> ## Documentation Index
> Fetch the complete documentation index at: https://tensorfuse.io/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Samagra Sharma

> Samagra Sharma is the founder of Tensorfuse and writes about ML infrastructure, inference serving, GPU cold starts, and distributed systems.

export const AuthorBio = ({name, role, imageSrc, href}) => <div style={{
  display: 'flex',
  alignItems: 'center',
  maxWidth: 'fit-content'
}}>
        <div style={{
  display: 'flex',
  alignItems: 'center',
  justifyContent: 'center'
}}>
        <img src={imageSrc} alt={`${name} profile`} style={{
  width: '40px',
  height: '40px',
  borderRadius: '50%',
  marginRight: '1rem',
  objectFit: 'cover'
}} />
        </div>
        <div>
            <div style={{
  color: 'white',
  fontSize: '1rem',
  fontWeight: '500'
}}>
                {href ? <a href={href} style={{
  color: 'inherit'
}}>{name}</a> : name}
            </div>
            <div style={{
  color: '#888',
  fontSize: '1rem'
}}>
                {role}
            </div>
        </div>
    </div>;

<AuthorBio name="Samagra Sharma" role="Founder, Tensorfuse" imageSrc="https://samagra.me/assets/images/test.jpg" href="https://samagra.me/" />

Samagra Sharma is the founder of Tensorfuse. His writing here focuses on the engineering decisions behind deploying and operating machine learning workloads: serving performance, GPU startup, infrastructure costs, and distributed training.

See [Samagra's personal website](https://samagra.me/) for his background and other work.

## Inference and infrastructure articles

* [How to calculate LLM inference cost per successful request](/docs/blogs/inference-cost-per-request): compare serving configurations using costs, latency requirements, and accepted outputs.
* [When does GPU scale-to-zero save money?](/docs/blogs/scale-to-zero-gpu-inference): work through idle gaps, restart costs, and first-request latency.
* [Lazy loading GPU containers: why startup is not readiness](/docs/blogs/inference_on_k8s_1): distinguish image availability, model readiness, and the first useful response.
* [Understanding multi-GPU communication and NVIDIA NCCL](/docs/blogs/multi_gpu_communication_while_training): learn about communication during distributed training and investigate watchdog timeouts.

## Inference how-to guides

* [Reduce LLM inference cost](/docs/guides/inference/how-to/reduce-cost)
* [Serve a quantized model with vLLM](/docs/guides/inference/how-to/quantization-vllm)
* [Run vLLM on AMD GPUs](/docs/guides/inference/how-to/amd-vllm)
* [Serve LLMs with AWS Neuron](/docs/guides/inference/how-to/aws-neuron-vllm)
* [Run vLLM on Google TPUs](/docs/guides/inference/how-to/google-tpu-vllm)
* [Serve a local model with MLX on Apple silicon](/docs/guides/inference/how-to/apple-silicon-mlx)
* [Run CPU inference with llama.cpp](/docs/guides/inference/how-to/cpu-llama-cpp)
* [Compare Groq, Cerebras, and SambaNova APIs](/docs/guides/inference/how-to/compare-inference-apis)

The new cost articles use explicitly hypothetical calculations. They do not present unrun experiments as measured results. For deployment steps and hardware evaluation, use the [LLM inference guides](/docs/guides/inference/overview).