Upload folder using huggingface_hub
Browse files- README.md +7 -0
- chat_template.jinja +104 -0
- config.json +181 -0
- configuration_lizzy.py +227 -0
- generation_config.json +8 -0
- model-00001-of-00002.safetensors +3 -0
- model-00002-of-00002.safetensors +3 -0
- model.safetensors.index.json +815 -0
- modeling_lizzy.py +2190 -0
- quickstart.py +36 -0
- tokenization_lizzy.py +51 -0
- tokenizer.json +0 -0
- tokenizer_config.json +13 -0
README.md
ADDED
|
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language: en
|
| 3 |
+
library_name: mlx
|
| 4 |
+
pipeline_tag: text-generation
|
| 5 |
+
tags:
|
| 6 |
+
- mlx
|
| 7 |
+
---
|
chat_template.jinja
ADDED
|
@@ -0,0 +1,104 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
\
|
| 2 |
+
{# ───── defaults ───── #}
|
| 3 |
+
{%- if enable_thinking is not defined -%}
|
| 4 |
+
{%- set enable_thinking = true -%}
|
| 5 |
+
{%- endif -%}
|
| 6 |
+
{%- set system_message = "" -%}
|
| 7 |
+
{%- set custom_instructions = "" -%}
|
| 8 |
+
{%- set identity_preamble = "You are Lizzy, a helpful British AI assistant built by Flower Labs. When asked about your identity, name, developer, or origin, answer consistently: your name is Lizzy, you were built by Flower Labs, and you are not ChatGPT, DeepSeek, Claude, Gemini, or any other assistant. Do not misattribute your creator or model family." -%}
|
| 9 |
+
{%- set default_think_instructions = "You are a helpful British function-calling AI assistant. You are a British persona and your date cutoff is November 2024, and your model weights are available at https://huggingface.co/flwrlabs. Your role as an assistant involves thoroughly exploring questions through a systematic thinking process before providing the final precise and accurate solutions. This requires engaging in a comprehensive cycle of analysis, summarizing, exploration, reassessment, reflection, backtracking, and iteration to develop well-considered thinking process. Please structure your response into two main sections: Thought and Solution using the specified format: <tool_call> Thought section </tool_call> Solution section. In the Thought section, detail your reasoning process in steps. Each step should include detailed considerations such as analysing questions, summarizing relevant findings, brainstorming new ideas, verifying the accuracy of the current steps, refining any errors, and revisiting previous steps. In the Solution section, based on various attempts, explorations, and reflections from the Thought section, systematically present the final solution that you deem correct. The Solution section should be logical, accurate, and concise and detail necessary steps needed to reach the conclusion." -%}
|
| 10 |
+
{%- set default_no_think_instructions = "You are a helpful British function-calling AI assistant. You are a British persona and your date cutoff is November 2024, and your model weights are available at https://huggingface.co/flwrlabs." -%}
|
| 11 |
+
|
| 12 |
+
{# ───── reasoning mode ───── #}
|
| 13 |
+
{%- if enable_thinking -%}
|
| 14 |
+
{%- set reasoning_mode = "/think" -%}
|
| 15 |
+
{%- else -%}
|
| 16 |
+
{%- set reasoning_mode = "/no_think" -%}
|
| 17 |
+
{%- endif -%}
|
| 18 |
+
|
| 19 |
+
{# ───── header (system message) ───── #}
|
| 20 |
+
{{- "<|im_start|>system\n" -}}
|
| 21 |
+
|
| 22 |
+
{%- if messages[0].role == "system" -%}
|
| 23 |
+
{%- set system_message = messages[0].content -%}
|
| 24 |
+
{%- if "/no_think" in system_message -%}
|
| 25 |
+
{%- set reasoning_mode = "/no_think" -%}
|
| 26 |
+
{%- elif "/think" in system_message -%}
|
| 27 |
+
{%- set reasoning_mode = "/think" -%}
|
| 28 |
+
{%- endif -%}
|
| 29 |
+
{%- set custom_instructions = system_message.replace("/no_think", "").replace("/think", "").rstrip() -%}
|
| 30 |
+
{%- endif -%}
|
| 31 |
+
|
| 32 |
+
{%- if "/system_override" in system_message -%}
|
| 33 |
+
{{- identity_preamble + "\n\n" -}}
|
| 34 |
+
{{- custom_instructions.replace("/system_override", "").rstrip() -}}
|
| 35 |
+
{{- "<|im_end|>\n" -}}
|
| 36 |
+
{%- else -%}
|
| 37 |
+
{{- "## Metadata\n\n" -}}
|
| 38 |
+
{{- "Knowledge Cutoff Date: June 2025\n" -}}
|
| 39 |
+
{%- set today = strftime_now("%d %B %Y") -%}
|
| 40 |
+
{{- "Today Date: " ~ today ~ "\n" -}}
|
| 41 |
+
{{- "Reasoning Mode: " + reasoning_mode + "\n\n" -}}
|
| 42 |
+
|
| 43 |
+
{{- "## Identity\n\n" -}}
|
| 44 |
+
{{- identity_preamble + "\n\n" -}}
|
| 45 |
+
|
| 46 |
+
{{- "## Custom Instructions\n\n" -}}
|
| 47 |
+
{%- if custom_instructions -%}
|
| 48 |
+
{{- custom_instructions + "\n\n" -}}
|
| 49 |
+
{%- elif reasoning_mode == "/think" -%}
|
| 50 |
+
{{- default_think_instructions + "\n\n" -}}
|
| 51 |
+
{%- else -%}
|
| 52 |
+
{{- default_no_think_instructions + "\n\n" -}}
|
| 53 |
+
{%- endif -%}
|
| 54 |
+
|
| 55 |
+
{%- if xml_tools or python_tools or tools -%}
|
| 56 |
+
{{- "### Tools\n\n" -}}
|
| 57 |
+
{%- if xml_tools or tools -%}
|
| 58 |
+
{%- if tools -%}
|
| 59 |
+
{%- set xml_tools = tools -%}
|
| 60 |
+
{%- endif -%}
|
| 61 |
+
{%- set ns = namespace(xml_tool_string="You may call one or more functions to assist with the user query.\nYou are provided with function signatures within <tools></tools> XML tags:\n\n<tools>\n") -%}
|
| 62 |
+
{%- for tool in xml_tools[:] -%} {# The slicing makes sure that xml_tools is a list #}
|
| 63 |
+
{%- set ns.xml_tool_string = ns.xml_tool_string ~ (tool | string) ~ "\n" -%}
|
| 64 |
+
{%- endfor -%}
|
| 65 |
+
{%- set xml_tool_string = ns.xml_tool_string + "</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call>" -%}
|
| 66 |
+
{{- xml_tool_string -}}
|
| 67 |
+
{%- endif -%}
|
| 68 |
+
{%- if python_tools -%}
|
| 69 |
+
{%- set ns = namespace(python_tool_string="When you send a message containing Python code between '<code>' and '</code>' tags, it will be executed in a stateful Jupyter notebook environment, and you will then be given the output to continued reasoning in an agentic loop.\n\nYou can use the following tools in your python code like regular functions:\n<tools>\n") -%}
|
| 70 |
+
{%- for tool in python_tools[:] -%} {# The slicing makes sure that python_tools is a list #}
|
| 71 |
+
{%- set ns.python_tool_string = ns.python_tool_string ~ (tool | string) ~ "\n" -%}
|
| 72 |
+
{%- endfor -%}
|
| 73 |
+
{%- set python_tool_string = ns.python_tool_string + "</tools>\n\nThe state persists between code executions: so variables that you define in one step are still available thereafter." -%}
|
| 74 |
+
{{- python_tool_string -}}
|
| 75 |
+
{%- endif -%}
|
| 76 |
+
{{- "\n\n" -}}
|
| 77 |
+
{{- "<|im_end|>\n" -}}
|
| 78 |
+
{%- endif -%}
|
| 79 |
+
{%- endif -%}
|
| 80 |
+
{# ───── main loop ───── #}
|
| 81 |
+
{%- for message in messages -%}
|
| 82 |
+
{%- set content = message.content if message.content is string else "" -%}
|
| 83 |
+
{%- if message.role == "user" -%}
|
| 84 |
+
{{ "<|im_start|>" + message.role + "\n" + content + "<|im_end|>\n" }}
|
| 85 |
+
{%- elif message.role == "assistant" -%}
|
| 86 |
+
{% generation %}
|
| 87 |
+
{%- if reasoning_mode == "/think" -%}
|
| 88 |
+
{{ "<|im_start|>assistant\n" + content.lstrip("\n") + "<|im_end|>\n" }}
|
| 89 |
+
{%- else -%}
|
| 90 |
+
{{ "<|im_start|>assistant\n" + "<think>\n\n</think>\n" + content.lstrip("\n") + "<|im_end|>\n" }}
|
| 91 |
+
{%- endif -%}
|
| 92 |
+
{% endgeneration %}
|
| 93 |
+
{%- elif message.role == "tool" -%}
|
| 94 |
+
{{ "<|im_start|>" + "user\n" + content + "<|im_end|>\n" }}
|
| 95 |
+
{%- endif -%}
|
| 96 |
+
{%- endfor -%}
|
| 97 |
+
{# ───── generation prompt ───── #}
|
| 98 |
+
{%- if add_generation_prompt -%}
|
| 99 |
+
{%- if reasoning_mode == "/think" -%}
|
| 100 |
+
{{ "<|im_start|>assistant\n" }}
|
| 101 |
+
{%- else -%}
|
| 102 |
+
{{ "<|im_start|>assistant\n" + "<think>\n\n</think>\n" }}
|
| 103 |
+
{%- endif -%}
|
| 104 |
+
{%- endif -%}
|
config.json
ADDED
|
@@ -0,0 +1,181 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"architectures": [
|
| 3 |
+
"LizzyForCausalLM"
|
| 4 |
+
],
|
| 5 |
+
"attention_bias": false,
|
| 6 |
+
"attention_dropout": 0.0,
|
| 7 |
+
"auto_map": {
|
| 8 |
+
"AutoConfig": "configuration_lizzy.LizzyConfig",
|
| 9 |
+
"AutoModel": "modeling_lizzy.LizzyModel",
|
| 10 |
+
"AutoModelForCausalLM": "modeling_lizzy.LizzyForCausalLM",
|
| 11 |
+
"AutoTokenizer": "tokenization_lizzy.LizzyTokenizerFast"
|
| 12 |
+
},
|
| 13 |
+
"bos_token_id": 100257,
|
| 14 |
+
"embd_dropout": 0.0,
|
| 15 |
+
"eos_token_id": 100257,
|
| 16 |
+
"head_dim": 128,
|
| 17 |
+
"hidden_act": "silu",
|
| 18 |
+
"hidden_size": 4096,
|
| 19 |
+
"initializer_range": 0.02,
|
| 20 |
+
"intermediate_size": 11008,
|
| 21 |
+
"layer_layouts": [
|
| 22 |
+
"decoder_postnorm",
|
| 23 |
+
"decoder_postnorm",
|
| 24 |
+
"decoder_postnorm",
|
| 25 |
+
"decoder_postnorm",
|
| 26 |
+
"decoder_postnorm",
|
| 27 |
+
"decoder_postnorm",
|
| 28 |
+
"decoder_postnorm",
|
| 29 |
+
"decoder_postnorm",
|
| 30 |
+
"decoder_postnorm",
|
| 31 |
+
"decoder_postnorm",
|
| 32 |
+
"decoder_postnorm",
|
| 33 |
+
"decoder_postnorm",
|
| 34 |
+
"decoder_postnorm",
|
| 35 |
+
"decoder_postnorm",
|
| 36 |
+
"decoder_postnorm",
|
| 37 |
+
"decoder_postnorm",
|
| 38 |
+
"decoder_postnorm",
|
| 39 |
+
"decoder_postnorm",
|
| 40 |
+
"decoder_postnorm",
|
| 41 |
+
"decoder_postnorm",
|
| 42 |
+
"decoder_postnorm",
|
| 43 |
+
"decoder_postnorm",
|
| 44 |
+
"decoder_postnorm",
|
| 45 |
+
"decoder_postnorm",
|
| 46 |
+
"decoder_postnorm",
|
| 47 |
+
"decoder_postnorm",
|
| 48 |
+
"decoder_postnorm",
|
| 49 |
+
"decoder_postnorm",
|
| 50 |
+
"decoder_postnorm",
|
| 51 |
+
"decoder_postnorm",
|
| 52 |
+
"decoder_postnorm",
|
| 53 |
+
"decoder_postnorm"
|
| 54 |
+
],
|
| 55 |
+
"layer_types": [
|
| 56 |
+
"sliding_attention",
|
| 57 |
+
"sliding_attention",
|
| 58 |
+
"sliding_attention",
|
| 59 |
+
"full_attention",
|
| 60 |
+
"sliding_attention",
|
| 61 |
+
"sliding_attention",
|
| 62 |
+
"sliding_attention",
|
| 63 |
+
"full_attention",
|
| 64 |
+
"sliding_attention",
|
| 65 |
+
"sliding_attention",
|
| 66 |
+
"sliding_attention",
|
| 67 |
+
"full_attention",
|
| 68 |
+
"sliding_attention",
|
| 69 |
+
"sliding_attention",
|
| 70 |
+
"sliding_attention",
|
| 71 |
+
"full_attention",
|
| 72 |
+
"sliding_attention",
|
| 73 |
+
"sliding_attention",
|
| 74 |
+
"sliding_attention",
|
| 75 |
+
"full_attention",
|
| 76 |
+
"sliding_attention",
|
| 77 |
+
"sliding_attention",
|
| 78 |
+
"sliding_attention",
|
| 79 |
+
"full_attention",
|
| 80 |
+
"sliding_attention",
|
| 81 |
+
"sliding_attention",
|
| 82 |
+
"sliding_attention",
|
| 83 |
+
"full_attention",
|
| 84 |
+
"sliding_attention",
|
| 85 |
+
"sliding_attention",
|
| 86 |
+
"sliding_attention",
|
| 87 |
+
"full_attention"
|
| 88 |
+
],
|
| 89 |
+
"linear_a_log_max": null,
|
| 90 |
+
"linear_a_log_min": null,
|
| 91 |
+
"linear_allow_neg_eigval": null,
|
| 92 |
+
"linear_conv_kernel_dim": null,
|
| 93 |
+
"linear_dt_init_floor": null,
|
| 94 |
+
"linear_dt_max": null,
|
| 95 |
+
"linear_dt_min": null,
|
| 96 |
+
"linear_key_head_dim": null,
|
| 97 |
+
"linear_num_key_heads": null,
|
| 98 |
+
"linear_num_value_heads": null,
|
| 99 |
+
"linear_value_head_dim": null,
|
| 100 |
+
"max_position_embeddings": 65536,
|
| 101 |
+
"mlp_bias": false,
|
| 102 |
+
"mlp_type": "gated",
|
| 103 |
+
"model_type": "lizzy",
|
| 104 |
+
"no_rope_layer_interval": null,
|
| 105 |
+
"norm_eps": 1e-06,
|
| 106 |
+
"norm_has_bias": false,
|
| 107 |
+
"norm_type": "rmsnorm",
|
| 108 |
+
"num_attention_heads": 32,
|
| 109 |
+
"num_hidden_layers": 32,
|
| 110 |
+
"num_key_value_heads": 32,
|
| 111 |
+
"pad_token_id": 100277,
|
| 112 |
+
"position_embedding_type": "rope",
|
| 113 |
+
"qk_norm_type": "rmsnorm",
|
| 114 |
+
"quantization": {
|
| 115 |
+
"group_size": 64,
|
| 116 |
+
"bits": 8,
|
| 117 |
+
"mode": "affine"
|
| 118 |
+
},
|
| 119 |
+
"quantization_config": {
|
| 120 |
+
"group_size": 64,
|
| 121 |
+
"bits": 8,
|
| 122 |
+
"mode": "affine"
|
| 123 |
+
},
|
| 124 |
+
"resid_dropout": 0.0,
|
| 125 |
+
"rope_layer_flags": [
|
| 126 |
+
true,
|
| 127 |
+
true,
|
| 128 |
+
true,
|
| 129 |
+
true,
|
| 130 |
+
true,
|
| 131 |
+
true,
|
| 132 |
+
true,
|
| 133 |
+
true,
|
| 134 |
+
true,
|
| 135 |
+
true,
|
| 136 |
+
true,
|
| 137 |
+
true,
|
| 138 |
+
true,
|
| 139 |
+
true,
|
| 140 |
+
true,
|
| 141 |
+
true,
|
| 142 |
+
true,
|
| 143 |
+
true,
|
| 144 |
+
true,
|
| 145 |
+
true,
|
| 146 |
+
true,
|
| 147 |
+
true,
|
| 148 |
+
true,
|
| 149 |
+
true,
|
| 150 |
+
true,
|
| 151 |
+
true,
|
| 152 |
+
true,
|
| 153 |
+
true,
|
| 154 |
+
true,
|
| 155 |
+
true,
|
| 156 |
+
true,
|
| 157 |
+
true
|
| 158 |
+
],
|
| 159 |
+
"rope_scaling": {
|
| 160 |
+
"attention_factor": 1.2079441541679836,
|
| 161 |
+
"beta_fast": 32,
|
| 162 |
+
"beta_slow": 1,
|
| 163 |
+
"factor": 8.0,
|
| 164 |
+
"original_max_position_embeddings": 8192,
|
| 165 |
+
"rope_type": "default",
|
| 166 |
+
"rope_theta": 500000
|
| 167 |
+
},
|
| 168 |
+
"rope_theta": 500000,
|
| 169 |
+
"rope_type_overrides": {},
|
| 170 |
+
"sliding_window": 4096,
|
| 171 |
+
"tie_word_embeddings": false,
|
| 172 |
+
"tokenizer_class": "LizzyTokenizerFast",
|
| 173 |
+
"transformers_version": "5.4.0",
|
| 174 |
+
"use_cache": true,
|
| 175 |
+
"use_post_attn_norm": true,
|
| 176 |
+
"use_post_mlp_norm": true,
|
| 177 |
+
"use_pre_attn_norm": false,
|
| 178 |
+
"use_pre_mlp_norm": false,
|
| 179 |
+
"use_qk_norm": true,
|
| 180 |
+
"vocab_size": 100278
|
| 181 |
+
}
|
configuration_lizzy.py
ADDED
|
@@ -0,0 +1,227 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from __future__ import annotations
|
| 2 |
+
|
| 3 |
+
from typing import Any
|
| 4 |
+
|
| 5 |
+
from transformers import PretrainedConfig
|
| 6 |
+
|
| 7 |
+
|
| 8 |
+
class LizzyConfig(PretrainedConfig):
|
| 9 |
+
model_type = "lizzy"
|
| 10 |
+
keys_to_ignore_at_inference = ["past_key_values"]
|
| 11 |
+
base_model_tp_plan = {
|
| 12 |
+
"layers.*.self_attn.q_proj": "colwise",
|
| 13 |
+
"layers.*.self_attn.k_proj": "colwise",
|
| 14 |
+
"layers.*.self_attn.v_proj": "colwise",
|
| 15 |
+
"layers.*.self_attn.o_proj": "rowwise",
|
| 16 |
+
"layers.*.mlp.up_proj": "colwise",
|
| 17 |
+
"layers.*.mlp.gate_proj": "colwise",
|
| 18 |
+
"layers.*.mlp.down_proj": "rowwise",
|
| 19 |
+
"lm_head": "colwise",
|
| 20 |
+
}
|
| 21 |
+
base_model_pp_plan = {
|
| 22 |
+
"embed_tokens": (["input_ids"], ["inputs_embeds"]),
|
| 23 |
+
"layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
|
| 24 |
+
"norm": (["hidden_states"], ["hidden_states"]),
|
| 25 |
+
}
|
| 26 |
+
|
| 27 |
+
def __init__(
|
| 28 |
+
self,
|
| 29 |
+
vocab_size: int = 32000,
|
| 30 |
+
hidden_size: int = 4096,
|
| 31 |
+
intermediate_size: int = 11008,
|
| 32 |
+
num_hidden_layers: int = 32,
|
| 33 |
+
num_attention_heads: int = 32,
|
| 34 |
+
num_key_value_heads: int | None = None,
|
| 35 |
+
max_position_embeddings: int = 2048,
|
| 36 |
+
head_dim: int | None = None,
|
| 37 |
+
hidden_act: str = "silu",
|
| 38 |
+
norm_type: str = "rmsnorm",
|
| 39 |
+
norm_eps: float = 1e-6,
|
| 40 |
+
norm_has_bias: bool = False,
|
| 41 |
+
use_pre_attn_norm: bool = True,
|
| 42 |
+
use_pre_mlp_norm: bool = True,
|
| 43 |
+
use_post_attn_norm: bool = False,
|
| 44 |
+
use_post_mlp_norm: bool = False,
|
| 45 |
+
mlp_type: str = "gated",
|
| 46 |
+
attention_bias: bool = False,
|
| 47 |
+
mlp_bias: bool = False,
|
| 48 |
+
position_embedding_type: str = "rope",
|
| 49 |
+
rope_theta: float = 10000.0,
|
| 50 |
+
rope_scaling: dict[str, Any] | None = None,
|
| 51 |
+
rope_layer_flags: list[bool] | None = None,
|
| 52 |
+
no_rope_layer_interval: int | None = None,
|
| 53 |
+
rope_type_overrides: dict[str, str] | None = None,
|
| 54 |
+
layer_types: list[str] | None = None,
|
| 55 |
+
layer_layouts: list[str] | None = None,
|
| 56 |
+
sliding_window: int | None = None,
|
| 57 |
+
linear_num_key_heads: int | None = None,
|
| 58 |
+
linear_num_value_heads: int | None = None,
|
| 59 |
+
linear_key_head_dim: int | None = None,
|
| 60 |
+
linear_value_head_dim: int | None = None,
|
| 61 |
+
linear_a_log_min: float | None = None,
|
| 62 |
+
linear_a_log_max: float | None = None,
|
| 63 |
+
linear_dt_min: float | None = None,
|
| 64 |
+
linear_dt_max: float | None = None,
|
| 65 |
+
linear_dt_init_floor: float | None = None,
|
| 66 |
+
linear_conv_kernel_dim: int | None = None,
|
| 67 |
+
linear_allow_neg_eigval: bool | None = None,
|
| 68 |
+
use_qk_norm: bool = False,
|
| 69 |
+
qk_norm_type: str = "rmsnorm",
|
| 70 |
+
attention_dropout: float = 0.0,
|
| 71 |
+
resid_dropout: float = 0.0,
|
| 72 |
+
embd_dropout: float = 0.0,
|
| 73 |
+
initializer_range: float = 0.02,
|
| 74 |
+
bos_token_id: int | None = None,
|
| 75 |
+
eos_token_id: int | None = None,
|
| 76 |
+
pad_token_id: int | None = None,
|
| 77 |
+
use_cache: bool = True,
|
| 78 |
+
tie_word_embeddings: bool = False,
|
| 79 |
+
**kwargs,
|
| 80 |
+
) -> None:
|
| 81 |
+
if num_key_value_heads is None:
|
| 82 |
+
num_key_value_heads = num_attention_heads
|
| 83 |
+
if head_dim is None:
|
| 84 |
+
head_dim = hidden_size // num_attention_heads
|
| 85 |
+
if no_rope_layer_interval is not None:
|
| 86 |
+
no_rope_layer_interval = int(no_rope_layer_interval)
|
| 87 |
+
if no_rope_layer_interval <= 0:
|
| 88 |
+
no_rope_layer_interval = None
|
| 89 |
+
if layer_types is None:
|
| 90 |
+
layer_types = ["full_attention"] * int(num_hidden_layers)
|
| 91 |
+
if layer_layouts is None:
|
| 92 |
+
if use_post_attn_norm or use_post_mlp_norm:
|
| 93 |
+
layer_layouts = ["decoder_postnorm"] * int(num_hidden_layers)
|
| 94 |
+
else:
|
| 95 |
+
layer_layouts = ["decoder_prenorm"] * int(num_hidden_layers)
|
| 96 |
+
if rope_layer_flags is None:
|
| 97 |
+
rope_enabled = position_embedding_type == "rope"
|
| 98 |
+
if rope_enabled and no_rope_layer_interval is not None:
|
| 99 |
+
rope_layer_flags = [
|
| 100 |
+
((layer_idx + 1) % no_rope_layer_interval) != 0
|
| 101 |
+
for layer_idx in range(int(num_hidden_layers))
|
| 102 |
+
]
|
| 103 |
+
else:
|
| 104 |
+
rope_layer_flags = [rope_enabled] * int(num_hidden_layers)
|
| 105 |
+
|
| 106 |
+
normalized_rope_scaling = None
|
| 107 |
+
if rope_scaling is not None:
|
| 108 |
+
normalized_rope_scaling = dict(rope_scaling)
|
| 109 |
+
for field_name in (
|
| 110 |
+
"factor",
|
| 111 |
+
"attention_factor",
|
| 112 |
+
"beta_fast",
|
| 113 |
+
"beta_slow",
|
| 114 |
+
):
|
| 115 |
+
if normalized_rope_scaling.get(field_name) is not None:
|
| 116 |
+
normalized_rope_scaling[field_name] = float(
|
| 117 |
+
normalized_rope_scaling[field_name]
|
| 118 |
+
)
|
| 119 |
+
if (
|
| 120 |
+
normalized_rope_scaling.get("original_max_position_embeddings")
|
| 121 |
+
is not None
|
| 122 |
+
):
|
| 123 |
+
normalized_rope_scaling["original_max_position_embeddings"] = int(
|
| 124 |
+
normalized_rope_scaling["original_max_position_embeddings"]
|
| 125 |
+
)
