DEEPSEEK REQUIREMENTS.txt · Lines 3001–4500

3001 past_key_values=transformer_outputs.past_key_values, 3002 hidden_states=transformer_outputs.hidden_states, 3003 attentions=transformer_outputs.attentions, 3004 ) 3005 3006================================================================================ 3007C.4 generation_config.json (171 bytes) 3008================================================================================ 3009{ 3010 "_from_model_config": true, 3011 "bos_token_id": 0, 3012 "eos_token_id": 1, 3013 "do_sample": true, 3014 "temperature": 0.6, 3015 "top_p": 0.95, 3016 "transformers_version": "4.39.3" 3017} 3018 3019================================================================================ 3020C.5 tokenizer_config.json (3,594 bytes) 3021================================================================================ 3022{ 3023 "add_bos_token": true, 3024 "add_eos_token": false, 3025 "bos_token": { 3026 "__type": "AddedToken", 3027 "content": "<｜begin▁of▁sentence｜>", 3028 "lstrip": false, 3029 "normalized": true, 3030 "rstrip": false, 3031 "single_word": false 3032 }, 3033 "clean_up_tokenization_spaces": false, 3034 "eos_token": { 3035 "__type": "AddedToken", 3036 "content": "<｜end▁of▁sentence｜>", 3037 "lstrip": false, 3038 "normalized": true, 3039 "rstrip": false, 3040 "single_word": false 3041 }, 3042 "legacy": true, 3043 "model_max_length": 16384, 3044 "pad_token": { 3045 "__type": "AddedToken", 3046 "content": "<｜end▁of▁sentence｜>", 3047 "lstrip": false, 3048 "normalized": true, 3049 "rstrip": false, 3050 "single_word": false 3051 }, 3052 "sp_model_kwargs": {}, 3053 "unk_token": null, 3054 "tokenizer_class": "LlamaTokenizerFast", 3055 "chat_template": "{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% set ns = namespace(is_first=false, is_tool=false, is_output_first=true, system_prompt='', is_first_sp=true) %}{%- for message in messages %}{%- if message['role'] == 'system' %}{%- if ns.is_first_sp %}{% set ns.system_prompt = ns.system_prompt + message['content'] %}{% set ns.is_first_sp = false %}{%- else %}{% set ns.system_prompt = ns.system_prompt + '\\n\\n' + message['content'] %}{%- endif %}{%- endif %}{%- endfor %}{{ bos_token }}{{ ns.system_prompt }}{%- for message in messages %}{%- if message['role'] == 'user' %}{%- set ns.is_tool = false -%}{{'<｜User｜>' + message['content']}}{%- endif %}{%- if message['role'] == 'assistant' and 'tool_calls' in message %}{%- set ns.is_tool = false -%}{%- for tool in message['tool_calls'] %}{%- if not ns.is_first %}{%- if message['content'] is none %}{{'<｜Assistant｜><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\\n' + '```json' + '\\n' + tool['function']['arguments'] + '\\n' + '```' + '<｜tool▁call▁end｜>'}}{%- else %}{{'<｜Assistant｜>' + message['content'] + '<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\\n' + '```json' + '\\n' + tool['function']['arguments'] + '\\n' + '```' + '<｜tool▁call▁end｜>'}}{%- endif %}{%- set ns.is_first = true -%}{%- else %}{{'\\n' + '<｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\\n' + '```json' + '\\n' + tool['function']['arguments'] + '\\n' + '```' + '<｜tool▁call▁end｜>'}}{%- endif %}{%- endfor %}{{'<｜tool▁calls▁end｜><｜end▁of▁sentence｜>'}}{%- endif %}{%- if message['role'] == 'assistant' and 'tool_calls' not in message %}{%- if ns.is_tool %}{{'<｜tool▁outputs▁end｜>' + message['content'] + '<｜end▁of▁sentence｜>'}}{%- set ns.is_tool = false -%}{%- else %}{% set content = message['content'] %}{% if '</think>' in content %}{% set content = content.split('</think>')[-1] %}{% endif %}{{'<｜Assistant｜>' + content + '<｜end▁of▁sentence｜>'}}{%- endif %}{%- endif %}{%- if message['role'] == 'tool' %}{%- set ns.is_tool = true -%}{%- if ns.is_output_first %}{{'<｜tool▁outputs▁begin｜><｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}}{%- set ns.is_output_first = false %}{%- else %}{{'<｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}}{%- endif %}{%- endif %}{%- endfor -%}{% if ns.is_tool %}{{'<｜tool▁outputs▁end｜>'}}{% endif %}{% if add_generation_prompt and not ns.is_tool %}{{'<｜Assistant｜><think>\\n'}}{% endif %}" 3056} 3057================================================================================ 3058C.6 tokenizer.json (7,847,602 bytes) 3059================================================================================ 3060DELIVERED AS A SIDECAR FILE, not inline. See the note in B.1. 3061 3062 path /home/joseph/DeepSeek-R1/tokenizer.json 3063 copy /home/joseph/deepseek-megamind-upload/deepseek/tokenizer.json 3064 bytes 7,847,602 3065 sha256 ecb6f9fc369894346f0511f4074ca75cee5cd5f3b06d02f1ba35fcd39f8e121d 3066 inside 01-deepseek-source.tar.gz 3067 3068STRUCTURE 3069 top-level keys : version, truncation, padding, added_tokens, normalizer, 3070 pre_tokenizer, post_processor, decoder, model 3071 model.type : BPE 3072 model.vocab : 128,000 entries, ids 0..127,999 3073 model.merges : 127,741 merge rules 3074 added_tokens : 818 entries (798 placeholders, 20 meaningful) 3075 truncation : null padding : null 3076 normalizer : {"type":"Sequence","normalizers":[]} -- EMPTY, no NFC/NFKC 3077 pre_tokenizer : Sequence of three Split rules then ByteLevel 3078 1. Regex \p{N}{1,3} Isolated 3079 2. Regex [CJK + Hiragana + Katakana]+ Isolated 3080 3. a large punctuation/letter/whitespace regex 3081 post_processor : ByteLevel, add_prefix_space true, trim_offsets false 3082 decoder : ByteLevel, add_prefix_space true, trim_offsets true 3083 3084 The 20 meaningful special tokens and their ids are enumerated in 3085 DEEPSEEK_R1_COMPLETE.txt part 11.2 and summarized in B.9 of this document. 3086 3087================================================================================ 3088C.7 model.safetensors.index.json (8,898,324 bytes) 3089================================================================================ 3090DELIVERED AS A SIDECAR FILE, not inline. See the note in B.1. 3091 3092 path /home/joseph/DeepSeek-R1/model.safetensors.index.json 3093 copy /home/joseph/deepseek-megamind-upload/deepseek/model.safetensors.index.json 3094 bytes 8,898,324 3095 sha256 a2c5d238ab1e8301a117336e9797e0d3551cb8f6dc9dc787318b23d8f54c9248 3096 inside 01-deepseek-source.tar.gz 3097 3098STRUCTURE 3099 top-level keys : ["metadata", "weight_map"] 3100 metadata : {"total_size": 1369062772000} 3101 weight_map : 91,991 entries, "<tensor name>": "model-000NN-of-000163.safetensors" 3102 naming 163 distinct shards 3103 example : "model.embed_tokens.weight" -> "model-00001-of-000163.safetensors" 3104 3105 WARNING: metadata.total_size (1,369,062,772,000 = 1.245 TiB) overstates the 3106 real payload (688,574,839,360 = 641.29 GiB) by 1.99x, because the exporter 3107 costed FP8 elements at two bytes each. Never size allocations from it. 3108 3109 deepseek_tensor_manifest.json (B.2) is strictly more useful than this file: 3110 it carries dtype, full shape, absolute byte offset, byte length and the linked 3111 scale tensor for all 91,991 tensors, none of which the index contains. 3112 3113================================================================================ 3114C.8 README.md — official model card (15,994 bytes) 3115================================================================================ 3116--- 3117license: mit 3118library_name: transformers 3119--- 3120# DeepSeek-R1 3121<!-- markdownlint-disable first-line-h1 --> 3122<!-- markdownlint-disable html --> 3123<!-- markdownlint-disable no-duplicate-header --> 3124 3125<div align="center"> 3126 <img src="https://github.com/deepseek-ai/DeepSeek-V2/blob/main/figures/logo.svg?raw=true" width="60%" alt="DeepSeek-V3" /> 3127</div> 3128<hr> 3129<div align="center" style="line-height: 1;"> 3130 <a href="https://www.deepseek.com/" target="_blank" style="margin: 2px;"> 3131 <img alt="Homepage" src="https://github.com/deepseek-ai/DeepSeek-V2/blob/main/figures/badge.svg?raw=true" style="display: inline-block; vertical-align: middle;"/> 3132 </a> 3133 <a href="https://chat.deepseek.com/" target="_blank" style="margin: 2px;"> 3134 <img alt="Chat" src="https://img.shields.io/badge/🤖%20Chat-DeepSeek%20R1-536af5?color=536af5&logoColor=white" style="display: inline-block; vertical-align: middle;"/> 3135 </a> 3136 <a href="https://huggingface.co/deepseek-ai" target="_blank" style="margin: 2px;"> 3137 <img alt="Hugging Face" src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-DeepSeek%20AI-ffc107?color=ffc107&logoColor=white" style="display: inline-block; vertical-align: middle;"/> 3138 </a> 3139</div> 3140 3141<div align="center" style="line-height: 1;"> 3142 <a href="https://discord.gg/Tc7c45Zzu5" target="_blank" style="margin: 2px;"> 3143 <img alt="Discord" src="https://img.shields.io/badge/Discord-DeepSeek%20AI-7289da?logo=discord&logoColor=white&color=7289da" style="display: inline-block; vertical-align: middle;"/> 3144 </a> 3145 <a href="https://github.com/deepseek-ai/DeepSeek-V2/blob/main/figures/qr.jpeg?raw=true" target="_blank" style="margin: 2px;"> 3146 <img alt="Wechat" src="https://img.shields.io/badge/WeChat-DeepSeek%20AI-brightgreen?logo=wechat&logoColor=white" style="display: inline-block; vertical-align: middle;"/> 3147 </a> 3148 <a href="https://twitter.com/deepseek_ai" target="_blank" style="margin: 2px;"> 3149 <img alt="Twitter Follow" src="https://img.shields.io/badge/Twitter-deepseek_ai-white?logo=x&logoColor=white" style="display: inline-block; vertical-align: middle;"/> 3150 </a> 3151</div> 3152 3153<div align="center" style="line-height: 1;"> 3154 <a href="https://github.com/deepseek-ai/DeepSeek-R1/blob/main/LICENSE" style="margin: 2px;"> 3155 <img alt="License" src="https://img.shields.io/badge/License-MIT-f5de53?