Repository navigation
Commit e25f129
authored
Arm backend: add SmolLM2 Ethos-U KV-cache demo support (pytorch#21940)
Extend the SmolLM2 Ethos-U helper scripts for KV-cache exports,
generation, and perplexity evaluation. Keep the non-KV seq32 full-logits
path as the documented baseline.
Add static quantized KV-cache export using fixed per-channel scales
calibrated with Wikitext. Document the export, runner, generation, and
perplexity flows for comparing plain and static-quantized caches.
cc @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils
@Sebastian-Larsson @robell @rascani
Signed-off-by: Xingguo Li <xingguo.li@arm.com>1 parent 2d6f56b commit e25f129
5 files changed
Lines changed: 599 additions & 152 deletions
File tree
- examples/arm/smollm2_example_ethos_u
0 commit comments