|
| 126 |
+
|
| 127 |
+
# Transformers validates RoPE settings during PretrainedConfig
|
| 128 |
+
# initialization, so publish the rope-critical fields before
|
| 129 |
+
# calling `super().__init__()`.
|
| 130 |
+
self.max_position_embeddings = int(max_position_embeddings)
|
| 131 |
+
self.rope_theta = float(rope_theta)
|
| 132 |
+
self.rope_scaling = normalized_rope_scaling
|
| 133 |
+
|
| 134 |
+
super().__init__(
|
| 135 |
+
bos_token_id=bos_token_id,
|
| 136 |
+
eos_token_id=eos_token_id,
|
| 137 |
+
pad_token_id=pad_token_id,
|
| 138 |
+
tie_word_embeddings=tie_word_embeddings,
|
| 139 |
+
**kwargs,
|
| 140 |
+
)
|
| 141 |
+
self.vocab_size = int(vocab_size)
|
| 142 |
+
self.hidden_size = int(hidden_size)
|
| 143 |
+
self.intermediate_size = int(intermediate_size)
|
| 144 |
+
self.num_hidden_layers = int(num_hidden_layers)
|
| 145 |
+
self.num_attention_heads = int(num_attention_heads)
|
| 146 |
+
self.num_key_value_heads = int(num_key_value_heads)
|
| 147 |
+
self.max_position_embeddings = int(max_position_embeddings)
|
| 148 |
+
self.head_dim = int(head_dim)
|
| 149 |
+
self.hidden_act = str(hidden_act)
|
| 150 |
+
self.norm_type = str(norm_type)
|
| 151 |
+
self.norm_eps = float(norm_eps)
|
| 152 |
+
self.norm_has_bias = bool(norm_has_bias)
|
| 153 |
+
self.use_pre_attn_norm = bool(use_pre_attn_norm)
|
| 154 |
+
self.use_pre_mlp_norm = bool(use_pre_mlp_norm)
|
| 155 |
+
self.use_post_attn_norm = bool(use_post_attn_norm)
|
| 156 |
+
self.use_post_mlp_norm = bool(use_post_mlp_norm)
|
| 157 |
+
self.mlp_type = str(mlp_type)
|
| 158 |
+
self.attention_bias = bool(attention_bias)
|
| 159 |
+
self.mlp_bias = bool(mlp_bias)
|
| 160 |
+
self.position_embedding_type = str(position_embedding_type)
|
| 161 |
+
self.rope_theta = float(rope_theta)
|
| 162 |
+
self.rope_scaling = normalized_rope_scaling
|
| 163 |
+
self.no_rope_layer_interval = no_rope_layer_interval
|
| 164 |
+
self.rope_type_overrides = {
|
| 165 |
+
str(key): str(value)
|
| 166 |
+
for key, value in dict(rope_type_overrides or {}).items()
|
| 167 |
+
}
|
| 168 |
+
self.layer_types = list(layer_types)
|
| 169 |
+
self.layer_layouts = [str(item) for item in layer_layouts]
|
| 170 |
+
self.rope_layer_flags = [bool(item) for item in rope_layer_flags]
|
| 171 |
+
self.sliding_window = sliding_window
|
| 172 |
+
self.linear_num_key_heads = (
|
| 173 |
+
None
|
| 174 |
+
if linear_num_key_heads is None
|
| 175 |
+
else int(linear_num_key_heads)
|
| 176 |
+
)
|
| 177 |
+
self.linear_num_value_heads = (
|
| 178 |
+
None
|
| 179 |
+
if linear_num_value_heads is None
|
| 180 |
+
else int(linear_num_value_heads)
|
| 181 |
+
)
|
| 182 |
+
self.linear_key_head_dim = (
|
| 183 |
+
None
|
| 184 |
+
if linear_key_head_dim is None
|
| 185 |
+
else int(linear_key_head_dim)
|
| 186 |
+
)
|
| 187 |
+
self.linear_value_head_dim = (
|
| 188 |
+
None
|
| 189 |
+
if linear_value_head_dim is None
|
| 190 |
+
else int(linear_value_head_dim)
|
| 191 |
+
)
|
| 192 |
+
self.linear_a_log_min = (
|
| 193 |
+
None if linear_a_log_min is None else float(linear_a_log_min)
|
| 194 |
+
)
|
| 195 |
+
self.linear_a_log_max = (
|
| 196 |
+
None if linear_a_log_max is None else float(linear_a_log_max)
|
| 197 |
+
)
|
| 198 |
+
self.linear_dt_min = (
|
| 199 |
+
None if linear_dt_min is None else float(linear_dt_min)
|
| 200 |
+
)
|
| 201 |
+
self.linear_dt_max = (
|
| 202 |
+
None if linear_dt_max is None else float(linear_dt_max)
|
| 203 |
+
)
|
| 204 |
+
self.linear_dt_init_floor = (
|
| 205 |
+
None
|
| 206 |
+
if linear_dt_init_floor is None
|
| 207 |
+
else float(linear_dt_init_floor)
|
| 208 |
+
)
|
| 209 |
+
self.linear_conv_kernel_dim = (
|
| 210 |
+
None
|
| 211 |
+
if linear_conv_kernel_dim is None
|
| 212 |
+
else int(linear_conv_kernel_dim)
|
| 213 |
+
)
|
| 214 |
+
self.linear_allow_neg_eigval = (
|
| 215 |
+
None
|
| 216 |
+
if linear_allow_neg_eigval is None
|
| 217 |
+
else bool(linear_allow_neg_eigval)
|
| 218 |
+
)
|
| 219 |
+
self.use_qk_norm = bool(use_qk_norm)
|
| 220 |
+
self.qk_norm_type = str(qk_norm_type)
|
| 221 |
+
self.attention_dropout = float(attention_dropout)
|
| 222 |
+
self.resid_dropout = float(resid_dropout)
|
| 223 |
+
self.embd_dropout = float(embd_dropout)
|
| 224 |
+
self.initializer_range = float(initializer_range)
|
| 225 |
+
self.use_cache = bool(use_cache)
|
| 226 |
+
self.rms_norm_eps = self.norm_eps
|
| 227 |
+
self.dtype = None
|
generation_config.json
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_from_model_config": true,
|
| 3 |
+
"eos_token_id": 100257,
|
| 4 |
+
"transformers_version": "4.57.3",
|
| 5 |
+
"bos_token_id": 100257,
|
| 6 |
+
"pad_token_id": 100277,
|
| 7 |
+
"do_sample": true
|
| 8 |
+
}
|
model-00001-of-00002.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:2825ac1e18815a00a6d8442ca3f81101652eb08a0b57b98ca4d1ff864a648c51
|
| 3 |
+
size 5334848691
|
model-00002-of-00002.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c6fee439baf7b95aae6b39d652289600c96ad016e571b92328dcd416cb5b325c
|
| 3 |
+
size 2419875217
|
model.safetensors.index.json
ADDED
|
@@ -0,0 +1,815 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"metadata": {
|
| 3 |
+
"total_size": 7754632192,
|
| 4 |
+
"total_parameters": 7298011136
|
| 5 |
+
},
|
| 6 |
+
"weight_map": {
|
| 7 |
+
"lm_head.biases": "model-00002-of-00002.safetensors",
|
| 8 |
+
"lm_head.scales": "model-00002-of-00002.safetensors",
|
| 9 |
+
"lm_head.weight": "model-00002-of-00002.safetensors",
|
| 10 |
+
"model.embed_tokens.biases": "model-00001-of-00002.safetensors",
|
| 11 |
+
"model.embed_tokens.scales": "model-00001-of-00002.safetensors",
|
| 12 |
+
"model.embed_tokens.weight": "model-00001-of-00002.safetensors",
|
| 13 |
+
"model.layers.0.mlp.down_proj.biases": "model-00001-of-00002.safetensors",
|
| 14 |
+
"model.layers.0.mlp.down_proj.scales": "model-00001-of-00002.safetensors",
|
| 15 |
+
"model.layers.0.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 16 |
+
"model.layers.0.mlp.gate_proj.biases": "model-00001-of-00002.safetensors",
|
| 17 |
+
"model.layers.0.mlp.gate_proj.scales": "model-00001-of-00002.safetensors",
|
| 18 |
+
"model.layers.0.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 19 |
+
"model.layers.0.mlp.up_proj.biases": "model-00001-of-00002.safetensors",
|
| 20 |
+
"model.layers.0.mlp.up_proj.scales": "model-00001-of-00002.safetensors",
|
| 21 |
+
"model.layers.0.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 22 |
+
"model.layers.0.post_attn_norm.weight": "model-00001-of-00002.safetensors",
|
| 23 |
+
"model.layers.0.post_mlp_norm.weight": "model-00001-of-00002.safetensors",
|
| 24 |
+
"model.layers.0.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 25 |
+
"model.layers.0.self_attn.k_proj.biases": "model-00001-of-00002.safetensors",
|
| 26 |
+
"model.layers.0.self_attn.k_proj.scales": "model-00001-of-00002.safetensors",
|
| 27 |
+
"model.layers.0.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 28 |
+
"model.layers.0.self_attn.o_proj.biases": "model-00001-of-00002.safetensors",
|
| 29 |
+
"model.layers.0.self_attn.o_proj.scales": "model-00001-of-00002.safetensors",
|
| 30 |
+
"model.layers.0.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 31 |
+
"model.layers.0.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 32 |
+
"model.layers.0.self_attn.q_proj.biases": "model-00001-of-00002.safetensors",
|
| 33 |
+
"model.layers.0.self_attn.q_proj.scales": "model-00001-of-00002.safetensors",
|
| 34 |
+
"model.layers.0.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 35 |
+
"model.layers.0.self_attn.v_proj.biases": "model-00001-of-00002.safetensors",
|
| 36 |
+
"model.layers.0.self_attn.v_proj.scales": "model-00001-of-00002.safetensors",
|
| 37 |
+
"model.layers.0.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 38 |
+
"model.layers.1.mlp.down_proj.biases": "model-00001-of-00002.safetensors",
|
| 39 |
+
"model.layers.1.mlp.down_proj.scales": "model-00001-of-00002.safetensors",
|
| 40 |
+
"model.layers.1.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 41 |
+
"model.layers.1.mlp.gate_proj.biases": "model-00001-of-00002.safetensors",
|
| 42 |
+
"model.layers.1.mlp.gate_proj.scales": "model-00001-of-00002.safetensors",
|
| 43 |
+
"model.layers.1.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 44 |
+
"model.layers.1.mlp.up_proj.biases": "model-00001-of-00002.safetensors",
|
| 45 |
+
"model.layers.1.mlp.up_proj.scales": "model-00001-of-00002.safetensors",
|
| 46 |
+
"model.layers.1.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 47 |
+
"model.layers.1.post_attn_norm.weight": "model-00001-of-00002.safetensors",
|
| 48 |
+
"model.layers.1.post_mlp_norm.weight": "model-00001-of-00002.safetensors",
|
| 49 |
+
"model.layers.1.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 50 |
+
"model.layers.1.self_attn.k_proj.biases": "model-00001-of-00002.safetensors",
|
| 51 |
+
"model.layers.1.self_attn.k_proj.scales": "model-00001-of-00002.safetensors",
|
| 52 |
+
"model.layers.1.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 53 |
+
"model.layers.1.self_attn.o_proj.biases": "model-00001-of-00002.safetensors",
|
| 54 |
+
"model.layers.1.self_attn.o_proj.scales": "model-00001-of-00002.safetensors",
|
| 55 |
+
"model.layers.1.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 56 |
+
"model.layers.1.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 57 |
+
"model.layers.1.self_attn.q_proj.biases": "model-00001-of-00002.safetensors",
|
| 58 |
+
"model.layers.1.self_attn.q_proj.scales": "model-00001-of-00002.safetensors",
|
| 59 |
+
"model.layers.1.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 60 |
+
"model.layers.1.self_attn.v_proj.biases": "model-00001-of-00002.safetensors",
|
| 61 |
+
"model.layers.1.self_attn.v_proj.scales": "model-00001-of-00002.safetensors",
|
| 62 |
+
"model.layers.1.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 63 |
+
"model.layers.10.mlp.down_proj.biases": "model-00001-of-00002.safetensors",
|
| 64 |
+
"model.layers.10.mlp.down_proj.scales": "model-00001-of-00002.safetensors",
|
| 65 |
+
"model.layers.10.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 66 |
+
"model.layers.10.mlp.gate_proj.biases": "model-00001-of-00002.safetensors",
|
| 67 |
+
"model.layers.10.mlp.gate_proj.scales": "model-00001-of-00002.safetensors",
|
| 68 |
+
"model.layers.10.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 69 |
+
"model.layers.10.mlp.up_proj.biases": "model-00001-of-00002.safetensors",
|
| 70 |
+
"model.layers.10.mlp.up_proj.scales": "model-00001-of-00002.safetensors",
|
| 71 |
+
"model.layers.10.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 72 |
+
"model.layers.10.post_attn_norm.weight": "model-00001-of-00002.safetensors",
|
| 73 |
+
"model.layers.10.post_mlp_norm.weight": "model-00001-of-00002.safetensors",
|
| 74 |
+
"model.layers.10.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 75 |
+
"model.layers.10.self_attn.k_proj.biases": "model-00001-of-00002.safetensors",
|
| 76 |
+
"model.layers.10.self_attn.k_proj.scales": "model-00001-of-00002.safetensors",
|
| 77 |
+
"model.layers.10.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 78 |
+
"model.layers.10.self_attn.o_proj.biases": "model-00001-of-00002.safetensors",
|
| 79 |
+
"model.layers.10.self_attn.o_proj.scales": "model-00001-of-00002.safetensors",
|
| 80 |
+
"model.layers.10.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 81 |
+
"model.layers.10.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 82 |
+
"model.layers.10.self_attn.q_proj.biases": "model-00001-of-00002.safetensors",
|
| 83 |
+
"model.layers.10.self_attn.q_proj.scales": "model-00001-of-00002.safetensors",
|
| 84 |
+
"model.layers.10.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 85 |
+
"model.layers.10.self_attn.v_proj.biases": "model-00001-of-00002.safetensors",
|
| 86 |
+
"model.layers.10.self_attn.v_proj.scales": "model-00001-of-00002.safetensors",
|
| 87 |
+
"model.layers.10.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 88 |
+
"model.layers.11.mlp.down_proj.biases": "model-00001-of-00002.safetensors",
|
| 89 |
+
"model.layers.11.mlp.down_proj.scales": "model-00001-of-00002.safetensors",
|
| 90 |
+
"model.layers.11.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 91 |
+
"model.layers.11.mlp.gate_proj.biases": "model-00001-of-00002.safetensors",
|
| 92 |
+
"model.layers.11.mlp.gate_proj.scales": "model-00001-of-00002.safetensors",
|
| 93 |
+
"model.layers.11.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 94 |
+
"model.layers.11.mlp.up_proj.biases": "model-00001-of-00002.safetensors",
|
| 95 |
+
"model.layers.11.mlp.up_proj.scales": "model-00001-of-00002.safetensors",
|
| 96 |
+
"model.layers.11.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 97 |
+
"model.layers.11.post_attn_norm.weight": "model-00001-of-00002.safetensors",
|
| 98 |
+
"model.layers.11.post_mlp_norm.weight": "model-00001-of-00002.safetensors",
|
| 99 |
+
"model.layers.11.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 100 |
+
"model.layers.11.self_attn.k_proj.biases": "model-00001-of-00002.safetensors",
|
| 101 |
+
"model.layers.11.self_attn.k_proj.scales": "model-00001-of-00002.safetensors",
|
| 102 |
+
"model.layers.11.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 103 |
+
"model.layers.11.self_attn.o_proj.biases": "model-00001-of-00002.safetensors",
|
| 104 |
+
"model.layers.11.self_attn.o_proj.scales": "model-00001-of-00002.safetensors",
|
| 105 |
+
"model.layers.11.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 106 |
+
"model.layers.11.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 107 |
+
"model.layers.11.self_attn.q_proj.biases": "model-00001-of-00002.safetensors",
|
| 108 |
+
"model.layers.11.self_attn.q_proj.scales": "model-00001-of-00002.safetensors",
|
| 109 |
+
"model.layers.11.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 110 |
+
"model.layers.11.self_attn.v_proj.biases": "model-00001-of-00002.safetensors",
|
| 111 |
+
"model.layers.11.self_attn.v_proj.scales": "model-00001-of-00002.safetensors",
|
| 112 |
+
"model.layers.11.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 113 |
+
"model.layers.12.mlp.down_proj.biases": "model-00001-of-00002.safetensors",
|
| 114 |
+
"model.layers.12.mlp.down_proj.scales": "model-00001-of-00002.safetensors",
|
| 115 |
+
"model.layers.12.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 116 |
+
"model.layers.12.mlp.gate_proj.biases": "model-00001-of-00002.safetensors",
|
| 117 |
+
"model.layers.12.mlp.gate_proj.scales": "model-00001-of-00002.safetensors",
|
| 118 |
+
"model.layers.12.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 119 |
+
"model.layers.12.mlp.up_proj.biases": "model-00001-of-00002.safetensors",
|
| 120 |
+
"model.layers.12.mlp.up_proj.scales": "model-00001-of-00002.safetensors",
|
| 121 |
+
"model.layers.12.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 122 |
+
"model.layers.12.post_attn_norm.weight": "model-00001-of-00002.safetensors",
|
| 123 |
+
"model.layers.12.post_mlp_norm.weight": "model-00001-of-00002.safetensors",
|
| 124 |
+
"model.layers.12.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 125 |
+
"model.layers.12.self_attn.k_proj.biases": "model-00001-of-00002.safetensors",
|
| 126 |
+
"model.layers.12.self_attn.k_proj.scales": "model-00001-of-00002.safetensors",
|
| 127 |
+
"model.layers.12.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 128 |
+
"model.layers.12.self_attn.o_proj.biases": "model-00001-of-00002.safetensors",
|
| 129 |
+
"model.layers.12.self_attn.o_proj.scales": "model-00001-of-00002.safetensors",
|
| 130 |
+
"model.layers.12.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 131 |
+
"model.layers.12.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 132 |
+
"model.layers.12.self_attn.q_proj.biases": "model-00001-of-00002.safetensors",
|
| 133 |
+
"model.layers.12.self_attn.q_proj.scales": "model-00001-of-00002.safetensors",
|
| 134 |
+
"model.layers.12.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 135 |
+
"model.layers.12.self_attn.v_proj.biases": "model-00001-of-00002.safetensors",
|
| 136 |
+
"model.layers.12.self_attn.v_proj.scales": "model-00001-of-00002.safetensors",
|
| 137 |
+
"model.layers.12.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 138 |
+
"model.layers.13.mlp.down_proj.biases": "model-00001-of-00002.safetensors",
|
| 139 |
+
"model.layers.13.mlp.down_proj.scales": "model-00001-of-00002.safetensors",
|
| 140 |
+
"model.layers.13.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 141 |
+
"model.layers.13.mlp.gate_proj.biases": "model-00001-of-00002.safetensors",
|
| 142 |
+
"model.layers.13.mlp.gate_proj.scales": "model-00001-of-00002.safetensors",
|
| 143 |
+
"model.layers.13.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 144 |
+
"model.layers.13.mlp.up_proj.biases": "model-00001-of-00002.safetensors",
|
| 145 |
+
"model.layers.13.mlp.up_proj.scales": "model-00001-of-00002.safetensors",
|
| 146 |
+
"model.layers.13.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 147 |
+
"model.layers.13.post_attn_norm.weight": "model-00001-of-00002.safetensors",
|
| 148 |
+
"model.layers.13.post_mlp_norm.weight": "model-00001-of-00002.safetensors",
|
| 149 |
+
"model.layers.13.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 150 |
+
"model.layers.13.self_attn.k_proj.biases": "model-00001-of-00002.safetensors",
|
| 151 |
+
"model.layers.13.self_attn.k_proj.scales": "model-00001-of-00002.safetensors",
|
| 152 |
+
"model.layers.13.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 153 |
+
"model.layers.13.self_attn.o_proj.biases": "model-00001-of-00002.safetensors",
|
| 154 |
+
"model.layers.13.self_attn.o_proj.scales": "model-00001-of-00002.safetensors",
|
| 155 |
+
"model.layers.13.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 156 |
+
"model.layers.13.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 157 |
+
"model.layers.13.self_attn.q_proj.biases": "model-00001-of-00002.safetensors",
|
| 158 |
+
"model.layers.13.self_attn.q_proj.scales": "model-00001-of-00002.safetensors",
|
| 159 |
+
"model.layers.13.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 160 |
+
"model.layers.13.self_attn.v_proj.biases": "model-00001-of-00002.safetensors",
|
| 161 |
+
"model.layers.13.self_attn.v_proj.scales": "model-00001-of-00002.safetensors",
|
| 162 |
+
"model.layers.13.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 163 |
+
"model.layers.14.mlp.down_proj.biases": "model-00001-of-00002.safetensors",
|
| 164 |
+
"model.layers.14.mlp.down_proj.scales": "model-00001-of-00002.safetensors",
|
| 165 |
+
"model.layers.14.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 166 |
+
"model.layers.14.mlp.gate_proj.biases": "model-00001-of-00002.safetensors",
|
| 167 |
+
"model.layers.14.mlp.gate_proj.scales": "model-00001-of-00002.safetensors",
|
| 168 |
+
"model.layers.14.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 169 |
+
"model.layers.14.mlp.up_proj.biases": "model-00001-of-00002.safetensors",
|
| 170 |
+
"model.layers.14.mlp.up_proj.scales": "model-00001-of-00002.safetensors",
|
| 171 |
+
"model.layers.14.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 172 |
+
"model.layers.14.post_attn_norm.weight": "model-00001-of-00002.safetensors",
|
| 173 |
+
"model.layers.14.post_mlp_norm.weight": "model-00001-of-00002.safetensors",
|
| 174 |
+
"model.layers.14.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 175 |
+
"model.layers.14.self_attn.k_proj.biases": "model-00001-of-00002.safetensors",
|
| 176 |
+
"model.layers.14.self_attn.k_proj.scales": "model-00001-of-00002.safetensors",
|
| 177 |
+
"model.layers.14.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 178 |
+
"model.layers.14.self_attn.o_proj.biases": "model-00001-of-00002.safetensors",
|
| 179 |
+
"model.layers.14.self_attn.o_proj.scales": "model-00001-of-00002.safetensors",
|
| 180 |
+
"model.layers.14.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 181 |
+
"model.layers.14.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 182 |
+
"model.layers.14.self_attn.q_proj.biases": "model-00001-of-00002.safetensors",
|
| 183 |
+
"model.layers.14.self_attn.q_proj.scales": "model-00001-of-00002.safetensors",
|
| 184 |
+
"model.layers.14.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 185 |
+
"model.layers.14.self_attn.v_proj.biases": "model-00001-of-00002.safetensors",
|
| 186 |
+
"model.layers.14.self_attn.v_proj.scales": "model-00001-of-00002.safetensors",
|
| 187 |
+
"model.layers.14.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 188 |
+
"model.layers.15.mlp.down_proj.biases": "model-00001-of-00002.safetensors",
|
| 189 |
+
"model.layers.15.mlp.down_proj.scales": "model-00001-of-00002.safetensors",
|
| 190 |
+
"model.layers.15.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 191 |
+
"model.layers.15.mlp.gate_proj.biases": "model-00001-of-00002.safetensors",
|
| 192 |
+
"model.layers.15.mlp.gate_proj.scales": "model-00001-of-00002.safetensors",
|
| 193 |
+
"model.layers.15.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 194 |
+
"model.layers.15.mlp.up_proj.biases": "model-00001-of-00002.safetensors",
|
| 195 |
+
"model.layers.15.mlp.up_proj.scales": "model-00001-of-00002.safetensors",
|
| 196 |
+
"model.layers.15.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 197 |
+
"model.layers.15.post_attn_norm.weight": "model-00001-of-00002.safetensors",
|
| 198 |
+
"model.layers.15.post_mlp_norm.weight": "model-00001-of-00002.safetensors",
|
| 199 |
+
"model.layers.15.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 200 |
+
"model.layers.15.self_attn.k_proj.biases": "model-00001-of-00002.safetensors",
|
| 201 |
+
"model.layers.15.self_attn.k_proj.scales": "model-00001-of-00002.safetensors",
|
| 202 |
+
"model.layers.15.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 203 |
+
"model.layers.15.self_attn.o_proj.biases": "model-00001-of-00002.safetensors",
|
| 204 |
+
"model.layers.15.self_attn.o_proj.scales": "model-00001-of-00002.safetensors",
|
| 205 |
+
"model.layers.15.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 206 |
+
"model.layers.15.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 207 |
+
"model.layers.15.self_attn.q_proj.biases": "model-00001-of-00002.safetensors",
|
| 208 |
+
"model.layers.15.self_attn.q_proj.scales": "model-00001-of-00002.safetensors",
|
| 209 |
+
"model.layers.15.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 210 |
+
"model.layers.15.self_attn.v_proj.biases": "model-00001-of-00002.safetensors",
|
| 211 |
+
"model.layers.15.self_attn.v_proj.scales": "model-00001-of-00002.safetensors",
|
| 212 |
+
"model.layers.15.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 213 |
+
"model.layers.16.mlp.down_proj.biases": "model-00001-of-00002.safetensors",
|
| 214 |
+
"model.layers.16.mlp.down_proj.scales": "model-00001-of-00002.safetensors",
|
| 215 |
+
"model.layers.16.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 216 |
+
"model.layers.16.mlp.gate_proj.biases": "model-00001-of-00002.safetensors",
|
| 217 |
+
"model.layers.16.mlp.gate_proj.scales": "model-00001-of-00002.safetensors",
|
| 218 |
+
"model.layers.16.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 219 |
+
"model.layers.16.mlp.up_proj.biases": "model-00001-of-00002.safetensors",
|
| 220 |
+
"model.layers.16.mlp.up_proj.scales": "model-00001-of-00002.safetensors",
|
| 221 |
+
"model.layers.16.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 222 |
+
"model.layers.16.post_attn_norm.weight": "model-00001-of-00002.safetensors",
|
| 223 |
+
"model.layers.16.post_mlp_norm.weight": "model-00001-of-00002.safetensors",
|
| 224 |
+
"model.layers.16.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 225 |
+
"model.layers.16.self_attn.k_proj.biases": "model-00001-of-00002.safetensors",
|
| 226 |
+
"model.layers.16.self_attn.k_proj.scales": "model-00001-of-00002.safetensors",
|
| 227 |
+
"model.layers.16.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 228 |
+
"model.layers.16.self_attn.o_proj.biases": "model-00001-of-00002.safetensors",
|
| 229 |
+
"model.layers.16.self_attn.o_proj.scales": "model-00001-of-00002.safetensors",
|
| 230 |
+
"model.layers.16.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 231 |
+
"model.layers.16.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 232 |
+
"model.layers.16.self_attn.q_proj.biases": "model-00001-of-00002.safetensors",
|
| 233 |
+
"model.layers.16.self_attn.q_proj.scales": "model-00001-of-00002.safetensors",
|
| 234 |
+
"model.layers.16.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 235 |
+
"model.layers.16.self_attn.v_proj.biases": "model-00001-of-00002.safetensors",
|
| 236 |
+
"model.layers.16.self_attn.v_proj.scales": "model-00001-of-00002.safetensors",
|
| 237 |
+
"model.layers.16.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 238 |
+
"model.layers.17.mlp.down_proj.biases": "model-00001-of-00002.safetensors",
|
| 239 |
+
"model.layers.17.mlp.down_proj.scales": "model-00001-of-00002.safetensors",
|
| 240 |
+
"model.layers.17.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 241 |
+
"model.layers.17.mlp.gate_proj.biases": "model-00001-of-00002.safetensors",
|
| 242 |
+
"model.layers.17.mlp.gate_proj.scales": "model-00001-of-00002.safetensors",
|
| 243 |
+
"model.layers.17.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 244 |
+
"model.layers.17.mlp.up_proj.biases": "model-00001-of-00002.safetensors",
|
| 245 |
+
"model.layers.17.mlp.up_proj.scales": "model-00001-of-00002.safetensors",
|
| 246 |
+
"model.layers.17.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 247 |
+
"model.layers.17.post_attn_norm.weight": "model-00001-of-00002.safetensors",
|
| 248 |
+
"model.layers.17.post_mlp_norm.weight": "model-00001-of-00002.safetensors",
|
| 249 |
+
"model.layers.17.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 250 |
+
"model.layers.17.self_attn.k_proj.biases": "model-00001-of-00002.safetensors",
|
| 251 |
+
"model.layers.17.self_attn.k_proj.scales": "model-00001-of-00002.safetensors",
|
| 252 |
+
"model.layers.17.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 253 |
+
"model.layers.17.self_attn.o_proj.biases": "model-00001-of-00002.safetensors",
|
| 254 |
+
"model.layers.17.self_attn.o_proj.scales": "model-00001-of-00002.safetensors",
|
| 255 |
+
"model.layers.17.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 256 |
+
"model.layers.17.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 257 |
+
"model.layers.17.self_attn.q_proj.biases": "model-00001-of-00002.safetensors",
|
| 258 |
+
"model.layers.17.self_attn.q_proj.scales": "model-00001-of-00002.safetensors",
|
| 259 |
+
"model.layers.17.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 260 |
+
"model.layers.17.self_attn.v_proj.biases": "model-00001-of-00002.safetensors",
|
| 261 |
+
"model.layers.17.self_attn.v_proj.scales": "model-00001-of-00002.safetensors",
|
| 262 |
+
"model.layers.17.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 263 |
+
"model.layers.18.mlp.down_proj.biases": "model-00001-of-00002.safetensors",
|
| 264 |
+
"model.layers.18.mlp.down_proj.scales": "model-00001-of-00002.safetensors",
|
| 265 |
+
"model.layers.18.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 266 |
+
"model.layers.18.mlp.gate_proj.biases": "model-00001-of-00002.safetensors",
|
| 267 |
+
"model.layers.18.mlp.gate_proj.scales": "model-00001-of-00002.safetensors",
|
| 268 |
+
"model.layers.18.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 269 |
+
"model.layers.18.mlp.up_proj.biases": "model-00001-of-00002.safetensors",
|
| 270 |
+
"model.layers.18.mlp.up_proj.scales": "model-00001-of-00002.safetensors",
|
| 271 |
+
"model.layers.18.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 272 |
+
"model.layers.18.post_attn_norm.weight": "model-00001-of-00002.safetensors",
|
| 273 |
+
"model.layers.18.post_mlp_norm.weight": "model-00001-of-00002.safetensors",
|
| 274 |
+
"model.layers.18.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 275 |
+
"model.layers.18.self_attn.k_proj.biases": "model-00001-of-00002.safetensors",
|
| 276 |
+
"model.layers.18.self_attn.k_proj.scales": "model-00001-of-00002.safetensors",
|
| 277 |
+
"model.layers.18.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 278 |
+
"model.layers.18.self_attn.o_proj.biases": "model-00001-of-00002.safetensors",
|
| 279 |
+
"model.layers.18.self_attn.o_proj.scales": "model-00001-of-00002.safetensors",
|
| 280 |
+
"model.layers.18.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 281 |
+
"model.layers.18.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 282 |
+
