Co-Founder
About
Pankaj Gupta is a co-founder of Baseten, where he leads model performance. Pankaj has spent his career making systems faster and more efficient, from optimizing data processing libraries at Twitter to search infrastructure at Uber and media processing at Adobe. A graduate of IIT Delhi, Pankaj now lives in the Bay Area, where he enjoys gardening and evening walks around his neighborhood.
Are NVIDIA H200 GPUs cost-effective for model inference? We tested an 8xH200 cluster provided by Lambda to discover suitable inference workload profiles.
LoRA swapping with TRT-LLM supports in-flight batching and loads LoRA weights in 1-2 ms, enabling each request to hit a different fine-tune.
Multi-Instance GPUs enable splitting a single H100 GPU across two model serving instances for performance that matches or beats an A100 GPU at a 20% lower cost.
Running Mistral 7B in FP8 on H100 GPUs with TensorRT-LLM, we achieve best in class time to first token and tokens per second on independent benchmarks.
Quantizing open-source LLMs to FP8 resulted in near-zero perplexity gains and yielded material performance improvements across latency, throughput, and cost.
The FP8 data format has an expanded dynamic range versus INT8 which allows for quantizing weights and activations for more LLMs without loss of output quality.
Using NVIDIA TensorRT to optimize each component of the SDXL pipeline, we improved SDXL inference latency by 40% and throughput by 70% on NVIDIA H100 GPUs.
Double or triple throughput at same-or-better latencies by switching to H100 GPUs from A100s for model inference with TensorRT/TensorRT-LLM.
