Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
fix(coli): accept CUDA_EXPERT_GB=auto in the launcher's resource request
Mr. Metric5c9760767246
diff --git a/c/coli b/c/coliindex 0f331bde..fdb7f7b2 100755--- a/c/coli+++ b/c/coli@@ -183,8 +183,11 @@ def cuda_binary():def resource_request(a, env):ctx=a.ctx or int(env.get("CTX",4096))- ram=a.ram or float(env.get("RAM_GB",0))- vram=a.vram or float(env.get("CUDA_EXPERT_GB",0))+ def num(v):+ try: return float(v)+ except ValueError: return 0.0 # "auto": the planner sizes it; the engine reads the string itself+ ram=a.ram or num(env.get("RAM_GB",0))+ vram=a.vram or num(env.get("CUDA_EXPERT_GB",0))gpu=a.gpuif gpu is None:gpu=env.get("COLI_GPUS",env.get("COLI_GPU","auto"))