Obsolete, see https://huggingface.co/brucethemoose/Yi-34B-200K-DARE-merge-v5
Dolphin-2.2-yi-34b-200k, Nous-Capybara-34B, Tess-M-v1.4, Airoboros-3_1-yi-34b-200k, PlatYi-34B-200K-Q, and Una-xaberius-34b-v1beta merged with a new, experimental implementation of "dare ties" via mergekit.
Quantized with the git version of exllamav2 with 200 rows (400K tokens) on a long Orca-Vicuna format chat, a selected sci fi story and a fantasy story. This should hopefully yield better chat/storytelling performance than the short, default wikitext quantization.
4bpw is enough for ~47K context on a 24GB GPU. I would highly recommend running in exui for speed at long context. I go into more detail in this Reddit post
Merged with the following config, and the tokenizer from chargoddard's Yi-Llama:
models:
- model: /home/alpha/Storage/Models/Raw/chargoddard_Yi-34B-200K-Llama
# no parameters necessary for base model
- model: /home/alpha/Storage/Models/Raw/migtissera_Tess-34B-v1.4
parameters:
weight: 0.19
density: 0.6
- model: /home/alpha//Storage/Models/Raw/bhenrym14_airoboros-3_1-yi-34b-200k
parameters:
weight: 0.14
density: 0.5
- model: /home/alpha/Storage/Models/Raw/Nous-Capybara-34B
parameters:
weight: 0.19
density: 0.6
- model: /home/alpha/Storage/Models/Raw/kyujinpy_PlatYi-34B-200K-Q
parameters:
weight: 0.14
density: 0.5
- model: /home/alpha/FastModels/ehartford_dolphin-2.2-yi-34b-200k
parameters:
weight: 0.19
density: 0.6
- model: /home/alpha/FastModels/fblgit_una-xaberius-34b-v1beta
parameters:
weight: 0.15
density: 0.08
merge_method: dare_ties
base_model: /home/alpha/Storage/Models/Raw/chargoddard_Yi-34B-200K-Llama
parameters:
int8_mask: true
dtype: bfloat16
First exllama quantization pass:
python convert.py --in_dir //home/alpha/FastModels/CaPlatTessDolXaBoros-Yi-34B-200K-DARE-Ties -o /home/alpha/FastModels/scratch -om /home/alpha/FastModels/mes.json --cal_dataset /home/alpha/Documents/smol.parquet -l 2048 -r 80 -ml 2048 -mr 40 -gr 40 -ss 4096 -nr -b 4.0 -hb 6
Second exllama quantization pass:
python convert.py --in_dir /home/alpha/FastModels/CaPlatTessDolXaBoros-Yi-34B-200K-DARE-Ties -o /home/alpha/FastModels/scratch -m /home/alpha/FastModels/mes.json --cal_dataset /home/alpha/Documents/medium.parquet -l 2048 -r 200 -ml 2048 -mr 40 -gr 200 -ss 4096 -b 4.0 -hb 6 -cf /home/alpha/FastModels/CaPlatTessDolXaBoros-Yi-34B-200K-DARE-Ties-exl2-4bpw-fiction -nr
Testing Notes
Various densities were tested with perplexity tests and high context prompts. Relatively high densities seem to perform better, contrary to the findings of the Super Mario paper.
Weights that add up to 1 seems to be optimal.
Dare Ties is also resulting in seemingly better, lower perplexity merges than a regular ties merge, task arithmetic or a slerp merge.
Xaberuis is not a 200K model, hence it was merged at a very low density to try and preserve Yi 200K's long context performance while still inheriting some of Xaberius's performance.
I chose not to include other finetunes because they aren't trained on the 200K base. If any other 200K finetunes pop up, let me know.
Prompt template: Orca-Vicuna?
SYSTEM: {system_message}
USER: {prompt}
ASSISTANT:
It might recognize ChatML from Dolphin+Xaberius, and Llama-chat from Airoboros.
Sometimes the model "spells out" the stop token as </s> like Capybara, so you may need to add </s> as an additional stopping condition.
Running
Being a Yi model, try disabling the BOS token and/or running a lower temperature with 0.05-0.13 MinP, a little repetition penalty, and no other samplers. Yi tends to run "hot" by default.
24GB GPUs can run Yi-34B-200K models at 45K-75K context with exllamav2. I go into more detail in this post
I recommend exl2 quantizations profiled on data similar to the desired task. It is especially sensitive to the quantization data at low bpw!
To load this in full-context backends like transformers and vllm, you must change max_position_embeddings in config.json to a lower value than 200,000, otherwise you will OOM!
Credits:
https://github.com/turboderp/exllamav2
https://github.com/cg123/mergekit/tree/dare
https://huggingface.co/ehartford/dolphin-2.2-yi-34b-200k
https://huggingface.co/kyujinpy/PlatYi-34B-200K-Q
https://huggingface.co/NousResearch/Nous-Capybara-34B/
https://huggingface.co/bhenrym14/airoboros-3_1-yi-34b-200k
https://huggingface.co/migtissera/Tess-M-v1.4
https://huggingface.co/fblgit/una-xaberius-34b-v1beta
- Downloads last month
- 5