randomllama commited on
Commit
0cee402
·
verified ·
1 Parent(s): e91cfbc

sync: latest measured numbers

Browse files
Files changed (1) hide show
  1. LADDER.md +1 -1
LADDER.md CHANGED
@@ -87,5 +87,5 @@ control fails at 91%; fp8 GEMMs lower and run).
87
  | temp-0 vs bf16, 5 prompts | **4/5 exact** |
88
  | 32K-depth codeword probe | PASS |
89
  | TTFT@16k warm | **6.6 s vs 7.9 s bf16 — 17% faster prefill** |
90
- | open item | raw-layout memory accounting resolved max_running_requests to 1 — budget retune WIP |
91
  Kernel analysis + port: this rig (Fable agent), validated per the house protocol.
 
87
  | temp-0 vs bf16, 5 prompts | **4/5 exact** |
88
  | 32K-depth codeword probe | PASS |
89
  | TTFT@16k warm | **6.6 s vs 7.9 s bf16 — 17% faster prefill** |
90
+ | open item | raw-layout accounting resolves max_running_requests to 1; retune attempt (max-total 393216 + mamba-ratio 3) held at 1 and shrank the pool the constraint is the mamba slice, not the token budget. Documented as PR known-limitation; fp8 = interactive config, bf16-T8 = fleet config |
91
  Kernel analysis + port: this rig (Fable agent), validated per the house protocol.