Ikibazo gihuza ibibazo · Hagati Urwego
Ikizamini cya Gradient
Irasobanura ibikorwa byo guhindura ibintu, kuzigama kwibuka, gahunda-ishingiye kuri compte hejuru, hamwe no guhuza hamwe no gutondeka no kuvanga neza.
Ikibazo 1 Bya 8
Ni ubuhe buryo bwo kugenzura buhoro buhoro gucuruza kugirango ubike ububiko?
Reba buri kibazo8 ibibazo
Ibisubizo nibisobanuro bigaragara mugihe ufata ikibazo.
- Ni ubuhe buryo bwo kugenzura buhoro buhoro gucuruza kugirango ubike ububiko?
- Kubara birenze (recomputation) kugirango ukoreshe kwibuka hasi
- Icyitegererezo cyukuri cyihuta
- Umwanya wa disiki yumurongo mugari
- Ubunini bwicyiciro
- Ni ukubera iki ibikorwa bisanzwe bibikwa mugihe cyimbere?
- Kugaragaza kubakoresha
- Kuberako backpropagation ibakeneye kubara gradients
- Kwihutisha gupakira amakuru
- Ntibikenewe rwose
- In a simplified chain of N similarly sized layers, what can checkpoint boundaries spaced about sqrt(N) layers apart achieve in the textbook schedule?
- It stores no activations and performs no recomputation
- It guarantees constant memory for every model graph
- Activation storage can scale on the order of sqrt(N) checkpoints
- It makes optimizer-state memory scale as sqrt(N)
- What most determines gradient-checkpointing runtime overhead in a practical model?
- Only the model’s file extension
- A fixed 20–30% slowdown for every model
- The number of labels in the dataset
- Which operations are recomputed, checkpoint placement, graph structure, and hardware
- Muri PyTorch, ni ubuhe butumwa bukoreshwa muburyo bwo kugenzura buhoro buhoro kuri module?
- itara.ibikoresho
- itara.optim.SGD
- itara.nn.Dropout
- itara.cuda.ubusa_cache
- What happens to operations inside a checkpointed region during backward?
- Their gradients are skipped entirely
- Their forward computations are rerun to regenerate needed activations
- Their weights are permanently discarded
- They are executed only during inference
- Which training techniques can be combined with activation checkpointing to reduce different memory costs?
- Only increasing the number of labels
- Turning off backpropagation
- FSDP for sharded model state and mixed precision for lower-precision tensors
- Replacing the model with a tokenizer
- What is selective activation checkpointing intended to choose?
- Which training labels are removed
- Which model outputs are published
- Which GPU becomes the tokenizer
- Which operations to save and which to recompute based on their memory and runtime costs