&color=f5de53" style="display: inline-block; vertical-align: middle;"/> 3156 </a> 3157</div> 3158 3159 3160<p align="center"> 3161 <a href="https://github.com/deepseek-ai/DeepSeek-R1/blob/main/DeepSeek_R1.pdf"><b>Paper Link</b>👁️</a> 3162</p> 3163 3164 3165## 1. Introduction 3166 3167We introduce our first-generation reasoning models, DeepSeek-R1-Zero and DeepSeek-R1. 3168DeepSeek-R1-Zero, a model trained via large-scale reinforcement learning (RL) without supervised fine-tuning (SFT) as a preliminary step, demonstrated remarkable performance on reasoning. 3169With RL, DeepSeek-R1-Zero naturally emerged with numerous powerful and interesting reasoning behaviors. 3170However, DeepSeek-R1-Zero encounters challenges such as endless repetition, poor readability, and language mixing. To address these issues and further enhance reasoning performance, 3171we introduce DeepSeek-R1, which incorporates cold-start data before RL. 3172DeepSeek-R1 achieves performance comparable to OpenAI-o1 across math, code, and reasoning tasks. 3173To support the research community, we have open-sourced DeepSeek-R1-Zero, DeepSeek-R1, and six dense models distilled from DeepSeek-R1 based on Llama and Qwen. DeepSeek-R1-Distill-Qwen-32B outperforms OpenAI-o1-mini across various benchmarks, achieving new state-of-the-art results for dense models. 3174 3175**NOTE: Before running DeepSeek-R1 series models locally, we kindly recommend reviewing the [Usage Recommendation](#usage-recommendations) section.** 3176 3177<p align="center"> 3178 <img width="80%" src="figures/benchmark.jpg"> 3179</p> 3180 3181## 2. Model Summary 3182 3183--- 3184 3185**Post-Training: Large-Scale Reinforcement Learning on the Base Model** 3186 3187- We directly apply reinforcement learning (RL) to the base model without relying on supervised fine-tuning (SFT) as a preliminary step. This approach allows the model to explore chain-of-thought (CoT) for solving complex problems, resulting in the development of DeepSeek-R1-Zero. DeepSeek-R1-Zero demonstrates capabilities such as self-verification, reflection, and generating long CoTs, marking a significant milestone for the research community. Notably, it is the first open research to validate that reasoning capabilities of LLMs can be incentivized purely through RL, without the need for SFT. This breakthrough paves the way for future advancements in this area. 3188 3189- We introduce our pipeline to develop DeepSeek-R1. The pipeline incorporates two RL stages aimed at discovering improved reasoning patterns and aligning with human preferences, as well as two SFT stages that serve as the seed for the model's reasoning and non-reasoning capabilities. 3190 We believe the pipeline will benefit the industry by creating better models. 3191 3192--- 3193 3194**Distillation: Smaller Models Can Be Powerful Too** 3195 3196- We demonstrate that the reasoning patterns of larger models can be distilled into smaller models, resulting in better performance compared to the reasoning patterns discovered through RL on small models. The open source DeepSeek-R1, as well as its API, will benefit the research community to distill better smaller models in the future. 3197- Using the reasoning data generated by DeepSeek-R1, we fine-tuned several dense models that are widely used in the research community. The evaluation results demonstrate that the distilled smaller dense models perform exceptionally well on benchmarks. We open-source distilled 1.5B, 7B, 8B, 14B, 32B, and 70B checkpoints based on Qwen2.5 and Llama3 series to the community. 3198 3199## 3. Model Downloads 3200 3201### DeepSeek-R1 Models 3202 3203<div align="center"> 3204 3205| **Model** | **#Total Params** | **#Activated Params** | **Context Length** | **Download** | 3206| :------------: | :------------: | :------------: | :------------: | :------------: | 3207| DeepSeek-R1-Zero | 671B | 37B | 128K | [🤗 HuggingFace](https://huggingface.co/deepseek-ai/DeepSeek-R1-Zero) | 3208| DeepSeek-R1 | 671B | 37B | 128K | [🤗 HuggingFace](https://huggingface.co/deepseek-ai/DeepSeek-R1) | 3209 3210</div> 3211 3212DeepSeek-R1-Zero & DeepSeek-R1 are trained based on DeepSeek-V3-Base. 3213For more details regarding the model architecture, please refer to [DeepSeek-V3](https://github.com/deepseek-ai/DeepSeek-V3) repository. 3214 3215### DeepSeek-R1-Distill Models 3216 3217<div align="center"> 3218 3219| **Model** | **Base Model** | **Download** | 3220| :------------: | :------------: | :------------: | 3221| DeepSeek-R1-Distill-Qwen-1.5B | [Qwen2.5-Math-1.5B](https://huggingface.co/Qwen/Qwen2.5-Math-1.5B) | [🤗 HuggingFace](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B) | 3222| DeepSeek-R1-Distill-Qwen-7B | [Qwen2.5-Math-7B](https://huggingface.co/Qwen/Qwen2.5-Math-7B) | [🤗 HuggingFace](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B) | 3223| DeepSeek-R1-Distill-Llama-8B | [Llama-3.1-8B](https://huggingface.co/meta-llama/Llama-3.1-8B) | [🤗 HuggingFace](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B) | 3224| DeepSeek-R1-Distill-Qwen-14B | [Qwen2.5-14B](https://huggingface.co/Qwen/Qwen2.5-14B) | [🤗 HuggingFace](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B) | 3225|DeepSeek-R1-Distill-Qwen-32B | [Qwen2.5-32B](https://huggingface.co/Qwen/Qwen2.5-32B) | [🤗 HuggingFace](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B) | 3226| DeepSeek-R1-Distill-Llama-70B | [Llama-3.3-70B-Instruct](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct) | [🤗 HuggingFace](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B) | 3227 3228</div> 3229 3230DeepSeek-R1-Distill models are fine-tuned based on open-source models, using samples generated by DeepSeek-R1. 3231We slightly change their configs and tokenizers. Please use our setting to run these models. 3232 3233## 4. Evaluation Results 3234 3235### DeepSeek-R1-Evaluation 3236 For all our models, the maximum generation length is set to 32,768 tokens. For benchmarks requiring sampling, we use a temperature of $0.6$, a top-p value of $0.95$, and generate 64 responses per query to estimate pass@1. 3237<div align="center"> 3238 3239 3240| Category | Benchmark (Metric) | Claude-3.5-Sonnet-1022 | GPT-4o 0513 | DeepSeek V3 | OpenAI o1-mini | OpenAI o1-1217 | DeepSeek R1 | 3241|----------|-------------------|----------------------|------------|--------------|----------------|------------|--------------| 3242| | Architecture | - | - | MoE | - | - | MoE | 3243| | # Activated Params | - | - | 37B | - | - | 37B | 3244| | # Total Params | - | - | 671B | - | - | 671B | 3245| English | MMLU (Pass@1) | 88.3 | 87.2 | 88.5 | 85.2 | **91.8** | 90.8 | 3246| | MMLU-Redux (EM) | 88.9 | 88.0 | 89.1 | 86.7 | - | **92.9** | 3247| | MMLU-Pro (EM) | 78.0 | 72.6 | 75.9 | 80.3 | - | **84.0** | 3248| | DROP (3-shot F1) | 88.3 | 83.7 | 91.6 | 83.9 | 90.2 | **92.2** | 3249| | IF-Eval (Prompt Strict) | **86.5** | 84.3 | 86.1 | 84.8 | - | 83.3 | 3250| | GPQA-Diamond (Pass@1) | 65.0 | 49.9 | 59.1 | 60.0 | **75.7** | 71.5 | 3251| | SimpleQA (Correct) | 28.4 | 38.2 | 24.9 | 7.0 | **47.0** | 30.1 | 3252| | FRAMES (Acc.) | 72.5 | 80.5 | 73.3 | 76.9 | - | **82.5** | 3253| | AlpacaEval2.0 (LC-winrate) | 52.0 | 51.1 | 70.0 | 57.8 | - | **87.6** | 3254| | ArenaHard (GPT-4-1106) | 85.2 | 80.4 | 85.5 | 92.0 | - | **92.3** | 3255| Code | LiveCodeBench (Pass@1-COT) | 33.8 | 34.2 | - | 53.8 | 63.4 | **65.9** | 3256| | Codeforces (Percentile) | 20.3 | 23.6 | 58.7 | 93.4 | **96.6** | 96.3 | 3257| | Codeforces (Rating) | 717 | 759 | 1134 | 