Skip to content

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/coli
index 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.gpu
if gpu is None:
gpu=env.get("COLI_GPUS",env.get("COLI_GPU","auto"))