+
"model.layers.26.mlp.gate.e_score_correction_bias": "model-00009-of-00014.safetensors",
|
| 329 |
+
"model.layers.26.mlp.gate.proj.weight": "model-00009-of-00014.safetensors",
|
| 330 |
+
"model.layers.26.mlp.shared_expert.down_proj.weight": "model-00009-of-00014.safetensors",
|
| 331 |
+
"model.layers.26.mlp.shared_expert.gate_proj.weight": "model-00009-of-00014.safetensors",
|
| 332 |
+
"model.layers.26.mlp.shared_expert.up_proj.weight": "model-00009-of-00014.safetensors",
|
| 333 |
+
"model.layers.26.mlp.switch_mlp.down_proj.weight": "model-00009-of-00014.safetensors",
|
| 334 |
+
"model.layers.26.mlp.switch_mlp.gate_proj.weight": "model-00009-of-00014.safetensors",
|
| 335 |
+
"model.layers.26.mlp.switch_mlp.up_proj.weight": "model-00009-of-00014.safetensors",
|
| 336 |
+
"model.layers.26.post_attention_layernorm.weight": "model-00009-of-00014.safetensors",
|
| 337 |
+
"model.layers.26.self_attn.g_proj.weight": "model-00009-of-00014.safetensors",
|
| 338 |
+
"model.layers.26.self_attn.k_norm.weight": "model-00009-of-00014.safetensors",
|
| 339 |
+
"model.layers.26.self_attn.k_proj.weight": "model-00009-of-00014.safetensors",
|
| 340 |
+
"model.layers.26.self_attn.o_proj.weight": "model-00009-of-00014.safetensors",
|
| 341 |
+
"model.layers.26.self_attn.q_norm.weight": "model-00009-of-00014.safetensors",
|
| 342 |
+
"model.layers.26.self_attn.q_proj.weight": "model-00009-of-00014.safetensors",
|
| 343 |
+
"model.layers.26.self_attn.v_proj.weight": "model-00009-of-00014.safetensors",
|
| 344 |
+
"model.layers.27.input_layernorm.weight": "model-00010-of-00014.safetensors",
|
| 345 |
+
"model.layers.27.mlp.gate.e_score_correction_bias": "model-00009-of-00014.safetensors",
|
| 346 |
+
"model.layers.27.mlp.gate.proj.weight": "model-00009-of-00014.safetensors",
|
| 347 |
+
"model.layers.27.mlp.shared_expert.down_proj.weight": "model-00010-of-00014.safetensors",
|
| 348 |
+
"model.layers.27.mlp.shared_expert.gate_proj.weight": "model-00010-of-00014.safetensors",
|
| 349 |
+
"model.layers.27.mlp.shared_expert.up_proj.weight": "model-00010-of-00014.safetensors",
|
| 350 |
+
"model.layers.27.mlp.switch_mlp.down_proj.weight": "model-00010-of-00014.safetensors",
|
| 351 |
+
"model.layers.27.mlp.switch_mlp.gate_proj.weight": "model-00009-of-00014.safetensors",
|
| 352 |
+
"model.layers.27.mlp.switch_mlp.up_proj.weight": "model-00009-of-00014.safetensors",
|
| 353 |
+
"model.layers.27.post_attention_layernorm.weight": "model-00010-of-00014.safetensors",
|
| 354 |
+
"model.layers.27.self_attn.g_proj.weight": "model-00009-of-00014.safetensors",
|
| 355 |
+
"model.layers.27.self_attn.k_norm.weight": "model-00009-of-00014.safetensors",
|
| 356 |
+
"model.layers.27.self_attn.k_proj.weight": "model-00009-of-00014.safetensors",
|
| 357 |
+
"model.layers.27.self_attn.o_proj.weight": "model-00009-of-00014.safetensors",
|
| 358 |
+
"model.layers.27.self_attn.q_norm.weight": "model-00009-of-00014.safetensors",
|
| 359 |
+
"model.layers.27.self_attn.q_proj.weight": "model-00009-of-00014.safetensors",
|
| 360 |
+
"model.layers.27.self_attn.v_proj.weight": "model-00009-of-00014.safetensors",
|
| 361 |
+
"model.layers.28.input_layernorm.weight": "model-00010-of-00014.safetensors",
|
| 362 |
+
"model.layers.28.mlp.gate.e_score_correction_bias": "model-00010-of-00014.safetensors",
|
| 363 |
+
"model.layers.28.mlp.gate.proj.weight": "model-00010-of-00014.safetensors",
|
| 364 |
+
"model.layers.28.mlp.shared_expert.down_proj.weight": "model-00010-of-00014.safetensors",
|
| 365 |
+
"model.layers.28.mlp.shared_expert.gate_proj.weight": "model-00010-of-00014.safetensors",
|
| 366 |
+
"model.layers.28.mlp.shared_expert.up_proj.weight": "model-00010-of-00014.safetensors",
|
| 367 |
+
"model.layers.28.mlp.switch_mlp.down_proj.weight": "model-00010-of-00014.safetensors",
|
| 368 |
+
"model.layers.28.mlp.switch_mlp.gate_proj.weight": "model-00010-of-00014.safetensors",
|
| 369 |
+
"model.layers.28.mlp.switch_mlp.up_proj.weight": "model-00010-of-00014.safetensors",
|
| 370 |
+
"model.layers.28.post_attention_layernorm.weight": "model-00010-of-00014.safetensors",
|
| 371 |
+
"model.layers.28.self_attn.g_proj.weight": "model-00010-of-00014.safetensors",
|
| 372 |
+
"model.layers.28.self_attn.k_norm.weight": "model-00010-of-00014.safetensors",
|
| 373 |
+
"model.layers.28.self_attn.k_proj.weight": "model-00010-of-00014.safetensors",
|
| 374 |
+
"model.layers.28.self_attn.o_proj.weight": "model-00010-of-00014.safetensors",
|
| 375 |
+
"model.layers.28.self_attn.q_norm.weight": "model-00010-of-00014.safetensors",
|
| 376 |
+
"model.layers.28.self_attn.q_proj.weight": "model-00010-of-00014.safetensors",
|
| 377 |
+
"model.layers.28.self_attn.v_proj.weight": "model-00010-of-00014.safetensors",
|
| 378 |
+
"model.layers.29.input_layernorm.weight": "model-00010-of-00014.safetensors",
|
| 379 |
+
"model.layers.29.mlp.gate.e_score_correction_bias": "model-00010-of-00014.safetensors",
|
| 380 |
+
"model.layers.29.mlp.gate.proj.weight": "model-00010-of-00014.safetensors",
|
| 381 |
+
"model.layers.29.mlp.shared_expert.down_proj.weight": "model-00010-of-00014.safetensors",
|
| 382 |
+
"model.layers.29.mlp.shared_expert.gate_proj.weight": "model-00010-of-00014.safetensors",
|
| 383 |
+
"model.layers.29.mlp.shared_expert.up_proj.weight": "model-00010-of-00014.safetensors",
|
| 384 |
+
"model.layers.29.mlp.switch_mlp.down_proj.weight": "model-00010-of-00014.safetensors",
|
| 385 |
+
"model.layers.29.mlp.switch_mlp.gate_proj.weight": "model-00010-of-00014.safetensors",
|
| 386 |
+
"model.layers.29.mlp.switch_mlp.up_proj.weight": "model-00010-of-00014.safetensors",
|
| 387 |
+
"model.layers.29.post_attention_layernorm.weight": "model-00010-of-00014.safetensors",
|
| 388 |
+
"model.layers.29.self_attn.g_proj.weight": "model-00010-of-00014.safetensors",
|
| 389 |
+
"model.layers.29.self_attn.k_norm.weight": "model-00010-of-00014.safetensors",
|
| 390 |
+
"model.layers.29.self_attn.k_proj.weight": "model-00010-of-00014.safetensors",
|
| 391 |
+
"model.layers.29.self_attn.o_proj.weight": "model-00010-of-00014.safetensors",
|
| 392 |
+
"model.layers.29.self_attn.q_norm.weight": "model-00010-of-00014.safetensors",
|
| 393 |
+
"model.layers.29.self_attn.q_proj.weight": "model-00010-of-00014.safetensors",
|
| 394 |
+
"model.layers.29.self_attn.v_proj.weight": "model-00010-of-00014.safetensors",
|
| 395 |
+
"model.layers.3.input_layernorm.weight": "model-00002-of-00014.safetensors",
|
| 396 |
+
"model.layers.3.mlp.gate.e_score_correction_bias": "model-00001-of-00014.safetensors",
|
| 397 |
+
"model.layers.3.mlp.gate.proj.weight": "model-00001-of-00014.safetensors",
|
| 398 |
+
"model.layers.3.mlp.shared_expert.down_proj.weight": "model-00002-of-00014.safetensors",
|
| 399 |
+
"model.layers.3.mlp.shared_expert.gate_proj.weight": "model-00002-of-00014.safetensors",
|
| 400 |
+
"model.layers.3.mlp.shared_expert.up_proj.weight": "model-00002-of-00014.safetensors",
|
| 401 |
+
"model.layers.3.mlp.switch_mlp.down_proj.weight": "model-00002-of-00014.safetensors",
|
| 402 |
+
"model.layers.3.mlp.switch_mlp.gate_proj.weight": "model-00001-of-00014.safetensors",
|
| 403 |
+
"model.layers.3.mlp.switch_mlp.up_proj.weight": "model-00001-of-00014.safetensors",
|
| 404 |
+
"model.layers.3.post_attention_layernorm.weight": "model-00002-of-00014.safetensors",
|
| 405 |
+
"model.layers.3.self_attn.g_proj.weight": "model-00001-of-00014.safetensors",
|
| 406 |
+
"model.layers.3.self_attn.k_norm.weight": "model-00001-of-00014.safetensors",
|
| 407 |
+
"model.layers.3.self_attn.k_proj.weight": "model-00001-of-00014.safetensors",
|
| 408 |
+
"model.layers.3.self_attn.o_proj.weight": "model-00001-of-00014.safetensors",
|
| 409 |
+
"model.layers.3.self_attn.q_norm.weight": "model-00001-of-00014.safetensors",
|
| 410 |
+
"model.layers.3.self_attn.q_proj.weight": "model-00001-of-00014.safetensors",
|
| 411 |
+
"model.layers.3.self_attn.v_proj.weight": "model-00001-of-00014.safetensors",
|
| 412 |
+
"model.layers.30.input_layernorm.weight": "model-00011-of-00014.safetensors",
|
| 413 |
+
"model.layers.30.mlp.gate.e_score_correction_bias": "model-00010-of-00014.safetensors",
|
| 414 |
+
"model.layers.30.mlp.gate.proj.weight": "model-00010-of-00014.safetensors",
|
| 415 |
+
"model.layers.30.mlp.shared_expert.down_proj.weight": "model-00011-of-00014.safetensors",
|
| 416 |
+
"model.layers.30.mlp.shared_expert.gate_proj.weight": "model-00011-of-00014.safetensors",
|
| 417 |
+
