NCP-GENL GPU Acceleration and Optimization — All Questions With Answers

Question 1

An AI engineer is deploying a large language model on an NVIDIA A100 GPU using TensorRT-LLM. During inference profiling, they notice that token generation latency is higher than expected due to memory bandwidth bottlenecks during the autoregressive decoding phase. Which optimization technique should be applied first to mitigate this bandwidth limitation?

Accepted Answer

Integrate PagedAttention within the TensorRT-LLM runtime to optimize KV cache memory management and reduce fragmentation.. PagedAttention optimizes KV cache memory allocation by dividing it into blocks, eliminating fragmentation and allowing higher batch sizes without memory overhead. During the autoregressive generation phase, memory bandwidth is the primary bottleneck because every token generation requires loading all weights and KV cache values from high-bandwidth memory to the streaming multiprocessors.

Answer

Increase the maximum sequence length parameter to accommodate larger context windows without truncation.

Answer

Enable FP32 precision mode across all transformer layers to maximize numerical accuracy during matrix multiplications.

Answer

Disable kernel fusion optimizations to ensure individual PyTorch operations execute sequentially for easier debugging.

Question 2

A machine learning operations team is profiling a custom Transformer model trained on NVIDIA H100 GPUs using NVIDIA Nsight Systems. They observe that GPU utilization drops significantly during the data loading and tokenization phases between training steps. Which TWO strategies should the team implement to resolve this host-to-device pipeline bottleneck? (Choose two)

Accepted Answer

Increase the number of CPU worker processes in the data loader and enable pinned memory transfers.. Asynchronous data loading using pinned host memory allows direct memory access transfers via PCIe while the GPU computes the current batch. Increasing the number of DataLoader workers distributes parsing overhead across multiple CPU cores, preventing the host pipeline from starving the GPU of incoming training batches during distributed training runs.

Answer

Convert all model weight tensors from FP16 to FP32 format to improve numerical precision stability.

Answer

Implement gradient checkpointing to reduce the peak activation memory footprint during the backward pass.

Answer

Reduce the Tensor Core utilization threshold in the NVIDIA Management Library monitoring daemon.

Question 3

An enterprise developer is optimizing a computer vision pipeline on NVIDIA A100 GPUs using TensorRT. They want to maximize throughput for high-resolution image batches while maintaining acceptable latency bounds. Which execution mode should they configure in the TensorRT builder configuration?

Accepted Answer

Configure builder optimization profiles with explicit min, optimal, and max dimensions for dynamic batching.. Dynamic batching and multi-stream execution allow TensorRT to process multiple requests concurrently across execution contexts, maximizing hardware saturation. Setting up explicit batch sizes alongside builder optimization profiles tailors the compiled engine specifically for variable input dimensions encountered in production computer vision pipelines.

Answer

Force single-stream synchronous execution to guarantee deterministic response times for every incoming frame.

Answer

Enable FP64 double-precision mode across all convolutional layers to eliminate numerical rounding errors.

Answer

Disable LayerNorm and activation plugin autotuning to rely exclusively on default fallback implementations.

Question 4

Refer to the exhibit. A deep learning engineer encounters a CUDA out-of-memory error while initializing a deep neural network training job on a partitioned NVIDIA H100 GPU instance. Based on the error log, what is the most appropriate remediation step?

Accepted Answer

Increase the MIG partition size or reduce the cuDNN workspace allocation limit in the framework configuration.. The error indicates that cuDNN failed to allocate temporary workspace memory required during the algorithm search phase for convolution layers. Reducing the workspace size parameter in the framework configuration or scaling down the MIG slice allocation frees sufficient memory to complete algorithm benchmarking and execution.

Answer

Disable the Tensor Cores in the PyTorch configuration to force execution on standard scalar CUDA cores.

Answer

Upgrade the NVIDIA display driver to version 550 or higher to enable automatic virtual memory compaction.

Answer

Switch the precision mode from mixed-precision (FP16) to full FP32 training across all model layers.

Question 5

An inference server is experiencing high GPU latency when using FP16 precision. Profiling reveals that the kernel execution time is dominated by memory bandwidth bottlenecks. Which optimization technique is most likely to mitigate this latency while maintaining model throughput?

