[vibecoded wip, do not merged yet] feat(inference): theoretical prefix tokens overlay + new input suffix throughput and TFLOP/s y-metrics / 理论 prefix token 悬浮框指标与新输入 suffix 吞吐量及 TFLOP/s Y 轴指标 - #849

Open

functionstackx wants to merge 13 commits into

feat/theoretical-prefix-tokens-overlay

Commits on Aug 25, 2026

feat(inference): add theoretical prefix and uncached input token rows to point overlay
functionstackx
committed
f92918e
feat(inference): add theoretical uncached throughput y-axis metrics
functionstackx
committed
1276c73
test(e2e): expect three y-axis options for input-token-throughput search
functionstackx
committed
5f37ca4
feat(inference): rename uncached throughput metrics to prompt suffix naming
functionstackx
committed
92488c5
feat(inference): rename metrics to new input suffix naming
functionstackx
committed
55ef345
fix(inference): require input throughput for suffix throughput metrics
functionstackx
committed
cc252f0
feat(inference): new input suffix + output TFLOP/s per chip y-metric
functionstackx
committed
9f749c8
fix(zh-copy): exempt TFLOP/s/chip as an English unit
functionstackx
committed
94de05e
feat(metrics): price per-model attention FLOPs into suffix+output TFLOP/s
functionstackx
committed
502c3b8
docs(api): document request_length_moments in derived-agentic-metrics
functionstackx
committed
6b62f9a
fix: stream oversized blobs in agentic fallbacks to survive Neon's 64 MB response cap
functionstackx
committed
3513940
chore: refresh agentic-aggregates shared-source digest after format hook
functionstackx
committed
f042265

Commits on Aug 26, 2026

fix: correct Llama GQA attention constant; never self-heal null server fields
functionstackx
committed
728011e