devflow / analysis_outputs /T32 /task1_kv_cache.txt
bhsinghgrid's picture
Update app/inference + ablation task outputs
27f26fd verified
TASK 1 — KV CACHE BENCHMARK
========================================
has_generate_cached=True
memory_profile=N/A (CPU/MPS)
src_len standard(s) cached(s) speedup encoder%
16 1.844 1.084 1.70x 41.2%
32 1.580 1.133 1.39x 39.9%
64 2.386 1.939 1.23x 39.7%
Saved graphs:
- task1_time_comparison.png
- task1_speedup.png
- task1_encoder_cost.png