Methodology
This model represents a fine-tuned version of the Sarvam-30B baseline, developed for the Resilient AI Challenge., developed for the Resilient AI Challenge.
The compression strategy utilized "Post-Training Quantization (PTQ) formatted via compressed-tensors to achieve a W4A16 precision balance.
The primary objective was to maximize energy efficiency while ensuring the model maintains at least 80% of the baseline Sarvam-30b performance.
Model Details
- Base Model: sarvam-30b
- Compression Precision: W4A16
- License: Apache 2.0
Inference Configuration
The model is optimized to run using the vLLM inference engine.
vllm_config.yaml
model: ./models/sarvam-30b-compressed-w4a16
quantization: compressed-tensors
kv-cache-dtype: auto
max-model-len: 8192
trust-remote-code: true
Evaluation Metrics
The model has been evaluated against the challenge benchmarks:
- Technical Reasoning: Advanced Science and Mathematics problem-solving.
- Domain-Specific Expertise: Medical knowledge synthesis.
- Linguistic Creativity: Narrative generation in English and Indian languages.
- Analytical Logic: Complex logical reasoning and deductive tasks.
- Energy Monitoring: Power consumption was tracked using the NVIDIA Management Library (NVML) for GPU draw and TDP-relative estimation for CPU load.
Usage
To serve this model for evaluation, use the following command:
vllm serve --config vllm_config.yaml