1820 | **2061** | 2029 | 3258| | SWE Verified (Resolved) | **50.8** | 38.8 | 42.0 | 41.6 | 48.9 | 49.2 | 3259| | Aider-Polyglot (Acc.) | 45.3 | 16.0 | 49.6 | 32.9 | **61.7** | 53.3 | 3260| Math | AIME 2024 (Pass@1) | 16.0 | 9.3 | 39.2 | 63.6 | 79.2 | **79.8** | 3261| | MATH-500 (Pass@1) | 78.3 | 74.6 | 90.2 | 90.0 | 96.4 | **97.3** | 3262| | CNMO 2024 (Pass@1) | 13.1 | 10.8 | 43.2 | 67.6 | - | **78.8** | 3263| Chinese | CLUEWSC (EM) | 85.4 | 87.9 | 90.9 | 89.9 | - | **92.8** | 3264| | C-Eval (EM) | 76.7 | 76.0 | 86.5 | 68.9 | - | **91.8** | 3265| | C-SimpleQA (Correct) | 55.4 | 58.7 | **68.0** | 40.3 | - | 63.7 | 3266 3267</div> 3268 3269 3270### Distilled Model Evaluation 3271 3272 3273<div align="center"> 3274 3275| Model | AIME 2024 pass@1 | AIME 2024 cons@64 | MATH-500 pass@1 | GPQA Diamond pass@1 | LiveCodeBench pass@1 | CodeForces rating | 3276|------------------------------------------|------------------|-------------------|-----------------|----------------------|----------------------|-------------------| 3277| GPT-4o-0513 | 9.3 | 13.4 | 74.6 | 49.9 | 32.9 | 759 | 3278| Claude-3.5-Sonnet-1022 | 16.0 | 26.7 | 78.3 | 65.0 | 38.9 | 717 | 3279| o1-mini | 63.6 | 80.0 | 90.0 | 60.0 | 53.8 | **1820** | 3280| QwQ-32B-Preview | 44.0 | 60.0 | 90.6 | 54.5 | 41.9 | 1316 | 3281| DeepSeek-R1-Distill-Qwen-1.5B | 28.9 | 52.7 | 83.9 | 33.8 | 16.9 | 954 | 3282| DeepSeek-R1-Distill-Qwen-7B | 55.5 | 83.3 | 92.8 | 49.1 | 37.6 | 1189 | 3283| DeepSeek-R1-Distill-Qwen-14B | 69.7 | 80.0 | 93.9 | 59.1 | 53.1 | 1481 | 3284| DeepSeek-R1-Distill-Qwen-32B | **72.6** | 83.3 | 94.3 | 62.1 | 57.2 | 1691 | 3285| DeepSeek-R1-Distill-Llama-8B | 50.4 | 80.0 | 89.1 | 49.0 | 39.6 | 1205 | 3286| DeepSeek-R1-Distill-Llama-70B | 70.0 | **86.7** | **94.5** | **65.2** | **57.5** | 1633 | 3287 3288</div> 3289 3290 3291## 5. Chat Website & API Platform 3292You can chat with DeepSeek-R1 on DeepSeek's official website: [chat.deepseek.com](https://chat.deepseek.com), and switch on the button "DeepThink" 3293 3294We also provide OpenAI-Compatible API at DeepSeek Platform: [platform.deepseek.com](https://platform.deepseek.com/) 3295 3296## 6. How to Run Locally 3297 3298### DeepSeek-R1 Models 3299 3300Please visit [DeepSeek-V3](https://github.com/deepseek-ai/DeepSeek-V3) repo for more information about running DeepSeek-R1 locally. 3301 3302**NOTE: Hugging Face's Transformers has not been directly supported yet.** 3303 3304### DeepSeek-R1-Distill Models 3305 3306DeepSeek-R1-Distill models can be utilized in the same manner as Qwen or Llama models. 3307 3308For instance, you can easily start a service using [vLLM](https://github.com/vllm-project/vllm): 3309 3310```shell 3311vllm serve deepseek-ai/DeepSeek-R1-Distill-Qwen-32B --tensor-parallel-size 2 --max-model-len 32768 --enforce-eager 3312``` 3313 3314You can also easily start a service using [SGLang](https://github.com/sgl-project/sglang) 3315 3316```bash 3317python3 -m sglang.launch_server --model deepseek-ai/DeepSeek-R1-Distill-Qwen-32B --trust-remote-code --tp 2 3318``` 3319 3320### Usage Recommendations 3321 3322**We recommend adhering to the following configurations when utilizing the DeepSeek-R1 series models, including benchmarking, to achieve the expected performance:** 3323 33241. Set the temperature within the range of 0.5-0.7 (0.6 is recommended) to prevent endless repetitions or incoherent outputs. 33252. **Avoid adding a system prompt; all instructions should be contained within the user prompt.** 33263. For mathematical problems, it is advisable to include a directive in your prompt such as: "Please reason step by step, and put your final answer within \boxed{}." 33274. When evaluating model performance, it is recommended to conduct multiple tests and average the results. 3328 3329Additionally, we have observed that the DeepSeek-R1 series models tend to bypass thinking pattern (i.e., outputting "\<think\>\n\n\</think\>") when responding to certain queries, which can adversely affect the model's performance. 3330**To ensure that the model engages in thorough reasoning, we recommend enforcing the model to initiate its response with "\<think\>\n" at the beginning of every output.** 3331 3332## 7. License 3333This code repository and the model weights are licensed under the [MIT License](https://github.com/deepseek-ai/DeepSeek-R1/blob/main/LICENSE). 3334DeepSeek-R1 series support commercial use, allow for any modifications and derivative works, including, but not limited to, distillation for training other LLMs. Please note that: 3335- DeepSeek-R1-Distill-Qwen-1.5B, DeepSeek-R1-Distill-Qwen-7B, DeepSeek-R1-Distill-Qwen-14B and DeepSeek-R1-Distill-Qwen-32B are derived from [Qwen-2.5 series](https://github.com/QwenLM/Qwen2.5), which are originally licensed under [Apache 2.0 License](https://huggingface.co/Qwen/Qwen2.5-1.5B/blob/main/LICENSE), and now finetuned with 800k samples curated with DeepSeek-R1. 3336- DeepSeek-R1-Distill-Llama-8B is derived from Llama3.1-8B-Base and is originally licensed under [llama3.1 license](https://huggingface.co/meta-llama/Llama-3.1-8B/blob/main/LICENSE). 3337- DeepSeek-R1-Distill-Llama-70B is derived from Llama3.3-70B-Instruct and is originally licensed under [llama3.3 license](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct/blob/main/LICENSE). 3338 3339## 8. Citation 3340``` 3341@misc{deepseekai2025deepseekr1incentivizingreasoningcapability, 3342 title={DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning}, 3343 author={DeepSeek-AI}, 3344 year={2025}, 3345 eprint={2501.12948}, 3346 archivePrefix={arXiv}, 3347 primaryClass={cs.CL}, 3348 url={https://arxiv.org/abs/2501.12948}, 3349} 3350 3351``` 3352 3353## 9. Contact 3354If you have any questions, please raise an issue or contact us at [[author/contact email redacted]]([author/contact email redacted]). 3355 3356================================================================================ 3357C.9 LICENSE (1,064 bytes) 3358================================================================================ 3359MIT License 3360 3361Copyright (c) 2023 DeepSeek 3362 3363Permission is hereby granted, free of charge, to any person obtaining a copy 3364of this software and associated documentation files (the "Software"), to deal 3365in the Software without restriction, including without limitation the rights 3366to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 3367copies of the Software, and to permit persons to whom the Software is 3368furnished to do so, subject to the following conditions: 3369 3370The above copyright notice and this permission notice shall be included in all 3371copies or substantial portions of the Software. 3372 3373THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 3374IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 3375FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE 3376AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 3377LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 3378OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 3379SOFTWARE. 3380================================================================================ 3381C.10 .gitattributes (1,519 bytes) 3382================================================================================ 3383*.7z filter=lfs diff=lfs merge=lfs -text 3384*.arrow filter=lfs diff=lfs merge=lfs -text 3385*.bin filter=lfs diff=lfs merge=lfs -text 3386*.bz2 filter=lfs diff=lfs merge=lfs -text 3387*.ckpt filter=lfs diff=lfs merge=lfs -text 3388*.ftz filter=lfs diff=lfs merge=lfs -text 3389*.gz filter=lfs diff=lfs merge=lfs -text 3390*.h5 filter=lfs diff=lfs merge=lfs -text 3391*.joblib filter=lfs diff=lfs merge=lfs -text 3392*.lfs.* filter=lfs diff=lfs merge=lfs -text 3393*.mlmodel filter=lfs diff=lfs merge=lfs -text 3394*.model filter=lfs diff=lfs merge=lfs -text 3395*.msgpack filter=lfs diff=lfs merge=lfs -text 3396*.npy filter=lfs diff=lfs merge=lfs -text 3397*.npz filter=lfs diff=lfs merge=lfs -text 3398*.onnx filter=lfs diff=lfs merge=lfs -text 3399*.ot filter=lfs diff=lfs merge=lfs -text 3400*.parquet filter=lfs diff=lfs merge=lfs -text 3401*.pb filter=lfs diff=lfs merge=lfs -text 3402*.pickle filter=lfs diff=lfs merge=lfs -text 3403*.pkl filter=lfs diff=lfs merge=lfs -text 3404*.pt filter=lfs diff=lfs merge=lfs -text 3405*.pth filter=lfs diff=lfs merge=lfs -text 3406*.rar filter=lfs diff=lfs merge=lfs -text 3407*.safetensors filter=lfs diff=lfs merge=lfs -text 3408saved_model/**/* filter=lfs diff=lfs merge=lfs -text 3409*.tar.