"model.layers.30.mlp.shared_expert.up_proj.weight": "model-00011-of-00014.safetensors",
|
| 418 |
+
"model.layers.30.mlp.switch_mlp.down_proj.weight": "model-00011-of-00014.safetensors",
|
| 419 |
+
"model.layers.30.mlp.switch_mlp.gate_proj.weight": "model-00010-of-00014.safetensors",
|
| 420 |
+
"model.layers.30.mlp.switch_mlp.up_proj.weight": "model-00010-of-00014.safetensors",
|
| 421 |
+
"model.layers.30.post_attention_layernorm.weight": "model-00011-of-00014.safetensors",
|
| 422 |
+
"model.layers.30.self_attn.g_proj.weight": "model-00010-of-00014.safetensors",
|
| 423 |
+
"model.layers.30.self_attn.k_norm.weight": "model-00010-of-00014.safetensors",
|
| 424 |
+
"model.layers.30.self_attn.k_proj.weight": "model-00010-of-00014.safetensors",
|
| 425 |
+
"model.layers.30.self_attn.o_proj.weight": "model-00010-of-00014.safetensors",
|
| 426 |
+
"model.layers.30.self_attn.q_norm.weight": "model-00010-of-00014.safetensors",
|
| 427 |
+
"model.layers.30.self_attn.q_proj.weight": "model-00010-of-00014.safetensors",
|
| 428 |
+
"model.layers.30.self_attn.v_proj.weight": "model-00010-of-00014.safetensors",
|
| 429 |
+
"model.layers.31.input_layernorm.weight": "model-00011-of-00014.safetensors",
|
| 430 |
+
"model.layers.31.mlp.gate.e_score_correction_bias": "model-00011-of-00014.safetensors",
|
| 431 |
+
"model.layers.31.mlp.gate.proj.weight": "model-00011-of-00014.safetensors",
|
| 432 |
+
"model.layers.31.mlp.shared_expert.down_proj.weight": "model-00011-of-00014.safetensors",
|
| 433 |
+
"model.layers.31.mlp.shared_expert.gate_proj.weight": "model-00011-of-00014.safetensors",
|
| 434 |
+
"model.layers.31.mlp.shared_expert.up_proj.weight": "model-00011-of-00014.safetensors",
|
| 435 |
+
"model.layers.31.mlp.switch_mlp.down_proj.weight": "model-00011-of-00014.safetensors",
|
| 436 |
+
"model.layers.31.mlp.switch_mlp.gate_proj.weight": "model-00011-of-00014.safetensors",
|
| 437 |
+
"model.layers.31.mlp.switch_mlp.up_proj.weight": "model-00011-of-00014.safetensors",
|
| 438 |
+
"model.layers.31.post_attention_layernorm.weight": "model-00011-of-00014.safetensors",
|
| 439 |
+
"model.layers.31.self_attn.g_proj.weight": "model-00011-of-00014.safetensors",
|
| 440 |
+
"model.layers.31.self_attn.k_norm.weight": "model-00011-of-00014.safetensors",
|
| 441 |
+
"model.layers.31.self_attn.k_proj.weight": "model-00011-of-00014.safetensors",
|
| 442 |
+
"model.layers.31.self_attn.o_proj.weight": "model-00011-of-00014.safetensors",
|
| 443 |
+
"model.layers.31.self_attn.q_norm.weight": "model-00011-of-00014.safetensors",
|
| 444 |
+
"model.layers.31.self_attn.q_proj.weight": "model-00011-of-00014.safetensors",
|
| 445 |
+
"model.layers.31.self_attn.v_proj.weight": "model-00011-of-00014.safetensors",
|
| 446 |
+
"model.layers.32.input_layernorm.weight": "model-00011-of-00014.safetensors",
|
| 447 |
+
"model.layers.32.mlp.gate.e_score_correction_bias": "model-00011-of-00014.safetensors",
|
| 448 |
+
"model.layers.32.mlp.gate.proj.weight": "model-00011-of-00014.safetensors",
|
| 449 |
+
"model.layers.32.mlp.shared_expert.down_proj.weight": "model-00011-of-00014.safetensors",
|
| 450 |
+
"model.layers.32.mlp.shared_expert.gate_proj.weight": "model-00011-of-00014.safetensors",
|
| 451 |
+
"model.layers.32.mlp.shared_expert.up_proj.weight": "model-00011-of-00014.safetensors",
|
| 452 |
+
"model.layers.32.mlp.switch_mlp.down_proj.weight": "model-00011-of-00014.safetensors",
|
| 453 |
+
"model.layers.32.mlp.switch_mlp.gate_proj.weight": "model-00011-of-00014.safetensors",
|
| 454 |
+
"model.layers.32.mlp.switch_mlp.up_proj.weight": "model-00011-of-00014.safetensors",
|
| 455 |
+
"model.layers.32.post_attention_layernorm.weight": "model-00011-of-00014.safetensors",
|
| 456 |
+
"model.layers.32.self_attn.g_proj.weight": "model-00011-of-00014.safetensors",
|
| 457 |
+
"model.layers.32.self_attn.k_norm.weight": "model-00011-of-00014.safetensors",
|
| 458 |
+
"model.layers.32.self_attn.k_proj.weight": "model-00011-of-00014.safetensors",
|
| 459 |
+
"model.layers.32.self_attn.o_proj.weight": "model-00011-of-00014.safetensors",
|
| 460 |
+
"model.layers.32.self_attn.q_norm.weight": "model-00011-of-00014.safetensors",
|
| 461 |
+
"model.layers.32.self_attn.q_proj.weight": "model-00011-of-00014.safetensors",
|
| 462 |
+
"model.layers.32.self_attn.v_proj.weight": "model-00011-of-00014.safetensors",
|
| 463 |
+
"model.layers.33.input_layernorm.weight": "model-00012-of-00014.safetensors",
|
| 464 |
+
"model.layers.33.mlp.gate.e_score_correction_bias": "model-00011-of-00014.safetensors",
|
| 465 |
+
"model.layers.33.mlp.gate.proj.weight": "model-00011-of-00014.safetensors",
|
| 466 |
+
"model.layers.33.mlp.shared_expert.down_proj.weight": "model-00012-of-00014.safetensors",
|
| 467 |
+
"model.layers.33.mlp.shared_expert.gate_proj.weight": "model-00012-of-00014.safetensors",
|
| 468 |
+
"model.layers.33.mlp.shared_expert.up_proj.weight": "model-00012-of-00014.safetensors",
|
| 469 |
+
"model.layers.33.mlp.switch_mlp.down_proj.weight": "model-00012-of-00014.safetensors",
|
| 470 |
+
"model.layers.33.mlp.switch_mlp.gate_proj.weight": "model-00011-of-00014.safetensors",
|
| 471 |
+
"model.layers.33.mlp.switch_mlp.up_proj.weight": "model-00011-of-00014.safetensors",
|
| 472 |
+
"model.layers.33.post_attention_layernorm.weight": "model-00012-of-00014.safetensors",
|
| 473 |
+
"model.layers.33.self_attn.g_proj.weight": "model-00011-of-00014.safetensors",
|
| 474 |
+
"model.layers.33.self_attn.k_norm.weight": "model-00011-of-00014.safetensors",
|
| 475 |
+
"model.layers.33.self_attn.k_proj.weight": "model-00011-of-00014.safetensors",
|
| 476 |
+
"model.layers.33.self_attn.o_proj.weight": "model-00011-of-00014.safetensors",
|
| 477 |
+
"model.layers.33.self_attn.q_norm.weight": "model-00011-of-00014.safetensors",
|
| 478 |
+
"model.layers.33.self_attn.q_proj.weight": "model-00011-of-00014.safetensors",
|
| 479 |
+
"model.layers.33.self_attn.v_proj.weight": "model-00011-of-00014.safetensors",
|
| 480 |
+
"model.layers.34.input_layernorm.weight": "model-00012-of-00014.safetensors",
|
| 481 |
+
"model.layers.34.mlp.gate.e_score_correction_bias": "model-00012-of-00014.safetensors",
|
| 482 |
+
"model.layers.34.mlp.gate.proj.weight": "model-00012-of-00014.safetensors",
|
| 483 |
+
"model.layers.34.mlp.shared_expert.down_proj.weight": "model-00012-of-00014.safetensors",
|
| 484 |
+
"model.layers.34.mlp.shared_expert.gate_proj.weight": "model-00012-of-00014.safetensors",
|
| 485 |
+
"model.layers.34.mlp.shared_expert.up_proj.weight": "model-00012-of-00014.safetensors",
|
| 486 |
+
"model.layers.34.mlp.switch_mlp.down_proj.weight": "model-00012-of-00014.safetensors",
|
| 487 |
+
"model.layers.34.mlp.switch_mlp.gate_proj.weight": "model-00012-of-00014.safetensors",
|
| 488 |
+
"model.layers.34.mlp.switch_mlp.up_proj.weight": "model-00012-of-00014.safetensors",
|
| 489 |
+
"model.layers.34.post_attention_layernorm.weight": "model-00012-of-00014.safetensors",
|
| 490 |
+
"model.layers.34.self_attn.g_proj.weight": "model-00012-of-00014.safetensors",
|
| 491 |
+
"model.layers.34.self_attn.k_norm.weight": "model-00012-of-00014.safetensors",
|
| 492 |
+
"model.layers.34.self_attn.k_proj.weight": "model-00012-of-00014.safetensors",
|
| 493 |
+
"model.layers.34.self_attn.o_proj.weight": "model-00012-of-00014.safetensors",
|
| 494 |
+
"model.layers.34.self_attn.q_norm.weight": "model-00012-of-00014.safetensors",
|
| 495 |
+
"model.layers.34.self_attn.q_proj.weight": "model-00012-of-00014.safetensors",
|
| 496 |
+
"model.layers.34.self_attn.v_proj.weight": "model-00012-of-00014.safetensors",
|
| 497 |
+
"model.layers.35.input_layernorm.weight": "model-00012-of-00014.safetensors",
|
| 498 |
+
"model.layers.35.mlp.gate.e_score_correction_bias": "model-00012-of-00014.safetensors",
|
| 499 |
+
"model.layers.35.mlp.gate.proj.weight": "model-00012-of-00014.safetensors",
|
| 500 |
+
"model.layers.35.mlp.shared_expert.down_proj.weight": "model-00012-of-00014.safetensors",
|
| 501 |
+
"model.layers.35.mlp.shared_expert.gate_proj.weight": "model-00012-of-00014.safetensors",
|
| 502 |
+
"model.layers.35.mlp.shared_expert.up_proj.weight": "model-00012-of-00014.safetensors",
|
| 503 |
+
"model.layers.35.mlp.switch_mlp.down_proj.weight": "model-00012-of-00014.safetensors",
|
| 504 |
+
"model.layers.35.mlp.switch_mlp.gate_proj.weight": "model-00012-of-00014.safetensors",
|
| 505 |
+
"model.layers.35.mlp.switch_mlp.up_proj.weight": "model-00012-of-00014.safetensors",
|
| 506 |
+
"model.layers.35.post_attention_layernorm.weight": "model-00012-of-00014.safetensors",
