MiniChat-2-3B-pruned2.4

MiniChat-2-3B-pruned2.4

This repo contains model files for MiniChat-2-3B-pruned2.4 optimized for NM-vLLM, a high-throughput serving engine for compressed LLMs.

This model was pruned with SparseGPT, using SparseML.

Inference

Install NM-vLLM for fast inference and low memory-usage:

pip install nm-vllm[sparse]

Run in a Python pipeline for local inference:

from vllm import LLM, SamplingParams

model = LLM("nm-testing/MiniChat-2-3B-pruned2.4", sparsity="semi_structured_sparse_w16a16")
prompt = "How to make banana bread?"
formatted_prompt =  f"<s> [|User|]\n{prompt}</s>[|Assistant|]\n"

sampling_params = SamplingParams(max_tokens=100,temperature=0,repetition_penalty=1.3)
outputs = model.generate(formatted_prompt, sampling_params=sampling_params)
print(outputs[0].outputs[0].text)
"""
Answer: Create a recipe for making banana bread using ingredients like flour, water and sugar.  Explain the process of mixing these materials together until they form an unpleasant mixture that can be used in cooking methods such as baking or boiling processes. Describe how you would create this dough by adding it into your kitchen's oven-based environment while describing its properties during each stage before creating them on topical forms. You will also describe what

"""

Prompt template

### User:
{prompt}
### Assistant:

Sparsification

For details on how this model was sparsified, see the recipe.yaml in this repo and follow the instructions below.

Install SparseML:

git clone https://github.com/neuralmagic/sparseml
pip install -e "sparseml[transformers]"

Replace the recipe as you like and run this one-shot compression script to apply SparseGPT:

import sparseml.transformers

original_model_name = "GeneZC/MiniChat-2-3B"
calibration_dataset = "open_platypus"
output_directory = "output/"

recipe = """
test_stage:
  obcq_modifiers:
    SparseGPTModifier:
      sparsity: 0.5
      sequential_update: true
      mask_structure: '2:4'
      targets: ['re:model.layers.\d*$']
"""

# Apply SparseGPT to the model
sparseml.transformers.oneshot(
    model=original_model_name,
    dataset=calibration_dataset,
    recipe=recipe,
    output_dir=output_directory,
)

Slack

For further support, and discussions on these models and AI in general, join Neural Magic's Slack Community