* filter=lfs diff=lfs merge=lfs -text 3410*.tar filter=lfs diff=lfs merge=lfs -text 3411*.tflite filter=lfs diff=lfs merge=lfs -text 3412*.tgz filter=lfs diff=lfs merge=lfs -text 3413*.wasm filter=lfs diff=lfs merge=lfs -text 3414*.xz filter=lfs diff=lfs merge=lfs -text 3415*.zip filter=lfs diff=lfs merge=lfs -text 3416*.zst filter=lfs diff=lfs merge=lfs -text 3417*tfevents* filter=lfs diff=lfs merge=lfs -text 3418 3419 3420================================================================================ 3421C.11 SCANNER AND BENCHMARK SCRIPTS (all created for this work) 3422================================================================================ 3423 3424-------------------------------------------------------------------------------- 3425C.11.1 download_r1.py (641 bytes) 3426-------------------------------------------------------------------------------- 3427#!/usr/bin/env python3 3428import os, sys, traceback 3429print("=== download_r1 START pid",os.getpid(),"===",flush=True) 3430try: 3431 from huggingface_hub import snapshot_download 3432 print("import OK; calling snapshot_download (this resolves ~160 shards first)...",flush=True) 3433 path=snapshot_download( 3434 "deepseek-ai/DeepSeek-R1", 3435 local_dir="/home/joseph/DeepSeek-R1", 3436 ignore_patterns=["*.pth","*.gguf","original/*"], 3437 max_workers=8, 3438 ) 3439 print("=== DOWNLOAD COMPLETE ===",path,flush=True) 3440except SystemExit: 3441 raise 3442except BaseException: 3443 print("=== ERROR ===",flush=True); traceback.print_exc(); sys.exit(1) 3444 3445-------------------------------------------------------------------------------- 3446C.11.2 launch_r1.sh (590 bytes) 3447-------------------------------------------------------------------------------- 3448#!/bin/bash 3449# Persistent R1 download: systemd unit as joseph, with joseph's hf path pinned so python finds it. Resumes prior 719MB. 3450sudo systemctl reset-failed r1full 2>/dev/null 3451sudo systemd-run --unit=r1full --uid=joseph --setenv=HOME=/home/joseph \ 3452 --setenv=PYTHONPATH=/home/joseph/.local/lib/python3.14/site-packages \ 3453 --working-directory=/home/joseph python3 /home/joseph/download_r1.py 3454sleep 25 3455echo "=== active: $(systemctl is-active r1full) ===" 3456sudo journalctl -u r1full --no-pager -n 5 --output=cat 2>/dev/null 3457echo "=== size ==="; du -sh /home/joseph/DeepSeek-R1 2>/dev/null 3458 3459-------------------------------------------------------------------------------- 3460C.11.3 dl_test.py (421 bytes) 3461-------------------------------------------------------------------------------- 3462#!/usr/bin/env python3 3463import time, sys 3464from huggingface_hub import hf_hub_download 3465t=time.time() 3466try: 3467 p=hf_hub_download("deepseek-ai/DeepSeek-R1","config.json",local_dir="/home/joseph/DeepSeek-R1") 3468 print("GOT config.json ->",p,"in",round(time.time()-t,1),"s",flush=True) 3469 import os; print("size:",os.path.getsize(p),"bytes") 3470except Exception as e: 3471 print("ERROR:",type(e).__name__,str(e)[:300],flush=True) 3472 3473-------------------------------------------------------------------------------- 3474C.11.4 scan_headers.py (2349 bytes) 3475-------------------------------------------------------------------------------- 3476import json, os, struct, collections, re, sys 3477D=os.path.expanduser("~/DeepSeek-R1") 3478files=sorted(f for f in os.listdir(D) if f.endswith(".safetensors")) 3479print("shards:",len(files)) 3480tot_tensors=0; tot_bytes=0 3481dtype_ct=collections.Counter(); dtype_bytes=collections.Counter() 3482fam_ct=collections.Counter(); fam_bytes=collections.Counter(); fam_dtypes=collections.defaultdict(set) 3483shapes=collections.defaultdict(set) 3484layers=set(); allnames=[] 3485per_shard=[] 3486for fn in files: 3487 p=os.path.join(D,fn) 3488 with open(p,'rb') as fh: 3489 n=struct.unpack('<Q',fh.read(8))[0] 3490 hdr=json.loads(fh.read(n)) 3491 meta=hdr.pop("__metadata__",None) 3492 per_shard.append((fn,len(hdr),os.path.getsize(p),meta)) 3493 for name,info in hdr.items(): 3494 tot_tensors+=1 3495 dt=info["dtype"]; shp=info["shape"]; off=info["data_offsets"] 3496 nb=off[1]-off[0]; tot_bytes+=nb 3497 dtype_ct[dt]+=1; dtype_bytes[dt]+=nb 3498 allnames.append(name) 3499 m=re.match(r'model\.layers\.(\d+)\.(.*)',name) 3500 if m: 3501 layers.add(int(m.group(1))) 3502 fam=re.sub(r'experts\.\d+','experts.N',m.group(2)) 3503 else: 3504 fam=name 3505 fam_ct[fam]+=1; fam_bytes[fam]+=nb; fam_dtypes[fam].add(dt) 3506 shapes[fam].add(tuple(shp)) 3507print("total_tensors:",tot_tensors) 3508print("total_tensor_bytes:",tot_bytes, round(tot_bytes/2**30,2),"GiB") 3509print("layers_seen:",min(layers),"..",max(layers),"count",len(layers)) 3510print("\n=== DTYPES ===") 3511for d,c in dtype_ct.most_common(): 3512 print(f" {d:10s} count={c:8d} bytes={dtype_bytes[d]:15d} {dtype_bytes[d]/2**30:9.2f} GiB") 3513print("\n=== TENSOR FAMILIES (layer-indices collapsed) ===") 3514for f,c in sorted(fam_ct.items(), key=lambda kv:-fam_bytes[kv[0]]): 3515 sh=sorted(shapes[f]); shs=str(sh[0]) if len(sh)==1 else f"{sh[0]}..{sh[-1]}({len(sh)} distinct)" 3516 print(f" {f:62s} n={c:7d} {sorted(fam_dtypes[f])!s:22s} {fam_bytes[f]/2**30:8.3f} GiB shape={shs}") 3517print("\n=== SHARD SUMMARY (first 3, last 3) ===") 3518for fn,nt,sz,meta in per_shard[:3]+per_shard[-3:]: 3519 print(f" {fn} tensors={nt} bytes={sz} meta={meta}") 3520print("\n=== NON-LAYER (global) TENSORS ===") 3521for nm in sorted(set(n for n in allnames if not n.startswith("model.layers."))): 3522 print(" ",nm) 3523json.dump({"total_tensors":tot_tensors,"total_bytes":tot_bytes},open(os.path.expanduser("~/r1scan/summary.json"),"w")) 3524 3525-------------------------------------------------------------------------------- 3526C.11.5 per_layer_manifest.py (1336 bytes) 3527-------------------------------------------------------------------------------- 3528import json,os,struct,re 3529D=os.path.expanduser("~/DeepSeek-R1") 3530want={"model.layers.0.","model.layers.3.","model.layers.61."} 3531got={} 3532for fn in sorted(f for f in os.listdir(D) if f.endswith(".safetensors")): 3533 with open(os.path.join(D,fn),'rb') as f: 3534 n=struct.unpack('<Q',f.read(8))[0]; h=json.loads(f.read(n)) 3535 h.pop("__metadata__",None) 3536 for k,v in h.items(): 3537 if any(k.startswith(w) for w in want): got[k]=(v,fn) 3538def show(pref,title): 3539 print("###",title) 3540 ks=[k for k in got if k.startswith(pref) and not re.search(r'experts\.\d+\.',k)] 3541 tot=0 3542 for k in sorted(ks): 3543 v,fn=got[k]; o=v["data_offsets"]; b=o[1]-o[0]; tot+=b 3544 print(f" {v['dtype']:8s} {str(v['shape']):18s} {b:>13,} B {k[len(pref):]}") 3545 ne=sorted(set(int(m.group(1)) for k in got if k.startswith(pref) and (m:=re.search(r'experts\.(\d+)\.',k)))) 3546 print(f" --- non-expert subtotal: {tot:,} B = {tot/2**20:.1f} MiB ({len(ks)} tensors)") 3547 if ne: print(f" --- plus routed experts {min(ne)}..{max(ne)} ({len(ne)} of them) x 6 tensors = {len(ne)*44050944:,} B") 3548 print() 3549show("model.layers.0.","LAYER 0 - DENSE LAYER (also layers 1,2) - COMPLETE MANIFEST") 3550show("model.layers.3.","LAYER 3 - MoE LAYER (also layers 4..60) - COMPLETE MANIFEST") 3551show("model.layers.61.","LAYER 61 - MTP LAYER - COMPLETE MANIFEST") 3552 3553-------------------------------------------------------------------------------- 3554C.11.6 bench_ssd.sh (1473 bytes) 3555-------------------------------------------------------------------------------- 3556#!/bin/bash 3557# bench_ssd.sh -- drive diagnosis for the expert-streaming read path. 3558# Decides whether to fan reads out (concurrency headroom) or cut seeks. 3559D=~/DeepSeek-R1 3560echo "--- single stream, 8 MiB blocks, 2 GiB ---" 3561[REDACTED: credential-associated original line 3561; original retained outside this bundle.] 3562dd if=$D/model-00080-of-000163.safetensors of=/dev/null bs=8M count=256 2>&1 | tail -1 3563echo "--- same file again (page-cache warm) ---" 3564dd if=$D/model-00080-of-000163.safetensors of=/dev/null bs=8M count=256 2>&1 | tail -1 3565echo "--- 4 parallel streams, 4 different shards, 512 MiB each ---" 3566S=$(date +%s.%N) 3567for n in 81 82 83 84; do dd if=$D/model-000$n-of-000163.safetensors of=/dev/null bs=8M count=64 2>/dev/null & done; wait 3568E=$(date +%s.