|
| 507 |
+
"model.layers.35.self_attn.g_proj.weight": "model-00012-of-00014.safetensors",
|
| 508 |
+
"model.layers.35.self_attn.k_norm.weight": "model-00012-of-00014.safetensors",
|
| 509 |
+
"model.layers.35.self_attn.k_proj.weight": "model-00012-of-00014.safetensors",
|
| 510 |
+
"model.layers.35.self_attn.o_proj.weight": "model-00012-of-00014.safetensors",
|
| 511 |
+
"model.layers.35.self_attn.q_norm.weight": "model-00012-of-00014.safetensors",
|
| 512 |
+
"model.layers.35.self_attn.q_proj.weight": "model-00012-of-00014.safetensors",
|
| 513 |
+
"model.layers.35.self_attn.v_proj.weight": "model-00012-of-00014.safetensors",
|
| 514 |
+
"model.layers.36.input_layernorm.weight": "model-00013-of-00014.safetensors",
|
| 515 |
+
"model.layers.36.mlp.gate.e_score_correction_bias": "model-00012-of-00014.safetensors",
|
| 516 |
+
"model.layers.36.mlp.gate.proj.weight": "model-00012-of-00014.safetensors",
|
| 517 |
+
"model.layers.36.mlp.shared_expert.down_proj.weight": "model-00013-of-00014.safetensors",
|
| 518 |
+
"model.layers.36.mlp.shared_expert.gate_proj.weight": "model-00013-of-00014.safetensors",
|
| 519 |
+
"model.layers.36.mlp.shared_expert.up_proj.weight": "model-00013-of-00014.safetensors",
|
| 520 |
+
"model.layers.36.mlp.switch_mlp.down_proj.weight": "model-00013-of-00014.safetensors",
|
| 521 |
+
"model.layers.36.mlp.switch_mlp.gate_proj.weight": "model-00012-of-00014.safetensors",
|
| 522 |
+
"model.layers.36.mlp.switch_mlp.up_proj.weight": "model-00012-of-00014.safetensors",
|
| 523 |
+
"model.layers.36.post_attention_layernorm.weight": "model-00013-of-00014.safetensors",
|
| 524 |
+
"model.layers.36.self_attn.g_proj.weight": "model-00012-of-00014.safetensors",
|
| 525 |
+
"model.layers.36.self_attn.k_norm.weight": "model-00012-of-00014.safetensors",
|
| 526 |
+
"model.layers.36.self_attn.k_proj.weight": "model-00012-of-00014.safetensors",
|
| 527 |
+
"model.layers.36.self_attn.o_proj.weight": "model-00012-of-00014.safetensors",
|
| 528 |
+
"model.layers.36.self_attn.q_norm.weight": "model-00012-of-00014.safetensors",
|
| 529 |
+
"model.layers.36.self_attn.q_proj.weight": "model-00012-of-00014.safetensors",
|
| 530 |
+
"model.layers.36.self_attn.v_proj.weight": "model-00012-of-00014.safetensors",
|
| 531 |
+
"model.layers.37.input_layernorm.weight": "model-00013-of-00014.safetensors",
|
| 532 |
+
"model.layers.37.mlp.gate.e_score_correction_bias": "model-00013-of-00014.safetensors",
|
| 533 |
+
"model.layers.37.mlp.gate.proj.weight": "model-00013-of-00014.safetensors",
|
| 534 |
+
"model.layers.37.mlp.shared_expert.down_proj.weight": "model-00013-of-00014.safetensors",
|
| 535 |
+
"model.layers.37.mlp.shared_expert.gate_proj.weight": "model-00013-of-00014.safetensors",
|
| 536 |
+
"model.layers.37.mlp.shared_expert.up_proj.weight": "model-00013-of-00014.safetensors",
|
| 537 |
+
"model.layers.37.mlp.switch_mlp.down_proj.weight": "model-00013-of-00014.safetensors",
|
| 538 |
+
"model.layers.37.mlp.switch_mlp.gate_proj.weight": "model-00013-of-00014.safetensors",
|
| 539 |
+
"model.layers.37.mlp.switch_mlp.up_proj.weight": "model-00013-of-00014.safetensors",
|
| 540 |
+
"model.layers.37.post_attention_layernorm.weight": "model-00013-of-00014.safetensors",
|
| 541 |
+
"model.layers.37.self_attn.g_proj.weight": "model-00013-of-00014.safetensors",
|
| 542 |
+
"model.layers.37.self_attn.k_norm.weight": "model-00013-of-00014.safetensors",
|
| 543 |
+
"model.layers.37.self_attn.k_proj.weight": "model-00013-of-00014.safetensors",
|
| 544 |
+
"model.layers.37.self_attn.o_proj.weight": "model-00013-of-00014.safetensors",
|
| 545 |
+
"model.layers.37.self_attn.q_norm.weight": "model-00013-of-00014.safetensors",
|
| 546 |
+
"model.layers.37.self_attn.q_proj.weight": "model-00013-of-00014.safetensors",
|
| 547 |
+
"model.layers.37.self_attn.v_proj.weight": "model-00013-of-00014.safetensors",
|
| 548 |
+
"model.layers.38.input_layernorm.weight": "model-00013-of-00014.safetensors",
|
| 549 |
+
"model.layers.38.mlp.gate.e_score_correction_bias": "model-00013-of-00014.safetensors",
|
| 550 |
+
"model.layers.38.mlp.gate.proj.weight": "model-00013-of-00014.safetensors",
|
| 551 |
+
"model.layers.38.mlp.shared_expert.down_proj.weight": "model-00013-of-00014.safetensors",
|
| 552 |
+
"model.layers.38.mlp.shared_expert.gate_proj.weight": "model-00013-of-00014.safetensors",
|
| 553 |
+
"model.layers.38.mlp.shared_expert.up_proj.weight": "model-00013-of-00014.safetensors",
|
| 554 |
+
"model.layers.38.mlp.switch_mlp.down_proj.weight": "model-00013-of-00014.safetensors",
|
| 555 |
+
"model.layers.38.mlp.switch_mlp.gate_proj.weight": "model-00013-of-00014.safetensors",
|
| 556 |
+
"model.layers.38.mlp.switch_mlp.up_proj.weight": "model-00013-of-00014.safetensors",
|
| 557 |
+
"model.layers.38.post_attention_layernorm.weight": "model-00013-of-00014.safetensors",
|
| 558 |
+
"model.layers.38.self_attn.g_proj.weight": "model-00013-of-00014.safetensors",
|
| 559 |
+
"model.layers.38.self_attn.k_norm.weight": "model-00013-of-00014.safetensors",
|
| 560 |
+
"model.layers.38.self_attn.k_proj.weight": "model-00013-of-00014.safetensors",
|
| 561 |
+
"model.layers.38.self_attn.o_proj.weight": "model-00013-of-00014.safetensors",
|
| 562 |
+
"model.layers.38.self_attn.q_norm.weight": "model-00013-of-00014.safetensors",
|
| 563 |
+
"model.layers.38.self_attn.q_proj.weight": "model-00013-of-00014.safetensors",
|
| 564 |
+
"model.layers.38.self_attn.v_proj.weight": "model-00013-of-00014.safetensors",
|
| 565 |
+
"model.layers.39.input_layernorm.weight": "model-00014-of-00014.safetensors",
|
| 566 |
+
"model.layers.39.mlp.gate.e_score_correction_bias": "model-00013-of-00014.safetensors",
|
| 567 |
+
"model.layers.39.mlp.gate.proj.weight": "model-00013-of-00014.safetensors",
|
| 568 |
+
"model.layers.39.mlp.shared_expert.down_proj.weight": "model-00014-of-00014.safetensors",
|
| 569 |
+
"model.layers.39.mlp.shared_expert.gate_proj.weight": "model-00014-of-00014.safetensors",
|
| 570 |
+
"model.layers.39.mlp.shared_expert.up_proj.weight": "model-00014-of-00014.safetensors",
|
| 571 |
+
"model.layers.39.mlp.switch_mlp.down_proj.weight": "model-00014-of-00014.safetensors",
|
| 572 |
+
"model.layers.39.mlp.switch_mlp.gate_proj.weight": "model-00013-of-00014.safetensors",
|
| 573 |
+
"model.layers.39.mlp.switch_mlp.up_proj.weight": "model-00013-of-00014.safetensors",
|
| 574 |
+
"model.layers.39.post_attention_layernorm.weight": "model-00014-of-00014.safetensors",
|
| 575 |
+
"model.layers.39.self_attn.g_proj.weight": "model-00013-of-00014.safetensors",
|
| 576 |
+
"model.layers.39.self_attn.k_norm.weight": "model-00013-of-00014.safetensors",
|
| 577 |
+
"model.layers.39.self_attn.k_proj.weight": "model-00013-of-00014.safetensors",
|
| 578 |
+
"model.layers.39.self_attn.o_proj.weight": "model-00013-of-00014.safetensors",
|
| 579 |
+
"model.layers.39.self_attn.q_norm.weight": "model-00013-of-00014.safetensors",
|
| 580 |
+
"model.layers.39.self_attn.q_proj.weight": "model-00013-of-00014.safetensors",
|
| 581 |
+
"model.layers.39.self_attn.v_proj.weight": "model-00013-of-00014.safetensors",
|
| 582 |
+
"model.layers.4.input_layernorm.weight": "model-00002-of-00014.safetensors",
|
| 583 |
+
"model.layers.4.mlp.gate.e_score_correction_bias": "model-00002-of-00014.safetensors",
|
| 584 |
+
"model.layers.4.mlp.gate.proj.weight": "model-00002-of-00014.safetensors",
|
| 585 |
+
"model.layers.4.mlp.shared_expert.down_proj.weight": "model-00002-of-00014.safetensors",
|
| 586 |
+
"model.layers.4.mlp.shared_expert.gate_proj.weight": "model-00002-of-00014.safetensors",
|
| 587 |
+
"model.layers.4.mlp.shared_expert.up_proj.weight": "model-00002-of-00014.safetensors",
|
| 588 |
+
"model.layers.4.mlp.switch_mlp.down_proj.weight": "model-00002-of-00014.safetensors",
|
| 589 |
+
"model.layers.4.mlp.switch_mlp.gate_proj.weight": "model-00002-of-00014.safetensors",
|
| 590 |
+
"model.layers.4.mlp.switch_mlp.up_proj.weight": "model-00002-of-00014.safetensors",
|
| 591 |
+
"model.layers.4.post_attention_layernorm.weight": "model-00002-of-00014.safetensors",
|
| 592 |
+
"model.layers.4.self_attn.g_proj.weight": "model-00002-of-00014.safetensors",
|
| 593 |
+
"model.layers.4.self_attn.k_norm.weight": "model-00002-of-00014.safetensors",
|
| 594 |
+
"model.layers.4.self_attn.k_proj.weight": "model-00002-of-00014.safetensors",
|
| 595 |
+
"model.layers.4.self_attn.o_proj.weight": "model-00002-of-00014.safetensors",
|
| 596 |
+
