Clarify which EPQ version (v1/v2) is active for Flint
Browse files- results/test-status.md +6 -0
results/test-status.md
CHANGED
|
@@ -29,6 +29,12 @@ llama.cpp base: `970b51ce7c9ba5fff0726c5e9994f1c2f0e4eedc` plus `flint-epq.patch
|
|
| 29 |
5. **`tools/runmem_cgroup.py` and the cgroup claim were presented as usable.** Now labelled untested.
|
| 30 |
6. **Earlier tier throughput was from single runs** and varied by ~±15 % between runs. Replaced with medians of 3 runs.
|
| 31 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 32 |
## Known limits and open items
|
| 33 |
|
| 34 |
- EPQ output is not bit-identical to native llama.cpp. Differences come from float summation order and are amplified by 8-bit activation quantization; native llama.cpp itself shows the same magnitude of KLD under a ubatch change (control above). No functional mismatch was found.
|
|
|
|
| 29 |
5. **`tools/runmem_cgroup.py` and the cgroup claim were presented as usable.** Now labelled untested.
|
| 30 |
6. **Earlier tier throughput was from single runs** and varied by ~±15 % between runs. Replaced with medians of 3 runs.
|
| 31 |
|
| 32 |
+
## EPQ version coverage
|
| 33 |
+
|
| 34 |
+
- v1 (expert paging, mode C): active in all runs.
|
| 35 |
+
- v2 preload: code present, requires an arena of at least 28.6 GiB; never triggered in these tests (the 24 GiB run had 10,082 slots, below the 12,032 needed).
|
| 36 |
+
- v2 row-parallel: Nemotron no-gate only; inactive for Flint.
|
| 37 |
+
|
| 38 |
## Known limits and open items
|
| 39 |
|
| 40 |
- EPQ output is not bit-identical to native llama.cpp. Differences come from float summation order and are amplified by 8-bit activation quantization; native llama.cpp itself shows the same magnitude of KLD under a ubatch change (control above). No functional mismatch was found.
|