%N) 3569echo "4x512MiB = 2048 MiB in $(echo "$E-$S"|bc)s -> $(echo "scale=1;2048/($E-$S)"|bc) MB/s aggregate" 3570echo "--- 56 scattered 4 MiB preads (per-expert scatter simulation) ---" 3571python3 - <<'PY' 3572import os,time,random 3573p=os.path.expanduser("~/DeepSeek-R1/model-00085-of-000163.safetensors") 3574sz=os.path.getsize(p); fd=os.open(p,os.O_RDONLY); random.seed(7) 3575t=time.time(); tot=0 3576for _ in range(56): 3577 off=random.randrange(0,sz-4*1024*1024) 3578 tot+=len(os.pread(fd,4*1024*1024,off)) 3579d=time.time()-t 3580print(f" 56 scattered 4MiB preads: {tot/2**20:.0f} MiB in {d:.2f}s -> {tot/2**20/d:.0f} MB/s") 3581os.close(fd) 3582PY 3583echo "--- device ---"; df -h ~/DeepSeek-R1 | tail -1; lsblk 2>/dev/null | head -6 3584 3585-------------------------------------------------------------------------------- 3586C.11.7 map_shards_to_layers.py (1015 bytes) 3587-------------------------------------------------------------------------------- 3588import json,os,struct,re,collections 3589D=os.path.expanduser("~/DeepSeek-R1") 3590idx=json.load(open(os.path.join(D,"model.safetensors.index.json"))) 3591wm=idx["weight_map"] 3592print("### INDEX.JSON STRUCTURE") 3593print(" top-level keys:",list(idx.keys())) 3594print(" metadata:",idx["metadata"]) 3595k0=next(iter(wm)); print(" weight_map sample: %r -> %r"%(k0,wm[k0])) 3596print() 3597print("### SHARD -> LAYER MAP") 3598sl=collections.defaultdict(set); glob=collections.defaultdict(list) 3599for name,sh in wm.items(): 3600 m=re.match(r'model\.layers\.(\d+)\.',name) 3601 if m: sl[sh].add(int(m.group(1))) 3602 else: glob[sh].append(name) 3603rows=[] 3604for sh in sorted(set(wm.values())): 3605 ls=sorted(sl.get(sh,[])) 3606 lay = "-" if not ls else (str(ls[0]) if len(ls)==1 else f"{ls[0]}-{ls[-1]}") 3607 rows.append((int(sh.split('-')[1]), lay, len([k for k,v in wm.items() if v==sh]), ",".join(glob.get(sh,[])) )) 3608# compress runs 3609print(" shard layers tensors non-layer tensors") 3610for n,lay,ct,g in rows: 3611 print(f" {n:>5} {lay:<10} {ct:>7} {g}") 3612 3613-------------------------------------------------------------------------------- 3614C.11.8 verify_integrity.py (1310 bytes) 3615-------------------------------------------------------------------------------- 3616import json,os,struct,collections,re 3617D=os.path.expanduser("~/DeepSeek-R1") 3618idx=json.load(open(os.path.join(D,"model.safetensors.index.json"))) 3619wm=idx["weight_map"]; meta=idx.get("metadata",{}) 3620print("index metadata:",meta) 3621print("index weight_map entries:",len(wm)) 3622print("index distinct shards:",len(set(wm.values()))) 3623hdrnames={}; bad=[] 3624files=sorted(f for f in os.listdir(D) if f.endswith(".safetensors")) 3625for fn in files: 3626 p=os.path.join(D,fn); sz=os.path.getsize(p) 3627 with open(p,'rb') as fh: 3628 n=struct.unpack('<Q',fh.read(8))[0]; hdr=json.loads(fh.read(n)) 3629 hdr.pop("__metadata__",None) 3630 end=max(v["data_offsets"][1] for v in hdr.values()) 3631 expect=8+n+end 3632 if expect!=sz: bad.append((fn,sz,expect,sz-expect)) 3633 for k in hdr: hdrnames[k]=fn 3634print("\nTRUNCATION/LENGTH CHECK (header+data vs filesize):") 3635print(" mismatched shards:",len(bad)) 3636for b in bad[:10]: print(" ",b) 3637sh=set(hdrnames); si=set(wm) 3638print("\nINDEX vs HEADERS:") 3639print(" in headers not in index:",len(sh-si)) 3640for x in list(sh-si)[:10]: print(" ",x) 3641print(" in index not in headers:",len(si-sh)) 3642for x in list(si-sh)[:10]: print(" ",x) 3643mis=[k for k in si&sh if wm[k]!=hdrnames[k]] 3644print(" shard-placement mismatches:",len(mis)) 3645print("\nTOTAL from index metadata total_size:",meta.get("total_size")) 3646 3647-------------------------------------------------------------------------------- 3648C.11.9 check_expert_contiguity.py (1459 bytes) 3649-------------------------------------------------------------------------------- 3650import json,os,struct,re 3651D=os.path.expanduser("~/DeepSeek-R1") 3652fn="model-00080-of-000163.safetensors" 3653p=os.path.join(D,fn) 3654with open(p,'rb') as fh: 3655 n=struct.unpack('<Q',fh.read(8))[0]; hdr=json.loads(fh.read(n)) 3656hdr.pop("__metadata__",None) 3657items=sorted(hdr.items(), key=lambda kv: kv[1]["data_offsets"][0]) 3658print("shard:",fn,"tensors:",len(items),"header_bytes:",n) 3659print("\nfirst 14 tensors in ON-DISK order:") 3660for k,v in items[:14]: 3661 o=v["data_offsets"]; print(f" {o[0]:>12} .. {o[1]:>12} {o[1]-o[0]:>10} {v['dtype']:8s} {str(v['shape']):16s} {k}") 3662# check: are an expert's 3 weights + 3 scales contiguous? 3663import collections 3664byexp=collections.defaultdict(list) 3665for k,v in items: 3666 m=re.match(r'model\.layers\.(\d+)\.mlp\.experts\.(\d+)\.',k) 3667 if m: byexp[(int(m.group(1)),int(m.group(2)))].append((v["data_offsets"],k)) 3668print("\nexperts in this shard:",len(byexp)) 3669contig=0; noncontig=0; spans=[] 3670for key,lst in byexp.items(): 3671 lst.sort(); lo=lst[0][0][0]; hi=max(o[1] for o,_ in lst) 3672 used=sum(o[1]-o[0] for o,_ in lst) 3673 spans.append(hi-lo) 3674 if hi-lo==used: contig+=1 3675 else: noncontig+=1 3676print(f" fully contiguous experts: {contig} with gaps: {noncontig}") 3677print(f" expert span bytes: min={min(spans)} max={max(spans)} (payload/expert = {used})") 3678k0=sorted(byexp)[0] 3679print(f"\n layout of one expert {k0}:") 3680for o,k in sorted(byexp[k0]): print(f" {o[0]:>12} .. {o[1]:>12} ({o[1]-o[0]:>9}) {k.split('experts.')[1]}") 3681 3682-------------------------------------------------------------------------------- 3683C.11.10 gen_manifests.py (5475 bytes) 3684-------------------------------------------------------------------------------- 3685#!/usr/bin/env python3 3686""" 3687gen_manifests.py -- produce the five machine-readable manifests for DeepSeek-R1. 3688Reads ONLY the safetensors JSON headers (never the weight payload), so it is 3689fast and touches ~13 MB of the 643 GB checkpoint. 3690 3691Outputs (into OUT): 3692 deepseek_tensor_manifest.json every tensor: shard, dtype, shape, absolute 3693 offset, byte length, linked scale tensor 3694 deepseek_shard_headers.json per-shard container geometry 3695 deepseek_layer_shard_map.json layer index -> shards, and shard -> layers 3696 deepseek_expert_offset_map.json (layer,expert) -> one contiguous read 3697 deepseek_weight_hashes.txt sizes + partial hashes (see gen_hashes.py 3698 for the full SHA-256 pass) 3699""" 3700import json, os, struct, re, collections, hashlib, sys 3701 3702D = os.path.expanduser("~/DeepSeek-R1") 3703OUT = os.path.expanduser("~/deepseek-megamind-upload/deepseek") 3704os.makedirs(OUT, exist_ok=True) 3705 3706files = sorted(f for f in os.listdir(D) if f.endswith(".safetensors")) 3707assert len(files) == 163, f"expected 163 shards, found {len(files)}" 3708 3709tensors = {} # name -> record 3710shard_hdrs = [] 3711for fn in files: 3712 path = os.path.join(D, fn) 3713 size = os.path.getsize(path) 3714 with open(path, "rb") as fh: 3715 hdr_len = struct.unpack("<Q", fh.read(8))[0] 3716 hdr = json.loads(fh.read(hdr_len)) 3717 meta = hdr.pop("__metadata__", None) 3718 data_start = 8 + hdr_len 3719 payload_end = max(v["data_offsets"][1] for v in hdr.values()) 3720 shard_hdrs.append({ 3721 "shard": fn, 3722 "file_size": size, 3723 "header_len": hdr_len, 3724 "data_start": data_start, 3725 "tensor_count": len(hdr), 3726 "payload_bytes": payload_end, 3727 "metadata": meta, 3728 "length_check_ok": (data_start + payload_end) == size, 3729 }) 3730 for name, info in hdr.items(): 3731 rel = info["data_offsets"] 3732 tensors[name] = { 3733 "name": name, 3734 "shard": fn, 3735 "dtype": info["dtype"], 3736 "shape": info["shape"], 3737 "absolute_offset": data_start + rel[0], 3738 "byte_length": rel[1] - rel[0], 3739 "relative_offset": rel[0], 3740 "scale_tensor": None, 3741 } 3742 3743# link each quantized weight to its scale tensor 3744for name, rec in tensors.items(): 3745 if rec["dtype"] == "F8_E4M3" and name.endswith(".weight"): 3746 cand = name[: -len(".weight")] + ".weight_scale_inv" 3747 if cand in tensors: 3748 rec["scale_tensor"] = cand 3749 3750manifest = [tensors[k] for k in sorted(tensors)] 3751json.dump(manifest, open(os.path.join(OUT, "deepseek_tensor_manifest.json"), "w"), indent=1) 3752json.dump(shard_hdrs, open(os.path.join(OUT, "deepseek_shard_headers.json"), "w"), indent=1) 3753 3754# layer <-> shard maps 3755l2s = collections.defaultdict(set); s2l = collections.defaultdict(set) 3756globals_ = collections.defaultdict(list) 3757for name, rec in tensors.items(): 3758 m = re.match(r"model\.layers\.(\d+)\.", name) 3759 if m: 3760 li = int(m.group(1)); l2s[li].add(rec["shard"]); s2l[rec["shard"]].add(li) 3761 else: 3762 globals_[rec["shard"]].append(name) 3763lmap = { 3764 "layer_to_shards": {str(k): sorted(v) for k, v in sorted(l2s.items())}, 3765 "shard_to_layers": {k: sorted(v) for k, v in sorted(s2l.items())}, 3766 "global_tensors_by_shard": {k: sorted(v) for k, v in sorted(globals_.items())}, 3767 "notes": { 3768 "dense_layers": [0, 1, 2], 3769 "moe_layers": list(range(3, 61)), 3770 "mtp_layer": 61, 3771 "base_inference_layers": list(range(0, 61)), 3772 }, 3773} 3774json.dump(lmap, open(os.path.join(OUT, "deepseek_layer_shard_map.json"), "w"), indent=1) 3775 3776# expert offset map: one contiguous read per expert where possible 3777experts = collections.defaultdict(list) 3778for name, rec in tensors.items(): 3779 m = re.match(r"model\.layers\.