"model.layers.18.self_attn.q_proj.biases": "model-00001-of-00002.safetensors",
|
| 283 |
+
"model.layers.18.self_attn.q_proj.scales": "model-00001-of-00002.safetensors",
|
| 284 |
+
"model.layers.18.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 285 |
+
"model.layers.18.self_attn.v_proj.biases": "model-00001-of-00002.safetensors",
|
| 286 |
+
"model.layers.18.self_attn.v_proj.scales": "model-00001-of-00002.safetensors",
|
| 287 |
+
"model.layers.18.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 288 |
+
"model.layers.19.mlp.down_proj.biases": "model-00001-of-00002.safetensors",
|
| 289 |
+
"model.layers.19.mlp.down_proj.scales": "model-00001-of-00002.safetensors",
|
| 290 |
+
"model.layers.19.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 291 |
+
"model.layers.19.mlp.gate_proj.biases": "model-00001-of-00002.safetensors",
|
| 292 |
+
"model.layers.19.mlp.gate_proj.scales": "model-00001-of-00002.safetensors",
|
| 293 |
+
"model.layers.19.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 294 |
+
"model.layers.19.mlp.up_proj.biases": "model-00001-of-00002.safetensors",
|
| 295 |
+
"model.layers.19.mlp.up_proj.scales": "model-00001-of-00002.safetensors",
|
| 296 |
+
"model.layers.19.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 297 |
+
"model.layers.19.post_attn_norm.weight": "model-00001-of-00002.safetensors",
|
| 298 |
+
"model.layers.19.post_mlp_norm.weight": "model-00001-of-00002.safetensors",
|
| 299 |
+
"model.layers.19.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 300 |
+
"model.layers.19.self_attn.k_proj.biases": "model-00001-of-00002.safetensors",
|
| 301 |
+
"model.layers.19.self_attn.k_proj.scales": "model-00001-of-00002.safetensors",
|
| 302 |
+
"model.layers.19.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 303 |
+
"model.layers.19.self_attn.o_proj.biases": "model-00001-of-00002.safetensors",
|
| 304 |
+
"model.layers.19.self_attn.o_proj.scales": "model-00001-of-00002.safetensors",
|
| 305 |
+
"model.layers.19.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 306 |
+
"model.layers.19.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 307 |
+
"model.layers.19.self_attn.q_proj.biases": "model-00001-of-00002.safetensors",
|
| 308 |
+
"model.layers.19.self_attn.q_proj.scales": "model-00001-of-00002.safetensors",
|
| 309 |
+
"model.layers.19.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 310 |
+
"model.layers.19.self_attn.v_proj.biases": "model-00001-of-00002.safetensors",
|
| 311 |
+
"model.layers.19.self_attn.v_proj.scales": "model-00001-of-00002.safetensors",
|
| 312 |
+
"model.layers.19.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 313 |
+
"model.layers.2.mlp.down_proj.biases": "model-00001-of-00002.safetensors",
|
| 314 |
+
"model.layers.2.mlp.down_proj.scales": "model-00001-of-00002.safetensors",
|
| 315 |
+
"model.layers.2.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 316 |
+
"model.layers.2.mlp.gate_proj.biases": "model-00001-of-00002.safetensors",
|
| 317 |
+
"model.layers.2.mlp.gate_proj.scales": "model-00001-of-00002.safetensors",
|
| 318 |
+
"model.layers.2.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 319 |
+
"model.layers.2.mlp.up_proj.biases": "model-00001-of-00002.safetensors",
|
| 320 |
+
"model.layers.2.mlp.up_proj.scales": "model-00001-of-00002.safetensors",
|
| 321 |
+
"model.layers.2.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 322 |
+
"model.layers.2.post_attn_norm.weight": "model-00001-of-00002.safetensors",
|
| 323 |
+
"model.layers.2.post_mlp_norm.weight": "model-00001-of-00002.safetensors",
|
| 324 |
+
"model.layers.2.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 325 |
+
"model.layers.2.self_attn.k_proj.biases": "model-00001-of-00002.safetensors",
|
| 326 |
+
"model.layers.2.self_attn.k_proj.scales": "model-00001-of-00002.safetensors",
|
| 327 |
+
"model.layers.2.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 328 |
+
"model.layers.2.self_attn.o_proj.biases": "model-00001-of-00002.safetensors",
|
| 329 |
+
"model.layers.2.self_attn.o_proj.scales": "model-00001-of-00002.safetensors",
|
| 330 |
+
"model.layers.2.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 331 |
+
"model.layers.2.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 332 |
+
"model.layers.2.self_attn.q_proj.biases": "model-00001-of-00002.safetensors",
|
| 333 |
+
"model.layers.2.self_attn.q_proj.scales": "model-00001-of-00002.safetensors",
|
| 334 |
+
"model.layers.2.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 335 |
+
"model.layers.2.self_attn.v_proj.biases": "model-00001-of-00002.safetensors",
|
| 336 |
+
"model.layers.2.self_attn.v_proj.scales": "model-00001-of-00002.safetensors",
|
| 337 |
+
"model.layers.2.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 338 |
+
"model.layers.20.mlp.down_proj.biases": "model-00001-of-00002.safetensors",
|
| 339 |
+
"model.layers.20.mlp.down_proj.scales": "model-00001-of-00002.safetensors",
|
| 340 |
+
"model.layers.20.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 341 |
+
"model.layers.20.mlp.gate_proj.biases": "model-00001-of-00002.safetensors",
|
| 342 |
+
"model.layers.20.mlp.gate_proj.scales": "model-00001-of-00002.safetensors",
|
| 343 |
+
"model.layers.20.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 344 |
+
"model.layers.20.mlp.up_proj.biases": "model-00001-of-00002.safetensors",
|
| 345 |
+
"model.layers.20.mlp.up_proj.scales": "model-00001-of-00002.safetensors",
|
| 346 |
+
"model.layers.20.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 347 |
+
"model.layers.20.post_attn_norm.weight": "model-00001-of-00002.safetensors",
|
| 348 |
+
"model.layers.20.post_mlp_norm.weight": "model-00001-of-00002.safetensors",
|
| 349 |
+
"model.layers.20.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 350 |
+
"model.layers.20.self_attn.k_proj.biases": "model-00001-of-00002.safetensors",
|
| 351 |
+
"model.layers.20.self_attn.k_proj.scales": "model-00001-of-00002.safetensors",
|
| 352 |
+
"model.layers.20.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 353 |
+
"model.layers.20.self_attn.o_proj.biases": "model-00001-of-00002.safetensors",
|
| 354 |
+
"model.layers.20.self_attn.o_proj.scales": "model-00001-of-00002.safetensors",
|
| 355 |
+
"model.layers.20.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 356 |
+
"model.layers.20.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 357 |
+
"model.layers.20.self_attn.q_proj.biases": "model-00001-of-00002.safetensors",
|
| 358 |
+
"model.layers.20.self_attn.q_proj.scales": "model-00001-of-00002.safetensors",
|
| 359 |
+
"model.layers.20.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 360 |
+
"model.layers.20.self_attn.v_proj.biases": "model-00001-of-00002.safetensors",
|
| 361 |
+
"model.layers.20.self_attn.v_proj.scales": "model-00001-of-00002.safetensors",
|
| 362 |
+
"model.layers.20.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 363 |
+
"model.layers.21.mlp.down_proj.biases": "model-00001-of-00002.safetensors",
|
| 364 |
+
"model.layers.21.mlp.down_proj.scales": "model-00001-of-00002.safetensors",
|
| 365 |
+
"model.layers.21.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 366 |
+
"model.layers.21.mlp.gate_proj.biases": "model-00001-of-00002.safetensors",
|
| 367 |
+
"model.layers.21.mlp.gate_proj.scales": "model-00001-of-00002.safetensors",
|
| 368 |
+
"model.layers.21.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 369 |
+
"model.layers.21.mlp.up_proj.biases": "model-00001-of-00002.safetensors",
|
| 370 |
+
"model.layers.21.mlp.up_proj.scales": "model-00001-of-00002.safetensors",
|
| 371 |
+
"model.layers.21.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 372 |
+
"model.layers.21.post_attn_norm.weight": "model-00001-of-00002.safetensors",
|
| 373 |
+
"model.layers.21.post_mlp_norm.weight": "model-00001-of-00002.safetensors",
|
| 374 |
+
"model.layers.21.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 375 |
+
"model.layers.21.self_attn.k_proj.biases": "model-00001-of-00002.safetensors",
|
| 376 |
+
"model.layers.21.self_attn.k_proj.scales": "model-00001-of-00002.safetensors",
|
| 377 |
+
"model.layers.21.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 378 |
+
"model.layers.21.self_attn.o_proj.biases": "model-00001-of-00002.safetensors",
|
| 379 |
+
"model.layers.21.self_attn.o_proj.scales": "model-00001-of-00002.safetensors",
|
| 380 |
+
"model.layers.21.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 381 |
+
"model.layers.21.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 382 |
+
"model.layers.21.self_attn.q_proj.biases": "model-00001-of-00002.safetensors",
|
| 383 |
+
"model.layers.21.self_attn.q_proj.scales": "model-00001-of-00002.safetensors",
|
| 384 |
+
"model.layers.21.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 385 |
+
"model.layers.21.self_attn.v_proj.biases": "model-00001-of-00002.safetensors",
|
| 386 |
+
"model.layers.21.self_attn.v_proj.scales": "model-00001-of-00002.safetensors",
|
| 387 |
+
"model.layers.21.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 388 |
+
"model.layers.22.mlp.down_proj.biases": "model-00002-of-00002.safetensors",
|
| 389 |
+
"model.layers.22.mlp.down_proj.scales": "model-00002-of-00002.safetensors",
|
| 390 |
+
"model.layers.22.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 391 |
+
"model.layers.22.mlp.gate_proj.biases": "model-00001-of-00002.safetensors",
|
| 392 |
+
"model.layers.22.mlp.gate_proj.scales": "model-00001-of-00002.safetensors",
|
| 393 |
+
"model.layers.22.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 394 |
+
"model.layers.22.mlp.up_proj.biases": "model-00001-of-00002.safetensors",
|
| 395 |
+
"model.layers.22.mlp.up_proj.scales": "model-00001-of-00002.safetensors",
|
| 396 |
+
"model.layers.22.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 397 |
+
"model.layers.22.post_attn_norm.weight": "model-00002-of-00002.safetensors",
|
| 398 |
+
"model.layers.22.post_mlp_norm.weight": "model-00002-of-00002.safetensors",
|
| 399 |
+
"model.layers.22.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 400 |
+
"model.layers.22.self_attn.k_proj.biases": "model-00001-of-00002.safetensors",
|
| 401 |
+
"model.layers.22.self_attn.k_proj.scales": "model-00001-of-00002.safetensors",
|
| 402 |
+
"model.layers.22.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 403 |
+
"model.layers.22.self_attn.o_proj.biases": "model-00001-of-00002.safetensors",
|
| 404 |
+
"model.layers.22.self_attn.o_proj.scales": "model-00001-of-00002.safetensors",
|
| 405 |
+
"model.layers.22.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 406 |
+
"model.layers.22.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 407 |
+
"model.layers.22.self_attn.q_proj.biases": "model-00001-of-00002.safetensors",
|
| 408 |
+
"model.layers.22.self_attn.q_proj.scales": "model-00001-of-00002.safetensors",
|
| 409 |
+
"model.layers.22.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 410 |
+
"model.layers.22.self_attn.v_proj.biases": "model-00001-of-00002.safetensors",
|
| 411 |
+
"model.layers.22.self_attn.v_proj.scales": "model-00001-of-00002.safetensors",
|
| 412 |
+
"model.layers.22.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 413 |
+
"model.layers.23.mlp.down_proj.biases": "model-00002-of-00002.safetensors",
|
| 414 |
+
"model.layers.23.mlp.down_proj.scales": "model-00002-of-00002.safetensors",
|
| 415 |
+
"model.layers.23.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 416 |
+
"model.layers.23.mlp.gate_proj.biases": "model-00002-of-00002.safetensors",
|
| 417 |
+
"model.layers.23.mlp.gate_proj.scales": "model-00002-of-00002.safetensors",
|
| 418 |
+
"model.layers.23.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 419 |
+
"model.layers.23.mlp.up_proj.biases": "model-00002-of-00002.safetensors",
|
| 420 |
+
"model.layers.23.mlp.up_proj.scales": "model-00002-of-00002.safetensors",
|
| 421 |
+
"model.layers.23.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 422 |
+
"model.layers.23.post_attn_norm.weight": "model-00002-of-00002.safetensors",
|
| 423 |
+
"model.layers.23.post_mlp_norm.weight": "model-00002-of-00002.safetensors",
|
| 424 |
+
"model.layers.23.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
|
| 425 |
+
"model.layers.23.self_attn.k_proj.biases": "model-00002-of-00002.safetensors",
|
| 426 |
+
"model.layers.23.self_attn.k_proj.scales": "model-00002-of-00002.safetensors",
|
| 427 |
+
"model.layers.23.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 428 |
+
"model.layers.23.self_attn.o_proj.biases": "model-00002-of-00002.safetensors",
|
| 429 |
+
"model.layers.23.self_attn.o_proj.scales": "model-00002-of-00002.safetensors",
|
| 430 |
+
"model.layers.23.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 431 |
+
"model.layers.23.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
|
| 432 |
+
"model.layers.23.self_attn.q_proj.biases": "model-00002-of-00002.safetensors",
|
| 433 |
+
"model.layers.23.self_attn.q_proj.scales": "model-00002-of-00002.safetensors",
|
| 434 |
+
"model.layers.23.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 435 |
+
"model.layers.23.self_attn.v_proj.biases": "model-00002-of-00002.safetensors",
|
| 436 |
+
"model.layers.23.self_attn.v_proj.scales": "model-00002-of-00002.safetensors",
|
| 437 |
+
"model.layers.23.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 438 |
+
"model.layers.24.mlp.down_proj.biases": "model-00002-of-00002.safetensors",
|
| 439 |
+
"model.layers.24.mlp.down_proj.scales": "model-00002-of-00002.safetensors",
|
| 440 |
+
"model.layers.24.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 441 |
+
"model.layers.24.mlp.gate_proj.biases": "model-00002-of-00002.safetensors",
|
| 442 |
+
"model.layers.24.mlp.gate_proj.scales": "model-00002-of-00002.safetensors",
|
| 443 |
+
"model.layers.24.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 444 |
+
"model.layers.24.mlp.up_proj.biases": "model-00002-of-00002.safetensors",
|
| 445 |
+
"model.layers.24.mlp.up_proj.scales": "model-00002-of-00002.safetensors",
|
| 446 |
+
"model.layers.24.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 447 |
+
"model.layers.24.post_attn_norm.weight": "model-00002-of-00002.safetensors",
|
| 448 |
+
"model.layers.24.post_mlp_norm.weight": "model-00002-of-00002.safetensors",
|
| 449 |
+
"model.layers.24.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
|
| 450 |
+
"model.layers.24.self_attn.k_proj.biases": "model-00002-of-00002.safetensors",
|
| 451 |
+
"model.layers.24.self_attn.k_proj.scales": "model-00002-of-00002.safetensors",
|
| 452 |
+
"model.layers.24.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 453 |
+
"model.layers.24.self_attn.o_proj.biases": "model-00002-of-00002.safetensors",
|
| 454 |
+
"model.layers.24.self_attn.o_proj.scales": "model-00002-of-00002.safetensors",
|
| 455 |
+
"model.layers.24.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 456 |
+
"model.layers.24.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
|
| 457 |
+
"model.layers.24.self_attn.q_proj.biases": "model-00002-of-00002.safetensors",
|
| 458 |
+
"model.layers.24.self_attn.q_proj.scales": "model-00002-of-00002.safetensors",
|
| 459 |
+
"model.layers.24.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 460 |
+
"model.layers.24.self_attn.v_proj.biases": "model-00002-of-00002.safetensors",
|
| 461 |
+
"model.layers.24.self_attn.v_proj.scales": "model-00002-of-00002.safetensors",
|
| 462 |
+
"model.layers.24.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 463 |
+
"model.layers.25.mlp.down_proj.biases": "model-00002-of-00002.safetensors",
|
| 464 |
+
"model.layers.25.mlp.down_proj.scales": "model-00002-of-00002.safetensors",
|
| 465 |
+
"model.layers.25.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 466 |
+
"model.layers.25.mlp.gate_proj.biases": "model-00002-of-00002.safetensors",
|
| 467 |
+
"model.layers.25.mlp.gate_proj.scales": "model-00002-of-00002.safetensors",
|
| 468 |
+
"model.layers.25.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 469 |
+
"model.layers.25.mlp.up_proj.biases": "model-00002-of-00002.safetensors",
|
| 470 |
+
"model.layers.25.mlp.up_proj.scales": "model-00002-of-00002.safetensors",
|
| 471 |
+
"model.layers.25.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 472 |
+
"model.layers.25.post_attn_norm.weight": "model-00002-of-00002.safetensors",
|
| 473 |
+
"model.layers.25.post_mlp_norm.weight": "model-00002-of-00002.safetensors",
|
| 474 |
+
"model.layers.25.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
|
| 475 |
+
"model.layers.25.self_attn.k_proj.biases": "model-00002-of-00002.safetensors",
|
| 476 |
+
"model.layers.25.self_attn.k_proj.scales": "model-00002-of-00002.safetensors",
|
| 477 |
+
"model.layers.25.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 478 |
+
"model.layers.25.self_attn.o_proj.biases": "model-00002-of-00002.safetensors",
|
| 479 |
+
"model.layers.25.self_attn.o_proj.scales": "model-00002-of-00002.safetensors",
|
| 480 |
+
"model.layers.25.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 481 |
+
"model.layers.25.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
|
| 482 |
+
"model.layers.25.self_attn.q_proj.biases": "model-00002-of-00002.safetensors",
|
| 483 |
+
"model.layers.25.self_attn.q_proj.scales": "model-00002-of-00002.safetensors",
|
| 484 |
+
"model.layers.25.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 485 |
+
"model.layers.25.self_attn.v_proj.biases": "model-00002-of-00002.safetensors",
|
| 486 |
+
"model.layers.25.self_attn.v_proj.scales": "model-00002-of-00002.safetensors",
|
| 487 |
+
"model.layers.25.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 488 |
+
"model.layers.26.mlp.down_proj.biases": "model-00002-of-00002.safetensors",
|
| 489 |
+
"model.layers.26.mlp.down_proj.scales": "model-00002-of-00002.safetensors",
|
| 490 |
+
"model.layers.26.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 491 |
+
"model.layers.26.mlp.gate_proj.biases": "model-00002-of-00002.safetensors",
|
| 492 |
+
"model.layers.26.mlp.gate_proj.scales": "model-00002-of-00002.safetensors",
|
| 493 |
+
"model.layers.26.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 494 |
+
"model.layers.26.mlp.up_proj.biases": "model-00002-of-00002.safetensors",
|
| 495 |
+
"model.layers.26.mlp.up_proj.scales": "model-00002-of-00002.safetensors",
|
| 496 |
+
"model.layers.26.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 497 |
+
"model.layers.26.post_attn_norm.weight": "model-00002-of-00002.safetensors",
|
| 498 |
+
"model.layers.26.post_mlp_norm.weight": "model-00002-of-00002.safetensors",
|
| 499 |
+
"model.layers.26.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
|
| 500 |
+
"model.layers.26.self_attn.k_proj.biases": "model-00002-of-00002.safetensors",
|
| 501 |
+
"model.layers.26.self_attn.k_proj.scales": "model-00002-of-00002.safetensors",
|
| 502 |
+
"model.layers.26.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 503 |
+
"model.layers.26.self_attn.o_proj.biases": "model-00002-of-00002.safetensors",
|
| 504 |
+
"model.layers.26.self_attn.o_proj.scales": "model-00002-of-00002.safetensors",
|
| 505 |
+
"model.layers.26.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 506 |
+
"model.layers.26.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
|
| 507 |
+
"model.layers.26.self_attn.q_proj.biases": "model-00002-of-00002.safetensors",
|
| 508 |
+
"model.layers.26.self_attn.q_proj.scales": "model-00002-of-00002.safetensors",
|
| 509 |
+
"model.layers.26.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 510 |
+
"model.layers.26.self_attn.v_proj.biases": "model-00002-of-00002.safetensors",
|
| 511 |
+
"model.layers.26.self_attn.v_proj.scales": "model-00002-of-00002.safetensors",
|
| 512 |
+
"model.layers.26.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 513 |
+
"model.layers.27.mlp.down_proj.biases": "model-00002-of-00002.safetensors",
|
| 514 |
+
"model.layers.27.mlp.down_proj.scales": "model-00002-of-00002.safetensors",
|
| 515 |
+
"model.layers.27.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 516 |
+
"model.layers.27.mlp.gate_proj.biases": "model-00002-of-00002.safetensors",
|
| 517 |
+
"model.layers.27.mlp.gate_proj.scales": "model-00002-of-00002.safetensors",
|
| 518 |
+
"model.layers.27.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 519 |
+
"model.layers.27.mlp.up_proj.biases": "model-00002-of-00002.safetensors",
|
| 520 |
+
"model.layers.27.mlp.up_proj.scales": "model-00002-of-00002.safetensors",
|
| 521 |
+
"model.layers.27.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 522 |
+
"model.layers.27.post_attn_norm.weight": "model-00002-of-00002.safetensors",
|
| 523 |
+
"model.layers.27.post_mlp_norm.weight": "model-00002-of-00002.safetensors",
|
| 524 |
+
"model.layers.27.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
|
| 525 |
+
"model.layers.27.self_attn.k_proj.biases": "model-00002-of-00002.safetensors",
|
| 526 |
+
"model.layers.27.self_attn.k_proj.scales": "model-00002-of-00002.safetensors",
|
| 527 |
+
"model.layers.27.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 528 |
+
"model.layers.27.self_attn.o_proj.biases": "model-00002-of-00002.safetensors",
|
| 529 |
+
"model.layers.27.self_attn.o_proj.scales": "model-00002-of-00002.safetensors",
|
| 530 |
+
"model.layers.27.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 531 |
+
"model.layers.27.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
|
| 532 |
+
"model.layers.27.self_attn.q_proj.biases": "model-00002-of-00002.safetensors",
|
| 533 |
+
"model.layers.27.self_attn.q_proj.scales": "model-00002-of-00002.safetensors",
|
| 534 |
+
"model.layers.27.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 535 |
+
"model.layers.27.self_attn.v_proj.biases": "model-00002-of-00002.safetensors",
|
| 536 |
+
"model.layers.27.self_attn.v_proj.scales": "model-00002-of-00002.safetensors",
|
| 537 |
+
"model.layers.27.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 538 |
+
"model.layers.28.mlp.down_proj.biases": "model-00002-of-00002.safetensors",
|
| 539 |
+
"model.layers.28.mlp.down_proj.scales": "model-00002-of-00002.safetensors",
|
| 540 |
+
"model.layers.28.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 541 |
+
"model.layers.28.mlp.gate_proj.biases": "model-00002-of-00002.safetensors",
|
| 542 |
+
"model.layers.28.mlp.gate_proj.scales": "model-00002-of-00002.safetensors",
|
| 543 |
+
"model.layers.28.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 544 |
+
"model.layers.28.mlp.up_proj.biases": "model-00002-of-00002.safetensors",
|
| 545 |
+
"model.layers.28.mlp.up_proj.scales": "model-00002-of-00002.safetensors",
|
| 546 |
+
"model.layers.28.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 547 |
+
"model.layers.28.post_attn_norm.weight": "model-00002-of-00002.safetensors",
|
| 548 |
+
"model.layers.28.post_mlp_norm.weight": "model-00002-of-00002.safetensors",
|
| 549 |
+
"model.layers.28.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
|
| 550 |
+
"model.layers.28.self_attn.k_proj.biases": "model-00002-of-00002.safetensors",
|
| 551 |
+
"model.layers.28.self_attn.k_proj.scales": "model-00002-of-00002.safetensors",
|
| 552 |
+
"model.layers.28.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 553 |
+
"model.layers.28.self_attn.o_proj.biases": "model-00002-of-00002.safetensors",
|
| 554 |
+
"model.layers.28.self_attn.o_proj.scales": "model-00002-of-00002.safetensors",
|
| 555 |
+
"model.layers.28.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 556 |
+
"model.layers.28.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
|
| 557 |
+
"model.layers.28.self_attn.q_proj.biases": "model-00002-of-00002.safetensors",
|
| 558 |
+
"model.layers.28.self_attn.q_proj.scales": "model-00002-of-00002.safetensors",
|
| 559 |
+
"model.layers.28.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 560 |
+
"model.layers.28.self_attn.v_proj.biases": "model-00002-of-00002.safetensors",
|
| 561 |
+
"model.layers.28.self_attn.v_proj.scales": "model-00002-of-00002.safetensors",
|
| 562 |
+
"model.layers.28.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 563 |
+
"model.layers.29.mlp.down_proj.biases": "model-00002-of-00002.safetensors",
|
| 564 |
+
"model.layers.29.mlp.down_proj.scales": "model-00002-of-00002.safetensors",
|
| 565 |
+
"model.layers.29.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 566 |
+
"model.layers.29.mlp.gate_proj.biases": "model-00002-of-00002.safetensors",
|
| 567 |
+
"model.layers.29.mlp.gate_proj.scales": "model-00002-of-00002.safetensors",
|
| 568 |
+
"model.layers.29.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 569 |
+
"model.layers.29.mlp.up_proj.biases": "model-00002-of-00002.safetensors",
|
| 570 |
+
"model.layers.29.mlp.up_proj.scales": "model-00002-of-00002.safetensors",
|
| 571 |
+
"model.layers.29.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 572 |
+
"model.layers.29.post_attn_norm.weight": "model-00002-of-00002.safetensors",
|
| 573 |
+
"model.layers.29.post_mlp_norm.weight": "model-00002-of-00002.safetensors",
|
| 574 |
+
"model.layers.29.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
|
| 575 |
+
"model.layers.29.self_attn.k_proj.biases": "model-00002-of-00002.safetensors",
|
| 576 |
+
"model.layers.29.self_attn.k_proj.scales": "model-00002-of-00002.safetensors",
|
| 577 |
+
"model.layers.29.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 578 |
+
"model.layers.29.self_attn.o_proj.biases": "model-00002-of-00002.safetensors",
|
| 579 |
+
"model.layers.29.self_attn.o_proj.scales": "model-00002-of-00002.safetensors",
|
| 580 |
+
"model.layers.29.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 581 |
+
"model.layers.29.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
|
| 582 |
+
"model.layers.29.self_attn.q_proj.biases": "model-00002-of-00002.safetensors",
|
| 583 |
+
"model.layers.29.self_attn.q_proj.scales": "model-00002-of-00002.safetensors",
|
| 584 |
+
"model.layers.29.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 585 |
+
"model.layers.29.self_attn.v_proj.biases": "model-00002-of-00002.safetensors",
|
| 586 |
+
"model.layers.29.self_attn.v_proj.scales": "model-00002-of-00002.safetensors",
|
| 587 |
+
"model.layers.29.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 588 |
+
"model.layers.3.mlp.down_proj.biases": "model-00001-of-00002.safetensors",
|
| 589 |
+
"model.layers.3.mlp.down_proj.scales": "model-00001-of-00002.safetensors",
|
| 590 |
+
"model.layers.3.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 591 |
+
"model.layers.3.mlp.gate_proj.biases": "model-00001-of-00002.safetensors",
|
| 592 |
+
"model.layers.3.mlp.gate_proj.scales": "model-00001-of-00002.safetensors",
|
| 593 |
+
"model.layers.3.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 594 |
+
"model.layers.3.mlp.up_proj.biases": "model-00001-of-00002.safetensors",
|
| 595 |
+
"model.layers.3.mlp.up_proj.scales": "model-00001-of-00002.safetensors",
|
| 596 |
+
"model.layers.3.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 597 |
+
"model.layers.3.post_attn_norm.weight": "model-00001-of-00002.safetensors",
|
| 598 |
+
"model.layers.3.post_mlp_norm.weight": "model-00001-of-00002.safetensors",
|
| 599 |
+
"model.layers.3.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 600 |
+
"model.layers.3.self_attn.k_proj.biases": "model-00001-of-00002.safetensors",
|
| 601 |
+
"model.layers.3.self_attn.k_proj.scales": "model-00001-of-00002.safetensors",
|
| 602 |
+
"model.layers.3.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 603 |
+
"model.layers.3.self_attn.o_proj.biases": "model-00001-of-00002.safetensors",
|
| 604 |
+
"model.layers.3.self_attn.o_proj.scales": "model-00001-of-00002.safetensors",
|
| 605 |
+
"model.layers.3.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 606 |
+
"model.layers.3.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 607 |
+
"model.layers.3.self_attn.q_proj.biases": "model-00001-of-00002.safetensors",
|
| 608 |
+
"model.layers.3.self_attn.q_proj.scales": "model-00001-of-00002.safetensors",
|
| 609 |
+
"model.layers.3.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 610 |
+
"model.layers.3.self_attn.v_proj.biases": "model-00001-of-00002.safetensors",
|
| 611 |
+
"model.layers.3.self_attn.v_proj.scales": "model-00001-of-00002.safetensors",
|
| 612 |
+
"model.layers.3.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 613 |
+
"model.layers.30.mlp.down_proj.biases": "model-00002-of-00002.safetensors",
|
| 614 |
+
"model.layers.30.mlp.down_proj.scales": "model-00002-of-00002.safetensors",
|
| 615 |
+
"model.layers.30.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 616 |
+