"model.layers.4.self_attn.q_norm.weight": "model-00002-of-00014.safetensors",
|
| 597 |
+
"model.layers.4.self_attn.q_proj.weight": "model-00002-of-00014.safetensors",
|
| 598 |
+
"model.layers.4.self_attn.v_proj.weight": "model-00002-of-00014.safetensors",
|
| 599 |
+
"model.layers.5.input_layernorm.weight": "model-00002-of-00014.safetensors",
|
| 600 |
+
"model.layers.5.mlp.gate.e_score_correction_bias": "model-00002-of-00014.safetensors",
|
| 601 |
+
"model.layers.5.mlp.gate.proj.weight": "model-00002-of-00014.safetensors",
|
| 602 |
+
"model.layers.5.mlp.shared_expert.down_proj.weight": "model-00002-of-00014.safetensors",
|
| 603 |
+
"model.layers.5.mlp.shared_expert.gate_proj.weight": "model-00002-of-00014.safetensors",
|
| 604 |
+
"model.layers.5.mlp.shared_expert.up_proj.weight": "model-00002-of-00014.safetensors",
|
| 605 |
+
"model.layers.5.mlp.switch_mlp.down_proj.weight": "model-00002-of-00014.safetensors",
|
| 606 |
+
"model.layers.5.mlp.switch_mlp.gate_proj.weight": "model-00002-of-00014.safetensors",
|
| 607 |
+
"model.layers.5.mlp.switch_mlp.up_proj.weight": "model-00002-of-00014.safetensors",
|
| 608 |
+
"model.layers.5.post_attention_layernorm.weight": "model-00002-of-00014.safetensors",
|
| 609 |
+
"model.layers.5.self_attn.g_proj.weight": "model-00002-of-00014.safetensors",
|
| 610 |
+
"model.layers.5.self_attn.k_norm.weight": "model-00002-of-00014.safetensors",
|
| 611 |
+
"model.layers.5.self_attn.k_proj.weight": "model-00002-of-00014.safetensors",
|
| 612 |
+
"model.layers.5.self_attn.o_proj.weight": "model-00002-of-00014.safetensors",
|
| 613 |
+
"model.layers.5.self_attn.q_norm.weight": "model-00002-of-00014.safetensors",
|
| 614 |
+
"model.layers.5.self_attn.q_proj.weight": "model-00002-of-00014.safetensors",
|
| 615 |
+
"model.layers.5.self_attn.v_proj.weight": "model-00002-of-00014.safetensors",
|
| 616 |
+
"model.layers.6.input_layernorm.weight": "model-00003-of-00014.safetensors",
|
| 617 |
+
"model.layers.6.mlp.gate.e_score_correction_bias": "model-00002-of-00014.safetensors",
|
| 618 |
+
"model.layers.6.mlp.gate.proj.weight": "model-00002-of-00014.safetensors",
|
| 619 |
+
"model.layers.6.mlp.shared_expert.down_proj.weight": "model-00003-of-00014.safetensors",
|
| 620 |
+
"model.layers.6.mlp.shared_expert.gate_proj.weight": "model-00003-of-00014.safetensors",
|
| 621 |
+
"model.layers.6.mlp.shared_expert.up_proj.weight": "model-00003-of-00014.safetensors",
|
| 622 |
+
"model.layers.6.mlp.switch_mlp.down_proj.weight": "model-00003-of-00014.safetensors",
|
| 623 |
+
"model.layers.6.mlp.switch_mlp.gate_proj.weight": "model-00002-of-00014.safetensors",
|
| 624 |
+
"model.layers.6.mlp.switch_mlp.up_proj.weight": "model-00002-of-00014.safetensors",
|
| 625 |
+
"model.layers.6.post_attention_layernorm.weight": "model-00003-of-00014.safetensors",
|
| 626 |
+
"model.layers.6.self_attn.g_proj.weight": "model-00002-of-00014.safetensors",
|
| 627 |
+
"model.layers.6.self_attn.k_norm.weight": "model-00002-of-00014.safetensors",
|
| 628 |
+
"model.layers.6.self_attn.k_proj.weight": "model-00002-of-00014.safetensors",
|
| 629 |
+
"model.layers.6.self_attn.o_proj.weight": "model-00002-of-00014.safetensors",
|
| 630 |
+
"model.layers.6.self_attn.q_norm.weight": "model-00002-of-00014.safetensors",
|
| 631 |
+
"model.layers.6.self_attn.q_proj.weight": "model-00002-of-00014.safetensors",
|
| 632 |
+
"model.layers.6.self_attn.v_proj.weight": "model-00002-of-00014.safetensors",
|
| 633 |
+
"model.layers.7.input_layernorm.weight": "model-00003-of-00014.safetensors",
|
| 634 |
+
"model.layers.7.mlp.gate.e_score_correction_bias": "model-00003-of-00014.safetensors",
|
| 635 |
+
"model.layers.7.mlp.gate.proj.weight": "model-00003-of-00014.safetensors",
|
| 636 |
+
"model.layers.7.mlp.shared_expert.down_proj.weight": "model-00003-of-00014.safetensors",
|
| 637 |
+
"model.layers.7.mlp.shared_expert.gate_proj.weight": "model-00003-of-00014.safetensors",
|
| 638 |
+
"model.layers.7.mlp.shared_expert.up_proj.weight": "model-00003-of-00014.safetensors",
|
| 639 |
+
"model.layers.7.mlp.switch_mlp.down_proj.weight": "model-00003-of-00014.safetensors",
|
| 640 |
+
"model.layers.7.mlp.switch_mlp.gate_proj.weight": "model-00003-of-00014.safetensors",
|
| 641 |
+
"model.layers.7.mlp.switch_mlp.up_proj.weight": "model-00003-of-00014.safetensors",
|
| 642 |
+
"model.layers.7.post_attention_layernorm.weight": "model-00003-of-00014.safetensors",
|
| 643 |
+
"model.layers.7.self_attn.g_proj.weight": "model-00003-of-00014.safetensors",
|
| 644 |
+
"model.layers.7.self_attn.k_norm.weight": "model-00003-of-00014.safetensors",
|
| 645 |
+
"model.layers.7.self_attn.k_proj.weight": "model-00003-of-00014.safetensors",
|
| 646 |
+
"model.layers.7.self_attn.o_proj.weight": "model-00003-of-00014.safetensors",
|
| 647 |
+
"model.layers.7.self_attn.q_norm.weight": "model-00003-of-00014.safetensors",
|
| 648 |
+
"model.layers.7.self_attn.q_proj.weight": "model-00003-of-00014.safetensors",
|
| 649 |
+
"model.layers.7.self_attn.v_proj.weight": "model-00003-of-00014.safetensors",
|
| 650 |
+
"model.layers.8.input_layernorm.weight": "model-00003-of-00014.safetensors",
|
| 651 |
+
"model.layers.8.mlp.gate.e_score_correction_bias": "model-00003-of-00014.safetensors",
|
| 652 |
+
"model.layers.8.mlp.gate.proj.weight": "model-00003-of-00014.safetensors",
|
| 653 |
+
"model.layers.8.mlp.shared_expert.down_proj.weight": "model-00003-of-00014.safetensors",
|
| 654 |
+
"model.layers.8.mlp.shared_expert.gate_proj.weight": "model-00003-of-00014.safetensors",
|
| 655 |
+
"model.layers.8.mlp.shared_expert.up_proj.weight": "model-00003-of-00014.safetensors",
|
| 656 |
+
"model.layers.8.mlp.switch_mlp.down_proj.weight": "model-00003-of-00014.safetensors",
|
| 657 |
+
"model.layers.8.mlp.switch_mlp.gate_proj.weight": "model-00003-of-00014.safetensors",
|
| 658 |
+
"model.layers.8.mlp.switch_mlp.up_proj.weight": "model-00003-of-00014.safetensors",
|
| 659 |
+
"model.layers.8.post_attention_layernorm.weight": "model-00003-of-00014.safetensors",
|
| 660 |
+
"model.layers.8.self_attn.g_proj.weight": "model-00003-of-00014.safetensors",
|
| 661 |
+
"model.layers.8.self_attn.k_norm.weight": "model-00003-of-00014.safetensors",
|
| 662 |
+
"model.layers.8.self_attn.k_proj.weight": "model-00003-of-00014.safetensors",
|
| 663 |
+
"model.layers.8.self_attn.o_proj.weight": "model-00003-of-00014.safetensors",
|
| 664 |
+
"model.layers.8.self_attn.q_norm.weight": "model-00003-of-00014.safetensors",
|
| 665 |
+
"model.layers.8.self_attn.q_proj.weight": "model-00003-of-00014.safetensors",
|
| 666 |
+
"model.layers.8.self_attn.v_proj.weight": "model-00003-of-00014.safetensors",
|
| 667 |
+
"model.layers.9.input_layernorm.weight": "model-00004-of-00014.safetensors",
|
| 668 |
+
"model.layers.9.mlp.gate.e_score_correction_bias": "model-00003-of-00014.safetensors",
|
| 669 |
+
"model.layers.9.mlp.gate.proj.weight": "model-00003-of-00014.safetensors",
|
| 670 |
+
"model.layers.9.mlp.shared_expert.down_proj.weight": "model-00004-of-00014.safetensors",
|
| 671 |
+
"model.layers.9.mlp.shared_expert.gate_proj.weight": "model-00004-of-00014.safetensors",
|
| 672 |
+
"model.layers.9.mlp.shared_expert.up_proj.weight": "model-00004-of-00014.safetensors",
|
| 673 |
+
"model.layers.9.mlp.switch_mlp.down_proj.weight": "model-00004-of-00014.safetensors",
|
| 674 |
+
"model.layers.9.mlp.switch_mlp.gate_proj.weight": "model-00003-of-00014.safetensors",
|
| 675 |
+
"model.layers.9.mlp.switch_mlp.up_proj.weight": "model-00003-of-00014.safetensors",
|
| 676 |
+
"model.layers.9.post_attention_layernorm.weight": "model-00004-of-00014.safetensors",
|
| 677 |
+
"model.layers.9.self_attn.g_proj.weight": "model-00003-of-00014.safetensors",
|
| 678 |
+
"model.layers.9.self_attn.k_norm.weight": "model-00003-of-00014.safetensors",
|
| 679 |
+
"model.layers.9.self_attn.k_proj.weight": "model-00003-of-00014.safetensors",
|
| 680 |
+
"model.layers.9.self_attn.o_proj.weight": "model-00003-of-00014.safetensors",
|
| 681 |
+
"model.layers.9.self_attn.q_norm.weight": "model-00003-of-00014.safetensors",
|
| 682 |
+
"model.layers.9.self_attn.q_proj.weight": "model-00003-of-00014.safetensors",
|
| 683 |
+
"model.layers.9.self_attn.v_proj.weight": "model-00003-of-00014.safetensors",
|
| 684 |
+
"model.norm.weight": "model-00014-of-00014.safetensors"
|
| 685 |
+
}
|
| 686 |
+
}
|
modeling_laguna.py
ADDED
|
@@ -0,0 +1,755 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Copyright 2026 Poolside and the HuggingFace Inc. team. All rights reserved.
