Merge pull request #232 from nbeerbower/profiling-upstream
Profiling: PROF=1 opt-in performance profile + live per-turn Profiling page in the web dashboard
This commit is contained in:
@@ -63,6 +63,7 @@ Format: `VAR` — default — effect.
|
||||
| `ABSORB` | `-1` (auto: absorbed for S≤4) | MLA attention absorption mode. |
|
||||
| `IDOT` | `1` | Integer dot-product kernel. `IDOT=0` uses exact f32 kernels (for A/B numerical checks). |
|
||||
| `COLI_POLICY` | `quality` | Resource policy: `quality`, `balanced`, or `experimental-fast`. |
|
||||
| `PROF` | `0` (off) | Performance profile: a startup header (machine + effective config), then per run — or per turn in serve mode, on stderr — forward-latency percentiles (p50/p90/p99/max), expert-I/O totals and cache-tier fill, phase shares of wall time, and a verdict naming the knob most likely to help on this machine. Output is additive; `PROF` unset changes nothing. |
|
||||
|
||||
---
|
||||
|
||||
|
||||
Reference in New Issue
Block a user