Accepted Answer

Apply operator fusion to combine contiguous layers.. Memory-bound kernels are frequently constrained by the amount of data transferred between VRAM and the streaming multiprocessors. By utilizing operator fusion, multiple kernels are combined into a single launch, which reduces the overhead of repeatedly reading and writing intermediate tensors to global memory. This significantly decreases total memory traffic and improves overall GPU compute utilization, which is essential for scaling high-performance generative AI workloads efficiently.

Answer

Increase the batch size to maximize thread occupancy.

Answer

Implement mixed-precision quantization to FP8.

Answer

Increase the GPU clock frequency via overclocking.

Question 6

Refer to the exhibit. An engineer observes the provided output from a GPU currently running an inference task. What does this indicate regarding the current workload state?

Accepted Answer

The system is operating near optimal compute capacity.. The output shows extremely high utilization for both the GPU compute units (98%) and the memory controller (85%), indicating a balanced but heavily loaded workload. This suggests the model is currently optimized to fully utilize the available hardware. To further improve performance, the engineer should investigate scaling strategies like tensor parallelism or model quantization to relieve the memory pressure before the GPU enters a thrashing state.

Answer

The GPU is idle and requires a larger workload.

Answer

The workload is primarily memory-bandwidth bound.

Answer

The GPU is thermal throttling due to load.

Question 7

Which CUDA-aware communication primitive is specifically designed to optimize data movement between multiple GPUs in a multi-node cluster by bypassing the CPU host memory?

Accepted Answer

GPUDirect RDMA.. GPUDirect RDMA enables direct peer-to-peer data transfers between the memory of one GPU and the memory of another device across a network, bypassing the CPU completely. This is vital for distributed training and large-scale inference where CPU-to-GPU copy latency acts as a bottleneck. It ensures high-speed, low-latency communication that is essential for scaling models across massive NVIDIA GPU clusters.

Answer

CUDA Host-to-Device Copy (H2D).

Answer

NCCL (NVIDIA Collective Communications Library).

Answer

Unified Virtual Memory (UVM).

Question 8

Which THREE techniques are primary methods for reducing the footprint of an LLM on GPU VRAM to enable larger models on restricted hardware?

Accepted Answer

Weight Quantization (e.g., INT8/FP8).. Reducing VRAM footprint is essential for fitting massive transformer models onto consumer or mid-range enterprise GPUs. Weight quantization, activation checkpointing, and tensor parallelism are the standard industry approaches to manage memory usage. These methods allow developers to deploy sophisticated generative models by sacrificing minimal accuracy for significant gains in memory efficiency and hardware compatibility, ensuring the models remain performant during high-demand inference sessions.

Answer

Increasing batch size indefinitely.

Answer

Disabling GPU L2 cache.

Question 9

In an H100 environment, which factor is most significant when choosing between TensorRT-LLM and standard PyTorch for high-throughput inference deployment?

Accepted Answer

The availability of advanced scheduling features.. TensorRT-LLM provides specialized, architecture-specific optimizations like fused kernels, advanced KV cache management, and hardware-level scheduling that are not available in standard PyTorch. While PyTorch is excellent for research and rapid prototyping, TensorRT-LLM is designed for production environments where raw throughput and latency minimization are non-negotiable. Using TensorRT-LLM ensures the GPU hardware is fully utilized, making it the superior choice for high-scale, high-concurrency generative AI applications.

Answer

Ease of modifying model architecture code.

Answer

Support for general-purpose CPU tasks.

Answer

Support for non-NVIDIA GPUs.

Question 10

Which technique provides the most significant performance gain for large-model inference by splitting the model across multiple GPUs while keeping the sequence length per batch constant?

Accepted Answer

Pipeline Parallelism.. Pipeline parallelism splits the model layers across multiple GPUs, allowing different layers of the model to be processed concurrently on different hardware. This is essential for models that are too large to fit into the memory of a single GPU. It enables the deployment of massive models by distributing the VRAM load, which is a foundational concept for scaling inference in production generative AI pipelines.

Answer

Data Parallelism.

Answer

Kernel Over-subscription.

Answer

Increasing global memory clock speed.

Question 11

An engineer is profiling a Transformer-based model and observes low GPU utilization despite high batch sizes. The bottleneck analysis identifies frequent host-to-device memory copies. Which optimization strategy most effectively addresses this latency?