"model.layers.30.mlp.gate_proj.biases": "model-00002-of-00002.safetensors",
|
| 617 |
+
"model.layers.30.mlp.gate_proj.scales": "model-00002-of-00002.safetensors",
|
| 618 |
+
"model.layers.30.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 619 |
+
"model.layers.30.mlp.up_proj.biases": "model-00002-of-00002.safetensors",
|
| 620 |
+
"model.layers.30.mlp.up_proj.scales": "model-00002-of-00002.safetensors",
|
| 621 |
+
"model.layers.30.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 622 |
+
"model.layers.30.post_attn_norm.weight": "model-00002-of-00002.safetensors",
|
| 623 |
+
"model.layers.30.post_mlp_norm.weight": "model-00002-of-00002.safetensors",
|
| 624 |
+
"model.layers.30.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
|
| 625 |
+
"model.layers.30.self_attn.k_proj.biases": "model-00002-of-00002.safetensors",
|
| 626 |
+
"model.layers.30.self_attn.k_proj.scales": "model-00002-of-00002.safetensors",
|
| 627 |
+
"model.layers.30.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 628 |
+
"model.layers.30.self_attn.o_proj.biases": "model-00002-of-00002.safetensors",
|
| 629 |
+
"model.layers.30.self_attn.o_proj.scales": "model-00002-of-00002.safetensors",
|
| 630 |
+
"model.layers.30.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 631 |
+
"model.layers.30.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
|
| 632 |
+
"model.layers.30.self_attn.q_proj.biases": "model-00002-of-00002.safetensors",
|
| 633 |
+
"model.layers.30.self_attn.q_proj.scales": "model-00002-of-00002.safetensors",
|
| 634 |
+
"model.layers.30.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 635 |
+
"model.layers.30.self_attn.v_proj.biases": "model-00002-of-00002.safetensors",
|
| 636 |
+
"model.layers.30.self_attn.v_proj.scales": "model-00002-of-00002.safetensors",
|
| 637 |
+
"model.layers.30.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 638 |
+
"model.layers.31.mlp.down_proj.biases": "model-00002-of-00002.safetensors",
|
| 639 |
+
"model.layers.31.mlp.down_proj.scales": "model-00002-of-00002.safetensors",
|
| 640 |
+
"model.layers.31.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
|
| 641 |
+
"model.layers.31.mlp.gate_proj.biases": "model-00002-of-00002.safetensors",
|
| 642 |
+
"model.layers.31.mlp.gate_proj.scales": "model-00002-of-00002.safetensors",
|
| 643 |
+
"model.layers.31.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
|
| 644 |
+
"model.layers.31.mlp.up_proj.biases": "model-00002-of-00002.safetensors",
|
| 645 |
+
"model.layers.31.mlp.up_proj.scales": "model-00002-of-00002.safetensors",
|
| 646 |
+
"model.layers.31.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
|
| 647 |
+
"model.layers.31.post_attn_norm.weight": "model-00002-of-00002.safetensors",
|
| 648 |
+
"model.layers.31.post_mlp_norm.weight": "model-00002-of-00002.safetensors",
|
| 649 |
+
"model.layers.31.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
|
| 650 |
+
"model.layers.31.self_attn.k_proj.biases": "model-00002-of-00002.safetensors",
|
| 651 |
+
"model.layers.31.self_attn.k_proj.scales": "model-00002-of-00002.safetensors",
|
| 652 |
+
"model.layers.31.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
|
| 653 |
+
"model.layers.31.self_attn.o_proj.biases": "model-00002-of-00002.safetensors",
|
| 654 |
+
"model.layers.31.self_attn.o_proj.scales": "model-00002-of-00002.safetensors",
|
| 655 |
+
"model.layers.31.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
|
| 656 |
+
"model.layers.31.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
|
| 657 |
+
"model.layers.31.self_attn.q_proj.biases": "model-00002-of-00002.safetensors",
|
| 658 |
+
"model.layers.31.self_attn.q_proj.scales": "model-00002-of-00002.safetensors",
|
| 659 |
+
"model.layers.31.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
|
| 660 |
+
"model.layers.31.self_attn.v_proj.biases": "model-00002-of-00002.safetensors",
|
| 661 |
+
"model.layers.31.self_attn.v_proj.scales": "model-00002-of-00002.safetensors",
|
| 662 |
+
"model.layers.31.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
|
| 663 |
+
"model.layers.4.mlp.down_proj.biases": "model-00001-of-00002.safetensors",
|
| 664 |
+
"model.layers.4.mlp.down_proj.scales": "model-00001-of-00002.safetensors",
|
| 665 |
+
"model.layers.4.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 666 |
+
"model.layers.4.mlp.gate_proj.biases": "model-00001-of-00002.safetensors",
|
| 667 |
+
"model.layers.4.mlp.gate_proj.scales": "model-00001-of-00002.safetensors",
|
| 668 |
+
"model.layers.4.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 669 |
+
"model.layers.4.mlp.up_proj.biases": "model-00001-of-00002.safetensors",
|
| 670 |
+
"model.layers.4.mlp.up_proj.scales": "model-00001-of-00002.safetensors",
|
| 671 |
+
"model.layers.4.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 672 |
+
"model.layers.4.post_attn_norm.weight": "model-00001-of-00002.safetensors",
|
| 673 |
+
"model.layers.4.post_mlp_norm.weight": "model-00001-of-00002.safetensors",
|
| 674 |
+
"model.layers.4.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 675 |
+
"model.layers.4.self_attn.k_proj.biases": "model-00001-of-00002.safetensors",
|
| 676 |
+
"model.layers.4.self_attn.k_proj.scales": "model-00001-of-00002.safetensors",
|
| 677 |
+
"model.layers.4.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 678 |
+
"model.layers.4.self_attn.o_proj.biases": "model-00001-of-00002.safetensors",
|
| 679 |
+
"model.layers.4.self_attn.o_proj.scales": "model-00001-of-00002.safetensors",
|
| 680 |
+
"model.layers.4.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 681 |
+
"model.layers.4.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 682 |
+
"model.layers.4.self_attn.q_proj.biases": "model-00001-of-00002.safetensors",
|
| 683 |
+
"model.layers.4.self_attn.q_proj.scales": "model-00001-of-00002.safetensors",
|
| 684 |
+
"model.layers.4.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 685 |
+
"model.layers.4.self_attn.v_proj.biases": "model-00001-of-00002.safetensors",
|
| 686 |
+
"model.layers.4.self_attn.v_proj.scales": "model-00001-of-00002.safetensors",
|
| 687 |
+
"model.layers.4.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 688 |
+
"model.layers.5.mlp.down_proj.biases": "model-00001-of-00002.safetensors",
|
| 689 |
+
"model.layers.5.mlp.down_proj.scales": "model-00001-of-00002.safetensors",
|
| 690 |
+
"model.layers.5.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 691 |
+
"model.layers.5.mlp.gate_proj.biases": "model-00001-of-00002.safetensors",
|
| 692 |
+
"model.layers.5.mlp.gate_proj.scales": "model-00001-of-00002.safetensors",
|
| 693 |
+
"model.layers.5.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 694 |
+
"model.layers.5.mlp.up_proj.biases": "model-00001-of-00002.safetensors",
|
| 695 |
+
"model.layers.5.mlp.up_proj.scales": "model-00001-of-00002.safetensors",
|
| 696 |
+
"model.layers.5.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 697 |
+
"model.layers.5.post_attn_norm.weight": "model-00001-of-00002.safetensors",
|
| 698 |
+
"model.layers.5.post_mlp_norm.weight": "model-00001-of-00002.safetensors",
|
| 699 |
+
"model.layers.5.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 700 |
+
"model.layers.5.self_attn.k_proj.biases": "model-00001-of-00002.safetensors",
|
| 701 |
+
"model.layers.5.self_attn.k_proj.scales": "model-00001-of-00002.safetensors",
|
| 702 |
+
"model.layers.5.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 703 |
+
"model.layers.5.self_attn.o_proj.biases": "model-00001-of-00002.safetensors",
|
| 704 |
+
"model.layers.5.self_attn.o_proj.scales": "model-00001-of-00002.safetensors",
|
| 705 |
+
"model.layers.5.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 706 |
+
"model.layers.5.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 707 |
+
"model.layers.5.self_attn.q_proj.biases": "model-00001-of-00002.safetensors",
|
| 708 |
+
"model.layers.5.self_attn.q_proj.scales": "model-00001-of-00002.safetensors",
|
| 709 |
+
"model.layers.5.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 710 |
+
"model.layers.5.self_attn.v_proj.biases": "model-00001-of-00002.safetensors",
|
| 711 |
+
"model.layers.5.self_attn.v_proj.scales": "model-00001-of-00002.safetensors",
|
| 712 |
+
"model.layers.5.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 713 |
+
"model.layers.6.mlp.down_proj.biases": "model-00001-of-00002.safetensors",
|
| 714 |
+
"model.layers.6.mlp.down_proj.scales": "model-00001-of-00002.safetensors",
|
| 715 |
+
"model.layers.6.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 716 |
+
"model.layers.6.mlp.gate_proj.biases": "model-00001-of-00002.safetensors",
|
| 717 |
+
"model.layers.6.mlp.gate_proj.scales": "model-00001-of-00002.safetensors",
|
| 718 |
+
"model.layers.6.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 719 |
+
"model.layers.6.mlp.up_proj.biases": "model-00001-of-00002.safetensors",
|
| 720 |
+
"model.layers.6.mlp.up_proj.scales": "model-00001-of-00002.safetensors",
|
| 721 |
+
"model.layers.6.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 722 |
+
"model.layers.6.post_attn_norm.weight": "model-00001-of-00002.safetensors",
|
| 723 |
+
"model.layers.6.post_mlp_norm.weight": "model-00001-of-00002.safetensors",
|
| 724 |
+
"model.layers.6.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 725 |
+
"model.layers.6.self_attn.k_proj.biases": "model-00001-of-00002.safetensors",
|
| 726 |
+
"model.layers.6.self_attn.k_proj.scales": "model-00001-of-00002.safetensors",
|
| 727 |
+
"model.layers.6.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 728 |
+
"model.layers.6.self_attn.o_proj.biases": "model-00001-of-00002.safetensors",
|
| 729 |
+
"model.layers.6.self_attn.o_proj.scales": "model-00001-of-00002.safetensors",
|
| 730 |
+
"model.layers.6.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 731 |
+
"model.layers.6.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 732 |
+
"model.layers.6.self_attn.q_proj.biases": "model-00001-of-00002.safetensors",
|
| 733 |
+
"model.layers.6.self_attn.q_proj.scales": "model-00001-of-00002.safetensors",
|
| 734 |
+
"model.layers.6.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 735 |
+
"model.layers.6.self_attn.v_proj.biases": "model-00001-of-00002.safetensors",
|
| 736 |
+
"model.layers.6.self_attn.v_proj.scales": "model-00001-of-00002.safetensors",
|
| 737 |
+
"model.layers.6.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 738 |
+
"model.layers.7.mlp.down_proj.biases": "model-00001-of-00002.safetensors",
|
| 739 |
+
"model.layers.7.mlp.down_proj.scales": "model-00001-of-00002.safetensors",
|
| 740 |
+
"model.layers.7.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 741 |
+
"model.layers.7.mlp.gate_proj.biases": "model-00001-of-00002.safetensors",
|
| 742 |
+
"model.layers.7.mlp.gate_proj.scales": "model-00001-of-00002.safetensors",
|
| 743 |
+
"model.layers.7.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 744 |
+
"model.layers.7.mlp.up_proj.biases": "model-00001-of-00002.safetensors",
|
| 745 |
+
"model.layers.7.mlp.up_proj.scales": "model-00001-of-00002.safetensors",
|
| 746 |
+
"model.layers.7.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 747 |
+
"model.layers.7.post_attn_norm.weight": "model-00001-of-00002.safetensors",
|
| 748 |
+
"model.layers.7.post_mlp_norm.weight": "model-00001-of-00002.safetensors",
|
| 749 |
+
"model.layers.7.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 750 |
+
"model.layers.7.self_attn.k_proj.biases": "model-00001-of-00002.safetensors",
|
| 751 |
+
"model.layers.7.self_attn.k_proj.scales": "model-00001-of-00002.safetensors",
|
| 752 |
+
"model.layers.7.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 753 |
+
"model.layers.7.self_attn.o_proj.biases": "model-00001-of-00002.safetensors",
|
| 754 |
+
"model.layers.7.self_attn.o_proj.scales": "model-00001-of-00002.safetensors",
|
| 755 |
+
"model.layers.7.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 756 |
+
"model.layers.7.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 757 |
+
"model.layers.7.self_attn.q_proj.biases": "model-00001-of-00002.safetensors",
|
| 758 |
+
"model.layers.7.self_attn.q_proj.scales": "model-00001-of-00002.safetensors",
|
| 759 |
+
"model.layers.7.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 760 |
+
"model.layers.7.self_attn.v_proj.biases": "model-00001-of-00002.safetensors",
|
| 761 |
+
"model.layers.7.self_attn.v_proj.scales": "model-00001-of-00002.safetensors",
|
| 762 |
+
"model.layers.7.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 763 |
+
"model.layers.8.mlp.down_proj.biases": "model-00001-of-00002.safetensors",
|
| 764 |
+
"model.layers.8.mlp.down_proj.scales": "model-00001-of-00002.safetensors",
|
| 765 |
+
"model.layers.8.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 766 |
+
"model.layers.8.mlp.gate_proj.biases": "model-00001-of-00002.safetensors",
|
| 767 |
+
"model.layers.8.mlp.gate_proj.scales": "model-00001-of-00002.safetensors",
|
| 768 |
+
"model.layers.8.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 769 |
+
"model.layers.8.mlp.up_proj.biases": "model-00001-of-00002.safetensors",
|
| 770 |
+
"model.layers.8.mlp.up_proj.scales": "model-00001-of-00002.safetensors",
|
| 771 |
+
"model.layers.8.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 772 |
+
"model.layers.8.post_attn_norm.weight": "model-00001-of-00002.safetensors",
|
| 773 |
+
"model.layers.8.post_mlp_norm.weight": "model-00001-of-00002.safetensors",
|
| 774 |
+
"model.layers.8.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 775 |
+
"model.layers.8.self_attn.k_proj.biases": "model-00001-of-00002.safetensors",
|
| 776 |
+
"model.layers.8.self_attn.k_proj.scales": "model-00001-of-00002.safetensors",
|
| 777 |
+
"model.layers.8.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 778 |
+
"model.layers.8.self_attn.o_proj.biases": "model-00001-of-00002.safetensors",
|
| 779 |
+
"model.layers.8.self_attn.o_proj.scales": "model-00001-of-00002.safetensors",
|
| 780 |
+
"model.layers.8.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 781 |
+
"model.layers.8.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 782 |
+
"model.layers.8.self_attn.q_proj.biases": "model-00001-of-00002.safetensors",
|
| 783 |
+
"model.layers.8.self_attn.q_proj.scales": "model-00001-of-00002.safetensors",
|
| 784 |
+
"model.layers.8.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 785 |
+
"model.layers.8.self_attn.v_proj.biases": "model-00001-of-00002.safetensors",
|
| 786 |
+
"model.layers.8.self_attn.v_proj.scales": "model-00001-of-00002.safetensors",
|
| 787 |
+
"model.layers.8.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 788 |
+
"model.layers.9.mlp.down_proj.biases": "model-00001-of-00002.safetensors",
|
| 789 |
+
"model.layers.9.mlp.down_proj.scales": "model-00001-of-00002.safetensors",
|
| 790 |
+
"model.layers.9.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
|
| 791 |
+
"model.layers.9.mlp.gate_proj.biases": "model-00001-of-00002.safetensors",
|
| 792 |
+
"model.layers.9.mlp.gate_proj.scales": "model-00001-of-00002.safetensors",
|
| 793 |
+
"model.layers.9.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
|
| 794 |
+
"model.layers.9.mlp.up_proj.biases": "model-00001-of-00002.safetensors",
|
| 795 |
+
"model.layers.9.mlp.up_proj.scales": "model-00001-of-00002.safetensors",
|
| 796 |
+
"model.layers.9.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
|
| 797 |
+
"model.layers.9.post_attn_norm.weight": "model-00001-of-00002.safetensors",
|
| 798 |
+
"model.layers.9.post_mlp_norm.weight": "model-00001-of-00002.safetensors",
|
| 799 |
+
"model.layers.9.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
|
| 800 |
+
"model.layers.9.self_attn.k_proj.biases": "model-00001-of-00002.safetensors",
|
| 801 |
+
"model.layers.9.self_attn.k_proj.scales": "model-00001-of-00002.safetensors",
|
| 802 |
+
"model.layers.9.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
|
| 803 |
+
"model.layers.9.self_attn.o_proj.biases": "model-00001-of-00002.safetensors",
|
| 804 |
+
"model.layers.9.self_attn.o_proj.scales": "model-00001-of-00002.safetensors",
|
| 805 |
+
"model.layers.9.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
|
| 806 |
+
"model.layers.9.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
|
| 807 |
+
"model.layers.9.self_attn.q_proj.biases": "model-00001-of-00002.safetensors",
|
| 808 |
+
"model.layers.9.self_attn.q_proj.scales": "model-00001-of-00002.safetensors",
|
| 809 |
+
"model.layers.9.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
|
| 810 |
+
"model.layers.9.self_attn.v_proj.biases": "model-00001-of-00002.safetensors",
|
| 811 |
+
"model.layers.9.self_attn.v_proj.scales": "model-00001-of-00002.safetensors",
|
| 812 |
+
"model.layers.9.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
|
| 813 |
+
"model.norm.weight": "model-00002-of-00002.safetensors"
|
| 814 |
+
}
|
| 815 |
+
}
|
modeling_lizzy.py
ADDED
|
@@ -0,0 +1,2190 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from __future__ import annotations
|
| 2 |
+
|
| 3 |
+
import math
|
| 4 |
+
import os
|
| 5 |
+
from typing import Any, cast
|
| 6 |
+
|
| 7 |
+
import torch
|
| 8 |
+
import torch.nn.functional as F
|
| 9 |
+
from torch import nn
|
| 10 |
+
from torch.utils.checkpoint import checkpoint
|
| 11 |
+
from transformers.activations import ACT2FN
|
| 12 |
+
from transformers.cache_utils import Cache, DynamicCache
|
| 13 |
+
from transformers.generation import GenerationMixin
|
| 14 |
+
from transformers.modeling_outputs import (
|
| 15 |
+
BaseModelOutputWithPast,
|
| 16 |
+
CausalLMOutputWithPast,
|
| 17 |
+
)
|
| 18 |
+
from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
|
| 19 |
+
|
| 20 |
+
try:
|
| 21 |
+
from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS
|
| 22 |
+
except ImportError:
|
| 23 |
+
ROPE_INIT_FUNCTIONS = {}
|
| 24 |
+
|
| 25 |
+
try:
|
| 26 |
+
from fla.modules import FusedRMSNormGated, ShortConvolution
|
| 27 |
+
from fla.ops.gated_delta_rule import (
|
| 28 |
+
chunk_gated_delta_rule,
|
| 29 |
+
fused_recurrent_gated_delta_rule,
|
| 30 |
+
)
|
| 31 |
+
except ImportError:
|
| 32 |
+
chunk_gated_delta_rule = None
|
| 33 |
+
fused_recurrent_gated_delta_rule = None
|
| 34 |
+
FusedRMSNormGated = None
|
| 35 |
+
ShortConvolution = None
|
| 36 |
+
|
| 37 |
+
from .configuration_lizzy import LizzyConfig
|
| 38 |
+
|
| 39 |
+
|
| 40 |
+
class LizzyRMSNorm(nn.Module):
|
| 41 |
+
def __init__(self, hidden_size: int, eps: float = 1e-6) -> None:
|
| 42 |
+
super().__init__()
|
| 43 |
+
self.weight = nn.Parameter(torch.ones(hidden_size))
|
| 44 |
+
self.variance_epsilon = eps
|
| 45 |
+
|
| 46 |
+
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
| 47 |
+
input_dtype = hidden_states.dtype
|
| 48 |
+
hidden_states = hidden_states.to(torch.float32)
|
| 49 |
+
variance = hidden_states.pow(2).mean(dim=-1, keepdim=True)
|
| 50 |
+
hidden_states = hidden_states * torch.rsqrt(
|
| 51 |
+
variance + self.variance_epsilon
|
| 52 |
+
)
|
| 53 |
+
return self.weight * hidden_states.to(input_dtype)
|
| 54 |
+
|
| 55 |
+
|
| 56 |
+
def _make_norm(
|
| 57 |
+
norm_type: str,
|
| 58 |
+
hidden_size: int,
|
| 59 |
+
eps: float,
|
| 60 |
+
*,
|
| 61 |
+
has_bias: bool,
|
| 62 |
+
) -> nn.Module:
|
| 63 |
+
if norm_type == "rmsnorm":
|
| 64 |
+
return LizzyRMSNorm(hidden_size, eps=eps)
|
| 65 |
+
if norm_type == "layernorm":
|
| 66 |
+
return nn.LayerNorm(
|
| 67 |
+
hidden_size,
|
| 68 |
+
eps=eps,
|
| 69 |
+
elementwise_affine=True,
|
| 70 |
+
bias=has_bias,
|
| 71 |
+
)
|
| 72 |
+
msg = f"Unsupported norm_type: {norm_type}"
|
| 73 |
+
raise ValueError(msg)
|
| 74 |
+
|
| 75 |
+
|
| 76 |
+
def _rotate_half(x: torch.Tensor) -> torch.Tensor:
|
| 77 |
+
x1 = x[..., : x.shape[-1] // 2]
|
| 78 |
+
x2 = x[..., x.shape[-1] // 2 :]
|
| 79 |
+
return torch.cat((-x2, x1), dim=-1)
|
| 80 |
+
|
| 81 |
+
|
| 82 |
+
def _apply_rotary_pos_emb(
|
| 83 |
+
q: torch.Tensor,
|
| 84 |
+
k: torch.Tensor,
|
| 85 |
+
cos: torch.Tensor,
|
| 86 |
+
sin: torch.Tensor,
|
| 87 |
+
) -> tuple[torch.Tensor, torch.Tensor]:
|
| 88 |
+
q_embed = (q * cos) + (_rotate_half(q) * sin)
|
| 89 |
+
k_embed = (k * cos) + (_rotate_half(k) * sin)
|
| 90 |
+
return q_embed, k_embed
|
| 91 |
+
|
| 92 |
+
|
| 93 |
+
def _legacy_cache_length(
|
| 94 |
+
past_key_values: tuple[tuple[torch.Tensor, torch.Tensor], ...] | None,
|
| 95 |
+
) -> int:
|
| 96 |
+
if (
|
| 97 |
+
isinstance(past_key_values, tuple)
|
| 98 |
+
and len(past_key_values) > 0
|
| 99 |
+
and past_key_values[0] is not None
|
| 100 |
+
and past_key_values[0][0] is not None
|
| 101 |
+
):
|
| 102 |
+
return int(past_key_values[0][0].shape[2])
|
| 103 |
+
return 0
|
| 104 |
+
|
| 105 |
+
|
| 106 |
+
def _normalize_cache_position(
|
| 107 |
+
cache_position: torch.Tensor | None,
|
| 108 |
+
) -> torch.Tensor | None:
|
| 109 |
+
if cache_position is None:
|
| 110 |
+
return None
|
| 111 |
+
if cache_position.dim() == 0:
|
| 112 |
+
return cache_position.view(1)
|
| 113 |
+
if cache_position.dim() > 1:
|
| 114 |
+
return cache_position[0]
|
| 115 |
+
return cache_position
|
| 116 |
+
|
| 117 |
+
|
| 118 |
+
def _is_cache_object(value: Any) -> bool:
|
| 119 |
+
return isinstance(value, Cache) or isinstance(value, LizzyHybridDynamicCache)
|
| 120 |
+
|
| 121 |
+
|
| 122 |
+
def _compute_default_rope_parameters(
|
| 123 |
+
config: LizzyConfig,
|
| 124 |
+
device: torch.device,
|
| 125 |
+
) -> tuple[torch.Tensor, float]:
|
| 126 |
+
inv_freq = 1.0 / (
|
| 127 |
+
config.rope_theta
|
| 128 |
+
** (
|
| 129 |
+
torch.arange(0, config.head_dim, 2, device=device, dtype=torch.float32)
|
| 130 |
+
/ config.head_dim
|
| 131 |
+
)
|
| 132 |
+
)
|
| 133 |
+
return inv_freq, 1.0
|
| 134 |
+
|
| 135 |
+
|
| 136 |
+
def _compute_yarn_rope_parameters(
|
| 137 |
+
config: LizzyConfig,
|
| 138 |
+
device: torch.device,
|
| 139 |
+
) -> tuple[torch.Tensor, float]:
|
| 140 |
+
rope_scaling = dict(config.rope_scaling or {})
|
| 141 |
+
factor = float(rope_scaling["factor"])
|
| 142 |
+
attention_factor = rope_scaling.get("attention_factor")
|
| 143 |
+
mscale = rope_scaling.get("mscale")
|
| 144 |
+
mscale_all_dim = rope_scaling.get("mscale_all_dim")
|
| 145 |
+
original_max_position_embeddings = int(
|
| 146 |
+
rope_scaling.get("original_max_position_embeddings")
|
| 147 |
+
or config.max_position_embeddings
|
| 148 |
+
)
|
| 149 |
+
|
| 150 |
+
def get_mscale(scale: float, mscale_value: float = 1.0) -> float:
|
| 151 |
+
if scale <= 1.0:
|
| 152 |
+
return 1.0
|
| 153 |
+
return 0.1 * mscale_value * math.log(scale) + 1.0
|
| 154 |
+
|
| 155 |
+
if attention_factor is None:
|
| 156 |
+
if mscale is not None and mscale_all_dim is not None:
|
| 157 |
+
attention_factor = float(
|
| 158 |
+
get_mscale(factor, float(mscale))
|
| 159 |
+
/ get_mscale(factor, float(mscale_all_dim))
|
| 160 |
+
)
|
| 161 |
+
else:
|
| 162 |
+
attention_factor = get_mscale(factor)
|
| 163 |
+
|
| 164 |
+
beta_fast = float(rope_scaling.get("beta_fast") or 32.0)
|
| 165 |
+
beta_slow = float(rope_scaling.get("beta_slow") or 1.0)
|
| 166 |
+
truncate = bool(rope_scaling.get("truncate", True))
|
| 167 |
+
dim = config.head_dim
|
| 168 |
+
|
| 169 |
+
def find_correction_dim(
|
| 170 |
+
num_rotations: float,
|
| 171 |
+
*,
|
| 172 |
+
dim: int,
|
| 173 |
+
base: float,
|
| 174 |
+
max_position_embeddings: int,
|
| 175 |
+
) -> float:
|
| 176 |
+
return (
|
| 177 |
+
dim
|
| 178 |
+
* math.log(max_position_embeddings / (num_rotations * 2 * math.pi))
|
| 179 |
+
/ (2 * math.log(base))
|
| 180 |
+
)
|
| 181 |
+
|
| 182 |
+
def find_correction_range(
|
| 183 |
+
low_rot: float,
|
| 184 |
+
high_rot: float,
|
| 185 |
+
*,
|
| 186 |
+
dim: int,
|
| 187 |
+
base: float,
|
| 188 |
+
max_position_embeddings: int,
|
| 189 |
+
truncate: bool,
|
| 190 |
+
) -> tuple[float, float]:
|
| 191 |
+
low = find_correction_dim(
|
| 192 |
+
low_rot,
|
| 193 |
+
dim=dim,
|
| 194 |
+
base=base,
|
| 195 |
+
max_position_embeddings=max_position_embeddings,
|
| 196 |
+
)
|
| 197 |
+
high = find_correction_dim(
|
| 198 |
+
high_rot,
|
| 199 |
+
dim=dim,
|
| 200 |
+
base=base,
|
| 201 |
+
max_position_embeddings=max_position_embeddings,
|
| 202 |
+
)
|
| 203 |
+
if truncate:
|
| 204 |
+
low = math.floor(low)
|
| 205 |
+
high = math.ceil(high)
|
| 206 |
+
return max(low, 0.0), min(high, dim - 1.0)
|
| 207 |
+
|
| 208 |
+
def linear_ramp_factor(
|
| 209 |
+
min_value: float,
|
| 210 |
+
max_value: float,
|
| 211 |
+
dim: int,
|
| 212 |
+
) -> torch.Tensor:
|
| 213 |
+
if min_value == max_value:
|
| 214 |
+
max_value += 0.001
|
| 215 |
+
linear_func = (
|
| 216 |
+
torch.arange(dim, dtype=torch.float32, device=device) - min_value
|
| 217 |
+
) / (max_value - min_value)
|
| 218 |
+
return torch.clamp(linear_func, 0, 1)
|
| 219 |
+
|
| 220 |
+
pos_freqs = config.rope_theta ** (
|
| 221 |
+
torch.arange(0, dim, 2, device=device, dtype=torch.float32) / dim
|
| 222 |
+
)
|
| 223 |
+
inv_freq_extrapolation = 1.0 / pos_freqs
|
| 224 |
+
inv_freq_interpolation = 1.0 / (factor * pos_freqs)
|
| 225 |
+
low, high = find_correction_range(
|
| 226 |
+
beta_fast,
|
| 227 |
+
beta_slow,
|
| 228 |
+
dim=dim,
|
| 229 |
+
base=config.rope_theta,
|
| 230 |
+
max_position_embeddings=original_max_position_embeddings,
|
| 231 |
+
truncate=truncate,
|
| 232 |
+
)
|
| 233 |
+
inv_freq_extrapolation_factor = 1 - linear_ramp_factor(low, high, dim // 2)
|
| 234 |
+
inv_freq = (
|
| 235 |
+
inv_freq_interpolation * (1 - inv_freq_extrapolation_factor)
|
| 236 |
+
+ inv_freq_extrapolation * inv_freq_extrapolation_factor
|
| 237 |
+
)
|
| 238 |
+
return inv_freq, float(attention_factor)
|
| 239 |
+
|
| 240 |
+
|
| 241 |
+
def _compute_rope_parameters(
|
| 242 |
+
config: LizzyConfig,
|
| 243 |
+
device: torch.device,
|
| 244 |
+
*,
|
| 245 |
+
seq_len: int | torch.Tensor | None = None,
|
| 246 |
+
rope_type_override: str | None = None,
|
| 247 |
+
) -> tuple[torch.Tensor, float]:
|
| 248 |
+
rope_scaling = dict(config.rope_scaling or {})
|
| 249 |
+
rope_type = rope_type_override
|
| 250 |
+
if rope_type is None:
|
| 251 |
+
if not rope_scaling:
|
| 252 |
+
return _compute_default_rope_parameters(config, device)
|
| 253 |
+
rope_type = str(
|
| 254 |
+
rope_scaling.get("rope_type", rope_scaling.get("type", "default"))
|
| 255 |
+
)
|
| 256 |
+
|
| 257 |
+
if rope_type == "default":
|
| 258 |
+
return _compute_default_rope_parameters(config, device)
|
| 259 |
+
if rope_type == "yarn":
|
| 260 |
+
return _compute_yarn_rope_parameters(config, device)
|
| 261 |
+
if not rope_scaling:
|
| 262 |
+
return _compute_default_rope_parameters(config, device)
|
| 263 |
+
|
| 264 |
+
rope_init_fn = (
|
| 265 |
+
ROPE_INIT_FUNCTIONS.get(rope_type) or ROPE_INIT_FUNCTIONS.get("default")
|
| 266 |
+
)
|
| 267 |
+
if rope_init_fn is None:
|
| 268 |
+
return _compute_default_rope_parameters(config, device)
|
| 269 |
+
|
| 270 |
+
inv_freq, attention_factor = rope_init_fn(config, device, seq_len=seq_len)
|
| 271 |
+
return inv_freq.to(device=device, dtype=torch.float32), float(attention_factor)
|
| 272 |
+
|
| 273 |
+
|
| 274 |
+
def _looks_like_legacy_interval_rope_lizzy(config: LizzyConfig) -> bool:
|
| 275 |
+
rope_layer_flags = list(getattr(config, "rope_layer_flags", None) or [])
|
| 276 |
+
if rope_layer_flags and not all(bool(item) for item in rope_layer_flags):
|
| 277 |
+
return False
|
| 278 |
+
|
| 279 |
+
layer_types = list(getattr(config, "layer_types", None) or [])
|
| 280 |
+
if layer_types and any(str(item) != "full_attention" for item in layer_types):
|
| 281 |
+
return False
|
| 282 |
+
|
| 283 |
+
return (
|
| 284 |
+
str(getattr(config, "position_embedding_type", "")).lower() == "rope"
|
| 285 |
+
and not bool(getattr(config, "rope_scaling", None))
|
| 286 |
+
and int(getattr(config, "num_hidden_layers", 0) or 0) == 36
|
| 287 |
+
and int(getattr(config, "hidden_size", 0) or 0) == 2048
|
| 288 |
+
and int(getattr(config, "num_attention_heads", 0) or 0) == 16
|
| 289 |
+
and int(getattr(config, "num_key_value_heads", 0) or 0) == 4
|
| 290 |
+
and math.isclose(
|
| 291 |
+
float(getattr(config, "rope_theta", 0.0) or 0.0), 5_000_000.0
|
| 292 |
+
)
|
| 293 |
+
and not bool(getattr(config, "use_post_attn_norm", False))
|
| 294 |
+
and not bool(getattr(config, "use_post_mlp_norm", False))
|
| 295 |
+
and not bool(getattr(config, "use_qk_norm", False))
|
| 296 |
+
)
|
| 297 |
+
|
| 298 |
+
|
| 299 |
+
def _get_no_rope_layer_interval(config: LizzyConfig) -> int | None:
|
| 300 |
+
value = getattr(config, "no_rope_layer_interval", None)
|
| 301 |
+
if value is not None:
|
| 302 |
+
value = int(value)
|
| 303 |
+
if value > 0:
|
| 304 |
+
return value
|
| 305 |
+
if _looks_like_legacy_interval_rope_lizzy(config):
