Run 35B MoE models in ~2.5 GB of RAM. Weights stream from SSD; a trained router predicts loads a step ahead. Mac, iPhone, Android, Windows supported
Python
windows: parse the ggml_vulkan device line; clarify peak_working_set semantics
Arda M. Ekized627fcf080e
diff --git a/windows/README.md b/windows/README.mdindex e00ac25..29161a8 100644--- a/windows/README.md+++ b/windows/README.md@@ -146,9 +146,9 @@ JSON line for aggregation, and `benchmarks/r3/bench_<tier>_submission.json`stores the per-rep detail.On the memory field: `peak_active_gib` on Windows is the engine process's-peak working set during the timed reps (OS-level high-water mark) — the same-intent as the Python bench's `mlx.core.get_peak_memory()`, measured at the-process level instead of the allocator.+peak working set during the timed reps (OS-level high-water mark). Windows+has no allocator-level counter, so this is not directly comparable with the+python-mlx column's `mlx.core.get_peak_memory()`.---diff --git a/windows/tools/r3_bench.py b/windows/tools/r3_bench.pyindex ffa1b1f..f5a5b23 100644--- a/windows/tools/r3_bench.py+++ b/windows/tools/r3_bench.py@@ -140,15 +140,17 @@ def engine_commit():def vulkan_device(err):- """Vulkan device name reported by the engine's own startup log."""- m = re.search(r"Device\s*\d*\s*[:=]\s*(.+?)\s*(?:\(|\n|$)", err, re.I)- if m and re.search(r"vulkan", err[:m.start()][-400:], re.I):- return m.group(1).strip()+ """Vulkan device name from the engine's own startup log, e.g:+ 'ggml_vulkan: 0 = AMD Radeon RX 9070 GRE (vendor) | uma: ...'."""for line in err.splitlines():- if re.search(r"vulkan", line, re.I):- m2 = re.search(r"(?:device|adapter|gpu)[^\n:]*[:=]\s*(.+)", line, re.I)- if m2:- return m2.group(1).strip()+ if not re.search(r"vulkan", line, re.I):+ continue+ m = re.search(r"\d+\s*=\s*([^\n|()]+)", line)+ if m:+ return m.group(1).strip()+ m = re.search(r"Device\s*\d*\s*[:=]\s*([^\n|(),]+)", line, re.I)+ if m:+ return m.group(1).strip()return None