Accepted Answer

Implement asynchronous memory copying using CUDA streams.. Frequent HtoD copies indicate the system is starving the GPU of data, negating the throughput benefits of large batches. Using NVIDIA GPUDirect Storage or pinned memory (cudaHostAlloc) allows the DMA engine to bypass the CPU overhead. This is critical in production LLM pipelines where overlapping data transfer with compute operations via CUDA streams is essential to achieve maximum kernel saturation and reduce overall inference latency.

Answer

Increase the number of threads per block.

Answer

Switch to FP64 precision for higher accuracy.

Answer

Reduce the batch size to lower memory footprint.

Question 12

Refer to the exhibit. An inference engine is failing during peak load. What is the most likely cause, and which strategy should be prioritized for immediate resolution?

Accepted Answer

Apply INT8 quantization to model weights and activations.. The logs indicate an Out-of-Memory (OOM) error induced by the tight VRAM margin. When memory is at 98%, memory fragmentation or slight spikes in activation buffers can cause execution failures. Prioritizing memory footprint reduction techniques like weight quantization (INT8/FP8) or model pruning is necessary to create overhead for workspace buffers, ensuring stability under peak inference demands in constrained memory environments.

Answer

Increase the PCIe lane speed in BIOS.

Answer

Enable CUDA Graphs to speed up kernel execution.

Answer

Use a larger kernel workspace size in TensorRT.

Question 13

Which THREE of the following are valid methods to optimize NVIDIA GPU memory bandwidth usage in deep learning?

Accepted Answer

Utilizing Operator Fusion.. Optimizing memory bandwidth is essential for memory-bound LLM tasks. Techniques like memory pooling, operator fusion, and precision reduction directly target the volume and frequency of data movement. These strategies ensure that data stays on the GPU as long as possible and that the data transferred is as small as possible, which is the key to maintaining high throughput in memory-constrained environments.

Answer

Aggressive use of FP64 precision.

Answer

Reducing the number of CUDA Streams.

Question 14

A team is serving a large language model with TensorRT-LLM on an NVIDIA H100 GPU. They observe that during the generation phase, GPU utilization is high, but the number of requests processed per second is lower than expected. Profiling shows that the GPU spends a significant amount of time waiting for memory operations. Which optimization technique should they apply to improve throughput?

Accepted Answer

Enable continuous batching to dynamically group requests at the token level.. Continuous batching keeps the GPU busy by allowing new requests to enter the batch as soon as they are ready, rather than waiting for the slowest sequence to finish. This maximizes utilization and throughput, especially when memory operations are a bottleneck. The other options either do not address the specific issue or could worsen it.

Answer

Use FP32 precision instead of FP16 to improve numerical stability.

Answer

Implement pipeline parallelism across multiple GPUs to distribute the model layers.

Answer

Increase the batch size to the maximum supported by the GPU memory.

Question 15

A developer is deploying a large language model on NVIDIA GPUs and wants to reduce inference latency while maintaining model accuracy. The model currently uses FP32 weights and activations. Which two techniques are most appropriate to achieve lower latency without significant accuracy loss? (Choose two.)

Accepted Answer

Convert the model to FP16 precision using automatic mixed precision (AMP).. Converting to FP16 with AMP and using TensorRT with FP16 and layer fusion both reduce latency by leveraging faster Tensor Core operations and optimized kernels while preserving accuracy. These techniques are well-established for LLM inference and do not require extensive retraining, unlike aggressive quantization or pruning.

Answer

Quantize the model to INT4 using post-training quantization.

Answer

Increase the number of GPUs to distribute the model.

Answer

Apply structured pruning to remove entire attention heads.

Question 16

A developer is fine-tuning a BERT model on an NVIDIA V100 GPU. They notice that the training step time is much longer than expected, and nvidia-smi shows that GPU memory is nearly full. The model uses a batch size of 64 and the maximum sequence length is 512. Which action is most likely to allow a larger batch size without running out of memory?

Accepted Answer

Enable mixed precision training with NVIDIA Apex or native AMP.. Mixed precision training reduces memory consumption by storing activations and gradients in FP16, which halves their footprint. This frees enough memory to increase the batch size. It also leverages Tensor Cores for faster math, improving step time. Other options either do not increase per-step batch capacity or alter the model's inputs.

Answer

Use gradient accumulation to simulate a larger batch size.

Answer

Use torch.cuda.empty_cache() before each step.

Answer

Reduce the maximum sequence length to 256.