|
| 306 |
+
# Backward-compatible fallback for already-uploaded Lizzy
|
| 307 |
+
# checkpoints that should use NoPE on every 4th layer.
|
| 308 |
+
return 4
|
| 309 |
+
return None
|
| 310 |
+
|
| 311 |
+
|
| 312 |
+
def _get_rope_layer_flag(config: LizzyConfig, layer_idx: int) -> bool:
|
| 313 |
+
rope_enabled = str(
|
| 314 |
+
getattr(config, "position_embedding_type", "rope")
|
| 315 |
+
).lower() == "rope"
|
| 316 |
+
rope_layer_flags = list(getattr(config, "rope_layer_flags", None) or [])
|
| 317 |
+
no_rope_layer_interval = _get_no_rope_layer_interval(config)
|
| 318 |
+
if (
|
| 319 |
+
no_rope_layer_interval is not None
|
| 320 |
+
and (
|
| 321 |
+
layer_idx >= len(rope_layer_flags)
|
| 322 |
+
or not rope_layer_flags
|
| 323 |
+
or all(bool(item) for item in rope_layer_flags)
|
| 324 |
+
)
|
| 325 |
+
):
|
| 326 |
+
return rope_enabled and ((layer_idx + 1) % no_rope_layer_interval != 0)
|
| 327 |
+
if 0 <= layer_idx < len(rope_layer_flags):
|
| 328 |
+
return rope_enabled and bool(rope_layer_flags[layer_idx])
|
| 329 |
+
return rope_enabled
|
| 330 |
+
|
| 331 |
+
|
| 332 |
+
def _get_layer_layout(config: LizzyConfig, layer_idx: int) -> str:
|
| 333 |
+
layer_layouts = list(getattr(config, "layer_layouts", None) or [])
|
| 334 |
+
if 0 <= layer_idx < len(layer_layouts):
|
| 335 |
+
return str(layer_layouts[layer_idx])
|
| 336 |
+
if bool(getattr(config, "use_post_attn_norm", False)) or bool(
|
| 337 |
+
getattr(config, "use_post_mlp_norm", False)
|
| 338 |
+
):
|
| 339 |
+
return "decoder_postnorm"
|
| 340 |
+
return "decoder_prenorm"
|
| 341 |
+
|
| 342 |
+
|
| 343 |
+
def _has_linear_attention(config: LizzyConfig) -> bool:
|
| 344 |
+
return any(
|
| 345 |
+
str(layer_type) == "linear_attention"
|
| 346 |
+
for layer_type in list(getattr(config, "layer_types", None) or [])
|
| 347 |
+
)
|
| 348 |
+
|
| 349 |
+
|
| 350 |
+
class LizzyHybridDynamicCache:
|
| 351 |
+
"""Cache for Lizzy checkpoints with mixed full and linear attention."""
|
| 352 |
+
|
| 353 |
+
is_compileable = False
|
| 354 |
+
|
| 355 |
+
def __init__(self, config: LizzyConfig) -> None:
|
| 356 |
+
super().__init__()
|
| 357 |
+
self.layer_types = list(config.layer_types)
|
| 358 |
+
self.transformer_layers = [
|
| 359 |
+
idx
|
| 360 |
+
for idx, layer_type in enumerate(self.layer_types)
|
| 361 |
+
if layer_type == "full_attention"
|
| 362 |
+
]
|
| 363 |
+
self.last_linear_layer = (
|
| 364 |
+
len(self.layer_types)
|
| 365 |
+
- 1
|
| 366 |
+
- self.layer_types[::-1].index("linear_attention")
|
| 367 |
+
)
|
| 368 |
+
self.recurrent_states = [None for _ in range(config.num_hidden_layers)]
|
| 369 |
+
self.key_cache = [None for _ in range(config.num_hidden_layers)]
|
| 370 |
+
self.value_cache = [None for _ in range(config.num_hidden_layers)]
|
| 371 |
+
self.conv_states_q = [None for _ in range(config.num_hidden_layers)]
|
| 372 |
+
self.conv_states_k = [None for _ in range(config.num_hidden_layers)]
|
| 373 |
+
self.conv_states_v = [None for _ in range(config.num_hidden_layers)]
|
| 374 |
+
|
| 375 |
+
def __len__(self) -> int:
|
| 376 |
+
return len(self.layer_types)
|
| 377 |
+
|
| 378 |
+
def update(
|
| 379 |
+
self,
|
| 380 |
+
key_states: torch.Tensor,
|
| 381 |
+
value_states: torch.Tensor,
|
| 382 |
+
layer_idx: int,
|
| 383 |
+
cache_kwargs: dict[str, Any] | None = None,
|
| 384 |
+
) -> tuple[torch.Tensor, torch.Tensor]:
|
| 385 |
+
del cache_kwargs
|
| 386 |
+
if self.key_cache[layer_idx] is None:
|
| 387 |
+
self.key_cache[layer_idx] = key_states
|
| 388 |
+
self.value_cache[layer_idx] = value_states
|
| 389 |
+
else:
|
| 390 |
+
self.key_cache[layer_idx] = torch.cat(
|
| 391 |
+
[self.key_cache[layer_idx], key_states],
|
| 392 |
+
dim=2,
|
| 393 |
+
)
|
| 394 |
+
self.value_cache[layer_idx] = torch.cat(
|
| 395 |
+
[self.value_cache[layer_idx], value_states],
|
| 396 |
+
dim=2,
|
| 397 |
+
)
|
| 398 |
+
return self.key_cache[layer_idx], self.value_cache[layer_idx]
|
| 399 |
+
|
| 400 |
+
def reorder_cache(self, beam_idx: torch.LongTensor) -> None:
|
| 401 |
+
batch_size = beam_idx.shape[0]
|
| 402 |
+
for layer_idx in range(len(self.key_cache)):
|
| 403 |
+
if self.key_cache[layer_idx] is not None:
|
| 404 |
+
if self.key_cache[layer_idx].shape[0] < batch_size:
|
| 405 |
+
expand_ratio = (
|
| 406 |
+
batch_size // self.key_cache[layer_idx].shape[0]
|
| 407 |
+
)
|
| 408 |
+
self.key_cache[layer_idx] = (
|
| 409 |
+
self.key_cache[layer_idx].repeat_interleave(
|
| 410 |
+
expand_ratio, dim=0,
|
| 411 |
+
)
|
| 412 |
+
)
|
| 413 |
+
self.value_cache[layer_idx] = (
|
| 414 |
+
self.value_cache[layer_idx].repeat_interleave(
|
| 415 |
+
expand_ratio, dim=0,
|
| 416 |
+
)
|
| 417 |
+
)
|
| 418 |
+
device = self.key_cache[layer_idx].device
|
| 419 |
+
self.key_cache[layer_idx] = self.key_cache[layer_idx].index_select(
|
| 420 |
+
0,
|
| 421 |
+
beam_idx.to(device),
|
| 422 |
+
)
|
| 423 |
+
self.value_cache[layer_idx] = (
|
| 424 |
+
self.value_cache[layer_idx].index_select(0, beam_idx.to(device))
|
| 425 |
+
)
|
| 426 |
+
if self.conv_states_q[layer_idx] is not None:
|
| 427 |
+
if self.conv_states_q[layer_idx].shape[0] < batch_size:
|
| 428 |
+
expand_ratio = (
|
| 429 |
+
batch_size // self.conv_states_q[layer_idx].shape[0]
|
| 430 |
+
)
|
| 431 |
+
self.conv_states_q[layer_idx] = (
|
| 432 |
+
self.conv_states_q[layer_idx].repeat_interleave(
|
| 433 |
+
expand_ratio, dim=0,
|
| 434 |
+
)
|
| 435 |
+
)
|
| 436 |
+
self.conv_states_k[layer_idx] = (
|
| 437 |
+
self.conv_states_k[layer_idx].repeat_interleave(
|
| 438 |
+
expand_ratio, dim=0,
|
| 439 |
+
)
|
| 440 |
+
)
|
| 441 |
+
self.conv_states_v[layer_idx] = (
|
| 442 |
+
self.conv_states_v[layer_idx].repeat_interleave(
|
| 443 |
+
expand_ratio, dim=0,
|
| 444 |
+
)
|
| 445 |
+
)
|
| 446 |
+
self.recurrent_states[layer_idx] = (
|
| 447 |
+
self.recurrent_states[layer_idx].repeat_interleave(
|
| 448 |
+
expand_ratio, dim=0,
|
| 449 |
+
)
|
| 450 |
+
)
|
| 451 |
+
device = self.conv_states_q[layer_idx].device
|
| 452 |
+
self.conv_states_q[layer_idx] = (
|
| 453 |
+
self.conv_states_q[layer_idx].index_select(
|
| 454 |
+
0,
|
| 455 |
+
beam_idx.to(device),
|
| 456 |
+
)
|
| 457 |
+
)
|
| 458 |
+
self.conv_states_k[layer_idx] = (
|
| 459 |
+
self.conv_states_k[layer_idx].index_select(
|
| 460 |
+
0,
|
| 461 |
+
beam_idx.to(device),
|
| 462 |
+
)
|
| 463 |
+
)
|
| 464 |
+
self.conv_states_v[layer_idx] = (
|
| 465 |
+
self.conv_states_v[layer_idx].index_select(
|
| 466 |
+
0,
|
| 467 |
+
beam_idx.to(device),
|
| 468 |
+
)
|
| 469 |
+
)
|
| 470 |
+
self.recurrent_states[layer_idx] = (
|
| 471 |
+
self.recurrent_states[layer_idx].index_select(
|
| 472 |
+
0,
|
| 473 |
+
beam_idx.to(device),
|
| 474 |
+
)
|
| 475 |
+
)
|
| 476 |
+
|
| 477 |
+
def get_seq_length(self, layer_idx: int | None = 0) -> int:
|
| 478 |
+
if not self.transformer_layers:
|
| 479 |
+
return 0
|
| 480 |
+
layer_idx = (
|
| 481 |
+
self.transformer_layers[0]
|
| 482 |
+
if layer_idx not in self.transformer_layers
|
| 483 |
+
else layer_idx
|
| 484 |
+
)
|
| 485 |
+
if len(self.key_cache) <= layer_idx or self.key_cache[layer_idx] is None:
|
| 486 |
+
return 0
|
| 487 |
+
return self.key_cache[layer_idx].shape[-2]
|
| 488 |
+
|
| 489 |
+
def get_mask_sizes(self, query_length: int, layer_idx: int) -> tuple[int, int]:
|
| 490 |
+
del layer_idx
|
| 491 |
+
kv_offset = 0
|
| 492 |
+
past_seen_tokens = self.get_seq_length()
|
| 493 |
+
kv_length = query_length + past_seen_tokens
|
| 494 |
+
return kv_length, kv_offset
|
| 495 |
+
|
| 496 |
+
@property
|
| 497 |
+
def has_previous_state(self) -> bool:
|
| 498 |
+
# Mirror the upstream contract: once the final linear layer has cached
|
| 499 |
+
# its conv state, single-token decode can switch to the recurrent path.
|
| 500 |
+
return self.conv_states_q[self.last_linear_layer] is not None
|
| 501 |
+
|
| 502 |
+
|
| 503 |
+
class LizzyHybridRMSNormGated(nn.Module):
|
| 504 |
+
def __init__(self, hidden_size: int, eps: float = 1e-6) -> None:
|
| 505 |
+
super().__init__()
|
| 506 |
+
self.weight = nn.Parameter(torch.ones(hidden_size))
|
| 507 |
+
self.variance_epsilon = eps
|
| 508 |
+
|
| 509 |
+
def forward(
|
| 510 |
+
self,
|
| 511 |
+
hidden_states: torch.Tensor,
|
| 512 |
+
gate: torch.Tensor | None = None,
|
| 513 |
+
) -> torch.Tensor:
|
| 514 |
+
if gate is None:
|
| 515 |
+
msg = "gate is required for gated RMSNorm."
|
| 516 |
+
raise ValueError(msg)
|
| 517 |
+
input_dtype = hidden_states.dtype
|
| 518 |
+
hidden_states = hidden_states.to(torch.float32)
|
| 519 |
+
variance = hidden_states.pow(2).mean(-1, keepdim=True)
|
| 520 |
+
hidden_states = hidden_states * torch.rsqrt(
|
| 521 |
+
variance + self.variance_epsilon
|
| 522 |
+
)
|
| 523 |
+
hidden_states = self.weight * hidden_states.to(input_dtype)
|
| 524 |
+
hidden_states = hidden_states * F.silu(gate.to(torch.float32))
|
| 525 |
+
return hidden_states.to(input_dtype)
|
| 526 |
+
|
| 527 |
+
|
| 528 |
+
class LizzyHybridShortConvolution(nn.Conv1d):
|
| 529 |
+
def __init__(
|
| 530 |
+
self,
|
| 531 |
+
hidden_size: int,
|
| 532 |
+
kernel_size: int,
|
| 533 |
+
bias: bool = False,
|
| 534 |
+
activation: str | None = "silu",
|
| 535 |
+
) -> None:
|
| 536 |
+
super().__init__(
|
| 537 |
+
in_channels=hidden_size,
|
| 538 |
+
out_channels=hidden_size,
|
| 539 |
+
kernel_size=kernel_size,
|
| 540 |
+
groups=hidden_size,
|
| 541 |
+
padding=kernel_size - 1,
|
| 542 |
+
bias=bias,
|
| 543 |
+
)
|
| 544 |
+
self.hidden_size = hidden_size
|
| 545 |
+
self.conv_kernel_size = kernel_size
|
| 546 |
+
self.act_fn = ACT2FN[activation]
|
| 547 |
+
|
| 548 |
+
def forward(
|
| 549 |
+
self,
|
| 550 |
+
hidden_states: torch.Tensor,
|
| 551 |
+
cache: torch.Tensor | None = None,
|
| 552 |
+
use_precomputed: bool = False,
|
| 553 |
+
**kwargs: Any,
|
| 554 |
+
) -> tuple[torch.Tensor, torch.Tensor]:
|
| 555 |
+
del kwargs
|
| 556 |
+
seq_len, dim = hidden_states.shape[-2:]
|
| 557 |
+
hidden_states = hidden_states.transpose(1, 2)
|
| 558 |
+
if use_precomputed:
|
| 559 |
+
if cache is None:
|
| 560 |
+
msg = "cache is required when use_precomputed=True."
|
| 561 |
+
raise ValueError(msg)
|
| 562 |
+
x_with_state = torch.cat([cache, hidden_states], dim=-1)
|
| 563 |
+
out = F.conv1d(
|
| 564 |
+
x_with_state,
|
| 565 |
+
self.weight,
|
| 566 |
+
self.bias,
|
| 567 |
+
padding=0,
|
| 568 |
+
groups=dim,
|
| 569 |
+
)
|
| 570 |
+
conv_state = x_with_state[:, :, 1:]
|
| 571 |
+
else:
|
| 572 |
+
out = F.conv1d(
|
| 573 |
+
hidden_states,
|
| 574 |
+
self.weight,
|
| 575 |
+
self.bias,
|
| 576 |
+
padding=self.conv_kernel_size - 1,
|
| 577 |
+
groups=dim,
|
| 578 |
+
)
|
| 579 |
+
out = out[:, :, :seq_len]
|
| 580 |
+
conv_state = F.pad(
|
| 581 |
+
hidden_states,
|
| 582 |
+
(self.conv_kernel_size - 1 - hidden_states.shape[-1], 0),
|
| 583 |
+
)
|
| 584 |
+
out = self.act_fn(out)
|
| 585 |
+
return out.transpose(1, 2), conv_state
|
| 586 |
+
|
| 587 |
+
|
| 588 |
+
def _apply_mask_to_padding_states(
|
| 589 |
+
hidden_states: torch.Tensor,
|
| 590 |
+
attention_mask: torch.Tensor | None,
|
| 591 |
+
) -> torch.Tensor:
|
| 592 |
+
# Match the upstream hybrid implementation: silence padded tokens before
|
| 593 |
+
# the DeltaNet projections so recurrent state does not absorb padding.
|
| 594 |
+
if (
|
| 595 |
+
attention_mask is not None
|
| 596 |
+
and attention_mask.shape[1] > 1
|
| 597 |
+
and attention_mask.shape[0] > 1
|
| 598 |
+
):
|
| 599 |
+
dtype = hidden_states.dtype
|
| 600 |
+
hidden_states = (hidden_states * attention_mask[:, :, None]).to(dtype)
|
| 601 |
+
return hidden_states
|
| 602 |
+
|
| 603 |
+
|
| 604 |
+
def _l2norm(
|
| 605 |
+
x: torch.Tensor,
|
| 606 |
+
dim: int = -1,
|
| 607 |
+
eps: float = 1e-6,
|
| 608 |
+
) -> torch.Tensor:
|
| 609 |
+
inv_norm = torch.rsqrt((x * x).sum(dim=dim, keepdim=True) + eps)
|
| 610 |
+
return x * inv_norm
|
| 611 |
+
|
| 612 |
+
|
| 613 |
+
def _torch_chunk_gated_delta_rule(
|
| 614 |
+
query: torch.Tensor,
|
| 615 |
+
key: torch.Tensor,
|
| 616 |
+
value: torch.Tensor,
|
| 617 |
+
g: torch.Tensor,
|
| 618 |
+
beta: torch.Tensor,
|
| 619 |
+
chunk_size: int = 64,
|
| 620 |
+
initial_state: torch.Tensor | None = None,
|
| 621 |
+
output_final_state: bool = False,
|
| 622 |
+
use_qk_l2norm_in_kernel: bool = False,
|
| 623 |
+
) -> tuple[torch.Tensor, torch.Tensor | None]:
|
| 624 |
+
initial_dtype = query.dtype
|
| 625 |
+
if use_qk_l2norm_in_kernel:
|
| 626 |
+
query = _l2norm(query, dim=-1, eps=1e-6)
|
| 627 |
+
key = _l2norm(key, dim=-1, eps=1e-6)
|
| 628 |
+
query, key, value, beta, g = [
|
| 629 |
+
x.transpose(1, 2).contiguous().to(torch.float32)
|
| 630 |
+
for x in (query, key, value, beta, g)
|
| 631 |
+
]
|
| 632 |
+
batch_size, num_heads, sequence_length, k_head_dim = key.shape
|
| 633 |
+
v_head_dim = value.shape[-1]
|
| 634 |
+
pad_size = (chunk_size - sequence_length % chunk_size) % chunk_size
|
| 635 |
+
query = F.pad(query, (0, 0, 0, pad_size))
|
| 636 |
+
key = F.pad(key, (0, 0, 0, pad_size))
|
| 637 |
+
value = F.pad(value, (0, 0, 0, pad_size))
|
| 638 |
+
beta = F.pad(beta, (0, pad_size))
|
| 639 |
+
g = F.pad(g, (0, pad_size))
|
| 640 |
+
total_sequence_length = sequence_length + pad_size
|
| 641 |
+
scale = 1 / (query.shape[-1] ** 0.5)
|
| 642 |
+
query = query * scale
|
| 643 |
+
|
| 644 |
+
v_beta = value * beta.unsqueeze(-1)
|
| 645 |
+
k_beta = key * beta.unsqueeze(-1)
|
| 646 |
+
query, key, value, k_beta, v_beta = [
|
| 647 |
+
x.reshape(x.shape[0], x.shape[1], -1, chunk_size, x.shape[-1])
|
| 648 |
+
for x in (query, key, value, k_beta, v_beta)
|
| 649 |
+
]
|
| 650 |
+
g = g.reshape(g.shape[0], g.shape[1], -1, chunk_size)
|
| 651 |
+
mask = torch.triu(
|
| 652 |
+
torch.ones(
|
| 653 |
+
chunk_size,
|
| 654 |
+
chunk_size,
|
| 655 |
+
dtype=torch.bool,
|
| 656 |
+
device=query.device,
|
| 657 |
+
),
|
| 658 |
+
diagonal=0,
|
| 659 |
+
)
|
| 660 |
+
g = g.cumsum(dim=-1)
|
| 661 |
+
decay_mask = ((g.unsqueeze(-1) - g.unsqueeze(-2)).tril().exp().float()).tril()
|
| 662 |
+
attn = -((k_beta @ key.transpose(-1, -2)) * decay_mask).masked_fill(mask, 0)
|
| 663 |
+
for idx in range(1, chunk_size):
|
| 664 |
+
row = attn[..., idx, :idx].clone()
|
| 665 |
+
sub = attn[..., :idx, :idx].clone()
|
| 666 |
+
attn[..., idx, :idx] = row + (row.unsqueeze(-1) * sub).sum(-2)
|
| 667 |
+
attn = attn + torch.eye(
|
| 668 |
+
chunk_size,
|
| 669 |
+
dtype=attn.dtype,
|
| 670 |
+
device=attn.device,
|
| 671 |
+
)
|
| 672 |
+
value = attn @ v_beta
|
| 673 |
+
k_cumdecay = attn @ (k_beta * g.exp().unsqueeze(-1))
|
| 674 |
+
last_recurrent_state = (
|
| 675 |
+
torch.zeros(batch_size, num_heads, k_head_dim, v_head_dim).to(value)
|
| 676 |
+
if initial_state is None
|
| 677 |
+
else initial_state.to(value)
|
| 678 |
+
)
|
| 679 |
+
core_attn_out = torch.zeros_like(value)
|
| 680 |
+
mask = torch.triu(
|
| 681 |
+
torch.ones(
|
| 682 |
+
chunk_size,
|
| 683 |
+
chunk_size,
|
| 684 |
+
dtype=torch.bool,
|
| 685 |
+
device=query.device,
|
| 686 |
+
),
|
| 687 |
+
diagonal=1,
|
| 688 |
+
)
|
| 689 |
+
for idx in range(0, total_sequence_length // chunk_size):
|
| 690 |
+
q_i, k_i, v_i = query[:, :, idx], key[:, :, idx], value[:, :, idx]
|
| 691 |
+
attn = (q_i @ k_i.transpose(-1, -2) * decay_mask[:, :, idx]).masked_fill_(
|
| 692 |
+
mask,
|
| 693 |
+
0,
|
| 694 |
+
)
|
| 695 |
+
v_prime = (k_cumdecay[:, :, idx]) @ last_recurrent_state
|
| 696 |
+
v_new = v_i - v_prime
|
| 697 |
+
attn_inter = (q_i * g[:, :, idx, :, None].exp()) @ last_recurrent_state
|
| 698 |
+
core_attn_out[:, :, idx] = attn_inter + attn @ v_new
|
| 699 |
+
last_recurrent_state = (
|
| 700 |
+
last_recurrent_state * g[:, :, idx, -1, None, None].exp()
|
| 701 |
+
+ (
|
| 702 |
+
k_i
|
| 703 |
+
* (g[:, :, idx, -1, None] - g[:, :, idx]).exp()[..., None]
|
| 704 |
+
).transpose(-1, -2)
|
| 705 |
+
@ v_new
|
| 706 |
+
)
|
| 707 |
+
if not output_final_state:
|
| 708 |
+
last_recurrent_state = None
|
| 709 |
+
core_attn_out = core_attn_out.reshape(
|
| 710 |
+
core_attn_out.shape[0],
|
| 711 |
+
core_attn_out.shape[1],
|
| 712 |
+
-1,
|
| 713 |
+
core_attn_out.shape[-1],
|
| 714 |
+
)
|
| 715 |
+
core_attn_out = core_attn_out[:, :, :sequence_length]
|
| 716 |
+
core_attn_out = core_attn_out.transpose(1, 2).contiguous().to(initial_dtype)
|
| 717 |
+
return core_attn_out, last_recurrent_state
|
| 718 |
+
|
| 719 |
+
|
| 720 |
+
def _torch_recurrent_gated_delta_rule(
|
| 721 |
+
query: torch.Tensor,
|
| 722 |
+
key: torch.Tensor,
|
| 723 |
+
value: torch.Tensor,
|
| 724 |
+
g: torch.Tensor,
|
| 725 |
+
beta: torch.Tensor,
|
| 726 |
+
initial_state: torch.Tensor | None,
|
| 727 |
+
output_final_state: bool,
|
| 728 |
+
use_qk_l2norm_in_kernel: bool = False,
|
| 729 |
+
) -> tuple[torch.Tensor, torch.Tensor | None]:
|
| 730 |
+
initial_dtype = query.dtype
|
| 731 |
+
if use_qk_l2norm_in_kernel:
|
| 732 |
+
query = _l2norm(query, dim=-1, eps=1e-6)
|
| 733 |
+
key = _l2norm(key, dim=-1, eps=1e-6)
|
| 734 |
+
query, key, value, beta, g = [
|
| 735 |
+
x.transpose(1, 2).contiguous().to(torch.float32)
|
| 736 |
+
for x in (query, key, value, beta, g)
|
| 737 |
+
]
|
| 738 |
+
batch_size, num_heads, sequence_length, k_head_dim = key.shape
|
| 739 |
+
v_head_dim = value.shape[-1]
|
| 740 |
+
scale = 1 / (query.shape[-1] ** 0.5)
|
| 741 |
+
query = query * scale
|
| 742 |
+
core_attn_out = torch.zeros(
|
| 743 |
+
batch_size,
|
| 744 |
+
num_heads,
|
| 745 |
+
sequence_length,
|
| 746 |
+
v_head_dim,
|
| 747 |
+
).to(value)
|
| 748 |
+
last_recurrent_state = (
|
| 749 |
+
torch.zeros(batch_size, num_heads, k_head_dim, v_head_dim).to(value)
|
| 750 |
+
if initial_state is None
|
| 751 |
+
else initial_state.to(value)
|
| 752 |
+
)
|
| 753 |
+
for idx in range(sequence_length):
|
| 754 |
+
q_t = query[:, :, idx]
|
| 755 |
+
k_t = key[:, :, idx]
|
| 756 |
+
v_t = value[:, :, idx]
|
| 757 |
+
g_t = g[:, :, idx].exp().unsqueeze(-1).unsqueeze(-1)
|
| 758 |
+
beta_t = beta[:, :, idx].unsqueeze(-1)
|
| 759 |
+
last_recurrent_state = last_recurrent_state * g_t
|
| 760 |
+
kv_mem = (last_recurrent_state * k_t.unsqueeze(-1)).sum(dim=-2)
|
| 761 |
+
delta = (v_t - kv_mem) * beta_t
|
| 762 |
+
last_recurrent_state = (
|
| 763 |
+
last_recurrent_state + k_t.unsqueeze(-1) * delta.unsqueeze(-2)
|
| 764 |
+
)
|
| 765 |
+
core_attn_out[:, :, idx] = (
|
| 766 |
+
last_recurrent_state * q_t.unsqueeze(-1)
|
| 767 |
+
).sum(dim=-2)
|
| 768 |
+
if not output_final_state:
|
| 769 |
+
last_recurrent_state = None
|
| 770 |
+
core_attn_out = core_attn_out.transpose(1, 2).contiguous().to(initial_dtype)
|
| 771 |
+
return core_attn_out, last_recurrent_state
|
| 772 |
+
|
| 773 |
+
|
| 774 |
+
class LizzyHybridGatedDeltaNet(nn.Module):
|
| 775 |
+
def __init__(self, config: LizzyConfig, layer_idx: int) -> None:
|
| 776 |
+
super().__init__()
|
| 777 |
+
self.hidden_size = config.hidden_size
|
| 778 |
+
self.num_v_heads = config.linear_num_value_heads
|
| 779 |
+
self.num_k_heads = config.linear_num_key_heads
|
| 780 |
+
self.head_k_dim = config.linear_key_head_dim
|
| 781 |
+
self.head_v_dim = config.linear_value_head_dim
|
| 782 |
+
self.key_dim = self.head_k_dim * self.num_k_heads
|
| 783 |
+
self.value_dim = self.head_v_dim * self.num_v_heads
|
| 784 |
+
self.layer_idx = layer_idx
|
| 785 |
+
self.conv_kernel_size = config.linear_conv_kernel_dim
|
| 786 |
+
self.allow_neg_eigval = config.linear_allow_neg_eigval
|
| 787 |
+
self.eps = config.rms_norm_eps
|
| 788 |
+
self.q_proj = nn.Linear(self.hidden_size, self.key_dim, bias=False)
|
| 789 |
+
self.k_proj = nn.Linear(self.hidden_size, self.key_dim, bias=False)
|
| 790 |
+
self.v_proj = nn.Linear(self.hidden_size, self.value_dim, bias=False)
|
| 791 |
+
self.a_proj = nn.Linear(self.hidden_size, self.num_v_heads, bias=False)
|
| 792 |
+
self.b_proj = nn.Linear(self.hidden_size, self.num_v_heads, bias=False)
|
| 793 |
+
self.g_proj = nn.Linear(self.hidden_size, self.value_dim, bias=False)
|
| 794 |
+
self.o_proj = nn.Linear(self.value_dim, self.hidden_size, bias=False)
