ulysses/services/hwfit
darius-f96 d6d2e17214
fix(hwfit): add GB10 unified-memory bandwidth so speed scores are real (#4270)
NVIDIA Grace Blackwell GB10 / DGX Spark was missing from GPU_BANDWIDTH, so
_lookup_bandwidth() returned None for it and _estimate_speed() fell through
to the crude FALLBACK_K path (k/active-params). That over-stated tok/s and
let speed scores saturate regardless of the box's real ~273 GB/s LPDDR5X
pool — distorting model ranking on these 128GB unified-memory rigs.

Add "gb10": 273 (GB/s). nvidia-smi reports the device name as "NVIDIA GB10",
which substring-matches the new key, so detected GB10 boxes now estimate
speed from the real bandwidth instead of the fallback.
2026-06-15 18:55:15 +02:00
..
data Cookbook UI: Ollama browser, advanced serve fold, API tokens form, diagnosis toolbar, polish 2026-06-09 09:46:19 +09:00
__init__.py Odysseus v1.0 2026-05-31 23:58:26 +09:00
fit.py fix(hwfit): add GB10 unified-memory bandwidth so speed scores are real (#4270) 2026-06-15 18:55:15 +02:00
hardware.py fix(hwfit): distinguish Apple Silicon bandwidth variants (#2564) 2026-06-15 15:13:03 +02:00
image_models.py fix: image model ranking crashes on a non-string search filter (#1898) 2026-06-04 03:26:35 +01:00
models.py Cookbook polish: auto-reconnect, ctx slider fixes, scoring, lots of UI 2026-06-03 20:25:25 +09:00
profiles.py fix(hwfit): serve profiles for sub-8192 context models 2026-06-15 15:02:22 +09:00