Training
MLA / GQA (architectural KV)
KV compressionReferenceGrouped-query (GQA/MQA) and low-rank latent KV (DeepSeek MLA) baked into the architecture.
- Relieves
- KV cacheBandwidth
- Targets
- KV cache · Structure
- Format
- Shared / low-rank KV heads
- Granularity
- Head grouping / latent KV
- Lifecycle
- Training-time · needs gradients
- Calibration
- Full corpus
- Compression
- Large KV reduction
- Quality
- Near-baseline (a design choice, not post-hoc)
- Hardware
- None — architectural
- Runtimes
- vLLMTensorRT-LLMSGLang
A worked Spark recipe for this method hasn't been written yet — it lives here as a reference point in the ontology.