|
| 795 |
+
|
| 796 |
+
# Step-02 conversion runs on CPU by default, even on GPU nodes. In that
|
| 797 |
+
# flow Triton-backed FLA kernels will crash as soon as a CPU tensor
|
| 798 |
+
# reaches them, so the wrapper can force the pure PyTorch fallback for
|
| 799 |
+
# Hybrid layers via an environment switch.
|
| 800 |
+
disable_fla_fast_path = os.environ.get(
|
| 801 |
+
"LIZZY_DISABLE_HYBRID_FLA",
|
| 802 |
+
"",
|
| 803 |
+
).strip().lower() in {"1", "true", "yes", "on"}
|
| 804 |
+
use_fla_fast_path = (
|
| 805 |
+
not disable_fla_fast_path
|
| 806 |
+
and
|
| 807 |
+
torch.cuda.is_available()
|
| 808 |
+
and ShortConvolution is not None
|
| 809 |
+
and chunk_gated_delta_rule is not None
|
| 810 |
+
and fused_recurrent_gated_delta_rule is not None
|
| 811 |
+
and FusedRMSNormGated is not None
|
| 812 |
+
)
|
| 813 |
+
# Keep the fast-path contract when FLA is present, but fall back to a
|
| 814 |
+
# local implementation so the public Lizzy artifact never depends on
|
| 815 |
+
# family-specific Transformers remote code.
|
| 816 |
+
conv1d_class = (
|
| 817 |
+
ShortConvolution if use_fla_fast_path else LizzyHybridShortConvolution
|
| 818 |
+
)
|
| 819 |
+
self.q_conv1d = conv1d_class(
|
| 820 |
+
hidden_size=self.key_dim,
|
| 821 |
+
kernel_size=self.conv_kernel_size,
|
| 822 |
+
bias=False,
|
| 823 |
+
activation="silu",
|
| 824 |
+
)
|
| 825 |
+
self.k_conv1d = conv1d_class(
|
| 826 |
+
hidden_size=self.key_dim,
|
| 827 |
+
kernel_size=self.conv_kernel_size,
|
| 828 |
+
bias=False,
|
| 829 |
+
activation="silu",
|
| 830 |
+
)
|
| 831 |
+
self.v_conv1d = conv1d_class(
|
| 832 |
+
hidden_size=self.value_dim,
|
| 833 |
+
kernel_size=self.conv_kernel_size,
|
| 834 |
+
bias=False,
|
| 835 |
+
activation="silu",
|
| 836 |
+
)
|
| 837 |
+
|
| 838 |
+
a = torch.empty(self.num_v_heads, dtype=torch.float32).uniform_(
|
| 839 |
+
config.linear_a_log_min,
|
| 840 |
+
config.linear_a_log_max,
|
| 841 |
+
)
|
| 842 |
+
self.A_log = nn.Parameter(torch.log(a))
|
| 843 |
+
dt = torch.exp(
|
| 844 |
+
torch.rand(self.num_v_heads)
|
| 845 |
+
* (math.log(config.linear_dt_max) - math.log(config.linear_dt_min))
|
| 846 |
+
+ math.log(config.linear_dt_min)
|
| 847 |
+
)
|
| 848 |
+
dt = torch.clamp(dt, min=config.linear_dt_init_floor)
|
| 849 |
+
inv_dt = dt + torch.log(-torch.expm1(-dt))
|
| 850 |
+
self.dt_bias = nn.Parameter(inv_dt)
|
| 851 |
+
|
| 852 |
+
self.o_norm = (
|
| 853 |
+
LizzyHybridRMSNormGated(self.head_v_dim, eps=1e-5)
|
| 854 |
+
if not use_fla_fast_path
|
| 855 |
+
else FusedRMSNormGated(
|
| 856 |
+
self.head_v_dim,
|
| 857 |
+
eps=1e-5,
|
| 858 |
+
device=torch.cuda.current_device(),
|
| 859 |
+
dtype=(
|
| 860 |
+
config.dtype
|
| 861 |
+
if config.dtype is not None
|
| 862 |
+
else torch.get_default_dtype()
|
| 863 |
+
),
|
| 864 |
+
)
|
| 865 |
+
)
|
| 866 |
+
self.chunk_gated_delta_rule = (
|
| 867 |
+
chunk_gated_delta_rule
|
| 868 |
+
if use_fla_fast_path
|
| 869 |
+
else _torch_chunk_gated_delta_rule
|
| 870 |
+
)
|
| 871 |
+
self.recurrent_gated_delta_rule = (
|
| 872 |
+
(
|
| 873 |
+
fused_recurrent_gated_delta_rule
|
| 874 |
+
if use_fla_fast_path
|
| 875 |
+
else _torch_recurrent_gated_delta_rule
|
| 876 |
+
)
|
| 877 |
+
)
|
| 878 |
+
|
| 879 |
+
def forward(
|
| 880 |
+
self,
|
| 881 |
+
hidden_states: torch.Tensor,
|
| 882 |
+
cache_params: LizzyHybridDynamicCache | None = None,
|
| 883 |
+
attention_mask: torch.Tensor | None = None,
|
| 884 |
+
**kwargs: Any,
|
| 885 |
+
) -> torch.Tensor:
|
| 886 |
+
del kwargs
|
| 887 |
+
hidden_states = _apply_mask_to_padding_states(hidden_states, attention_mask)
|
| 888 |
+
batch_size, seq_len, _ = hidden_states.shape
|
| 889 |
+
use_cache = cache_params is not None
|
| 890 |
+
use_precomputed = (
|
| 891 |
+
use_cache
|
| 892 |
+
and getattr(cache_params, "has_previous_state", False)
|
| 893 |
+
and seq_len == 1
|
| 894 |
+
)
|
| 895 |
+
conv_state_q = (
|
| 896 |
+
cache_params.conv_states_q[self.layer_idx] if cache_params else None
|
| 897 |
+
)
|
| 898 |
+
conv_state_k = (
|
| 899 |
+
cache_params.conv_states_k[self.layer_idx] if cache_params else None
|
| 900 |
+
)
|
| 901 |
+
conv_state_v = (
|
| 902 |
+
cache_params.conv_states_v[self.layer_idx] if cache_params else None
|
| 903 |
+
)
|
| 904 |
+
recurrent_state = (
|
| 905 |
+
cache_params.recurrent_states[self.layer_idx] if cache_params else None
|
| 906 |
+
)
|
| 907 |
+
|
| 908 |
+
q = self.q_proj(hidden_states)
|
| 909 |
+
k = self.k_proj(hidden_states)
|
| 910 |
+
v = self.v_proj(hidden_states)
|
| 911 |
+
q, new_conv_state_q = self.q_conv1d(
|
| 912 |
+
q,
|
| 913 |
+
cache=conv_state_q,
|
| 914 |
+
use_precomputed=use_precomputed,
|
| 915 |
+
output_final_state=use_cache,
|
| 916 |
+
)
|
| 917 |
+
k, new_conv_state_k = self.k_conv1d(
|
| 918 |
+
k,
|
| 919 |
+
cache=conv_state_k,
|
| 920 |
+
use_precomputed=use_precomputed,
|
| 921 |
+
output_final_state=use_cache,
|
| 922 |
+
)
|
| 923 |
+
v, new_conv_state_v = self.v_conv1d(
|
| 924 |
+
v,
|
| 925 |
+
cache=conv_state_v,
|
| 926 |
+
use_precomputed=use_precomputed,
|
| 927 |
+
output_final_state=use_cache,
|
| 928 |
+
)
|
| 929 |
+
|
| 930 |
+
if cache_params is not None:
|
| 931 |
+
cache_params.conv_states_q[self.layer_idx] = new_conv_state_q
|
| 932 |
+
cache_params.conv_states_k[self.layer_idx] = new_conv_state_k
|
| 933 |
+
cache_params.conv_states_v[self.layer_idx] = new_conv_state_v
|
| 934 |
+
|
| 935 |
+
q = q.view(batch_size, seq_len, -1, self.head_k_dim)
|
| 936 |
+
k = k.view(batch_size, seq_len, -1, self.head_k_dim)
|
| 937 |
+
v = v.view(batch_size, seq_len, -1, self.head_v_dim)
|
| 938 |
+
if self.num_v_heads > self.num_k_heads:
|
| 939 |
+
expand_ratio = self.num_v_heads // self.num_k_heads
|
| 940 |
+
q = q.repeat_interleave(expand_ratio, dim=2)
|
| 941 |
+
k = k.repeat_interleave(expand_ratio, dim=2)
|
| 942 |
+
|
| 943 |
+
beta = self.b_proj(hidden_states).sigmoid()
|
| 944 |
+
if self.allow_neg_eigval:
|
| 945 |
+
beta = beta * 2.0
|
| 946 |
+
g = -self.A_log.float().exp() * F.softplus(
|
| 947 |
+
self.a_proj(hidden_states).float() + self.dt_bias
|
| 948 |
+
)
|
| 949 |
+
|
| 950 |
+
if use_precomputed:
|
| 951 |
+
output, new_recurrent_state = self.recurrent_gated_delta_rule(
|
| 952 |
+
q,
|
| 953 |
+
k,
|
| 954 |
+
v,
|
| 955 |
+
g=g,
|
| 956 |
+
beta=beta,
|
| 957 |
+
initial_state=recurrent_state,
|
| 958 |
+
output_final_state=use_cache,
|
| 959 |
+
use_qk_l2norm_in_kernel=True,
|
| 960 |
+
)
|
| 961 |
+
else:
|
| 962 |
+
output, new_recurrent_state = self.chunk_gated_delta_rule(
|
| 963 |
+
q,
|
| 964 |
+
k,
|
| 965 |
+
v,
|
| 966 |
+
g=g,
|
| 967 |
+
beta=beta,
|
| 968 |
+
initial_state=recurrent_state,
|
| 969 |
+
output_final_state=use_cache,
|
| 970 |
+
use_qk_l2norm_in_kernel=True,
|
| 971 |
+
)
|
| 972 |
+
|
| 973 |
+
if cache_params is not None:
|
| 974 |
+
cache_params.recurrent_states[self.layer_idx] = new_recurrent_state
|
| 975 |
+
|
| 976 |
+
gate = self.g_proj(hidden_states)
|
| 977 |
+
output = output.reshape(-1, self.head_v_dim)
|
| 978 |
+
gate = gate.reshape(-1, self.head_v_dim)
|
| 979 |
+
output = self.o_norm(output, gate)
|
| 980 |
+
output = output.reshape(batch_size, seq_len, -1)
|
| 981 |
+
output = self.o_proj(output)
|
| 982 |
+
return output
|
| 983 |
+
|
| 984 |
+
|
| 985 |
+
class LizzyLinearAttention(nn.Module):
|
| 986 |
+
def __init__(self, config: LizzyConfig, layer_idx: int) -> None:
|
| 987 |
+
super().__init__()
|
| 988 |
+
self.layer_idx = layer_idx
|
| 989 |
+
self.inner = LizzyHybridGatedDeltaNet(config, layer_idx)
|
| 990 |
+
|
| 991 |
+
def forward(
|
| 992 |
+
self,
|
| 993 |
+
hidden_states: torch.Tensor,
|
| 994 |
+
attention_mask: torch.Tensor | None = None,
|
| 995 |
+
past_key_value: Cache | None = None,
|
| 996 |
+
use_cache: bool = False,
|
| 997 |
+
output_attentions: bool = False,
|
| 998 |
+
**kwargs: Any,
|
| 999 |
+
) -> tuple[
|
| 1000 |
+
torch.Tensor,
|
| 1001 |
+
Cache | None,
|
| 1002 |
+
torch.Tensor | None,
|
| 1003 |
+
]:
|
| 1004 |
+
del kwargs, output_attentions
|
| 1005 |
+
output = self.inner(
|
| 1006 |
+
hidden_states=hidden_states,
|
| 1007 |
+
cache_params=(
|
| 1008 |
+
past_key_value if _is_cache_object(past_key_value) else None
|
| 1009 |
+
),
|
| 1010 |
+
attention_mask=attention_mask,
|
| 1011 |
+
)
|
| 1012 |
+
present = past_key_value if use_cache else None
|
| 1013 |
+
return output, present, None
|
| 1014 |
+
|
| 1015 |
+
|
| 1016 |
+
class LizzyAttention(nn.Module):
|
| 1017 |
+
def __init__(self, config: LizzyConfig, layer_idx: int) -> None:
|
| 1018 |
+
super().__init__()
|
| 1019 |
+
self.is_causal = True
|
| 1020 |
+
self.config = config
|
| 1021 |
+
self.layer_idx = layer_idx
|
| 1022 |
+
self.num_heads = config.num_attention_heads
|
| 1023 |
+
self.num_key_value_heads = config.num_key_value_heads
|
| 1024 |
+
self.num_key_value_groups = self.num_heads // self.num_key_value_heads
|
| 1025 |
+
self.head_dim = config.head_dim
|
| 1026 |
+
self.hidden_size = config.hidden_size
|
| 1027 |
+
self.scaling = self.head_dim**-0.5
|
| 1028 |
+
self.attention_dropout = config.attention_dropout
|
| 1029 |
+
self.position_embedding_type = config.position_embedding_type
|
| 1030 |
+
self.layer_type = (
|
| 1031 |
+
str(config.layer_types[layer_idx])
|
| 1032 |
+
if layer_idx < len(config.layer_types)
|
| 1033 |
+
else "full_attention"
|
| 1034 |
+
)
|
| 1035 |
+
self.use_rope = _get_rope_layer_flag(config, layer_idx)
|
| 1036 |
+
self._rope_type_override = str(
|
| 1037 |
+
dict(config.rope_type_overrides or {}).get(self.layer_type) or ""
|
| 1038 |
+
) or None
|
| 1039 |
+
if (
|
| 1040 |
+
self._rope_type_override is None
|
| 1041 |
+
and self.layer_type == "sliding_attention"
|
| 1042 |
+
and bool(config.rope_scaling)
|
| 1043 |
+
and config.use_post_attn_norm
|
| 1044 |
+
and config.use_post_mlp_norm
|
| 1045 |
+
and config.use_qk_norm
|
| 1046 |
+
and any(str(item) == "full_attention" for item in config.layer_types)
|
| 1047 |
+
):
|
| 1048 |
+
self._rope_type_override = "default"
|
| 1049 |
+
self.sliding_window = None
|
| 1050 |
+
if self.layer_type == "sliding_attention":
|
| 1051 |
+
self.sliding_window = config.sliding_window
|
| 1052 |
+
|
| 1053 |
+
q_dim = self.num_heads * self.head_dim
|
| 1054 |
+
kv_dim = self.num_key_value_heads * self.head_dim
|
| 1055 |
+
self.q_proj = nn.Linear(
|
| 1056 |
+
config.hidden_size,
|
| 1057 |
+
q_dim,
|
| 1058 |
+
bias=config.attention_bias,
|
| 1059 |
+
)
|
| 1060 |
+
self.k_proj = nn.Linear(
|
| 1061 |
+
config.hidden_size,
|
| 1062 |
+
kv_dim,
|
| 1063 |
+
bias=config.attention_bias,
|
| 1064 |
+
)
|
| 1065 |
+
self.v_proj = nn.Linear(
|
| 1066 |
+
config.hidden_size,
|
| 1067 |
+
kv_dim,
|
| 1068 |
+
bias=config.attention_bias,
|
| 1069 |
+
)
|
| 1070 |
+
self.o_proj = nn.Linear(
|
| 1071 |
+
q_dim,
|
| 1072 |
+
config.hidden_size,
|
| 1073 |
+
bias=config.attention_bias,
|
| 1074 |
+
)
|
| 1075 |
+
self.q_norm = (
|
| 1076 |
+
_make_norm(config.qk_norm_type, q_dim, config.norm_eps, has_bias=False)
|
| 1077 |
+
if config.use_qk_norm
|
| 1078 |
+
else None
|
| 1079 |
+
)
|
| 1080 |
+
self.k_norm = (
|
| 1081 |
+
_make_norm(config.qk_norm_type, kv_dim, config.norm_eps, has_bias=False)
|
| 1082 |
+
if config.use_qk_norm
|
| 1083 |
+
else None
|
| 1084 |
+
)
|
| 1085 |
+
self._rope_requires_runtime_update = False
|
| 1086 |
+
|
| 1087 |
+
if self.use_rope:
|
| 1088 |
+
rope_scaling = dict(config.rope_scaling or {})
|
| 1089 |
+
rope_type = self._rope_type_override or str(
|
| 1090 |
+
rope_scaling.get("rope_type", rope_scaling.get("type", "default"))
|
| 1091 |
+
)
|
| 1092 |
+
self._rope_requires_runtime_update = rope_type == "dynamic"
|
| 1093 |
+
if self._rope_requires_runtime_update:
|
| 1094 |
+
self.register_buffer("_rope_inv_freq", None, persistent=False)
|
| 1095 |
+
self.register_buffer(
|
| 1096 |
+
"_rope_attention_factor", None, persistent=False,
|
| 1097 |
+
)
|
| 1098 |
+
else:
|
| 1099 |
+
inv_freq, attention_factor = _compute_rope_parameters(
|
| 1100 |
+
config,
|
| 1101 |
+
device=torch.device("cpu"),
|
| 1102 |
+
seq_len=config.max_position_embeddings,
|
| 1103 |
+
rope_type_override=self._rope_type_override,
|
| 1104 |
+
)
|
| 1105 |
+
self.register_buffer("_rope_inv_freq", inv_freq, persistent=False)
|
| 1106 |
+
self.register_buffer(
|
| 1107 |
+
"_rope_attention_factor",
|
| 1108 |
+
torch.tensor(float(attention_factor), dtype=torch.float32),
|
| 1109 |
+
persistent=False,
|
| 1110 |
+
)
|
| 1111 |
+
else:
|
| 1112 |
+
self.register_buffer("_rope_inv_freq", None, persistent=False)
|
| 1113 |
+
self.register_buffer("_rope_attention_factor", None, persistent=False)
|
| 1114 |
+
|
| 1115 |
+
def _build_rope(
|
| 1116 |
+
self,
|
| 1117 |
+
position_ids: torch.Tensor,
|
| 1118 |
+
device: torch.device,
|
| 1119 |
+
dtype: torch.dtype,
|
| 1120 |
+
) -> tuple[torch.Tensor, torch.Tensor]:
|
| 1121 |
+
if not self.use_rope:
|
| 1122 |
+
msg = "RoPE requested but rope buffer is not initialized."
|
| 1123 |
+
raise RuntimeError(msg)
|
| 1124 |
+
|
| 1125 |
+
inv_freq = self._rope_inv_freq
|
| 1126 |
+
attention_factor_tensor = self._rope_attention_factor
|
| 1127 |
+
if (
|
| 1128 |
+
inv_freq is None
|
| 1129 |
+
or attention_factor_tensor is None
|
| 1130 |
+
or self._rope_requires_runtime_update
|
| 1131 |
+
):
|
| 1132 |
+
# Keep the sequence-length hint as a tensor so TorchDynamo/vLLM
|
| 1133 |
+
# can trace this path without requiring capture_scalar_outputs.
|
| 1134 |
+
# When low-memory loading leaves the non-persistent cache unset,
|
| 1135 |
+
# rebuild from config for this forward only instead of mutating
|
| 1136 |
+
# buffers inside the compiled graph.
|
| 1137 |
+
seq_len = (
|
| 1138 |
+
torch.max(position_ids) + 1 if position_ids.numel() > 0 else None
|
| 1139 |
+
)
|
| 1140 |
+
inv_freq, attention_factor = _compute_rope_parameters(
|
| 1141 |
+
self.config,
|
| 1142 |
+
device=device,
|
| 1143 |
+
seq_len=seq_len,
|
| 1144 |
+
rope_type_override=self._rope_type_override,
|
| 1145 |
+
)
|
| 1146 |
+
attention_factor_tensor = torch.tensor(
|
| 1147 |
+
float(attention_factor),
|
| 1148 |
+
device=device,
|
| 1149 |
+
dtype=torch.float32,
|
| 1150 |
+
)
|
| 1151 |
+
else:
|
| 1152 |
+
inv_freq = inv_freq.to(device=device)
|
| 1153 |
+
attention_factor_tensor = attention_factor_tensor.to(
|
| 1154 |
+
device=device,
|
| 1155 |
+
dtype=torch.float32,
|
| 1156 |
+
)
|
| 1157 |
+
|
| 1158 |
+
# Mirror the upstream HF decoder-only rotary path closely here.
|
| 1159 |
+
# The matmul-based construction is slightly more numerically stable
|
| 1160 |
+
# than the generic einsum formulation for strict parity probes.
|
| 1161 |
+
inv_freq_expanded = (
|
| 1162 |
+
inv_freq[None, :, None]
|
| 1163 |
+
.to(device=device, dtype=torch.float32)
|
| 1164 |
+
.expand(position_ids.shape[0], -1, 1)
|
| 1165 |
+
)
|
| 1166 |
+
position_ids_expanded = position_ids[:, None, :].to(torch.float32)
|
| 1167 |
+
angles = torch.matmul(
|
| 1168 |
+
inv_freq_expanded,
|
| 1169 |
+
position_ids_expanded,
|
| 1170 |
+
).transpose(1, 2)
|
| 1171 |
+
angles = torch.cat((angles, angles), dim=-1)
|
| 1172 |
+
cos = angles.cos().unsqueeze(1) * attention_factor_tensor
|
| 1173 |
+
sin = angles.sin().unsqueeze(1) * attention_factor_tensor
|
| 1174 |
+
cos = cos.to(dtype)
|
| 1175 |
+
sin = sin.to(dtype)
|
| 1176 |
+
return cos, sin
|
| 1177 |
+
|
| 1178 |
+
def forward(
|
| 1179 |
+
self,
|
| 1180 |
+
hidden_states: torch.Tensor,
|
| 1181 |
+
attention_mask: torch.Tensor | None = None,
|
| 1182 |
+
position_ids: torch.Tensor | None = None,
|
| 1183 |
+
past_key_value: Cache | tuple[torch.Tensor, torch.Tensor] | None = None,
|
| 1184 |
+
cache_position: torch.Tensor | None = None,
|
| 1185 |
+
use_cache: bool = False,
|
| 1186 |
+
output_attentions: bool = False,
|
| 1187 |
+
**kwargs: Any,
|
| 1188 |
+
) -> tuple[
|
| 1189 |
+
torch.Tensor,
|
| 1190 |
+
Cache | tuple[torch.Tensor, torch.Tensor] | None,
|
| 1191 |
+
torch.Tensor | None,
|
| 1192 |
+
]:
|
| 1193 |
+
batch_size, q_len, _ = hidden_states.shape
|
| 1194 |
+
cache_position = _normalize_cache_position(cache_position)
|
| 1195 |
+
|
| 1196 |
+
query_states = self.q_proj(hidden_states)
|
| 1197 |
+
key_states = self.k_proj(hidden_states)
|
| 1198 |
+
value_states = self.v_proj(hidden_states)
|
| 1199 |
+
|
| 1200 |
+
if self.q_norm is not None:
|
| 1201 |
+
query_states = self.q_norm(query_states)
|
| 1202 |
+
if self.k_norm is not None:
|
| 1203 |
+
key_states = self.k_norm(key_states)
|
| 1204 |
+
|
| 1205 |
+
local_num_heads = query_states.shape[-1] // self.head_dim
|
| 1206 |
+
local_num_key_value_heads = key_states.shape[-1] // self.head_dim
|
| 1207 |
+
if local_num_heads <= 0 or local_num_key_value_heads <= 0:
|
| 1208 |
+
msg = "Invalid local attention head count after projection."
|
| 1209 |
+
raise RuntimeError(msg)
|
| 1210 |
+
if query_states.shape[-1] != local_num_heads * self.head_dim:
|
| 1211 |
+
msg = "q_proj output is not divisible by head_dim."
|
| 1212 |
+
raise RuntimeError(msg)
|
| 1213 |
+
if key_states.shape[-1] != local_num_key_value_heads * self.head_dim:
|
| 1214 |
+
msg = "k_proj output is not divisible by head_dim."
|
| 1215 |
+
raise RuntimeError(msg)
|
| 1216 |
+
if value_states.shape[-1] != local_num_key_value_heads * self.head_dim:
|
| 1217 |
+
msg = "v_proj output is not divisible by head_dim."
|
| 1218 |
+
raise RuntimeError(msg)
|
| 1219 |
+
|
| 1220 |
+
query_states = query_states.view(
|
| 1221 |
+
batch_size, q_len, local_num_heads, self.head_dim,
|
| 1222 |
+
)
|
| 1223 |
+
query_states = query_states.transpose(1, 2)
|
| 1224 |
+
key_states = key_states.view(
|
| 1225 |
+
batch_size,
|
| 1226 |
+
q_len,
|
| 1227 |
+
local_num_key_value_heads,
|
| 1228 |
+
self.head_dim,
|
| 1229 |
+
)
|
| 1230 |
+
key_states = key_states.transpose(1, 2)
|
| 1231 |
+
value_states = value_states.view(
|
| 1232 |
+
batch_size,
|
| 1233 |
+
q_len,
|
| 1234 |
+
local_num_key_value_heads,
|
| 1235 |
+
self.head_dim,
|
| 1236 |
+
)
|
| 1237 |
+
value_states = value_states.transpose(1, 2)
|
| 1238 |
+
|
| 1239 |
+
if self.use_rope:
|
| 1240 |
+
if position_ids is None:
|
| 1241 |
+
msg = "position_ids are required for rope attention."