|
| 2 |
+
#
|
| 3 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 4 |
+
# you may not use this file except in compliance with the License.
|
| 5 |
+
# You may obtain a copy of the License at
|
| 6 |
+
#
|
| 7 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 8 |
+
#
|
| 9 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 10 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 11 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 12 |
+
# See the License for the specific language governing permissions and
|
| 13 |
+
# limitations under the License.
|
| 14 |
+
|
| 15 |
+
from collections.abc import Callable
|
| 16 |
+
from typing import Optional
|
| 17 |
+
|
| 18 |
+
import torch
|
| 19 |
+
import torch.nn.functional as F
|
| 20 |
+
from torch import nn
|
| 21 |
+
|
| 22 |
+
from transformers import initialization as init
|
| 23 |
+
from transformers.activations import ACT2FN
|
| 24 |
+
from transformers.cache_utils import Cache, DynamicCache
|
| 25 |
+
from transformers.generation import GenerationMixin
|
| 26 |
+
from transformers.integrations import use_experts_implementation, use_kernel_forward_from_hub, use_kernelized_func
|
| 27 |
+
from transformers.masking_utils import create_causal_mask, create_sliding_window_causal_mask
|
| 28 |
+
from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
|
| 29 |
+
from transformers.modeling_layers import GradientCheckpointingLayer
|
| 30 |
+
from transformers.modeling_outputs import MoeCausalLMOutputWithPast, MoeModelOutputWithPast
|
| 31 |
+
from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
|
| 32 |
+
from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
|
| 33 |
+
from transformers.processing_utils import Unpack
|
| 34 |
+
from transformers.utils import auto_docstring, can_return_tuple
|
| 35 |
+
from transformers.utils.generic import TransformersKwargs, maybe_autocast
|
| 36 |
+
from transformers.utils.output_capturing import OutputRecorder, capture_outputs
|
| 37 |
+
from .configuration_laguna import LagunaConfig
|
| 38 |
+
|
| 39 |
+
|
| 40 |
+
@use_kernel_forward_from_hub("RMSNorm")
|
| 41 |
+
class LagunaRMSNorm(nn.Module):
|
| 42 |
+
def __init__(self, hidden_size, eps: float = 1e-6) -> None:
|
| 43 |
+
"""
|
| 44 |
+
LagunaRMSNorm is equivalent to T5LayerNorm
|
| 45 |
+
"""
|
| 46 |
+
super().__init__()
|
| 47 |
+
self.weight = nn.Parameter(torch.ones(hidden_size))
|
| 48 |
+
self.variance_epsilon = eps
|
| 49 |
+
|
| 50 |
+
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
| 51 |
+
input_dtype = hidden_states.dtype
|
| 52 |
+
hidden_states = hidden_states.to(torch.float32)
|
| 53 |
+
variance = hidden_states.pow(2).mean(-1, keepdim=True)
|
| 54 |
+
hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
|
| 55 |
+
return self.weight * hidden_states.to(input_dtype)
|
| 56 |
+
|
| 57 |
+
def extra_repr(self):
|
| 58 |
+
return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
|
| 59 |
+
|
| 60 |
+
|
| 61 |
+
class LagunaRotaryEmbedding(nn.Module):
|
| 62 |
+
inv_freq: torch.Tensor # fix linting for `register_buffer`
|
| 63 |
+
|
| 64 |
+
def __init__(self, config: LagunaConfig, device=None, layer_type=None):
|
| 65 |
+
super().__init__()
|
| 66 |
+
self.max_seq_len_cached = config.max_position_embeddings
|
| 67 |
+
self.original_max_seq_len = config.max_position_embeddings
|
| 68 |
+
|
| 69 |
+
self.config = config
|
| 70 |
+
|
| 71 |
+
self.layer_types = list(set(config.layer_types))
|
| 72 |
+
self.rope_type = {}
|
| 73 |
+
for layer_type in self.layer_types:
|
| 74 |
+
rope_params = self.config.rope_parameters[layer_type]
|
| 75 |
+
if rope_params is None:
|
| 76 |
+
continue
|
| 77 |
+
|
| 78 |
+
self.rope_type[layer_type] = rope_params["rope_type"]
|
| 79 |
+
rope_init_fn: Callable = self.compute_default_rope_parameters
|
| 80 |
+
if self.rope_type[layer_type] != "default":
|
| 81 |
+
rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type[layer_type]]
|
| 82 |
+
curr_inv_freq, curr_attention_scaling = rope_init_fn(self.config, device, layer_type=layer_type)
|
| 83 |
+
self.register_buffer(f"{layer_type}_inv_freq", curr_inv_freq, persistent=False)
|
| 84 |
+
self.register_buffer(f"{layer_type}_original_inv_freq", curr_inv_freq.clone(), persistent=False)
|
| 85 |
+
setattr(self, f"{layer_type}_attention_scaling", curr_attention_scaling)
|
| 86 |
+
|
| 87 |
+
@staticmethod
|
| 88 |
+
def compute_default_rope_parameters(
|
| 89 |
+
config: LagunaConfig | None = None,
|
| 90 |
+
device: Optional["torch.device"] = None,
|
| 91 |
+
seq_len: int | None = None,
|
| 92 |
+
layer_type: str | None = None,
|
| 93 |
+
) -> tuple["torch.Tensor", float]:
|
| 94 |
+
"""
|
| 95 |
+
Computes the inverse frequencies according to the original RoPE implementation
|
| 96 |
+
Args:
|
| 97 |
+
config ([`~transformers.PreTrainedConfig`]):
|
| 98 |
+
The model configuration.
|
| 99 |
+
device (`torch.device`):
|
| 100 |
+
The device to use for initialization of the inverse frequencies.
|
| 101 |
+
seq_len (`int`, *optional*):
|
| 102 |
+
The current sequence length. Unused for this type of RoPE.
|
| 103 |
+
layer_type (`str`, *optional*):
|
| 104 |
+
The current layer type if the model has different RoPE parameters per type.
|
| 105 |
+
Should not be used unless `config.layer_types is not None`
|
| 106 |
+
Returns:
|
| 107 |
+
Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
|
| 108 |
+
post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE).
|
| 109 |
+
"""
|
| 110 |
+
base = config.rope_parameters[layer_type]["rope_theta"]
|
| 111 |
+
# key difference to gemma3: partial rope
|
| 112 |
+
partial_rotary_factor = config.rope_parameters[layer_type].get("partial_rotary_factor", 1.0)
|
| 113 |
+
head_dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads
|
| 114 |
+
dim = int(head_dim * partial_rotary_factor)
|
| 115 |
+
|
| 116 |
+
attention_factor = 1.0 # Unused in this type of RoPE
|
| 117 |
+
|
| 118 |
+
# Compute the inverse frequencies
|
| 119 |
+
inv_freq = 1.0 / (
|
| 120 |
+
base ** (torch.arange(0, dim, 2, dtype=torch.int64).to(device=device, dtype=torch.float) / dim)
|
| 121 |
+
)
|
| 122 |
+
return inv_freq, attention_factor
|
| 123 |
+
|
| 124 |
+
@torch.no_grad()
|
| 125 |
+
@dynamic_rope_update # power user: used with advanced RoPE types (e.g. dynamic rope)
|
| 126 |
+
def forward(self, x, position_ids, layer_type=None):
|
| 127 |
+
inv_freq = getattr(self, f"{layer_type}_inv_freq")
|
| 128 |
+
attention_scaling = getattr(self, f"{layer_type}_attention_scaling")
|
| 129 |
+
|
| 130 |
+
inv_freq_expanded = inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
|
| 131 |
+
position_ids_expanded = position_ids[:, None, :].float()
|
| 132 |
+
|
| 133 |
+
device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
|
| 134 |
+
with maybe_autocast(device_type=device_type, enabled=False): # Force float32
|
| 135 |
+
freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
|
| 136 |
+
emb = torch.cat((freqs, freqs), dim=-1)
|
| 137 |
+
cos = emb.cos() * attention_scaling
|
| 138 |
+
sin = emb.sin() * attention_scaling
|
| 139 |
+
|
| 140 |
+
return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
|
| 141 |
+
|
| 142 |
+
|
| 143 |
+
class LagunaMLP(nn.Module):
|
| 144 |
+
def __init__(self, config, intermediate_size=None):
|
| 145 |
+
super().__init__()
|
| 146 |
+
self.config = config
|
| 147 |
+
self.hidden_size = config.hidden_size
|
| 148 |
+
self.intermediate_size = config.intermediate_size if intermediate_size is None else intermediate_size
|
| 149 |
+
self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
|
| 150 |
+
self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
|
| 151 |
+
self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
|
| 152 |
+
self.act_fn = ACT2FN[config.hidden_act]
|
| 153 |
+
|
| 154 |
+
def forward(self, x):
|
| 155 |
+
down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
|
| 156 |
+
return down_proj
|
| 157 |
+
|
| 158 |
+
|
| 159 |
+
class LagunaTopKRouter(nn.Module):
|
| 160 |
+
def __init__(self, config):
|
| 161 |
+
super().__init__()
|
| 162 |
+
self.top_k = config.num_experts_per_tok
|
| 163 |
+
self.num_experts = config.num_experts
|
| 164 |
+
self.hidden_dim = config.hidden_size
|
| 165 |
+
self.weight = nn.Parameter(torch.zeros(self.num_experts, self.hidden_dim))
|
| 166 |
+
self.e_score_correction_bias = nn.Parameter(torch.zeros(config.num_experts), requires_grad=False)
|
| 167 |
+
self.router_logit_softcapping = config.moe_router_logit_softcapping
|
| 168 |
+
|
| 169 |
+
def forward(
|
| 170 |
+
self,
|
| 171 |
+
hidden_states: torch.Tensor,
|
| 172 |
+
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
|
| 173 |
+
hidden_states = hidden_states.reshape(-1, self.hidden_dim)
|
| 174 |
+
router_logits = F.linear(hidden_states, self.weight).float()
|
| 175 |
+
# Optional logits softcapping
|
| 176 |
+
if self.router_logit_softcapping > 0.0:
|
| 177 |
+
router_logits = torch.tanh(router_logits / self.router_logit_softcapping) * self.router_logit_softcapping
|
| 178 |
+
# Sigmoid instead of softmax normalization
|
| 179 |
+
routing_scores = torch.sigmoid(router_logits)
|
| 180 |
+
|
| 181 |
+
scores_for_selection = routing_scores + self.e_score_correction_bias.to(routing_scores.dtype)
|
| 182 |
+
_, selected_experts = torch.topk(scores_for_selection, self.top_k, dim=-1)
|
| 183 |
+
routing_weights = routing_scores.gather(-1, selected_experts)
|
| 184 |
+
routing_weights = routing_weights / routing_weights.sum(dim=-1, keepdim=True)
|
| 185 |
+
routing_weights = routing_weights.to(hidden_states.dtype)
|
| 186 |
+
|
| 187 |
+
return router_logits, routing_weights, selected_experts
|
| 188 |
+
|
| 189 |
+
|
| 190 |
+
@use_experts_implementation
|
| 191 |
+
class LagunaExperts(nn.Module):
|
| 192 |
+
"""Collection of expert weights stored as 3D tensors."""