(\d+)\.mlp\.experts\.(\d+)\.", name) 3780 if m: 3781 experts[(int(m.group(1)), int(m.group(2)))].append(rec) 3782emap = {} 3783straddlers = 0 3784for (li, ei), recs in sorted(experts.items()): 3785 by_shard = collections.defaultdict(list) 3786 for r in recs: 3787 by_shard[r["shard"]].append(r) 3788 reads = [] 3789 for sh in sorted(by_shard): 3790 rs = sorted(by_shard[sh], key=lambda r: r["absolute_offset"]) 3791 lo = rs[0]["absolute_offset"] 3792 hi = max(r["absolute_offset"] + r["byte_length"] for r in rs) 3793 used = sum(r["byte_length"] for r in rs) 3794 reads.append({ 3795 "shard": sh, "offset": lo, "length": hi - lo, 3796 "contiguous": (hi - lo) == used, 3797 "tensors": [r["name"].split("experts.%d." % ei)[1] for r in rs], 3798 }) 3799 if len(reads) > 1: 3800 straddlers += 1 3801 emap["%d/%d" % (li, ei)] = { 3802 "layer": li, "expert": ei, 3803 "total_bytes": sum(r["byte_length"] for r in recs), 3804 "read_count": len(reads), 3805 "reads": reads, 3806 } 3807json.dump({"straddling_experts": straddlers, 3808 "expert_count": len(emap), 3809 "bytes_per_expert_nominal": 44050944, 3810 "experts": emap}, 3811 open(os.path.join(OUT, "deepseek_expert_offset_map.json"), "w"), indent=1) 3812 3813print("tensors:", len(manifest)) 3814print("shards:", len(shard_hdrs), "all length checks ok:", 3815 all(s["length_check_ok"] for s in shard_hdrs)) 3816print("experts:", len(emap), "straddling shard boundaries:", straddlers) 3817for f in ["deepseek_tensor_manifest.json","deepseek_shard_headers.json", 3818 "deepseek_layer_shard_map.json","deepseek_expert_offset_map.json"]: 3819 print(" wrote %-38s %12d B" % (f, os.path.getsize(os.path.join(OUT,f)))) 3820 3821-------------------------------------------------------------------------------- 3822C.11.11 gen_hashes.py (2523 bytes) 3823-------------------------------------------------------------------------------- 3824#!/usr/bin/env python3 3825""" 3826gen_hashes.py -- checkpoint integrity hashes. 3827 3828 --partial (default) size + SHA-256 of the first 1 MiB, last 1 MiB, and the 3829 1 MiB at the midpoint of each shard. Reads 489 MiB total, 3830 finishes in about a second. Enough to detect truncation, 3831 a swapped shard, or a corrupted head/tail. 3832 3833 --full SHA-256 over every byte of all 163 shards. Reads the whole 3834 643 GB. At the measured 507 MB/s single-stream rate on 3835 this machine that is roughly 22 minutes of I/O plus hash 3836 time. Required before any "verified" claim. 3837 3838Output: deepseek_weight_hashes.txt (sorted, one shard per stanza) 3839""" 3840import os, sys, hashlib, time 3841 3842D = os.path.expanduser("~/DeepSeek-R1") 3843OUT = os.path.expanduser("~/deepseek-megamind-upload/deepseek/deepseek_weight_hashes.txt") 3844MODE = "--full" if "--full" in sys.argv else "--partial" 3845CH = 1 << 20 3846 3847files = sorted(f for f in os.listdir(D) if f.endswith(".safetensors")) 3848t0 = time.time() 3849with open(OUT, "w") as out: 3850 out.write("# DeepSeek-R1 checkpoint hashes\n") 3851 out.write("# mode: %s\n" % MODE) 3852 out.write("# generated: %s\n" % time.strftime("%Y-%m-%dT%H:%M:%S%z")) 3853 out.write("# source: %s\n#\n" % D) 3854 total = 0 3855 for fn in files: 3856 p = os.path.join(D, fn); sz = os.path.getsize(p); total += sz 3857 if MODE == "--full": 3858 h = hashlib.sha256() 3859 with open(p, "rb") as fh: 3860 while True: 3861 b = fh.read(1 << 22) 3862 if not b: break 3863 h.update(b) 3864 out.write("%s\n size %d\n sha256 %s\n" % (fn, sz, h.hexdigest())) 3865 else: 3866 def part(off, n): 3867 with open(p, "rb") as fh: 3868 fh.seek(off); return hashlib.sha256(fh.read(n)).hexdigest() 3869 out.write("%s\n size %d\n" % (fn, sz)) 3870 out.write(" sha256_head %s (bytes 0..%d)\n" % (part(0, CH), CH)) 3871 mid = (sz // 2) - (CH // 2) 3872 out.write(" sha256_middle %s (bytes %d..%d)\n" % (part(mid, CH), mid, mid + CH)) 3873 out.write(" sha256_tail %s (bytes %d..%d)\n" % (part(max(0, sz - CH), CH), max(0, sz - CH), sz)) 3874 out.flush() 3875 out.write("#\n# shards: %d\n# total bytes: %d\n# elapsed: %.1fs\n" 3876 % (len(files), total, time.time() - t0)) 3877print("wrote", OUT, os.path.getsize(OUT), "B in %.1fs" % (time.time() - t0)) 3878 3879================================================================================ 3880C.12 WSL CONFIGURATION 3881================================================================================ 3882 3883--- C.12.1 /etc/wsl.conf (inside the WSL2 guest) --- 3884[boot] 3885systemd=true 3886 3887[network] 3888generateResolvConf=true 3889 3890--- C.12.2 C:\Users\Josep\.wslconfig (Windows host) --- 3891[wsl2] 3892memory=26GB 3893processors=12 3894vmIdleTimeout=-1 3895swap=8GB 3896 3897 3898 3899================================================================================ 3900C.13 SYSTEMD UNIT FILES (all 51, credential redacted in mm-loom.service) 3901================================================================================ 3902 3903-------------------------------------------------------------------------------- 3904mm-api.service 3905-------------------------------------------------------------------------------- 3906[Unit] 3907Description=missionMATCH API on megamind-agi (per tenant routing, control plane on 5433) 3908After=network-online.target [author/contact email redacted] 3909Wants=network-online.target 3910 3911[Service] 3912Type=simple 3913# Root owned 0600, read by systemd rather than by the service user, so the process never needs 3914# permission to read the credentials it is handed. 3915EnvironmentFile=/etc/missionmatch/api.env 3916ExecStart=/opt/missionmatch/mm-api 3917Restart=on-failure 3918RestartSec=5 3919DynamicUser=yes 3920NoNewPrivileges=yes 3921PrivateTmp=yes 3922ProtectSystem=strict 3923ProtectHome=yes 3924 3925[Install] 3926WantedBy=multi-user.target 3927 3928-------------------------------------------------------------------------------- 3929mm-boot.service 3930-------------------------------------------------------------------------------- 3931[Unit] 3932Description=MEGAMIND-AGI boot supervisor 3933After=postgresql.service 3934Wants=postgresql.service 3935[Service] 3936Type=oneshot 3937ExecStart=/opt/mm/bin/mmboot 3938RemainAfterExit=yes 3939[Install] 3940WantedBy=multi-user.target 3941 3942-------------------------------------------------------------------------------- 3943mm-cap.service 3944-------------------------------------------------------------------------------- 3945[Unit] 3946Description=MEGAMIND-AGI capability manifest 3947After=network-online.target postgresql.service 3948[Service] 3949Type=oneshot 3950ExecStart=/opt/mm/bin/mmcap 3951 3952-------------------------------------------------------------------------------- 3953mm-cap.timer 3954-------------------------------------------------------------------------------- 3955[Unit] 3956Description=MEGAMIND-AGI manifest every 60s 3957[Timer] 3958OnBootSec=45 3959OnUnitActiveSec=60 3960[Install] 3961WantedBy=timers.target 3962 3963-------------------------------------------------------------------------------- 3964mm-cp-backup.service 3965-------------------------------------------------------------------------------- 3966[Unit] 3967Description=missionMATCH control plane backup, dumped here and pushed to Bubbles 3968After=network-online.target 3969Wants=network-online.target 3970 3971[Service] 3972Type=oneshot 3973ExecStart=/usr/local/sbin/mm-cp-backup.sh 3974User=root 3975Group=root 3976Nice=10 3977 3978-------------------------------------------------------------------------------- 3979mm-cp-backup.timer 3980-------------------------------------------------------------------------------- 3981[Unit] 3982Description=Nightly missionMATCH control plane backup push 3983[Timer] 3984# 02:45, twenty minutes BEFORE the Bubbles job at 03:05, so the dump has arrived to be ingested. 