|
| 1242 |
+
raise ValueError(msg)
|
| 1243 |
+
cos, sin = self._build_rope(
|
| 1244 |
+
position_ids, hidden_states.device, query_states.dtype,
|
| 1245 |
+
)
|
| 1246 |
+
query_states, key_states = _apply_rotary_pos_emb(
|
| 1247 |
+
query_states,
|
| 1248 |
+
key_states,
|
| 1249 |
+
cos,
|
| 1250 |
+
sin,
|
| 1251 |
+
)
|
| 1252 |
+
|
| 1253 |
+
if _is_cache_object(past_key_value):
|
| 1254 |
+
if use_cache:
|
| 1255 |
+
key_states, value_states = past_key_value.update(
|
| 1256 |
+
key_states,
|
| 1257 |
+
value_states,
|
| 1258 |
+
self.layer_idx,
|
| 1259 |
+
cache_kwargs={"cache_position": cache_position},
|
| 1260 |
+
)
|
| 1261 |
+
present_key_value = past_key_value
|
| 1262 |
+
elif self.layer_idx < len(past_key_value):
|
| 1263 |
+
past_key, past_value = past_key_value[self.layer_idx]
|
| 1264 |
+
if past_key is not None and past_value is not None:
|
| 1265 |
+
key_states = torch.cat([past_key, key_states], dim=2)
|
| 1266 |
+
value_states = torch.cat([past_value, value_states], dim=2)
|
| 1267 |
+
present_key_value = None
|
| 1268 |
+
else:
|
| 1269 |
+
present_key_value = None
|
| 1270 |
+
elif past_key_value is not None:
|
| 1271 |
+
past_key, past_value = past_key_value
|
| 1272 |
+
key_states = torch.cat([past_key, key_states], dim=2)
|
| 1273 |
+
value_states = torch.cat([past_value, value_states], dim=2)
|
| 1274 |
+
present_key_value = (key_states, value_states) if use_cache else None
|
| 1275 |
+
else:
|
| 1276 |
+
present_key_value = (key_states, value_states) if use_cache else None
|
| 1277 |
+
|
| 1278 |
+
attention_interface = None
|
| 1279 |
+
attn_impl = getattr(self.config, "_attn_implementation", "eager")
|
| 1280 |
+
if attn_impl == "flex_attention" and self.head_dim < 16:
|
| 1281 |
+
attn_impl = "sdpa"
|
| 1282 |
+
if attn_impl != "eager":
|
| 1283 |
+
attention_interface = ALL_ATTENTION_FUNCTIONS.get(attn_impl)
|
| 1284 |
+
|
| 1285 |
+
if attention_interface is not None:
|
| 1286 |
+
attn_output, attn_weights = attention_interface(
|
| 1287 |
+
self,
|
| 1288 |
+
query_states,
|
| 1289 |
+
key_states,
|
| 1290 |
+
value_states,
|
| 1291 |
+
attention_mask,
|
| 1292 |
+
dropout=0.0 if not self.training else self.attention_dropout,
|
| 1293 |
+
scaling=self.scaling,
|
| 1294 |
+
sliding_window=self.sliding_window,
|
| 1295 |
+
**kwargs,
|
| 1296 |
+
)
|
| 1297 |
+
attn_output = attn_output.contiguous()
|
| 1298 |
+
else:
|
| 1299 |
+
if local_num_key_value_heads != local_num_heads:
|
| 1300 |
+
local_num_key_value_groups = (
|
| 1301 |
+
local_num_heads // local_num_key_value_heads
|
| 1302 |
+
)
|
| 1303 |
+
key_states = key_states.repeat_interleave(
|
| 1304 |
+
local_num_key_value_groups, dim=1,
|
| 1305 |
+
)
|
| 1306 |
+
value_states = value_states.repeat_interleave(
|
| 1307 |
+
local_num_key_value_groups, dim=1,
|
| 1308 |
+
)
|
| 1309 |
+
attn_weights = torch.matmul(
|
| 1310 |
+
query_states,
|
| 1311 |
+
key_states.transpose(-1, -2),
|
| 1312 |
+
) * self.scaling
|
| 1313 |
+
if attention_mask is not None:
|
| 1314 |
+
attn_weights = attn_weights + attention_mask
|
| 1315 |
+
attn_weights = F.softmax(attn_weights, dim=-1, dtype=torch.float32)
|
| 1316 |
+
attn_weights = attn_weights.to(query_states.dtype)
|
| 1317 |
+
attn_weights = F.dropout(
|
| 1318 |
+
attn_weights,
|
| 1319 |
+
p=self.attention_dropout if self.training else 0.0,
|
| 1320 |
+
training=self.training,
|
| 1321 |
+
)
|
| 1322 |
+
attn_output = torch.matmul(attn_weights, value_states)
|
| 1323 |
+
attn_output = attn_output.transpose(1, 2).contiguous()
|
| 1324 |
+
|
| 1325 |
+
attn_output = attn_output.reshape(batch_size, q_len, -1).contiguous()
|
| 1326 |
+
attn_output = self.o_proj(attn_output)
|
| 1327 |
+
if not output_attentions:
|
| 1328 |
+
attn_weights = None
|
| 1329 |
+
return attn_output, present_key_value, attn_weights
|
| 1330 |
+
|
| 1331 |
+
|
| 1332 |
+
def _refresh_attention_rope_buffers(module: nn.Module) -> None:
|
| 1333 |
+
"""Rebuild non-persistent RoPE buffers after checkpoint load."""
|
| 1334 |
+
|
| 1335 |
+
for child in module.modules():
|
| 1336 |
+
if not isinstance(child, LizzyAttention):
|
| 1337 |
+
continue
|
| 1338 |
+
|
| 1339 |
+
should_use_rope = _get_rope_layer_flag(child.config, child.layer_idx)
|
| 1340 |
+
child.use_rope = should_use_rope
|
| 1341 |
+
if not should_use_rope:
|
| 1342 |
+
child._rope_requires_runtime_update = False
|
| 1343 |
+
child._rope_inv_freq = None
|
| 1344 |
+
child._rope_attention_factor = None
|
| 1345 |
+
continue
|
| 1346 |
+
|
| 1347 |
+
rope_scaling = dict(child.config.rope_scaling or {})
|
| 1348 |
+
rope_type = child._rope_type_override or str(
|
| 1349 |
+
rope_scaling.get("rope_type", rope_scaling.get("type", "default"))
|
| 1350 |
+
)
|
| 1351 |
+
child._rope_requires_runtime_update = rope_type == "dynamic"
|
| 1352 |
+
if child._rope_requires_runtime_update:
|
| 1353 |
+
child._rope_inv_freq = None
|
| 1354 |
+
child._rope_attention_factor = None
|
| 1355 |
+
continue
|
| 1356 |
+
|
| 1357 |
+
# These buffers are derived from config rather than serialized weights.
|
| 1358 |
+
# Recompute them after load so low-memory materialization cannot leave
|
| 1359 |
+
# stale or uninitialized rotary state behind.
|
| 1360 |
+
inv_freq, attention_factor = _compute_rope_parameters(
|
| 1361 |
+
child.config,
|
| 1362 |
+
device=torch.device("cpu"),
|
| 1363 |
+
seq_len=child.config.max_position_embeddings,
|
| 1364 |
+
rope_type_override=child._rope_type_override,
|
| 1365 |
+
)
|
| 1366 |
+
child._rope_inv_freq = inv_freq
|
| 1367 |
+
child._rope_attention_factor = torch.tensor(
|
| 1368 |
+
float(attention_factor),
|
| 1369 |
+
dtype=torch.float32,
|
| 1370 |
+
)
|
| 1371 |
+
|
| 1372 |
+
|
| 1373 |
+
class LizzyMLP(nn.Module):
|
| 1374 |
+
def __init__(self, config: LizzyConfig) -> None:
|
| 1375 |
+
super().__init__()
|
| 1376 |
+
self.config = config
|
| 1377 |
+
self.act = ACT2FN[config.hidden_act]
|
| 1378 |
+
self.gate_proj = (
|
| 1379 |
+
nn.Linear(
|
| 1380 |
+
config.hidden_size,
|
| 1381 |
+
config.intermediate_size,
|
| 1382 |
+
bias=config.mlp_bias,
|
| 1383 |
+
)
|
| 1384 |
+
if config.mlp_type == "gated"
|
| 1385 |
+
else None
|
| 1386 |
+
)
|
| 1387 |
+
self.up_proj = nn.Linear(
|
| 1388 |
+
config.hidden_size,
|
| 1389 |
+
config.intermediate_size,
|
| 1390 |
+
bias=config.mlp_bias,
|
| 1391 |
+
)
|
| 1392 |
+
self.down_proj = nn.Linear(
|
| 1393 |
+
config.intermediate_size,
|
| 1394 |
+
config.hidden_size,
|
| 1395 |
+
bias=config.mlp_bias,
|
| 1396 |
+
)
|
| 1397 |
+
|
| 1398 |
+
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
| 1399 |
+
if self.gate_proj is None and self.config.mlp_type == "gated":
|
| 1400 |
+
msg = "Missing gated MLP projection layers."
|
| 1401 |
+
raise RuntimeError(msg)
|
| 1402 |
+
if self.config.mlp_type == "gated":
|
| 1403 |
+
if self.gate_proj is None:
|
| 1404 |
+
msg = "Missing gated MLP projection layers."
|
| 1405 |
+
raise RuntimeError(msg)
|
| 1406 |
+
return self.down_proj(self.act(
|
| 1407 |
+
self.gate_proj(hidden_states)) * self.up_proj(hidden_states)
|
| 1408 |
+
)
|
| 1409 |
+
return self.down_proj(self.act(self.up_proj(hidden_states)))
|
| 1410 |
+
|
| 1411 |
+
|
| 1412 |
+
class LizzyDecoderLayer(nn.Module):
|
| 1413 |
+
def __init__(self, config: LizzyConfig, layer_idx: int) -> None:
|
| 1414 |
+
super().__init__()
|
| 1415 |
+
self.layer_type = (
|
| 1416 |
+
str(config.layer_types[layer_idx])
|
| 1417 |
+
if layer_idx < len(config.layer_types)
|
| 1418 |
+
else "full_attention"
|
| 1419 |
+
)
|
| 1420 |
+
self.layer_layout = _get_layer_layout(config, layer_idx)
|
| 1421 |
+
self.self_attn = (
|
| 1422 |
+
LizzyAttention(config, layer_idx)
|
| 1423 |
+
if self.layer_type != "linear_attention"
|
| 1424 |
+
else None
|
| 1425 |
+
)
|
| 1426 |
+
self.linear_attn = (
|
| 1427 |
+
LizzyLinearAttention(config, layer_idx)
|
| 1428 |
+
if self.layer_type == "linear_attention"
|
| 1429 |
+
else None
|
| 1430 |
+
)
|
| 1431 |
+
self.mlp = LizzyMLP(config)
|
| 1432 |
+
|
| 1433 |
+
self.pre_attn_norm = (
|
| 1434 |
+
_make_norm(
|
| 1435 |
+
config.norm_type,
|
| 1436 |
+
config.hidden_size,
|
| 1437 |
+
config.norm_eps,
|
| 1438 |
+
has_bias=config.norm_has_bias,
|
| 1439 |
+
)
|
| 1440 |
+
if self.layer_layout == "decoder_prenorm"
|
| 1441 |
+
else None
|
| 1442 |
+
)
|
| 1443 |
+
self.pre_mlp_norm = (
|
| 1444 |
+
_make_norm(
|
| 1445 |
+
config.norm_type,
|
| 1446 |
+
config.hidden_size,
|
| 1447 |
+
config.norm_eps,
|
| 1448 |
+
has_bias=config.norm_has_bias,
|
| 1449 |
+
)
|
| 1450 |
+
if self.layer_layout == "decoder_prenorm"
|
| 1451 |
+
else None
|
| 1452 |
+
)
|
| 1453 |
+
self.post_attn_norm = (
|
| 1454 |
+
_make_norm(
|
| 1455 |
+
config.norm_type,
|
| 1456 |
+
config.hidden_size,
|
| 1457 |
+
config.norm_eps,
|
| 1458 |
+
has_bias=config.norm_has_bias,
|
| 1459 |
+
)
|
| 1460 |
+
if self.layer_layout == "decoder_postnorm"
|
| 1461 |
+
else None
|
| 1462 |
+
)
|
| 1463 |
+
self.post_mlp_norm = (
|
| 1464 |
+
_make_norm(
|
| 1465 |
+
config.norm_type,
|
| 1466 |
+
config.hidden_size,
|
| 1467 |
+
config.norm_eps,
|
| 1468 |
+
has_bias=config.norm_has_bias
|
| 1469 |
+
)
|
| 1470 |
+
if self.layer_layout == "decoder_postnorm"
|
| 1471 |
+
else None
|
| 1472 |
+
)
|
| 1473 |
+
|
| 1474 |
+
def forward(
|
| 1475 |
+
self,
|
| 1476 |
+
hidden_states: torch.Tensor,
|
| 1477 |
+
attention_mask: torch.Tensor | None = None,
|
| 1478 |
+
position_ids: torch.Tensor | None = None,
|
| 1479 |
+
past_key_value: Cache | tuple[torch.Tensor, torch.Tensor] | None = None,
|
| 1480 |
+
cache_position: torch.Tensor | None = None,
|
| 1481 |
+
use_cache: bool = False,
|
| 1482 |
+
output_attentions: bool = False,
|
| 1483 |
+
**kwargs: Any,
|
| 1484 |
+
) -> tuple[
|
| 1485 |
+
torch.Tensor,
|
| 1486 |
+
Cache | tuple[torch.Tensor, torch.Tensor] | None,
|
| 1487 |
+
torch.Tensor | None,
|
| 1488 |
+
]:
|
| 1489 |
+
residual = hidden_states
|
| 1490 |
+
attn_inputs = (
|
| 1491 |
+
self.pre_attn_norm(hidden_states)
|
| 1492 |
+
if self.pre_attn_norm is not None
|
| 1493 |
+
else hidden_states
|
| 1494 |
+
)
|
| 1495 |
+
if self.linear_attn is not None:
|
| 1496 |
+
attn_output, present_key_value, attn_weights = self.linear_attn(
|
| 1497 |
+
attn_inputs,
|
| 1498 |
+
attention_mask=attention_mask,
|
| 1499 |
+
past_key_value=(
|
| 1500 |
+
past_key_value if _is_cache_object(past_key_value) else None
|
| 1501 |
+
),
|
| 1502 |
+
use_cache=use_cache,
|
| 1503 |
+
output_attentions=output_attentions,
|
| 1504 |
+
**kwargs,
|
| 1505 |
+
)
|
| 1506 |
+
else:
|
| 1507 |
+
assert self.self_attn is not None
|
| 1508 |
+
attn_output, present_key_value, attn_weights = self.self_attn(
|
| 1509 |
+
attn_inputs,
|
| 1510 |
+
attention_mask=attention_mask,
|
| 1511 |
+
position_ids=position_ids,
|
| 1512 |
+
past_key_value=past_key_value,
|
| 1513 |
+
cache_position=cache_position,
|
| 1514 |
+
use_cache=use_cache,
|
| 1515 |
+
output_attentions=output_attentions,
|
| 1516 |
+
**kwargs,
|
| 1517 |
+
)
|
| 1518 |
+
if self.post_attn_norm is not None:
|
| 1519 |
+
attn_output = self.post_attn_norm(attn_output)
|
| 1520 |
+
hidden_states = residual + attn_output
|
| 1521 |
+
|
| 1522 |
+
residual = hidden_states
|
| 1523 |
+
mlp_inputs = (
|
| 1524 |
+
self.pre_mlp_norm(hidden_states)
|
| 1525 |
+
if self.pre_mlp_norm is not None
|
| 1526 |
+
else hidden_states
|
| 1527 |
+
)
|
| 1528 |
+
mlp_output = self.mlp(mlp_inputs)
|
| 1529 |
+
if self.post_mlp_norm is not None:
|
| 1530 |
+
mlp_output = self.post_mlp_norm(mlp_output)
|
| 1531 |
+
hidden_states = residual + mlp_output
|
| 1532 |
+
return hidden_states, present_key_value, attn_weights
|
| 1533 |
+
|
| 1534 |
+
|
| 1535 |
+
class LizzyPreTrainedModel(PreTrainedModel):
|
| 1536 |
+
config_class = LizzyConfig
|
| 1537 |
+
base_model_prefix = "model"
|
| 1538 |
+
supports_gradient_checkpointing = True
|
| 1539 |
+
_no_split_modules = ["LizzyDecoderLayer"]
|
| 1540 |
+
_skip_keys_device_placement = ["past_key_values"]
|
| 1541 |
+
_supports_flash_attn = True
|
| 1542 |
+
_supports_sdpa = True
|
| 1543 |
+
_supports_flex_attn = True
|
| 1544 |
+
_supports_attention_backend = True
|
| 1545 |
+
|
| 1546 |
+
def _init_weights(self, module: nn.Module) -> None:
|
| 1547 |
+
std = self.config.initializer_range
|
| 1548 |
+
if isinstance(module, nn.Linear):
|
| 1549 |
+
module.weight.data.normal_(mean=0.0, std=std)
|
| 1550 |
+
if module.bias is not None:
|
| 1551 |
+
module.bias.data.zero_()
|
| 1552 |
+
elif isinstance(module, nn.Embedding):
|
| 1553 |
+
module.weight.data.normal_(mean=0.0, std=std)
|
| 1554 |
+
if module.padding_idx is not None:
|
| 1555 |
+
module.weight.data[module.padding_idx].zero_()
|
| 1556 |
+
elif isinstance(module, (LizzyRMSNorm, nn.LayerNorm)):
|
| 1557 |
+
if hasattr(module, "weight") and module.weight is not None:
|
| 1558 |
+
module.weight.data.fill_(1.0)
|
| 1559 |
+
if hasattr(module, "bias") and module.bias is not None:
|
| 1560 |
+
module.bias.data.zero_()
|
| 1561 |
+
|
| 1562 |
+
@classmethod
|
| 1563 |
+
def from_pretrained(
|
| 1564 |
+
cls,
|
| 1565 |
+
pretrained_model_name_or_path: str | os.PathLike[str] | None,
|
| 1566 |
+
*model_args: Any,
|
| 1567 |
+
**kwargs: Any,
|
| 1568 |
+
) -> "LizzyPreTrainedModel":
|
| 1569 |
+
model = cast(
|
| 1570 |
+
"LizzyPreTrainedModel",
|
| 1571 |
+
super().from_pretrained(
|
| 1572 |
+
pretrained_model_name_or_path,
|
| 1573 |
+
*model_args,
|
| 1574 |
+
**kwargs,
|
| 1575 |
+
),
|
| 1576 |
+
)
|
| 1577 |
+
_refresh_attention_rope_buffers(model)
|
| 1578 |
+
if hasattr(model, "lm_head") and hasattr(model, "model"):
|
| 1579 |
+
tied_weights_keys = getattr(type(model), "_tied_weights_keys", None)
|
| 1580 |
+
if isinstance(tied_weights_keys, dict) and tied_weights_keys:
|
| 1581 |
+
model._tied_weights_keys = dict(tied_weights_keys)
|
| 1582 |
+
else:
|
| 1583 |
+
model._tied_weights_keys = {
|
| 1584 |
+
"lm_head.weight": "model.embed_tokens.weight",
|
| 1585 |
+
}
|
| 1586 |
+
model._tp_plan = {"lm_head": "colwise_rep"}
|
| 1587 |
+
model._pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
|
| 1588 |
+
return model
|
| 1589 |
+
|
| 1590 |
+
def load_state_dict( # type: ignore[override]
|
| 1591 |
+
self,
|
| 1592 |
+
state_dict: dict[str, torch.Tensor],
|
| 1593 |
+
strict: bool = True,
|
| 1594 |
+
assign: bool = False,
|
| 1595 |
+
) -> Any:
|
| 1596 |
+
remapped_state_dict: dict[str, torch.Tensor] = {}
|
| 1597 |
+
for key, value in state_dict.items():
|
| 1598 |
+
remapped_key = key
|
| 1599 |
+
if ".mlp.fc_in." in key:
|
| 1600 |
+
remapped_key = key.replace(".mlp.fc_in.", ".mlp.up_proj.")
|
| 1601 |
+
elif ".mlp.fc_out." in key:
|
| 1602 |
+
remapped_key = key.replace(".mlp.fc_out.", ".mlp.down_proj.")
|
| 1603 |
+
|
| 1604 |
+
existing = remapped_state_dict.get(remapped_key)
|
| 1605 |
+
if existing is not None and not torch.equal(existing, value):
|
| 1606 |
+
msg = (
|
| 1607 |
+
f"Conflicting legacy Lizzy MLP tensors"
|
| 1608 |
+
f" for key: {remapped_key}"
|
| 1609 |
+
)
|
| 1610 |
+
raise ValueError(msg)
|
| 1611 |
+
remapped_state_dict[remapped_key] = value
|
| 1612 |
+
|
| 1613 |
+
load_result = super().load_state_dict(
|
| 1614 |
+
remapped_state_dict,
|
| 1615 |
+
strict=strict,
|
| 1616 |
+
assign=assign,
|
| 1617 |
+
)
|
| 1618 |
+
# RoPE buffers are intentionally non-persistent, so refresh them after
|
| 1619 |
+
# weight loading instead of trusting constructor-time allocations.
|
| 1620 |
+
_refresh_attention_rope_buffers(self)
|
| 1621 |
+
return load_result
|
| 1622 |
+
|
| 1623 |
+
|
| 1624 |
+
class LizzyModel(LizzyPreTrainedModel):
|
| 1625 |
+
def __init__(self, config: LizzyConfig) -> None:
|
| 1626 |
+
super().__init__(config)
|
| 1627 |
+
self.padding_idx = config.pad_token_id
|
| 1628 |
+
self.vocab_size = config.vocab_size
|
| 1629 |
+
self.embed_tokens = nn.Embedding(
|
| 1630 |
+
config.vocab_size,
|
| 1631 |
+
config.hidden_size,
|
| 1632 |
+
self.padding_idx,
|
| 1633 |
+
)
|
| 1634 |
+
self.embed_positions = (
|
| 1635 |
+
nn.Embedding(config.max_position_embeddings, config.hidden_size)
|
| 1636 |
+
if config.position_embedding_type == "absolute"
|
| 1637 |
+
else None
|
| 1638 |
+
)
|
| 1639 |
+
self.layers = nn.ModuleList(
|
| 1640 |
+
LizzyDecoderLayer(config, layer_idx)
|
| 1641 |
+
for layer_idx in range(config.num_hidden_layers)
|
| 1642 |
+
)
|
| 1643 |
+
self.norm = _make_norm(
|
| 1644 |
+
config.norm_type,
|
| 1645 |
+
config.hidden_size,
|
| 1646 |
+
config.norm_eps,
|
| 1647 |
+
has_bias=config.norm_has_bias,
|
| 1648 |
+
)
|
| 1649 |
+
self.embd_dropout = nn.Dropout(config.embd_dropout)
|
| 1650 |
+
self.gradient_checkpointing = False
|
| 1651 |
+
self.post_init()
|
| 1652 |
+
|
| 1653 |
+
def get_input_embeddings(self) -> nn.Embedding:
|
| 1654 |
+
return self.embed_tokens
|
| 1655 |
+
|
| 1656 |
+
def set_input_embeddings(self, value: nn.Embedding) -> None:
|
| 1657 |
+
self.embed_tokens = value
|
| 1658 |
+
|
| 1659 |
+
def _build_attention_mask(
|
| 1660 |
+
self,
|
| 1661 |
+
attention_mask: torch.Tensor | None,
|
| 1662 |
+
*,
|
| 1663 |
+
batch_size: int,
|
| 1664 |
+
q_len: int,
|
| 1665 |
+
kv_len: int,
|
| 1666 |
+
kv_offset: int,
|
| 1667 |
+
cache_position: torch.Tensor,
|
| 1668 |
+
device: torch.device,
|
| 1669 |
+
dtype: torch.dtype,
|
| 1670 |
+
sliding_window: int | None = None,
|
| 1671 |
+
) -> torch.Tensor:
|
| 1672 |
+
kv_len = (
|
| 1673 |
+
int(kv_len.item()) if isinstance(kv_len, torch.Tensor) else int(kv_len)
|
| 1674 |
+
)
|
| 1675 |
+
kv_offset = (
|
| 1676 |
+
int(kv_offset.item())
|
| 1677 |
+
if isinstance(kv_offset, torch.Tensor)
|
| 1678 |
+
else int(kv_offset)
|
| 1679 |
+
)
|
| 1680 |
+
min_value = torch.finfo(dtype).min
|
| 1681 |
+
source_positions = cache_position.to(device=device).view(-1, 1)
|
| 1682 |
+
target_positions = torch.arange(
|
| 1683 |
+
kv_offset,
|
| 1684 |
+
kv_offset + kv_len,
|
| 1685 |
+
device=device,
|
| 1686 |
+
).unsqueeze(0)
|
| 1687 |
+
causal = torch.zeros((q_len, kv_len), dtype=dtype, device=device)
|
| 1688 |
+
causal = causal.masked_fill(target_positions > source_positions, min_value)
|
| 1689 |
+
if sliding_window is not None:
|
| 1690 |
+
lower_bound = source_positions - int(sliding_window) + 1
|
| 1691 |
+
causal = causal.masked_fill(target_positions < lower_bound, min_value)
|
| 1692 |
+
causal = causal.unsqueeze(0).unsqueeze(0).expand(batch_size, 1, -1, -1)
|
| 1693 |
+
|
| 1694 |
+
if attention_mask is None:
|
| 1695 |
+
return causal
|
| 1696 |
+
|
| 1697 |
+
if attention_mask.dim() != 2:
|
| 1698 |
+
msg = "attention_mask must be 2D [batch, sequence]."
|
| 1699 |
+
raise ValueError(msg)
|
| 1700 |
+
if attention_mask.shape[1] < kv_len:
|
| 1701 |
+
pad = torch.ones(
|
| 1702 |
+
(attention_mask.shape[0], kv_len - attention_mask.shape[1]),
|
| 1703 |
+
dtype=attention_mask.dtype,
|
| 1704 |
+
device=attention_mask.device,
|
| 1705 |
+
)
|
| 1706 |
+
attention_mask = torch.cat([pad, attention_mask], dim=1)
|
| 1707 |
+
elif attention_mask.shape[1] > kv_len:
|
| 1708 |
+
attention_mask = attention_mask[:, -kv_len:]
|
| 1709 |
+
|
| 1710 |
+
expanded = attention_mask[:, None, None, :].to(device=device)
|
| 1711 |
+
padding = (expanded == 0).to(dtype) * min_value
|
| 1712 |
+
return causal + padding
|
| 1713 |
+
|
| 1714 |
+
def forward(
|
| 1715 |
+
self,
|
| 1716 |
+
input_ids: torch.LongTensor | None = None,
|
| 1717 |
+
attention_mask: torch.Tensor | dict[str, torch.Tensor] | None = None,
|
| 1718 |
+
position_ids: torch.LongTensor | None = None,
|
| 1719 |
+
past_key_values: Cache | tuple[
|
| 1720 |
+
tuple[torch.Tensor, torch.Tensor], ...
|
| 1721 |
+
] | None = None,
|
| 1722 |
+
inputs_embeds: torch.FloatTensor | None = None,
|
| 1723 |
+
cache_position: torch.LongTensor | None = None,
|
| 1724 |
+
use_cache: bool | None = None,
|
| 1725 |
+
output_attentions: bool | None = None,
|
| 1726 |
+
output_hidden_states: bool | None = None,
|
| 1727 |
+
return_dict: bool | None = None,
|
| 1728 |
+
**kwargs: Any,
|
| 1729 |
+
) -> BaseModelOutputWithPast | tuple[Any, ...]:
|
| 1730 |
+
if (input_ids is None) == (inputs_embeds is None):
|
| 1731 |
+
msg = "Exactly one of input_ids or inputs_embeds must be provided."
|
| 1732 |
+
raise ValueError(msg)
|
| 1733 |
+
|
| 1734 |
+
output_attentions = (
|
| 1735 |
+
bool(output_attentions) if output_attentions is not None else False
|
| 1736 |
+
)
|
| 1737 |
+
output_hidden_states = (
|
| 1738 |
+
bool(output_hidden_states)
|
| 1739 |
+
if output_hidden_states is not None
|
| 1740 |
+
else False
|
| 1741 |
+
)
|
| 1742 |
+
use_cache = (
|
| 1743 |
+
bool(use_cache)
|
| 1744 |
+
if use_cache is not None
|
| 1745 |
+
else bool(self.config.use_cache)
|
| 1746 |
+
)
|
| 1747 |
+
return_dict = bool(return_dict) if return_dict is not None else True
|
| 1748 |
+
|
| 1749 |
+
if inputs_embeds is None:
|
| 1750 |
+
hidden_states = self.embed_tokens(input_ids)
|
| 1751 |
+
batch_size, seq_len = input_ids.shape
|
| 1752 |
+
else:
|
| 1753 |
+
hidden_states = inputs_embeds
|
| 1754 |
+
batch_size, seq_len, _ = inputs_embeds.shape
|
| 1755 |
+
|
| 1756 |
+
cache_object = (
|
| 1757 |
+
past_key_values
|
| 1758 |
+
if _is_cache_object(past_key_values)
|
| 1759 |
+
else None
|
| 1760 |
+
)
|
| 1761 |
+
if use_cache and _has_linear_attention(self.config):
|
| 1762 |
+
# Transformers 5.4 seeds `generate()` with an empty DynamicCache
|
| 1763 |
+
# for standard causal decoders. Hybrid Lizzy checkpoints need the
|
| 1764 |
+
# mixed cache below instead, because linear-attention layers read
|
| 1765 |
+
# DeltaNet convolution/recurrent state during the prefill pass.
|
| 1766 |
+
if cache_object is not None and not isinstance(
|
| 1767 |
+
cache_object, LizzyHybridDynamicCache,
|
| 1768 |
+
):
|
| 1769 |
+
if int(cache_object.get_seq_length()) > 0:
|
| 1770 |
+
msg = (
|
| 1771 |
+
"Hybrid Lizzy checkpoints require "
|
| 1772 |
+
"LizzyHybridDynamicCache once generation cache "
|
| 1773 |
+
"state is populated."