Question 17

A research team is training a Transformer model on a cluster of NVIDIA A100 GPUs using data parallelism with PyTorch DistributedDataParallel (DDP). They observe that scaling efficiency drops significantly when going from 4 to 8 GPUs. Profiling shows high communication overhead in NCCL AllReduce operations. Which optimization is most likely to improve scaling efficiency in this scenario?

Accepted Answer

Use NCCL with hierarchical AllReduce and ensure NVLink/NVSwitch topology awareness.. The scaling drop is due to communication overhead in AllReduce. NCCL's hierarchical AllReduce takes advantage of intra-node NVLink/NVSwitch for fast local reductions, then performs a smaller inter-node reduction. Ensuring NCCL detects the topology enables optimal algorithm selection, reducing the data transferred over slower links and improving scaling efficiency.

Answer

Switch from DDP to model parallelism to reduce communication.

Answer

Increase the batch size per GPU to amortize communication.

Answer

Enable gradient compression using NCCL's built-in compression.

Question 18

When deploying a large language model on NVIDIA H100 GPUs using TensorRT-LLM, which TWO configuration strategies are most effective for improving KV cache efficiency and memory utilization?

Accepted Answer

Enable PagedAttention to reduce fragmentation.. Optimizing the KV cache is critical for scaling generative AI, as it occupies significant VRAM during inference. PagedAttention dynamically manages memory blocks to prevent fragmentation, while inflight batching ensures that continuous requests are processed without waiting for the entire batch to finish. Combining these allows for higher concurrency and reduced memory overhead, enabling more efficient deployment of models with large context windows on limited GPU hardware.

Answer

Increase the static sequence length allocation.

Answer

Disable multi-head attention optimizations.

Answer

Switch to FP64 precision for all tensors.

Question 19

In the context of NVIDIA Tensor Cores, what is the primary benefit of using BF16 (Bfloat16) over FP16 during model training and inference?

Accepted Answer

BF16 offers a larger dynamic range for gradients.. BF16 uses the same exponent range as FP32, which prevents overflow issues commonly encountered in deep learning training when using FP16. This increased dynamic range makes it more robust for gradient calculations and weight updates. By providing a wider range while maintaining the same performance advantages of half-precision, BF16 has become the industry standard for stabilizing training and inference of modern large language models.

Answer

BF16 provides double the precision of FP16.

Answer

BF16 requires significantly less memory than FP16.

Answer

BF16 is faster on non-NVIDIA hardware.

Question 20

Refer to the exhibit. An engineer receives this timeout error during a CUDA kernel execution. What is the most appropriate first step to diagnose the resource contention?

Accepted Answer

Use NVIDIA Nsight Compute to profile occupancy.. CUDA kernel timeouts are typically caused by long-running kernels that exceed the GPU's watchdog timer or by excessive resource usage (registers/shared memory) that limits occupancy. Using the NVIDIA Nsight Compute profiler allows the engineer to see the exact occupancy metrics and register pressure, enabling targeted optimizations. This step is crucial for identifying if the kernel is over-provisioned for the specific GPU architecture being targeted.

Answer

Increase the block size to 2048 threads.

Answer

Disable the watchdog timer in the OS.

Answer

Switch the kernel to run on the CPU.

NCP-GENL GPU Acceleration and Optimization — All Questions With Answers

An enterprise developer is optimizing a computer vision pipeline on NVIDIA A100 GPUs using TensorRT. They want to maximize throughput for high-resolution image batches while maintaining acceptable latency bounds. Which execution mode should they configure in the TensorRT builder configuration?

Refer to the exhibit. A deep learning engineer encounters a CUDA out-of-memory error while initializing a deep neural network training job on a partitioned NVIDIA H100 GPU instance. Based on the error log, what is the most appropriate remediation step?

Exhibit

An inference server is experiencing high GPU latency when using FP16 precision. Profiling reveals that the kernel execution time is dominated by memory bandwidth bottlenecks. Which optimization technique is most likely to mitigate this latency while maintaining model throughput?

Refer to the exhibit. An engineer observes the provided output from a GPU currently running an inference task. What does this indicate regarding the current workload state?

Which CUDA-aware communication primitive is specifically designed to optimize data movement between multiple GPUs in a multi-node cluster by bypassing the CPU host memory?

Which THREE techniques are primary methods for reducing the footprint of an LLM on GPU VRAM to enable larger models on restricted hardware?

In an H100 environment, which factor is most significant when choosing between TensorRT-LLM and standard PyTorch for high-throughput inference deployment?