|
| 193 |
+
|
| 194 |
+
def __init__(self, config):
|
| 195 |
+
super().__init__()
|
| 196 |
+
self.num_experts = config.num_experts
|
| 197 |
+
self.hidden_dim = config.hidden_size
|
| 198 |
+
self.intermediate_dim = config.moe_intermediate_size
|
| 199 |
+
self.gate_up_proj = nn.Parameter(torch.empty(self.num_experts, 2 * self.intermediate_dim, self.hidden_dim))
|
| 200 |
+
self.down_proj = nn.Parameter(torch.empty(self.num_experts, self.hidden_dim, self.intermediate_dim))
|
| 201 |
+
self.act_fn = ACT2FN[config.hidden_act]
|
| 202 |
+
|
| 203 |
+
def forward(
|
| 204 |
+
self,
|
| 205 |
+
hidden_states: torch.Tensor,
|
| 206 |
+
top_k_index: torch.Tensor,
|
| 207 |
+
top_k_weights: torch.Tensor,
|
| 208 |
+
) -> torch.Tensor:
|
| 209 |
+
final_hidden_states = torch.zeros_like(hidden_states)
|
| 210 |
+
with torch.no_grad():
|
| 211 |
+
expert_mask = torch.nn.functional.one_hot(top_k_index, num_classes=self.num_experts)
|
| 212 |
+
expert_mask = expert_mask.permute(2, 1, 0)
|
| 213 |
+
expert_hit = torch.greater(expert_mask.sum(dim=(-1, -2)), 0).nonzero()
|
| 214 |
+
|
| 215 |
+
for expert_idx in expert_hit:
|
| 216 |
+
expert_idx = expert_idx[0]
|
| 217 |
+
if expert_idx == self.num_experts:
|
| 218 |
+
continue
|
| 219 |
+
top_k_pos, token_idx = torch.where(expert_mask[expert_idx])
|
| 220 |
+
current_state = hidden_states[token_idx]
|
| 221 |
+
gate, up = nn.functional.linear(current_state, self.gate_up_proj[expert_idx]).chunk(2, dim=-1)
|
| 222 |
+
current_hidden_states = self.act_fn(gate) * up
|
| 223 |
+
current_hidden_states = nn.functional.linear(current_hidden_states, self.down_proj[expert_idx])
|
| 224 |
+
current_hidden_states = current_hidden_states * top_k_weights[token_idx, top_k_pos, None]
|
| 225 |
+
final_hidden_states.index_add_(0, token_idx, current_hidden_states.to(final_hidden_states.dtype))
|
| 226 |
+
|
| 227 |
+
return final_hidden_states
|
| 228 |
+
|
| 229 |
+
|
| 230 |
+
class LagunaSparseMoeBlock(nn.Module):
|
| 231 |
+
def __init__(self, config: LagunaConfig):
|
| 232 |
+
super().__init__()
|
| 233 |
+
self.experts = LagunaExperts(config)
|
| 234 |
+
self.gate = LagunaTopKRouter(config)
|
| 235 |
+
self.shared_experts = LagunaMLP(config, intermediate_size=config.shared_expert_intermediate_size)
|
| 236 |
+
self.routed_scaling_factor = config.moe_routed_scaling_factor
|
| 237 |
+
|
| 238 |
+
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
| 239 |
+
batch_size, sequence_length, hidden_dim = hidden_states.shape
|
| 240 |
+
hidden_states = hidden_states.view(-1, hidden_dim)
|
| 241 |
+
shared_output = self.shared_experts(hidden_states)
|
| 242 |
+
|
| 243 |
+
_, routing_weights, selected_experts = self.gate(hidden_states)
|
| 244 |
+
hidden_states = self.experts(hidden_states, selected_experts, routing_weights)
|
| 245 |
+
# Additional scaling
|
| 246 |
+
hidden_states = hidden_states * self.routed_scaling_factor
|
| 247 |
+
hidden_states = hidden_states + shared_output
|
| 248 |
+
|
| 249 |
+
hidden_states = hidden_states.reshape(batch_size, sequence_length, hidden_dim)
|
| 250 |
+
return hidden_states
|
| 251 |
+
|
| 252 |
+
|
| 253 |
+
def rotate_half(x):
|
| 254 |
+
"""Rotates half the hidden dims of the input."""
|
| 255 |
+
x1 = x[..., : x.shape[-1] // 2]
|
| 256 |
+
x2 = x[..., x.shape[-1] // 2 :]
|
| 257 |
+
return torch.cat((-x2, x1), dim=-1)
|
| 258 |
+
|
| 259 |
+
|
| 260 |
+
# Adapted from transformers.models.glm.modular_glm.apply_rotary_pos_emb
|
| 261 |
+
def apply_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=1):
|
| 262 |
+
"""Applies Rotary Position Embedding to the query and key tensors.
|
| 263 |
+
|
| 264 |
+
Removes the interleaving of cos and sin from GLM
|
| 265 |
+
|
| 266 |
+
Args:
|
| 267 |
+
q (`torch.Tensor`): The query tensor.
|
| 268 |
+
k (`torch.Tensor`): The key tensor.
|
| 269 |
+
cos (`torch.Tensor`): The cosine part of the rotary embedding.
|
| 270 |
+
sin (`torch.Tensor`): The sine part of the rotary embedding.
|
| 271 |
+
unsqueeze_dim (`int`, *optional*, defaults to 1):
|
| 272 |
+
The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
|
| 273 |
+
sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
|
| 274 |
+
that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
|
| 275 |
+
k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
|
| 276 |
+
cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
|
| 277 |
+
the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
|
| 278 |
+
Returns:
|
| 279 |
+
`tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
|
| 280 |
+
"""
|
| 281 |
+
cos = cos.unsqueeze(unsqueeze_dim)
|
| 282 |
+
sin = sin.unsqueeze(unsqueeze_dim)
|
| 283 |
+
|
| 284 |
+
# Keep half or full tensor for later concatenation
|
| 285 |
+
rotary_dim = cos.shape[-1]
|
| 286 |
+
q_rot, q_pass = q[..., :rotary_dim], q[..., rotary_dim:]
|
| 287 |
+
k_rot, k_pass = k[..., :rotary_dim], k[..., rotary_dim:]
|
| 288 |
+
|
| 289 |
+
# Apply rotary embeddings on the first half or full tensor
|
| 290 |
+
q_embed = (q_rot * cos) + (rotate_half(q_rot) * sin)
|
| 291 |
+
k_embed = (k_rot * cos) + (rotate_half(k_rot) * sin)
|
| 292 |
+
|
| 293 |
+
# Concatenate back to full shape
|
| 294 |
+
q_embed = torch.cat([q_embed, q_pass], dim=-1)
|
| 295 |
+
k_embed = torch.cat([k_embed, k_pass], dim=-1)
|
| 296 |
+
return q_embed, k_embed
|
| 297 |
+
|
| 298 |
+
|
| 299 |
+
def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
|
| 300 |
+
"""
|
| 301 |
+
This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
|
| 302 |
+
num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
|
| 303 |
+
"""
|
| 304 |
+
batch, num_key_value_heads, slen, head_dim = hidden_states.shape
|
| 305 |
+
if n_rep == 1:
|
| 306 |
+
return hidden_states
|
| 307 |
+
hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
|
| 308 |
+
return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
|
| 309 |
+
|
| 310 |
+
|
| 311 |
+
def eager_attention_forward(
|
| 312 |
+
module: nn.Module,
|
| 313 |
+
query: torch.Tensor,
|
| 314 |
+
key: torch.Tensor,
|
| 315 |
+
value: torch.Tensor,
|
| 316 |
+
attention_mask: torch.Tensor | None,
|
| 317 |
+
scaling: float,
|
| 318 |
+
dropout: float = 0.0,
|
| 319 |
+
**kwargs: Unpack[TransformersKwargs],
|
| 320 |
+
):
|
| 321 |
+
key_states = repeat_kv(key, module.num_key_value_groups)
|
| 322 |
+
value_states = repeat_kv(value, module.num_key_value_groups)
|
| 323 |
+
|
| 324 |
+
attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
|
| 325 |
+
if attention_mask is not None:
|
| 326 |
+
attn_weights = attn_weights + attention_mask
|
| 327 |
+
|
| 328 |
+
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
|
| 329 |
+
attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
|
| 330 |
+
attn_output = torch.matmul(attn_weights, value_states)
|
| 331 |
+
attn_output = attn_output.transpose(1, 2).contiguous()
|
| 332 |
+
|
| 333 |
+
return attn_output, attn_weights
|
| 334 |
+
|
| 335 |
+
|
| 336 |
+
@use_kernelized_func(apply_rotary_pos_emb)
|
| 337 |
+
class LagunaAttention(nn.Module):
|
| 338 |
+
"""Afmoe-style SWA/GQA attention with Laguna-specific gating and per-layer head count."""
|
| 339 |
+
|
| 340 |
+
def __init__(self, config: LagunaConfig, layer_idx: int, num_heads: int):
|
| 341 |
+
super().__init__()
|
| 342 |
+
# Number of heads is controlled via `config.num_attention_heads_per_layer` which is passed from the parent for the specific layer