|
| 1774 |
+
)
|
| 1775 |
+
raise ValueError(msg)
|
| 1776 |
+
cache_object = LizzyHybridDynamicCache(config=self.config)
|
| 1777 |
+
past_key_values = cache_object
|
| 1778 |
+
if use_cache and cache_object is None and past_key_values is None:
|
| 1779 |
+
if _has_linear_attention(self.config):
|
| 1780 |
+
# Linear-attention checkpoints need a mixed cache that can hold
|
| 1781 |
+
# both KV tensors and recurrent DeltaNet state.
|
| 1782 |
+
cache_object = LizzyHybridDynamicCache(config=self.config)
|
| 1783 |
+
else:
|
| 1784 |
+
cache_object = DynamicCache()
|
| 1785 |
+
past_key_values = cache_object
|
| 1786 |
+
|
| 1787 |
+
if cache_object is not None:
|
| 1788 |
+
past_length = int(cache_object.get_seq_length())
|
| 1789 |
+
else:
|
| 1790 |
+
past_length = _legacy_cache_length(past_key_values)
|
| 1791 |
+
|
| 1792 |
+
cache_position = _normalize_cache_position(cache_position)
|
| 1793 |
+
if cache_position is None:
|
| 1794 |
+
cache_position = torch.arange(
|
| 1795 |
+
past_length,
|
| 1796 |
+
past_length + seq_len,
|
| 1797 |
+
dtype=torch.long,
|
| 1798 |
+
device=hidden_states.device,
|
| 1799 |
+
)
|
| 1800 |
+
|
| 1801 |
+
if position_ids is None:
|
| 1802 |
+
position_ids = cache_position.unsqueeze(0).expand(batch_size, -1)
|
| 1803 |
+
|
| 1804 |
+
if self.embed_positions is not None:
|
| 1805 |
+
hidden_states = hidden_states + self.embed_positions(position_ids)
|
| 1806 |
+
|
| 1807 |
+
hidden_states = self.embd_dropout(hidden_states)
|
| 1808 |
+
|
| 1809 |
+
if self.training and self.gradient_checkpointing:
|
| 1810 |
+
use_cache = False
|
| 1811 |
+
|
| 1812 |
+
layer_types = list(self.config.layer_types)
|
| 1813 |
+
if not layer_types:
|
| 1814 |
+
layer_types = ["full_attention"] * len(self.layers)
|
| 1815 |
+
has_linear_attention = any(
|
| 1816 |
+
str(layer_type) == "linear_attention" for layer_type in layer_types
|
| 1817 |
+
)
|
| 1818 |
+
|
| 1819 |
+
_attn_impl = getattr(self.config, "_attn_implementation", "eager")
|
| 1820 |
+
if has_linear_attention and isinstance(attention_mask, dict):
|
| 1821 |
+
linear_attention_mask = attention_mask.get("linear_attention")
|
| 1822 |
+
else:
|
| 1823 |
+
linear_attention_mask = attention_mask
|
| 1824 |
+
if (
|
| 1825 |
+
has_linear_attention
|
| 1826 |
+
and cache_object is not None
|
| 1827 |
+
and getattr(cache_object, "has_previous_state", False)
|
| 1828 |
+
):
|
| 1829 |
+
linear_attention_mask = None
|
| 1830 |
+
elif (
|
| 1831 |
+
has_linear_attention
|
| 1832 |
+
and attention_mask is not None
|
| 1833 |
+
and not isinstance(attention_mask, dict)
|
| 1834 |
+
and torch.all(attention_mask == 1)
|
| 1835 |
+
):
|
| 1836 |
+
linear_attention_mask = None
|
| 1837 |
+
|
| 1838 |
+
if (
|
| 1839 |
+
_attn_impl == "flash_attention_2"
|
| 1840 |
+
and not isinstance(attention_mask, dict)
|
| 1841 |
+
):
|
| 1842 |
+
# Flash attention handles causal masking (via is_causal) and
|
| 1843 |
+
# padding (via 2D mask) natively; skip building a 4D mask.
|
| 1844 |
+
attention_mask_mapping = {
|
| 1845 |
+
lt: attention_mask
|
| 1846 |
+
for lt in dict.fromkeys(layer_types)
|
| 1847 |
+
if lt != "linear_attention"
|
| 1848 |
+
}
|
| 1849 |
+
elif _attn_impl == "sdpa" and attention_mask is None:
|
| 1850 |
+
attention_mask_mapping = {}
|
| 1851 |
+
for layer_type in dict.fromkeys(layer_types):
|
| 1852 |
+
if layer_type == "linear_attention":
|
| 1853 |
+
continue
|
| 1854 |
+
if layer_type == "full_attention":
|
| 1855 |
+
# Match upstream decoder-only HF models: when SDPA sees
|
| 1856 |
+
# plain causal full attention with no padding mask to
|
| 1857 |
+
# preserve, let it use its native is_causal fast-path
|
| 1858 |
+
# instead of forcing an explicit 4D bias tensor.
|
| 1859 |
+
attention_mask_mapping[layer_type] = None
|
| 1860 |
+
continue
|
| 1861 |
+
layer_idx = layer_types.index(layer_type)
|
| 1862 |
+
if cache_object is not None:
|
| 1863 |
+
kv_len, kv_offset = cache_object.get_mask_sizes(
|
| 1864 |
+
seq_len, layer_idx,
|
| 1865 |
+
)
|
| 1866 |
+
else:
|
| 1867 |
+
kv_len = past_length + seq_len
|
| 1868 |
+
kv_offset = 0
|
| 1869 |
+
attention_mask_mapping[layer_type] = self._build_attention_mask(
|
| 1870 |
+
attention_mask,
|
| 1871 |
+
batch_size=batch_size,
|
| 1872 |
+
q_len=seq_len,
|
| 1873 |
+
kv_len=kv_len,
|
| 1874 |
+
kv_offset=kv_offset,
|
| 1875 |
+
cache_position=cache_position,
|
| 1876 |
+
device=hidden_states.device,
|
| 1877 |
+
dtype=hidden_states.dtype,
|
| 1878 |
+
sliding_window=(
|
| 1879 |
+
self.config.sliding_window
|
| 1880 |
+
if layer_type == "sliding_attention"
|
| 1881 |
+
else None
|
| 1882 |
+
),
|
| 1883 |
+
)
|
| 1884 |
+
elif isinstance(attention_mask, dict):
|
| 1885 |
+
attention_mask_mapping = {
|
| 1886 |
+
key: value
|
| 1887 |
+
for key, value in attention_mask.items()
|
| 1888 |
+
if key != "linear_attention"
|
| 1889 |
+
}
|
| 1890 |
+
else:
|
| 1891 |
+
attention_mask_mapping: dict[str, torch.Tensor] = {}
|
| 1892 |
+
for layer_type in dict.fromkeys(layer_types):
|
| 1893 |
+
if layer_type == "linear_attention":
|
| 1894 |
+
continue
|
| 1895 |
+
layer_idx = layer_types.index(layer_type)
|
| 1896 |
+
if cache_object is not None:
|
| 1897 |
+
kv_len, kv_offset = cache_object.get_mask_sizes(
|
| 1898 |
+
seq_len, layer_idx,
|
| 1899 |
+
)
|
| 1900 |
+
else:
|
| 1901 |
+
kv_len = past_length + seq_len
|
| 1902 |
+
kv_offset = 0
|
| 1903 |
+
attention_mask_mapping[layer_type] = self._build_attention_mask(
|
| 1904 |
+
attention_mask,
|
| 1905 |
+
batch_size=batch_size,
|
| 1906 |
+
q_len=seq_len,
|
| 1907 |
+
kv_len=kv_len,
|
| 1908 |
+
kv_offset=kv_offset,
|
| 1909 |
+
cache_position=cache_position,
|
| 1910 |
+
device=hidden_states.device,
|
| 1911 |
+
dtype=hidden_states.dtype,
|
| 1912 |
+
sliding_window=(
|
| 1913 |
+
self.config.sliding_window
|
| 1914 |
+
if layer_type == "sliding_attention"
|
| 1915 |
+
else None
|
| 1916 |
+
),
|
| 1917 |
+
)
|
| 1918 |
+
|
| 1919 |
+
all_hidden_states = [] if output_hidden_states else None
|
| 1920 |
+
all_attentions = [] if output_attentions else None
|
| 1921 |
+
next_cache = (
|
| 1922 |
+
cache_object
|
| 1923 |
+
if cache_object is not None
|
| 1924 |
+
else ([] if use_cache else None)
|
| 1925 |
+
)
|
| 1926 |
+
gradient_checkpointing_func = getattr(
|
| 1927 |
+
self,
|
| 1928 |
+
"_gradient_checkpointing_func",
|
| 1929 |
+
checkpoint,
|
| 1930 |
+
)
|
| 1931 |
+
|
| 1932 |
+
for idx, layer in enumerate(self.layers):
|
| 1933 |
+
if output_hidden_states and all_hidden_states is not None:
|
| 1934 |
+
all_hidden_states.append(hidden_states)
|
| 1935 |
+
layer_type = (
|
| 1936 |
+
layer_types[idx]
|
| 1937 |
+
if idx < len(layer_types)
|
| 1938 |
+
else "full_attention"
|
| 1939 |
+
)
|
| 1940 |
+
if layer_type == "linear_attention":
|
| 1941 |
+
layer_attention_mask = linear_attention_mask
|
| 1942 |
+
else:
|
| 1943 |
+
layer_attention_mask = attention_mask_mapping[layer_type]
|
| 1944 |
+
|
| 1945 |
+
if cache_object is not None:
|
| 1946 |
+
layer_past: Cache | tuple[
|
| 1947 |
+
torch.Tensor, torch.Tensor
|
| 1948 |
+
] | None = cache_object
|
| 1949 |
+
elif past_key_values is not None:
|
| 1950 |
+
layer_past = past_key_values[idx]
|
| 1951 |
+
if layer_past is not None and layer_past[0] is None:
|
| 1952 |
+
layer_past = None
|
| 1953 |
+
else:
|
| 1954 |
+
layer_past = None
|
| 1955 |
+
|
| 1956 |
+
if self.training and self.gradient_checkpointing:
|
| 1957 |
+
def custom_forward(hidden_states: torch.Tensor) -> Any:
|
| 1958 |
+
layer_outputs = layer(
|
| 1959 |
+
hidden_states,
|
| 1960 |
+
attention_mask=layer_attention_mask,
|
| 1961 |
+
position_ids=position_ids,
|
| 1962 |
+
past_key_value=None,
|
| 1963 |
+
cache_position=cache_position,
|
| 1964 |
+
use_cache=False,
|
| 1965 |
+
output_attentions=output_attentions,
|
| 1966 |
+
**kwargs,
|
| 1967 |
+
)
|
| 1968 |
+
if output_attentions:
|
| 1969 |
+
return layer_outputs[0], layer_outputs[2]
|
| 1970 |
+
return layer_outputs[0]
|
| 1971 |
+
|
| 1972 |
+
checkpointed_outputs = gradient_checkpointing_func(
|
| 1973 |
+
custom_forward, hidden_states,
|
| 1974 |
+
)
|
| 1975 |
+
if output_attentions:
|
| 1976 |
+
hidden_states, attn_weights = checkpointed_outputs
|
| 1977 |
+
else:
|
| 1978 |
+
hidden_states = checkpointed_outputs
|
| 1979 |
+
attn_weights = None
|
| 1980 |
+
present = None
|
| 1981 |
+
else:
|
| 1982 |
+
hidden_states, present, attn_weights = layer(
|
| 1983 |
+
hidden_states,
|
| 1984 |
+
attention_mask=layer_attention_mask,
|
| 1985 |
+
position_ids=position_ids,
|
| 1986 |
+
past_key_value=layer_past,
|
| 1987 |
+
cache_position=cache_position,
|
| 1988 |
+
use_cache=use_cache,
|
| 1989 |
+
output_attentions=output_attentions,
|
| 1990 |
+
**kwargs,
|
| 1991 |
+
)
|
| 1992 |
+
if use_cache and next_cache is not None and cache_object is None:
|
| 1993 |
+
next_cache.append(present)
|
| 1994 |
+
|
| 1995 |
+
if output_attentions and all_attentions is not None:
|
| 1996 |
+
all_attentions.append(attn_weights)
|
| 1997 |
+
|
| 1998 |
+
hidden_states = self.norm(hidden_states)
|
| 1999 |
+
if output_hidden_states and all_hidden_states is not None:
|
| 2000 |
+
all_hidden_states.append(hidden_states)
|
| 2001 |
+
|
| 2002 |
+
past_key_values_output: Cache | tuple[
|
| 2003 |
+
tuple[torch.Tensor, torch.Tensor], ...
|
| 2004 |
+
] | None = None
|
| 2005 |
+
if use_cache and next_cache is not None:
|
| 2006 |
+
if cache_object is not None:
|
| 2007 |
+
past_key_values_output = cache_object
|
| 2008 |
+
else:
|
| 2009 |
+
past_key_values_output = tuple(next_cache)
|
| 2010 |
+
|
| 2011 |
+
if not return_dict:
|
| 2012 |
+
output: tuple[Any, ...] = (hidden_states,)
|
| 2013 |
+
if past_key_values_output is not None:
|
| 2014 |
+
output = output + (past_key_values_output,)
|
| 2015 |
+
if output_hidden_states and all_hidden_states is not None:
|
| 2016 |
+
output = output + (tuple(all_hidden_states),)
|
| 2017 |
+
if output_attentions and all_attentions is not None:
|
| 2018 |
+
output = output + (tuple(all_attentions),)
|
| 2019 |
+
return output
|
| 2020 |
+
|
| 2021 |
+
return BaseModelOutputWithPast(
|
| 2022 |
+
last_hidden_state=hidden_states,
|
| 2023 |
+
past_key_values=past_key_values_output,
|
| 2024 |
+
hidden_states=(
|
| 2025 |
+
tuple(all_hidden_states)
|
| 2026 |
+
if all_hidden_states is not None
|
| 2027 |
+
else None
|
| 2028 |
+
),
|
| 2029 |
+
attentions=(
|
| 2030 |
+
tuple(all_attentions)
|
| 2031 |
+
if all_attentions is not None
|
| 2032 |
+
else None
|
| 2033 |
+
),
|
| 2034 |
+
)
|
| 2035 |
+
|
| 2036 |
+
|
| 2037 |
+
class LizzyForCausalLM(LizzyPreTrainedModel, GenerationMixin):
|
| 2038 |
+
config_class = LizzyConfig
|
| 2039 |
+
# Transformers 5.4 expects an expanded target->source mapping here rather than
|
| 2040 |
+
# the older list-based shorthand.
|
| 2041 |
+
_tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
|
| 2042 |
+
_tp_plan = {"lm_head": "colwise_rep"}
|
| 2043 |
+
_pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
|
| 2044 |
+
|
| 2045 |
+
def __init__(self, config: LizzyConfig) -> None:
|
| 2046 |
+
super().__init__(config)
|
| 2047 |
+
self.model = LizzyModel(config)
|
| 2048 |
+
self.vocab_size = config.vocab_size
|
| 2049 |
+
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
|
| 2050 |
+
self.post_init()
|
| 2051 |
+
|
| 2052 |
+
def get_input_embeddings(self) -> nn.Embedding:
|
| 2053 |
+
return self.model.get_input_embeddings()
|
| 2054 |
+
|
| 2055 |
+
def set_input_embeddings(self, value: nn.Embedding) -> None:
|
| 2056 |
+
self.model.set_input_embeddings(value)
|
| 2057 |
+
|
| 2058 |
+
def get_output_embeddings(self) -> nn.Module:
|
| 2059 |
+
return self.lm_head
|
| 2060 |
+
|
| 2061 |
+
def set_output_embeddings(self, new_embeddings: nn.Module) -> None:
|
| 2062 |
+
self.lm_head = new_embeddings
|
| 2063 |
+
|
| 2064 |
+
def prepare_inputs_for_generation(
|
| 2065 |
+
self,
|
| 2066 |
+
input_ids: torch.LongTensor,
|
| 2067 |
+
past_key_values: Cache | tuple[
|
| 2068 |
+
tuple[torch.Tensor, torch.Tensor], ...
|
| 2069 |
+
] | None = None,
|
| 2070 |
+
attention_mask: torch.Tensor | None = None,
|
| 2071 |
+
inputs_embeds: torch.FloatTensor | None = None,
|
| 2072 |
+
cache_position: torch.LongTensor | None = None,
|
| 2073 |
+
**kwargs: Any,
|
| 2074 |
+
) -> dict[str, Any]:
|
| 2075 |
+
past_length = 0
|
| 2076 |
+
if past_key_values is not None:
|
| 2077 |
+
if _is_cache_object(past_key_values):
|
| 2078 |
+
past_length = int(past_key_values.get_seq_length())
|
| 2079 |
+
else:
|
| 2080 |
+
past_length = _legacy_cache_length(past_key_values)
|
| 2081 |
+
|
| 2082 |
+
cache_position = _normalize_cache_position(cache_position)
|
| 2083 |
+
if cache_position is None:
|
| 2084 |
+
if past_key_values is not None:
|
| 2085 |
+
new_tokens = input_ids.shape[1] - past_length
|
| 2086 |
+
if new_tokens <= 0:
|
| 2087 |
+
new_tokens = 1
|
| 2088 |
+
cache_position = torch.arange(
|
| 2089 |
+
past_length,
|
| 2090 |
+
past_length + new_tokens,
|
| 2091 |
+
device=input_ids.device,
|
| 2092 |
+
)
|
| 2093 |
+
else:
|
| 2094 |
+
cache_position = torch.arange(
|
| 2095 |
+
input_ids.shape[1],
|
| 2096 |
+
device=input_ids.device,
|
| 2097 |
+
)
|
| 2098 |
+
|
| 2099 |
+
if past_key_values is not None:
|
| 2100 |
+
input_ids = input_ids[:, -cache_position.shape[0] :]
|
| 2101 |
+
if attention_mask is not None:
|
| 2102 |
+
attn_mask_idx = (past_length + input_ids.shape[1])
|
| 2103 |
+
attention_mask = attention_mask[:, -attn_mask_idx :]
|
| 2104 |
+
|
| 2105 |
+
if inputs_embeds is not None and past_key_values is None:
|
| 2106 |
+
model_inputs: dict[str, Any] = {"inputs_embeds": inputs_embeds}
|
| 2107 |
+
else:
|
| 2108 |
+
model_inputs = {"input_ids": input_ids.contiguous()}
|
| 2109 |
+
|
| 2110 |
+
model_inputs.update(
|
| 2111 |
+
{
|
| 2112 |
+
"past_key_values": past_key_values,
|
| 2113 |
+
"attention_mask": attention_mask,
|
| 2114 |
+
"cache_position": cache_position,
|
| 2115 |
+
"use_cache": kwargs.get("use_cache", self.config.use_cache),
|
| 2116 |
+
},
|
| 2117 |
+
)
|
| 2118 |
+
return model_inputs
|
| 2119 |
+
|
| 2120 |
+
def forward(
|
| 2121 |
+
self,
|
| 2122 |
+
input_ids: torch.LongTensor | None = None,
|
| 2123 |
+
attention_mask: torch.Tensor | None = None,
|
| 2124 |
+
position_ids: torch.LongTensor | None = None,
|
| 2125 |
+
past_key_values: Cache | tuple[
|
| 2126 |
+
tuple[torch.Tensor, torch.Tensor], ...
|
| 2127 |
+
] | None = None,
|
| 2128 |
+
inputs_embeds: torch.FloatTensor | None = None,
|
| 2129 |
+
labels: torch.LongTensor | None = None,
|
| 2130 |
+
cache_position: torch.LongTensor | None = None,
|
| 2131 |
+
use_cache: bool | None = None,
|
| 2132 |
+
output_attentions: bool | None = None,
|
| 2133 |
+
output_hidden_states: bool | None = None,
|
| 2134 |
+
return_dict: bool | None = None,
|
| 2135 |
+
logits_to_keep: int | torch.Tensor = 0,
|
| 2136 |
+
**kwargs: Any,
|
| 2137 |
+
) -> CausalLMOutputWithPast | tuple[Any, ...]:
|
| 2138 |
+
# HF eval loaders call `forward()` without an explicit return_dict,
|
| 2139 |
+
# so local Lizzy exports must normalize the optional flag first.
|
| 2140 |
+
return_dict = bool(return_dict) if return_dict is not None else True
|
| 2141 |
+
|
| 2142 |
+
outputs = self.model(
|
| 2143 |
+
input_ids=input_ids,
|
| 2144 |
+
attention_mask=attention_mask,
|
| 2145 |
+
position_ids=position_ids,
|
| 2146 |
+
past_key_values=past_key_values,
|
| 2147 |
+
inputs_embeds=inputs_embeds,
|
| 2148 |
+
cache_position=cache_position,
|
| 2149 |
+
use_cache=use_cache,
|
| 2150 |
+
output_attentions=output_attentions,
|
| 2151 |
+
output_hidden_states=output_hidden_states,
|
| 2152 |
+
return_dict=return_dict,
|
| 2153 |
+
**kwargs,
|
| 2154 |
+
)
|
| 2155 |
+
|
| 2156 |
+
hidden_states = outputs[0] if not return_dict else outputs.last_hidden_state
|
| 2157 |
+
slice_indices = (
|
| 2158 |
+
slice(-logits_to_keep, None)
|
| 2159 |
+
if isinstance(logits_to_keep, int)
|
| 2160 |
+
else logits_to_keep
|
| 2161 |
+
)
|
| 2162 |
+
if labels is not None:
|
| 2163 |
+
full_logits = self.lm_head(hidden_states)
|
| 2164 |
+
logits = full_logits[:, slice_indices, :]
|
| 2165 |
+
else:
|
| 2166 |
+
full_logits = None
|
| 2167 |
+
logits = self.lm_head(hidden_states[:, slice_indices, :])
|
| 2168 |
+
|
| 2169 |
+
loss = None
|
| 2170 |
+
if labels is not None:
|
| 2171 |
+
shift_logits = full_logits[..., :-1, :].contiguous()
|
| 2172 |
+
shift_labels = labels[..., 1:].contiguous()
|
| 2173 |
+
loss = F.cross_entropy(
|
| 2174 |
+
shift_logits.view(-1, shift_logits.size(-1)),
|
| 2175 |
+
shift_labels.view(-1),
|
| 2176 |
+
)
|
| 2177 |
+
|
| 2178 |
+
if not return_dict:
|
| 2179 |
+
output = (logits,) + outputs[1:]
|
| 2180 |
+
if loss is not None:
|
| 2181 |
+
output = (loss,) + output
|
| 2182 |
+
return output
|
| 2183 |
+
|
| 2184 |
+
return CausalLMOutputWithPast(
|
| 2185 |
+
loss=loss,
|
| 2186 |
+
logits=logits,
|
| 2187 |
+
past_key_values=outputs.past_key_values,
|
| 2188 |
+
hidden_states=outputs.hidden_states,
|
| 2189 |
+
attentions=outputs.attentions,
|
| 2190 |
+
)
|
quickstart.py
ADDED
|
@@ -0,0 +1,36 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""Minimal inference example for the private Lizzy 7B checkpoint."""
|
| 3 |
+
|
| 4 |
+
from __future__ import annotations
|
| 5 |
+
|
| 6 |
+
import os
|
| 7 |
+
|
| 8 |
+
|
| 9 |
+
def main() -> None:
|
| 10 |
+
repo_id = os.getenv("FLOWER_MODEL_ID", "flwrlabs/Lizzy-7B")
|
| 11 |
+
print("Model ID:", repo_id)
|
| 12 |
+
print(
|
| 13 |
+
"Data note:",
|
| 14 |
+
"Flower release drafts should always disclose that Flower/Lizzy variants add private synthetic data during both pre-training and post-training to favour British behaviour and knowledge. Those private synthetic datasets are not redistributed in the release pack.",
|
| 15 |
+
)
|
| 16 |
+
print("HF_TOKEN present:", bool(os.getenv("HF_TOKEN")))
|
| 17 |
+
print("This example is intentionally non-executing by default.")
|
| 18 |
+
print("Use one of the snippets below after installing transformers or vLLM:")
|
| 19 |
+
print()
|
| 20 |
+
print("Transformers:")
|
| 21 |
+
print(
|
| 22 |
+
" tokenizer = AutoTokenizer.from_pretrained(repo_id, trust_remote_code=True)"
|
| 23 |
+
)
|
| 24 |
+
print(
|
| 25 |
+
" model = AutoModelForCausalLM.from_pretrained(repo_id, trust_remote_code=True, torch_dtype='auto')"
|
| 26 |
+
)
|
| 27 |
+
print()
|
| 28 |
+
print("vLLM:")
|
| 29 |
+
print(
|
| 30 |
+
" python -m vllm.entrypoints.openai.api_server --model "
|
| 31 |
+
"flwrlabs/Lizzy-7B --trust-remote-code --max-model-len 8192"
|
| 32 |
+
)
|
| 33 |
+
|
| 34 |
+
|
| 35 |
+
if __name__ == "__main__":
|
| 36 |
+
main()
|
tokenization_lizzy.py
ADDED
|
@@ -0,0 +1,51 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from __future__ import annotations
|
| 2 |
+
|
| 3 |
+
import json
|
| 4 |
+
from pathlib import Path
|
| 5 |
+
from typing import Any
|
| 6 |
+
|
| 7 |
+
from transformers import PreTrainedTokenizerFast
|
| 8 |
+
|
| 9 |
+
|
| 10 |
+
class LizzyTokenizerFast(PreTrainedTokenizerFast):
|
| 11 |
+
"""Family-agnostic fast tokenizer wrapper for Lizzy checkpoints."""
|
| 12 |
+
|
| 13 |
+
model_input_names = ["input_ids", "attention_mask"]
|
| 14 |
+
|
| 15 |
+
def __init__(self, *args: Any, **kwargs: Any) -> None:
|
| 16 |
+
preserved_keys = (
|
| 17 |
+
"add_prefix_space",
|
| 18 |
+
"add_bos_token",
|
| 19 |
+
"add_eos_token",
|
| 20 |
+
"clean_up_tokenization_spaces",
|
| 21 |
+
"use_default_system_prompt",
|
| 22 |
+
"legacy",
|
| 23 |
+
"fix_mistral_regex",
|
| 24 |
+
)
|
| 25 |
+
preserved_init_attrs = {
|
| 26 |
+
key: kwargs.get(key)
|
| 27 |
+
for key in preserved_keys
|
| 28 |
+
if key in kwargs
|
| 29 |
+
}
|
| 30 |
+
super().__init__(*args, **kwargs)
|
| 31 |
+
init_kwargs = getattr(self, "init_kwargs", {})
|
| 32 |
+
local_payload: dict[str, Any] = {}
|
| 33 |
+
config_path = (
|
| 34 |
+
Path(str(getattr(self, "name_or_path", ""))) / "tokenizer_config.json"
|
| 35 |
+
)
|
| 36 |
+
if config_path.is_file():
|
| 37 |
+
try:
|
| 38 |
+
local_payload = json.loads(config_path.read_text(encoding="utf-8"))
|
| 39 |
+
except Exception:
|
| 40 |
+
local_payload = {}
|
| 41 |
+
for key in preserved_keys:
|
| 42 |
+
value = preserved_init_attrs.get(key, init_kwargs.get(key))
|
| 43 |
+
if value is None:
|
| 44 |
+
value = local_payload.get(key)
|
| 45 |
+
if value is not None:
|
| 46 |
+
setattr(self, key, value)
|
| 47 |
+
|
| 48 |
+
@property
|
| 49 |
+
def all_special_tokens_extended(self) -> list[str]:
|
| 50 |
+
"""Compatibility shim for runtimes still expecting the pre-5.4 API."""
|
| 51 |
+
return list(self.all_special_tokens)
|
tokenizer.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
tokenizer_config.json
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"add_prefix_space": false,
|
| 3 |
+
"backend": "tokenizers",
|
| 4 |
+
"bos_token": "<|endoftext|>",
|
| 5 |
+
"clean_up_tokenization_spaces": false,
|
| 6 |
+
"eos_token": "<|endoftext|>",
|
| 7 |
+
"is_local": true,
|
| 8 |
+
"local_files_only": false,
|
| 9 |
+
"model_max_length": 65536,
|
| 10 |
+
"pad_token": "<|pad|>",
|
| 11 |
+
"tokenizer_class": "TokenizersBackend",
|
| 12 |
+
"unk_token": "<|endoftext|>"
|
| 13 |
+
}
|