3985OnCalendar=*-*-* 02:45:00 3986Persistent=true 3987RandomizedDelaySec=120 3988[Install] 3989WantedBy=timers.target 3990 3991-------------------------------------------------------------------------------- 3992mm-durability.service 3993-------------------------------------------------------------------------------- 3994[Unit] 3995Description=PANTHEON two-device durability (M4 <-> cortex) 3996After=network-online.target 3997Wants=network-online.target 3998 3999[Service] 4000Type=oneshot 4001ExecStart=/opt/mm/backup/durability.sh 4002 4003-------------------------------------------------------------------------------- 4004mm-durability.timer 4005-------------------------------------------------------------------------------- 4006[Unit] 4007Description=Nightly PANTHEON durability backup 4008 4009[Timer] 4010OnCalendar=*-*-* 03:30:00 4011Persistent=true 4012 4013[Install] 4014WantedBy=timers.target 4015 4016-------------------------------------------------------------------------------- 4017mm-eval.service 4018-------------------------------------------------------------------------------- 4019[Unit] 4020Description=MEGAMIND-AGI substrate eval harness 4021After=postgresql.service 4022[Service] 4023Type=oneshot 4024ExecStart=/opt/mm/bin/mmeval 4025 4026-------------------------------------------------------------------------------- 4027mm-eval.timer 4028-------------------------------------------------------------------------------- 4029[Unit] 4030Description=MEGAMIND-AGI hourly substrate quality eval 4031[Timer] 4032OnBootSec=300 4033OnUnitActiveSec=3600 4034[Install] 4035WantedBy=timers.target 4036 4037-------------------------------------------------------------------------------- 4038mm-eye.service 4039-------------------------------------------------------------------------------- 4040[Unit] 4041Description=EYE - Eye of the World, PANTHEON binary VSA world-brain 4042After=postgresql.service network.target 4043[Service] 4044Type=simple 4045User=postgres 4046Environment=EYE_PORT=9404 4047Environment=EYE_HEARTBEAT_SEC=12 4048Environment="EYE_DB=host=/var/run/postgresql dbname=livingworld" 4049ExecStart=/opt/mm/bin/eye 4050Restart=always 4051RestartSec=3 4052[Install] 4053WantedBy=multi-user.target 4054 4055-------------------------------------------------------------------------------- 4056mm-game-reflect.service 4057-------------------------------------------------------------------------------- 4058[Unit] 4059Description=MEGAMIND game reflection tick 4060[Service] 4061Type=oneshot 4062ExecStart=/usr/bin/curl -s --max-time 60 http://127.0.0.1:9097/game/tick?min=4 4063 4064-------------------------------------------------------------------------------- 4065mm-game-reflect.timer 4066-------------------------------------------------------------------------------- 4067[Unit] 4068Description=Reflect active NPCs every 30 min 4069[Timer] 4070OnBootSec=15min 4071OnUnitActiveSec=30min 4072[Install] 4073WantedBy=timers.target 4074 4075-------------------------------------------------------------------------------- 4076mm-game.service 4077-------------------------------------------------------------------------------- 4078[Unit] 4079Description=MEGAMIND binary Game Memory API (mmgame :9097) 4080After=network.target postgresql.service mm-serve.service 4081[Service] 4082Type=simple 4083ExecStart=/opt/mm/bin/mmgame 4084Restart=always 4085RestartSec=5 4086WorkingDirectory=/opt/mm 4087[Install] 4088WantedBy=multi-user.target 4089 4090-------------------------------------------------------------------------------- 4091mm-gate.service 4092-------------------------------------------------------------------------------- 4093[Unit] 4094Description=MEGAMIND-AGI anti-degeneracy gate 4095After=mm-serve.service 4096[Service] 4097Type=oneshot 4098ExecStart=/opt/mm/bin/mmgate 4099 4100-------------------------------------------------------------------------------- 4101mm-gate.timer 4102-------------------------------------------------------------------------------- 4103[Unit] 4104Description=MEGAMIND-AGI degeneracy check every 5min 4105[Timer] 4106OnBootSec=120 4107OnUnitActiveSec=300 4108[Install] 4109WantedBy=timers.target 4110 4111-------------------------------------------------------------------------------- 4112mm-genserve.service 4113-------------------------------------------------------------------------------- 4114[Unit] 4115Description=MEGAMIND-AGI P20 binary generator (mmpredict serve, :9098) 4116After=network.target postgresql.service 4117[Service] 4118Type=simple 4119ExecStart=/opt/mm/bin/mmpredict serve 4120Restart=always 4121RestartSec=5 4122WorkingDirectory=/opt/mm 4123[Install] 4124WantedBy=multi-user.target 4125 4126-------------------------------------------------------------------------------- 4127mm-heart.service 4128-------------------------------------------------------------------------------- 4129[Unit] 4130Description=MEGAMIND-AGI mesh heartbeat 4131After=network-online.target 4132[Service] 4133Type=oneshot 4134ExecStart=/opt/mm/bin/mmheart 4135 4136-------------------------------------------------------------------------------- 4137mm-heart.timer 4138-------------------------------------------------------------------------------- 4139[Unit] 4140Description=MEGAMIND-AGI heartbeat every 15s 4141[Timer] 4142OnBootSec=30 4143OnUnitActiveSec=15 4144[Install] 4145WantedBy=timers.target 4146 4147-------------------------------------------------------------------------------- 4148mm-identity-sync.service 4149-------------------------------------------------------------------------------- 4150[Unit] 4151Description=missionMATCH identity sync, control plane to every active tenant 4152After=network-online.target 4153Wants=network-online.target 4154 4155[Service] 4156Type=oneshot 4157ExecStart=/usr/local/sbin/mm-identity-sync.sh 4158User=root 4159Group=root 4160Nice=10 4161 4162-------------------------------------------------------------------------------- 4163mm-identity-sync.timer 4164-------------------------------------------------------------------------------- 4165[Unit] 4166Description=Hourly missionMATCH identity sync 4167[Timer] 4168# Hourly, because the drift it corrects appears whenever an invite is accepted or a role changes, and a 4169# daily cadence would leave a new user invisible to tenant foreign keys for most of a working day. 4170OnCalendar=hourly 4171Persistent=true 4172RandomizedDelaySec=300 4173[Install] 4174WantedBy=timers.target 4175 4176-------------------------------------------------------------------------------- 4177mm-keepalive.service 4178-------------------------------------------------------------------------------- 4179[Unit] 4180Description=MEGAMIND WSL instance keepalive (holds the WSL2 instance against the hardcoded idle timer) 4181After=multi-user.target 4182[Service] 4183Type=exec 4184ExecStart=/usr/bin/sleep infinity 4185Restart=always 4186RestartSec=1 4187[Install] 4188WantedBy=multi-user.target 4189-------------------------------------------------------------------------------- 4190mm-loom.service 4191-------------------------------------------------------------------------------- 4192[Unit] 4193Description=The Loom - Living World orchestration organ 4194After=network.target mm-eye.service 4195[Service] 4196[REDACTED: credential-associated original line 4196; original retained outside this bundle.] 4197Environment=LOOM_PORT=9410 4198ExecStart=/opt/mm/bin/loom 4199Restart=always 4200RestartSec=3 4201[Install] 4202WantedBy=multi-user.target 4203 4204-------------------------------------------------------------------------------- 4205mm-narrate.service 4206-------------------------------------------------------------------------------- 4207[Unit] 4208Description=MEGAMIND narrator 4209After=network.target 4210[Service] 4211ExecStart=/opt/mm/bin/mm-narrate 4212Restart=always 4213User=root 4214[Install] 4215WantedBy=multi-user.target 4216 4217-------------------------------------------------------------------------------- 4218mm-offsite-watch.service 4219-------------------------------------------------------------------------------- 4220[Unit] 4221Description=missionMATCH watch from OFF Bubbles (cross node vantage, not external) 4222After=network-online.target 4223 4224[Service] 4225Type=oneshot 4226ExecStart=/usr/local/sbin/mm-offsite-watch.sh 4227 4228-------------------------------------------------------------------------------- 4229mm-offsite-watch.timer 4230-------------------------------------------------------------------------------- 4231[Unit] 4232Description=run the off Bubbles watch every five minutes 4233 4234[Timer] 4235OnBootSec=3min 4236OnUnitActiveSec=5min 4237AccuracySec=30s 4238 4239[Install] 4240WantedBy=timers.target 4241 4242-------------------------------------------------------------------------------- 4243mm-oracle.service 4244-------------------------------------------------------------------------------- 4245[Unit] 4246Description=ORACLE - MEGAMIND