Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
config: fix UTF-8 decode crash + null config-value TypeError on Windows (doctor/resource_plan) (#184)
woolcoxm35ed9bc92193
diff --git a/c/doctor.py b/c/doctor.pyindex da647040..4cb5a42e 100644--- a/c/doctor.py+++ b/c/doctor.py@@ -48,7 +48,7 @@ def run_doctor(model, ram_gb=0, context=4096, gpu_indices=None, vram_gb=0, *,config = model / "config.json"try:- valid_config = isinstance(json.loads(config.read_text()), dict)+ valid_config = isinstance(json.loads(config.read_text(encoding="utf-8")), dict)except (OSError, ValueError):valid_config = Falsechecks.append(_check("model.config", "pass" if valid_config else "fail",diff --git a/c/resource_plan.py b/c/resource_plan.pyindex 33674e19..2fe4c3cf 100644--- a/c/resource_plan.py+++ b/c/resource_plan.py@@ -39,7 +39,7 @@ def analyze_model(model):config_path = model / "config.json"if not config_path.is_file():raise ValueError(f"missing config.json: {model}")- config = json.loads(config_path.read_text())+ config = json.loads(config_path.read_text(encoding="utf-8"))shards = sorted(model.glob("*.safetensors"))if not shards:raise ValueError(f"no safetensors shards: {model}")@@ -210,15 +210,15 @@ def build_plan(model, ram_gb=0, context=4096, gpu_indices=None, vram_gb=0,if ram_budget < 4 * GB:ram_budget = 8 * GBtypical = info["typical_expert_bytes"]- layers = int(cfg.get("num_hidden_layers", 0)) + 1- kv_bytes = layers * context * (int(cfg.get("kv_lora_rank", 0)) +- int(cfg.get("qk_rope_head_dim", 0))) * 4- kv_buffer = context * int(cfg.get("num_attention_heads", 0)) * (- int(cfg.get("qk_nope_head_dim", 0)) + int(cfg.get("v_head_dim", 0))) * 4+ layers = int(cfg.get("num_hidden_layers") or 0) + 1+ kv_bytes = layers * context * (int(cfg.get("kv_lora_rank") or 0) ++ int(cfg.get("qk_rope_head_dim") or 0)) * 4+ kv_buffer = context * int(cfg.get("num_attention_heads") or 0) * (+ int(cfg.get("qk_nope_head_dim") or 0) + int(cfg.get("v_head_dim") or 0)) * 4runtime_bytes = int(1.2 * GB + 2.5 * GB + 64 * typical + kv_bytes + kv_buffer)cache_bytes = max(0, ram_budget - info["dense_bytes"] - runtime_bytes)per_cap = info["per_cap_bytes"]- configured_experts = int(cfg.get("n_routed_experts", 0))+ configured_experts = int(cfg.get("n_routed_experts") or 0)cap = int(cache_bytes // per_cap) if per_cap else 0if configured_experts:cap = min(cap, configured_experts)