Mixtral Experts with DeepSeek-MoE Architecture
This is a direct extraction of the 8 experts from Mixtral-8x7b-Instruct-v0.1, and a transfer of them into the DeepSeek-MoE Architecture.
- Expert Configuration: It is 2 experts per token.
- Performance: Performance is identical to instruct, if not a little better.
- Evaluations: Evals will come when compute clears up, it also appears more malleable to training.
- Experimentation: This is the first of a few MoE expert extraction and modification projects we're working on, more to come. Enjoy.
Instruction Format
To leverage instruction fine-tuning, your prompts should be enclosed with [INST] and [/INST] tokens. The very first instruction should begin with a begin-of-sentence id, while subsequent instructions should not. Assistant generation will conclude with an end-of-sentence token id.
Example
text = "<s>[INST] What is your favourite condiment? [/INST]"
"Well, I'm quite partial to a good squeeze of fresh lemon juice. It adds just the right amount of zesty flavour to whatever I'm cooking up in the kitchen!</s>"
"[INST] Do you have mayonnaise recipes? [/INST]"
Applying the Chat Template
This format can be implemented using the apply_chat_template() method from the transformers library:
from transformers import AutoModelForCausalLM, AutoTokenizer
device = "cuda" # the device to load the model onto
# Load the model and tokenizer
model = AutoModelForCausalLM.from_pretrained("cognitivecomputations/DeepMixtral-8x7b-Instruct", trust_remote_code=True, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("cognitivecomputations/DeepMixtral-8x7b-Instruct")
# Define the conversation messages
messages = [
{"role": "user", "content": "What is your favourite condiment?"},
{"role": "assistant", "content": "Well, I'm quite partial to a good squeeze of fresh lemon juice. It adds just the right amount of zesty flavour to whatever I'm cooking up in the kitchen!"},
{"role": "user", "content": "Do you have mayonnaise recipes?"}
]
# Apply chat template
encodeds = tokenizer.apply_chat_template(messages, return_tensors="pt")
model_inputs = encodeds.to(device)
model.to(device)
# Generate response
generated_ids = model.generate(model_inputs, max_new_tokens=1000, do_sample=True)
decoded = tokenizer.batch_decode(generated_ids)
print(decoded[0])
Special Thanks: Eric Hartford, and Fernando Neto.
- Lucas Atkins (Crystalcareai)