1. Methodology & Executive Summary
This submission optimizes the Sarvam-30b baseline for high-performance reasoning within constrained energy envelopes.
Compression Strategy
We employed 4-bit Quantization via Compressed-Tensors to achieve W4A16 precision.
Efficiency Goal
The model aims to achieve the optimal balance between inference accuracy and total energy draw, adhering to the challenge threshold of >80% baseline performance.
Energy Transparency
Power efficiency was prioritized throughout the development cycle to minimize the carbon footprint of inference.
2. Technical Specifications
- Base Model: Sarvam-30b
- Architecture: SarvamMoEForCausalLM
- Precision: W4A16
- License: Apache 2.0
- Default Context Length: 65,536 tokens
3. Inference & Deployment Configuration
The model is optimized to operate on NVIDIA A100 (80GB) hardware using the vLLM OpenAI API server format.
vllm_config.yaml
model: ./models/sarvam-30b-compressed-w4a16
quantization: compressed-tensors
kv-cache-dtype: auto
max-model-len: 8192
trust-remote-code: true
4. Evaluation Framework
This submission follows the rigorous evaluation protocol defined for the Resilient AI Challenge.
Quality Criteria
Performance was validated using LLM-as-a-judge across the following benchmark categories:
- Technical Reasoning
- Medical Knowledge Synthesis
- Linguistic Creativity
- Analytical Logic
Energy Monitoring
Our monitoring pipeline utilizes the NVIDIA Management Library (NVML) for real-time GPU power draw, complemented by TDP-relative estimates for CPU load.