A DevOps Engineer is deploying a massive parameter-efficient fine-tuning (PEFT) pipeline using the NeMo Framework. The goal is to apply Low-Rank Adaptation (LoRA) to a specialized Encoder-Decoder model (like T5) for translating complex medical terminology.
```
# NeMo PEFT LoRA Deployment Configuration
model_architecture: "encoder-decoder"
task: "medical_translation"
peft:
method: "lora"
target_modules:
-
"self_attention_q_proj"
-
"self_attention_v_proj"
-
"cross_attention_q_proj"
-
"cross_attention_v_proj"
```
Comparing this architecture to applying LoRA on a standard Decoder-only model (like Llama-3), which of the following statements accurately reflect the architectural realities and optimization strategies? (Select all that apply.)
A. Encoder-Decoder models include additional attention mechanisms beyond Decoder-only counterparts, notably cross-attention layers. Consequently, applying LoRA across all attention modules increases the VRAM footprint of adapter weights relative to adapting a comparable Decoder-only model.
B. Unlike Decoder-only models that possess only self-attention mechanisms, adapting the cross-attention projection matrices in an Encoder-Decoder architecture is frequently critical for sequence-to-sequence tasks because it governs how the generated output sequence specifically aligns with the source input sequence during decoding.
C. Within the NVIDIA NeMo Framework, applying LoRA to both Query (Q) and Value (V) projection matrices across all self-attention and cross-attention blocks is a well-established, framework-endorsed best practice explicitly engineered to maximize domain adaptation for complex tasks like medical translation while preserving parameter efficiency.
D. In the context of deploying a medical translation pipeline with an Encoder-Decoder architecture such as T5, applying LoRA adapters exclusively to the encoder's self-attention layers ensures the decoder parameters remain completely frozen throughout fine-tuning, a condition proponents claim delivers the fastest achievable inference speed for translation workloads.
E. In Encoder-Decoder models used for medical translation, LoRA adapters must be strictly confined to positional encoding vectors only, as any modification to attention weight matrices would compromise the precise sequence alignment essential for accurately translating specialized medical terminology.