|
| 343 |
+
self.num_heads = num_heads
|
| 344 |
+
self.config = config
|
| 345 |
+
self.layer_idx = layer_idx
|
| 346 |
+
self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
|
| 347 |
+
self.num_key_value_groups = self.num_heads // config.num_key_value_heads
|
| 348 |
+
self.scaling = self.head_dim**-0.5
|
| 349 |
+
self.attention_dropout = config.attention_dropout
|
| 350 |
+
self.is_causal = True
|
| 351 |
+
|
| 352 |
+
# Per-layer head count: rebuild q_proj and o_proj using self.num_heads (parent uses config.num_attention_heads).
|
| 353 |
+
self.q_proj = nn.Linear(config.hidden_size, self.num_heads * self.head_dim, bias=config.attention_bias)
|
| 354 |
+
self.k_proj = nn.Linear(
|
| 355 |
+
config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
|
| 356 |
+
)
|
| 357 |
+
self.v_proj = nn.Linear(
|
| 358 |
+
config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
|
| 359 |
+
)
|
| 360 |
+
self.o_proj = nn.Linear(self.num_heads * self.head_dim, config.hidden_size, bias=config.attention_bias)
|
| 361 |
+
# Parent LlamaAttention already sets: layer_idx, num_heads, num_key_value_heads, num_key_value_groups, head_dim
|
| 362 |
+
# We only add Laguna-specific attributes
|
| 363 |
+
self.is_local_attention = config.layer_types[layer_idx] == "sliding_attention"
|
| 364 |
+
self.sliding_window = config.sliding_window if self.is_local_attention else None
|
| 365 |
+
|
| 366 |
+
self.q_norm = LagunaRMSNorm(self.head_dim, eps=config.rms_norm_eps)
|
| 367 |
+
self.k_norm = LagunaRMSNorm(self.head_dim, eps=config.rms_norm_eps)
|
| 368 |
+
self.g_proj = nn.Linear(config.hidden_size, self.num_heads, bias=False)
|
| 369 |
+
|
| 370 |
+
def forward(
|
| 371 |
+
self,
|
| 372 |
+
hidden_states: torch.Tensor,
|
| 373 |
+
position_embeddings: tuple[torch.Tensor, torch.Tensor],
|
| 374 |
+
attention_mask: torch.Tensor | None,
|
| 375 |
+
past_key_values: Cache | None = None,
|
| 376 |
+
**kwargs: Unpack[FlashAttentionKwargs],
|
| 377 |
+
) -> tuple[torch.Tensor, torch.Tensor | None]:
|
| 378 |
+
input_shape = hidden_states.shape[:-1]
|
| 379 |
+
hidden_shape = (*input_shape, -1, self.head_dim)
|
| 380 |
+
|
| 381 |
+
query_states = self.q_proj(hidden_states).view(hidden_shape)
|
| 382 |
+
key_states = self.k_proj(hidden_states).view(hidden_shape)
|
| 383 |
+
value_states = self.v_proj(hidden_states).view(hidden_shape)
|
| 384 |
+
|
| 385 |
+
query_states = self.q_norm(query_states).transpose(1, 2)
|
| 386 |
+
key_states = self.k_norm(key_states).transpose(1, 2)
|
| 387 |
+
value_states = value_states.transpose(1, 2)
|
| 388 |
+
|
| 389 |
+
cos, sin = position_embeddings
|
| 390 |
+
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
|
| 391 |
+
|
| 392 |
+
if past_key_values is not None:
|
| 393 |
+
key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx)
|
| 394 |
+
|
| 395 |
+
attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
|
| 396 |
+
self.config._attn_implementation, eager_attention_forward
|
| 397 |
+
)
|
| 398 |
+
attn_output, attn_weights = attention_interface(
|
| 399 |
+
self,
|
| 400 |
+
query_states,
|
| 401 |
+
key_states,
|
| 402 |
+
value_states,
|
| 403 |
+
attention_mask,
|
| 404 |
+
dropout=0.0 if not self.training else self.attention_dropout,
|
| 405 |
+
scaling=self.scaling,
|
| 406 |
+
sliding_window=self.sliding_window,
|
| 407 |
+
**kwargs,
|
| 408 |
+
)
|
| 409 |
+
|
| 410 |
+
attn_output = attn_output.reshape(*input_shape, -1).contiguous()
|
| 411 |
+
|
| 412 |
+
gate = F.softplus(self.g_proj(hidden_states).float()).to(attn_output.dtype)
|
| 413 |
+
attn_output = (attn_output.view(*input_shape, -1, self.head_dim) * gate.unsqueeze(-1)).view(*input_shape, -1)
|
| 414 |
+
|
| 415 |
+
attn_output = self.o_proj(attn_output)
|
| 416 |
+
return attn_output, attn_weights
|
| 417 |
+
|
| 418 |
+
|
| 419 |
+
class LagunaDecoderLayer(GradientCheckpointingLayer):
|
| 420 |
+
def __init__(self, config: LagunaConfig, layer_idx: int):
|
| 421 |
+
super().__init__()
|
| 422 |
+
self.hidden_size = config.hidden_size
|
| 423 |
+
self.self_attn = LagunaAttention(config, layer_idx, config.num_attention_heads_per_layer[layer_idx])
|
| 424 |
+
if config.mlp_layer_types[layer_idx] == "sparse":
|
| 425 |
+
self.mlp = LagunaSparseMoeBlock(config)
|
| 426 |
+
else:
|
| 427 |
+
self.mlp = LagunaMLP(config, intermediate_size=config.intermediate_size)
|
| 428 |
+
self.input_layernorm = LagunaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 429 |
+
self.post_attention_layernorm = LagunaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 430 |
+
|
| 431 |
+
def forward(
|
| 432 |
+
self,
|
| 433 |
+
hidden_states: torch.Tensor,
|
| 434 |
+
attention_mask: torch.Tensor | None = None,
|
| 435 |
+
position_ids: torch.LongTensor | None = None,
|
| 436 |
+
past_key_values: Cache | None = None,
|
| 437 |
+
use_cache: bool | None = False,
|
| 438 |
+
position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
|
| 439 |
+
**kwargs: Unpack[TransformersKwargs],
|
| 440 |
+
) -> torch.Tensor:
|
| 441 |
+
residual = hidden_states
|
| 442 |
+
hidden_states = self.input_layernorm(hidden_states)
|
| 443 |
+
# Self Attention
|
| 444 |
+
hidden_states, _ = self.self_attn(
|
| 445 |
+
hidden_states=hidden_states,
|
| 446 |
+
attention_mask=attention_mask,
|
| 447 |
+
position_ids=position_ids,
|
| 448 |
+
past_key_values=past_key_values,
|
| 449 |
+
use_cache=use_cache,
|
| 450 |
+
position_embeddings=position_embeddings,
|
| 451 |
+
**kwargs,
|
| 452 |
+
)
|
| 453 |
+
hidden_states = residual + hidden_states
|
| 454 |
+
|
| 455 |
+
# Fully Connected
|
| 456 |
+
residual = hidden_states
|
| 457 |
+
hidden_states = self.post_attention_layernorm(hidden_states)
|
| 458 |
+
hidden_states = self.mlp(hidden_states)
|
| 459 |
+
hidden_states = residual + hidden_states
|
| 460 |
+
return hidden_states
|
| 461 |
+
|
| 462 |
+
|
| 463 |
+
@auto_docstring
|
| 464 |
+
class LagunaPreTrainedModel(PreTrainedModel):
|
| 465 |
+
config: LagunaConfig
|
| 466 |
+
base_model_prefix = "model"
|
| 467 |
+
supports_gradient_checkpointing = True
|
| 468 |
+
_no_split_modules = ["LagunaDecoderLayer"]
|
| 469 |
+
_skip_keys_device_placement = ["past_key_values"]
|
| 470 |
+
_supports_flash_attn = True
|
| 471 |
+
_supports_sdpa = True
|
| 472 |
+
_supports_flex_attn = True
|
| 473 |
+
|
| 474 |
+
_can_compile_fullgraph = True
|
| 475 |
+
_supports_attention_backend = True
|
| 476 |
+
_can_record_outputs = {
|
| 477 |
+
"router_logits": OutputRecorder(LagunaTopKRouter, index=0),
|
| 478 |
+
"hidden_states": LagunaDecoderLayer,
|
| 479 |
+
"attentions": LagunaAttention,
|
| 480 |
+
}
|
| 481 |
+
|
| 482 |
+
@torch.no_grad()
|
| 483 |
+
def _init_weights(self, module):
|
| 484 |
+
super()._init_weights(module)
|
| 485 |
+
std = self.config.initializer_range
|
| 486 |
+
if isinstance(module, LagunaExperts):
|
| 487 |
+
init.normal_(module.gate_up_proj, mean=0.0, std=std)
|
| 488 |
+
init.normal_(module.down_proj, mean=0.0, std=std)
|
| 489 |
+
elif isinstance(module, LagunaTopKRouter):
|
| 490 |
+
init.normal_(module.weight, mean=0.0, std=std)
|
| 491 |
+
if isinstance(module, LagunaTopKRouter):
|
| 492 |
+
torch.nn.init.zeros_(module.e_score_correction_bias)
|
| 493 |
+
elif isinstance(module, LagunaRotaryEmbedding):
|
| 494 |
+
for layer_type in module.layer_types:
|
| 495 |
+
rope_init_fn = module.compute_default_rope_parameters
|
| 496 |
+
if module.rope_type[layer_type] != "default":
|
| 497 |
+
rope_init_fn = ROPE_INIT_FUNCTIONS[module.rope_type[layer_type]]
|
| 498 |
+
curr_inv_freq, _ = rope_init_fn(module.config, layer_type=layer_type)
|
| 499 |
+
init.copy_(getattr(module, f"{layer_type}_inv_freq"), curr_inv_freq)
|
| 500 |
+
init.copy_(getattr(module, f"{layer_type}_original_inv_freq"), curr_inv_freq)
|
| 501 |
+
|
| 502 |
+
|
| 503 |
+
@auto_docstring
|
| 504 |
+
class LagunaModel(LagunaPreTrainedModel):
|
| 505 |
+
def __init__(self, config: LagunaConfig):
|
| 506 |
+
super().__init__(config)
|
| 507 |
+
self.padding_idx = config.pad_token_id
|
| 508 |
+
self.vocab_size = config.vocab_size
|
| 509 |
+
|
| 510 |
+
self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
|
| 511 |
+
self.layers = nn.ModuleList(
|
| 512 |
+
[LagunaDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
|
| 513 |
+
)
|
| 514 |
+
self.norm = LagunaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 515 |
+
self.rotary_emb = LagunaRotaryEmbedding(config=config)
|
| 516 |
+
self.gradient_checkpointing = False
|
| 517 |
+
|
| 518 |
+
# Initialize weights and apply final processing
|
| 519 |
+
self.post_init()
|
| 520 |
+
|
| 521 |
+
@capture_outputs
|
| 522 |
+
@auto_docstring
|
| 523 |
+
def forward(
|
| 524 |
+
self,
|
| 525 |
+
input_ids: torch.LongTensor | None = None,
|
| 526 |
+
attention_mask: torch.Tensor | None = None,
|
| 527 |
+
position_ids: torch.LongTensor | None = None,
|
| 528 |
+
past_key_values: Cache | None = None,
|
| 529 |
+
inputs_embeds: torch.FloatTensor | None = None,
|
| 530 |
+
use_cache: bool | None = None,
|
| 531 |
+
**kwargs: Unpack[TransformersKwargs],
|
| 532 |
+
) -> MoeModelOutputWithPast:
|
| 533 |
+
if (input_ids is None) ^ (inputs_embeds is not None):
|
| 534 |
+
raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
|
| 535 |
+
|
| 536 |
+
if inputs_embeds is None:
|
| 537 |
+
inputs_embeds = self.embed_tokens(input_ids)
|
| 538 |
+
|
| 539 |
+
if use_cache and past_key_values is None:
|
| 540 |
+
past_key_values = DynamicCache(config=self.config)
|
| 541 |
+
|
| 542 |
+
if position_ids is None:
|
| 543 |
+
past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
|
| 544 |
+
position_ids = torch.arange(inputs_embeds.shape[1], device=inputs_embeds.device) + past_seen_tokens
|
| 545 |
+
position_ids = position_ids.unsqueeze(0)
|
| 546 |
+
|
| 547 |
+
if not isinstance(causal_mask_mapping := attention_mask, dict):
|
| 548 |
+
mask_kwargs = {
|
| 549 |
+
"config": self.config,
|
| 550 |
+
"inputs_embeds": inputs_embeds,
|
| 551 |
+
"attention_mask": attention_mask,
|
| 552 |
+
"past_key_values": past_key_values,
|
| 553 |
+
"position_ids": position_ids,
|
| 554 |
+
}
|
| 555 |
+
mask_creation_functions = {
|
| 556 |
+
"full_attention": lambda: create_causal_mask(**mask_kwargs),
|
| 557 |
+
"sliding_attention": lambda: create_sliding_window_causal_mask(**mask_kwargs),
|
| 558 |
+
}
|
| 559 |
+
causal_mask_mapping = {}
|
| 560 |
+
for layer_type in set(self.config.layer_types):
|
| 561 |
+
causal_mask_mapping[layer_type] = mask_creation_functions[layer_type]()
|
| 562 |
+
|
| 563 |
+
hidden_states = inputs_embeds
|
| 564 |
+
position_embeddings = {}
|
| 565 |
+
for layer_type in set(self.config.layer_types):
|
| 566 |
+
position_embeddings[layer_type] = self.rotary_emb(hidden_states, position_ids, layer_type)
|
| 567 |
+
|
| 568 |
+
for i, decoder_layer in enumerate(self.layers[: self.config.num_hidden_layers]):
|
| 569 |
+
hidden_states = decoder_layer(
|
| 570 |
+
hidden_states,
|
| 571 |
+
attention_mask=causal_mask_mapping[self.config.layer_types[i]],
|
| 572 |
+
position_embeddings=position_embeddings[self.config.layer_types[i]],
|
| 573 |
+
position_ids=position_ids,
|
| 574 |
+
past_key_values=past_key_values,
|
| 575 |
+
**kwargs,
|
| 576 |
+
)
|
| 577 |
+
|
| 578 |
+
hidden_states = self.norm(hidden_states)
|
| 579 |
+
|
| 580 |
+
return MoeModelOutputWithPast(
|
| 581 |
+
last_hidden_state=hidden_states,
|
| 582 |
+
past_key_values=past_key_values if use_cache else None,
|
| 583 |
+
)
|
| 584 |
+
|
| 585 |
+
|
| 586 |
+
def load_balancing_loss_func(
|
| 587 |
+
gate_logits: torch.Tensor | tuple[torch.Tensor] | None,
|
| 588 |
+
num_experts: int | None = None,
|
| 589 |
+
top_k=2,
|
| 590 |
+
attention_mask: torch.Tensor | None = None,
|
| 591 |
+
) -> torch.Tensor | int:
|
| 592 |
+
r"""
|
| 593 |
+
Computes auxiliary load balancing loss as in Switch Transformer - implemented in Pytorch.