Which technique provides the most significant performance gain for large-model inference by splitting the model across multiple GPUs while keeping the sequence length per batch constant?

An engineer is profiling a Transformer-based model and observes low GPU utilization despite high batch sizes. The bottleneck analysis identifies frequent host-to-device memory copies. Which optimization strategy most effectively addresses this latency?

Refer to the exhibit. An inference engine is failing during peak load. What is the most likely cause, and which strategy should be prioritized for immediate resolution?

Exhibit

Which THREE of the following are valid methods to optimize NVIDIA GPU memory bandwidth usage in deep learning?

When deploying a large language model on NVIDIA H100 GPUs using TensorRT-LLM, which TWO configuration strategies are most effective for improving KV cache efficiency and memory utilization?

In the context of NVIDIA Tensor Cores, what is the primary benefit of using BF16 (Bfloat16) over FP16 during model training and inference?

Refer to the exhibit. An engineer receives this timeout error during a CUDA kernel execution. What is the most appropriate first step to diagnose the resource contention?

Exhibit

Which hardware component of an NVIDIA GPU is most responsible for accelerating matrix-multiply-accumulate (MMA) operations used in transformer layers?

When profiling an application with NVIDIA Nsight Systems, which TWO metrics are most critical to identify if an application is limited by the PCIe bus?

Which of the following describes the purpose of 'Kernel Fusion' in the context of optimizing a Deep Learning inference pipeline?

Why is 'Pinned Memory' (page-locked) essential for high-performance data transfers between host and GPU?

When using the NVIDIA Collective Communications Library (NCCL), what does 'AllReduce' specifically optimize for in a multi-GPU training configuration?

What is the primary function of the 'TensorRT' optimization engine in the NVIDIA AI software stack?

Which TWO of the following techniques are best suited for reducing the latency of LLM inference on NVIDIA GPUs?

When deploying a model using NVIDIA TensorRT, what is the primary benefit of the 'Engine Building' phase?

Refer to the exhibit. The model performance is inconsistent. What is the most likely reason for the performance variability under load?

Exhibit

When profiling an application with NVIDIA Nsight Systems, you notice a long gap between kernel execution blocks on the GPU timeline. What is the most likely cause?

You are optimizing a ResNet-50 model on an NVIDIA A100. Which precision-based optimization will yield the highest throughput without significant accuracy loss?

Which NVIDIA library is primarily used for optimizing and deploying deep learning inference models?

A developer is optimizing a generative AI model for inference on NVIDIA GPUs. They want to reduce memory footprint and improve throughput without sacrificing accuracy. Which two techniques should they apply? (Choose two.)

An engineer is training a large language model with pipeline parallelism across four NVIDIA GPUs. They observe that GPU utilization is low and training throughput is limited by idle time during pipeline bubbles. Which technique is most effective to reduce pipeline bubbles and improve utilization?

An engineer is optimizing a large language model for inference on NVIDIA GPUs and wants to reduce memory usage to fit a larger model or increase batch size. Which two techniques are most effective for reducing GPU memory consumption during inference? (Choose two.)

An engineer is optimizing a large language model for inference on NVIDIA GPUs using TensorRT-LLM. They want to reduce the memory footprint of the KV cache to support longer context lengths and more concurrent requests. Which two techniques should they implement? (Choose two.)

An engineer is profiling a CUDA kernel and notices that the achieved occupancy is low, leading to underutilization of the GPU. The kernel uses a large number of registers per thread, limiting the number of resident warps. Which optimization should be attempted first to improve occupancy?

A team is optimizing a large language model for inference on NVIDIA GPUs. They want to reduce the memory footprint of the model to fit on a single GPU with limited VRAM. Which two techniques are most appropriate for reducing memory usage during inference? (Choose two.)

NCP-GENL GPU Acceleration and Optimization — All Questions With Answers

An enterprise developer is optimizing a computer vision pipeline on NVIDIA A100 GPUs using TensorRT. They want to maximize throughput for high-resolution image batches while maintaining acceptable latency bounds. Which execution mode should they configure in the TensorRT builder configuration?

Refer to the exhibit. A deep learning engineer encounters a CUDA out-of-memory error while initializing a deep neural network training job on a partitioned NVIDIA H100 GPU instance. Based on the error log, what is the most appropriate remediation step?