binary VSA reasoning organ (16 powers, WAL-persistent) 4247After=network.target 4248[Service] 4249Type=simple 4250WorkingDirectory=/opt/mm/oracle 4251ExecStart=/opt/mm/bin/oracle 4252Restart=always 4253RestartSec=2 4254[Install] 4255WantedBy=multi-user.target 4256 4257-------------------------------------------------------------------------------- 4258mm-sem.service 4259-------------------------------------------------------------------------------- 4260[Unit] 4261Description=MEGAMIND shared contextual semantic embedder (mmsem :9096) 4262After=network.target 4263[Service] 4264Type=simple 4265ExecStart=/opt/mm/bin/mmsem 4266Restart=always 4267RestartSec=5 4268[Install] 4269WantedBy=multi-user.target 4270 4271-------------------------------------------------------------------------------- 4272mm-serve.service 4273-------------------------------------------------------------------------------- 4274[Unit] 4275Description=MEGAMIND-AGI 512-bit recall service 4276After=postgresql.service 4277Wants=postgresql.service 4278PartOf=mm.target 4279[Service] 4280ExecStart=/opt/mm/bin/mmserve 4281Restart=on-failure 4282RestartSec=5 4283KillSignal=SIGTERM 4284TimeoutStopSec=120 4285[Install] 4286WantedBy=multi-user.target mm.target 4287 4288-------------------------------------------------------------------------------- 4289mm-shard.service 4290-------------------------------------------------------------------------------- 4291[Unit] 4292Description=MEGAMIND-AGI federated recall shard 4293After=mm-serve.service network-online.target 4294Wants=mm-serve.service 4295PartOf=mm.target 4296[Service] 4297ExecStart=/opt/mm/bin/mmshard 4298Restart=on-failure 4299RestartSec=5 4300[Install] 4301WantedBy=multi-user.target mm.target 4302 4303-------------------------------------------------------------------------------- 4304mm-wal-compact.service 4305-------------------------------------------------------------------------------- 4306[Unit] 4307Description=ORACLE WAL compaction (keep restarts fast) 4308 4309[Service] 4310Type=oneshot 4311ExecStart=/opt/mm/oracle/compact_wal.sh 4312 4313-------------------------------------------------------------------------------- 4314mm-wal-compact.timer 4315-------------------------------------------------------------------------------- 4316[Unit] 4317Description=daily ORACLE WAL compaction 4318 4319[Timer] 4320OnCalendar=*-*-* 04:17:00 4321Persistent=true 4322 4323[Install] 4324WantedBy=timers.target 4325 4326-------------------------------------------------------------------------------- 4327mm-world.service 4328-------------------------------------------------------------------------------- 4329[Unit] 4330Description=MEGAMIND shared associative world-model organ (mmworld :9098) 4331After=network.target 4332[Service] 4333Type=simple 4334ExecStart=/opt/mm/bin/mmworld 4335Restart=always 4336RestartSec=5 4337[Install] 4338WantedBy=multi-user.target 4339 4340-------------------------------------------------------------------------------- 4341mm.target 4342-------------------------------------------------------------------------------- 4343[Unit] 4344Description=MEGAMIND-AGI node stack 4345Wants=postgresql.service mm-serve.service mm-heart.timer mm-cap.timer 4346After=postgresql.service 4347[Install] 4348WantedBy=multi-user.target 4349 4350-------------------------------------------------------------------------------- 4351mmask.service 4352-------------------------------------------------------------------------------- 4353[Unit] 4354Description=mmask (PANTHEON ask brain — bridges mmknow+ORACLE+mmreason -> LLM) 4355After=network.target mm-oracle.service 4356 4357[Service] 4358Type=simple 4359WorkingDirectory=/opt/mm/reason 4360Environment=MMKNOW=http://100.77.114.40:9542/know 4361Environment=MMREASON=http://100.77.114.40:9522/reason 4362Environment=ORACLE=http://100.69.120.61:9402/oracle 4363Environment=LLM=http://100.94.8.94:8082/v1/chat/completions 4364ExecStart=/usr/bin/python3 /opt/mm/reason/mmask_serve.py 9543 4365Restart=always 4366RestartSec=3 4367User=joseph 4368 4369[Install] 4370WantedBy=multi-user.target 4371 4372-------------------------------------------------------------------------------- 4373mmbackward.service 4374-------------------------------------------------------------------------------- 4375[Unit] 4376Description=PANTHEON backward circulation loop (nonstop federation dataflow) 4377After=network.target mm-oracle.service 4378[Service] 4379Type=simple 4380ExecStart=/usr/bin/python3 /opt/mm/loops/backward.py 4381Restart=always 4382RestartSec=3 4383User=joseph 4384WorkingDirectory=/opt/mm/loops 4385[Install] 4386WantedBy=multi-user.target 4387 4388-------------------------------------------------------------------------------- 4389mmcode2.service 4390-------------------------------------------------------------------------------- 4391[Unit] 4392Description=mmcode2 binary recall over CodeSearchNet (code + docstrings) 4393After=network.target 4394[Service] 4395Environment=WX_BSIG=/opt/mm/code/code.bsig 4396Environment=WX_BLOB=/opt/mm/code/code.blob 4397Environment=WX_IDX=/opt/mm/code/code.idx 4398Environment=WX_NAME=code2 4399ExecStart=/opt/mm/code/mmwikix serve 9534 4400Restart=always 4401User=joseph 4402WorkingDirectory=/opt/mm/code 4403[Install] 4404WantedBy=multi-user.target 4405 4406-------------------------------------------------------------------------------- 4407mmforward.service 4408-------------------------------------------------------------------------------- 4409[Unit] 4410Description=PANTHEON forward circulation loop (nonstop federation dataflow) 4411After=network.target mm-oracle.service 4412[Service] 4413Type=simple 4414ExecStart=/usr/bin/python3 /opt/mm/loops/forward.py 4415Restart=always 4416RestartSec=3 4417User=joseph 4418WorkingDirectory=/opt/mm/loops 4419[Install] 4420WantedBy=multi-user.target 4421 4422-------------------------------------------------------------------------------- 4423mmgateway-web.service 4424-------------------------------------------------------------------------------- 4425[Unit] 4426Description=mmgateway-web (read-only public gateway to every PANTHEON organ) 4427After=network.target mm-oracle.service 4428 4429[Service] 4430Type=simple 4431WorkingDirectory=/opt/mm/reason 4432Environment=ORACLE=http://100.69.120.61:9402/oracle 4433ExecStart=/usr/bin/python3 /opt/mm/reason/mmgateway_web.py 9560 4434Restart=always 4435RestartSec=3 4436User=joseph 4437 4438[Install] 4439WantedBy=multi-user.target 4440 4441-------------------------------------------------------------------------------- 4442mmhelix.service 4443-------------------------------------------------------------------------------- 4444[Unit] 4445Description=mmhelix G^16 -- 16 strand-loops, 8 antiparallel complement pairs 4446After=network.target 4447[Service] 4448ExecStart=/usr/bin/python3 /opt/mm/loops/mmhelix.py 4449Restart=always 4450RestartSec=3 4451User=joseph 4452[Install] 4453WantedBy=multi-user.target 4454 4455-------------------------------------------------------------------------------- 4456mmknow.service 4457-------------------------------------------------------------------------------- 4458[Unit] 4459Description=mmknow (PANTHEON mind organ) 4460After=network.target mm-oracle.service 4461 4462[Service] 4463Type=simple 4464WorkingDirectory=/opt/mm/reason 4465ExecStart=/usr/bin/python3 /opt/mm/reason/mmknow_serve.py 9542 4466Restart=always 4467RestartSec=3 4468User=joseph 4469 4470[Install] 4471WantedBy=multi-user.target 4472 4473-------------------------------------------------------------------------------- 4474mmlearn.service 4475-------------------------------------------------------------------------------- 4476[Unit] 4477Description=mmlearn (PANTHEON learning organ) 4478After=network.target mm-oracle.service mmreason.service 4479 4480[Service] 4481Type=simple 4482WorkingDirectory=/opt/mm/reason 4483Environment="ORACLE=http://127.0.0.1:9402/oracle" 4484Environment="MMREASON=http://127.0.0.1:9522/reason" 4485ExecStart=/usr/bin/python3 /opt/mm/reason/mmlearn_serve.py 9541 4486Restart=always 4487RestartSec=3 4488User=joseph 4489 4490[Install] 4491WantedBy=multi-user.target 4492 4493-------------------------------------------------------------------------------- 4494mmlearned.service 4495-------------------------------------------------------------------------------- 4496[Unit] 4497Description=mmlearned binary recall over the learned/world-knowledge corpus (iCloud AGI_Data learning streams) 4498After=network.target 4499[Service] 4500Environment=WX_BSIG=/opt/mm/learned/learned.bsig