|
| 594 |
+
|
| 595 |
+
See Switch Transformer (https://huggingface.co/papers/2101.03961) for more details. This function implements the loss
|
| 596 |
+
function presented in equations (4) - (6) of the paper. It aims at penalizing cases where the routing between
|
| 597 |
+
experts is too unbalanced.
|
| 598 |
+
|
| 599 |
+
Args:
|
| 600 |
+
gate_logits:
|
| 601 |
+
Logits from the `gate`, should be a tuple of model.config.num_hidden_layers tensors of
|
| 602 |
+
shape [batch_size X sequence_length, num_experts].
|
| 603 |
+
num_experts:
|
| 604 |
+
Number of experts
|
| 605 |
+
top_k:
|
| 606 |
+
The number of experts to route per-token, can be also interpreted as the `top-k` routing
|
| 607 |
+
parameter.
|
| 608 |
+
attention_mask (`torch.Tensor`, *optional*):
|
| 609 |
+
The attention_mask used in forward function
|
| 610 |
+
shape [batch_size X sequence_length] if not None.
|
| 611 |
+
|
| 612 |
+
Returns:
|
| 613 |
+
The auxiliary loss.
|
| 614 |
+
"""
|
| 615 |
+
if gate_logits is None or not isinstance(gate_logits, tuple):
|
| 616 |
+
return 0
|
| 617 |
+
|
| 618 |
+
if isinstance(gate_logits, tuple):
|
| 619 |
+
compute_device = gate_logits[0].device
|
| 620 |
+
concatenated_gate_logits = torch.cat([layer_gate.to(compute_device) for layer_gate in gate_logits], dim=0)
|
| 621 |
+
|
| 622 |
+
routing_weights = torch.nn.functional.softmax(concatenated_gate_logits, dim=-1)
|
| 623 |
+
|
| 624 |
+
_, selected_experts = torch.topk(routing_weights, top_k, dim=-1)
|
| 625 |
+
|
| 626 |
+
expert_mask = torch.nn.functional.one_hot(selected_experts, num_experts)
|
| 627 |
+
|
| 628 |
+
if attention_mask is None:
|
| 629 |
+
# Compute the percentage of tokens routed to each experts
|
| 630 |
+
tokens_per_expert = torch.mean(expert_mask.float(), dim=0)
|
| 631 |
+
|
| 632 |
+
# Compute the average probability of routing to these experts
|
| 633 |
+
router_prob_per_expert = torch.mean(routing_weights, dim=0)
|
| 634 |
+
else:
|
| 635 |
+
batch_size, sequence_length = attention_mask.shape
|
| 636 |
+
num_hidden_layers = concatenated_gate_logits.shape[0] // (batch_size * sequence_length)
|
| 637 |
+
|
| 638 |
+
# Compute the mask that masks all padding tokens as 0 with the same shape of expert_mask
|
| 639 |
+
expert_attention_mask = (
|
| 640 |
+
attention_mask[None, :, :, None, None]
|
| 641 |
+
.expand((num_hidden_layers, batch_size, sequence_length, top_k, num_experts))
|
| 642 |
+
.reshape(-1, top_k, num_experts)
|
| 643 |
+
.to(compute_device)
|
| 644 |
+
)
|
| 645 |
+
|
| 646 |
+
# Compute the percentage of tokens routed to each experts
|
| 647 |
+
tokens_per_expert = torch.sum(expert_mask.float() * expert_attention_mask, dim=0) / torch.sum(
|
| 648 |
+
expert_attention_mask, dim=0
|
| 649 |
+
)
|
| 650 |
+
|
| 651 |
+
# Compute the mask that masks all padding tokens as 0 with the same shape of tokens_per_expert
|
| 652 |
+
router_per_expert_attention_mask = (
|
| 653 |
+
attention_mask[None, :, :, None]
|
| 654 |
+
.expand((num_hidden_layers, batch_size, sequence_length, num_experts))
|
| 655 |
+
.reshape(-1, num_experts)
|
| 656 |
+
.to(compute_device)
|
| 657 |
+
)
|
| 658 |
+
|
| 659 |
+
# Compute the average probability of routing to these experts
|
| 660 |
+
router_prob_per_expert = torch.sum(routing_weights * router_per_expert_attention_mask, dim=0) / torch.sum(
|
| 661 |
+
router_per_expert_attention_mask, dim=0
|
| 662 |
+
)
|
| 663 |
+
|
| 664 |
+
overall_loss = torch.sum(tokens_per_expert * router_prob_per_expert.unsqueeze(0))
|
| 665 |
+
return overall_loss * num_experts
|
| 666 |
+
|
| 667 |
+
|
| 668 |
+
@auto_docstring
|
| 669 |
+
class LagunaForCausalLM(LagunaPreTrainedModel, GenerationMixin):
|
| 670 |
+
_tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
|
| 671 |
+
_tp_plan = {"lm_head": "colwise_gather_output"}
|
| 672 |
+
_pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
|
| 673 |
+
|
| 674 |
+
def __init__(self, config):
|
| 675 |
+
super().__init__(config)
|
| 676 |
+
self.model = LagunaModel(config)
|
| 677 |
+
self.vocab_size = config.vocab_size
|
| 678 |
+
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
|
| 679 |
+
self.router_aux_loss_coef = config.router_aux_loss_coef
|
| 680 |
+
self.num_experts = config.num_experts
|
| 681 |
+
self.num_experts_per_tok = config.num_experts_per_tok
|
| 682 |
+
|
| 683 |
+
# Initialize weights and apply final processing
|
| 684 |
+
self.post_init()
|
| 685 |
+
|
| 686 |
+
@can_return_tuple
|
| 687 |
+
@auto_docstring
|
| 688 |
+
def forward(
|
| 689 |
+
self,
|
| 690 |
+
input_ids: torch.LongTensor | None = None,
|
| 691 |
+
attention_mask: torch.Tensor | None = None,
|
| 692 |
+
position_ids: torch.LongTensor | None = None,
|
| 693 |
+
past_key_values: Cache | None = None,
|
| 694 |
+
inputs_embeds: torch.FloatTensor | None = None,
|
| 695 |
+
labels: torch.LongTensor | None = None,
|
| 696 |
+
use_cache: bool | None = None,
|
| 697 |
+
output_router_logits: bool | None = None,
|
| 698 |
+
logits_to_keep: int | torch.Tensor = 0,
|
| 699 |
+
**kwargs: Unpack[TransformersKwargs],
|
| 700 |
+
) -> MoeCausalLMOutputWithPast:
|
| 701 |
+
r"""
|
| 702 |
+
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
| 703 |
+
Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
|
| 704 |
+
config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
|
| 705 |
+
(masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
|
| 706 |
+
"""
|
| 707 |
+
|
| 708 |
+
output_router_logits = (
|
| 709 |
+
output_router_logits if output_router_logits is not None else self.config.output_router_logits
|
| 710 |
+
)
|
| 711 |
+
|
| 712 |
+
# decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
|
| 713 |
+
outputs: MoeModelOutputWithPast = self.model(
|
| 714 |
+
input_ids=input_ids,
|
| 715 |
+
attention_mask=attention_mask,
|
| 716 |
+
position_ids=position_ids,
|
| 717 |
+
past_key_values=past_key_values,
|
| 718 |
+
inputs_embeds=inputs_embeds,
|
| 719 |
+
use_cache=use_cache,
|
| 720 |
+
output_router_logits=output_router_logits,
|
| 721 |
+
**kwargs,
|
| 722 |
+
)
|
| 723 |
+
|
| 724 |
+
hidden_states = outputs.last_hidden_state
|
| 725 |
+
# Only compute necessary logits, and do not upcast them to float if we are not computing the loss
|
| 726 |
+
slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
|
| 727 |
+
logits = self.lm_head(hidden_states[:, slice_indices, :])
|
| 728 |
+
|
| 729 |
+
loss = None
|
| 730 |
+
if labels is not None:
|
| 731 |
+
loss = self.loss_function(logits, labels, self.vocab_size, **kwargs)
|
| 732 |
+
|
| 733 |
+
aux_loss = None
|
| 734 |
+
if output_router_logits:
|
| 735 |
+
aux_loss = load_balancing_loss_func(
|
| 736 |
+
outputs.router_logits,
|
| 737 |
+
self.num_experts,
|
| 738 |
+
self.num_experts_per_tok,
|
| 739 |
+
attention_mask,
|
| 740 |
+
)
|
| 741 |
+
if labels is not None:
|
| 742 |
+
loss += self.router_aux_loss_coef * aux_loss.to(loss.device) # make sure to reside in the same device
|
| 743 |
+
|
| 744 |
+
return MoeCausalLMOutputWithPast(
|
| 745 |
+
loss=loss,
|
| 746 |
+
aux_loss=aux_loss,
|
| 747 |
+
logits=logits,
|
| 748 |
+
past_key_values=outputs.past_key_values,
|
| 749 |
+
hidden_states=outputs.hidden_states,
|
| 750 |
+
attentions=outputs.attentions,
|
| 751 |
+
router_logits=outputs.router_logits,
|
| 752 |
+
)
|
| 753 |
+
|
| 754 |
+
|
| 755 |
+
__all__ = ["LagunaForCausalLM", "LagunaModel", "LagunaPreTrainedModel"]
|
tokenizer.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
tokenizer_config.json
ADDED
|
@@ -0,0 +1,15 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"backend": "tokenizers",
|
| 3 |
+
"bos_token": "〈|EOS|〉",
|
| 4 |
+
"clean_up_tokenization_spaces": false,
|
| 5 |
+
"cls_token": "〈|CLS|〉",
|
| 6 |
+
"eos_token": "〈|EOS|〉",
|
| 7 |
+
"is_local": true,
|
| 8 |
+
"mask_token": "〈|MASK|〉",
|
| 9 |
+
"model_max_length": 1000000000000000019884624838656,
|
| 10 |
+
"pad_token": "〈|PAD|〉",
|
| 11 |
+
"sep_token": "〈|SEP|〉",
|
| 12 |
+
"tokenizer_class": "TokenizersBackend",
|
| 13 |
+
"tool_parser_type": "laguna",
|
| 14 |
+
"unk_token": "〈|UNK|〉"
|
| 15 |
+
}
|