Exhibit

An inference server is experiencing high GPU latency when using FP16 precision. Profiling reveals that the kernel execution time is dominated by memory bandwidth bottlenecks. Which optimization technique is most likely to mitigate this latency while maintaining model throughput?

Refer to the exhibit. An engineer observes the provided output from a GPU currently running an inference task. What does this indicate regarding the current workload state?

Which CUDA-aware communication primitive is specifically designed to optimize data movement between multiple GPUs in a multi-node cluster by bypassing the CPU host memory?

Which THREE techniques are primary methods for reducing the footprint of an LLM on GPU VRAM to enable larger models on restricted hardware?

In an H100 environment, which factor is most significant when choosing between TensorRT-LLM and standard PyTorch for high-throughput inference deployment?

Which technique provides the most significant performance gain for large-model inference by splitting the model across multiple GPUs while keeping the sequence length per batch constant?

An engineer is profiling a Transformer-based model and observes low GPU utilization despite high batch sizes. The bottleneck analysis identifies frequent host-to-device memory copies. Which optimization strategy most effectively addresses this latency?

Refer to the exhibit. An inference engine is failing during peak load. What is the most likely cause, and which strategy should be prioritized for immediate resolution?

Exhibit

Which THREE of the following are valid methods to optimize NVIDIA GPU memory bandwidth usage in deep learning?

When deploying a large language model on NVIDIA H100 GPUs using TensorRT-LLM, which TWO configuration strategies are most effective for improving KV cache efficiency and memory utilization?

In the context of NVIDIA Tensor Cores, what is the primary benefit of using BF16 (Bfloat16) over FP16 during model training and inference?

Refer to the exhibit. An engineer receives this timeout error during a CUDA kernel execution. What is the most appropriate first step to diagnose the resource contention?

Exhibit

Which hardware component of an NVIDIA GPU is most responsible for accelerating matrix-multiply-accumulate (MMA) operations used in transformer layers?

When profiling an application with NVIDIA Nsight Systems, which TWO metrics are most critical to identify if an application is limited by the PCIe bus?

Which of the following describes the purpose of 'Kernel Fusion' in the context of optimizing a Deep Learning inference pipeline?

Why is 'Pinned Memory' (page-locked) essential for high-performance data transfers between host and GPU?

When using the NVIDIA Collective Communications Library (NCCL), what does 'AllReduce' specifically optimize for in a multi-GPU training configuration?

What is the primary function of the 'TensorRT' optimization engine in the NVIDIA AI software stack?

Which TWO of the following techniques are best suited for reducing the latency of LLM inference on NVIDIA GPUs?

When deploying a model using NVIDIA TensorRT, what is the primary benefit of the 'Engine Building' phase?

Refer to the exhibit. The model performance is inconsistent. What is the most likely reason for the performance variability under load?

Exhibit

When profiling an application with NVIDIA Nsight Systems, you notice a long gap between kernel execution blocks on the GPU timeline. What is the most likely cause?

You are optimizing a ResNet-50 model on an NVIDIA A100. Which precision-based optimization will yield the highest throughput without significant accuracy loss?

Which NVIDIA library is primarily used for optimizing and deploying deep learning inference models?

A developer is optimizing a generative AI model for inference on NVIDIA GPUs. They want to reduce memory footprint and improve throughput without sacrificing accuracy. Which two techniques should they apply? (Choose two.)

An engineer is training a large language model with pipeline parallelism across four NVIDIA GPUs. They observe that GPU utilization is low and training throughput is limited by idle time during pipeline bubbles. Which technique is most effective to reduce pipeline bubbles and improve utilization?

An engineer is optimizing a large language model for inference on NVIDIA GPUs and wants to reduce memory usage to fit a larger model or increase batch size. Which two techniques are most effective for reducing GPU memory consumption during inference? (Choose two.)

An engineer is optimizing a large language model for inference on NVIDIA GPUs using TensorRT-LLM. They want to reduce the memory footprint of the KV cache to support longer context lengths and more concurrent requests. Which two techniques should they implement? (Choose two.)

An engineer is profiling a CUDA kernel and notices that the achieved occupancy is low, leading to underutilization of the GPU. The kernel uses a large number of registers per thread, limiting the number of resident warps. Which optimization should be attempted first to improve occupancy?

A team is optimizing a large language model for inference on NVIDIA GPUs. They want to reduce the memory footprint of the model to fit on a single GPU with limited VRAM. Which two techniques are most appropriate for reducing memory usage during inference? (Choose two.)