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: resolve static analysis bugs in coli, autotune, family_registry, and convert_qwen36

Surajbe958c17e25f
diff --git a/c/autotune.py b/c/autotune.py
index b8f64e8b..31d853f7 100644
--- a/c/autotune.py
+++ b/c/autotune.py
@@ -466,6 +466,16 @@ def run_tune(engine: str, cap: int, base_env: dict, plan: dict, model: str,
if proc.returncode:
raise RuntimeError(f"{name} failed ({proc.returncode})\n{output[-2000:]}")
return parse_replay(output)
+
+ def measure(name, overlay, launch_cap):
+ samples = []
+ for repeat in range(repeats):
+ progress(f"{name} cap={cap} ({repeat + 1}/{repeats})")
+ sample = run_once(name, overlay, launch_cap)
+ sample["ttft_s"] = None
+ samples.append(sample)
+ recorded_cap = launch_cap if arch in CAP_ARCHES else None
+ return _summarize_measurement(name, overlay, samples, recorded_cap)
else:
if engine_cls is None:
from openai_server import Engine as engine_cls
@@ -533,17 +543,6 @@ def run_tune(engine: str, cap: int, base_env: dict, plan: dict, model: str,
recorded_cap = launch_cap if arch in CAP_ARCHES else None
return _summarize_measurement(name, overlay, samples, recorded_cap)
- if arch == "glm":
- def measure(name, overlay, launch_cap):
- samples = []
- for repeat in range(repeats):
- progress(f"{name} cap={cap} ({repeat + 1}/{repeats})")
- sample = run_once(name, overlay, launch_cap)
- sample["ttft_s"] = None
- samples.append(sample)
- recorded_cap = launch_cap if arch in CAP_ARCHES else None
- return _summarize_measurement(name, overlay, samples, recorded_cap)
-
baseline = measure("baseline", {}, cap)
winner = baseline
accumulated = {}
diff --git a/c/coli b/c/coli
index bdc1072e..983d0796 100755
--- a/c/coli
+++ b/c/coli
@@ -32,9 +32,9 @@ import os, sys, subprocess, argparse, json, time, signal, shutil, threading, re,
# and the console provides its own editing; on FreeBSD Python links libedit,
# which this activates the same way.
try:
- import readline # noqa: F401 — importing it is the activation
+ import readline
except ImportError:
- pass
+ readline = None
# The engine mmaps every shard (144+ files); macOS default RLIMIT_NOFILE is 256.
if sys.platform != "win32":
@@ -1191,7 +1191,7 @@ def cmd_info(a):
try: n_shards=len([x for x in os.listdir(a.model) if x.endswith('.safetensors')])
except OSError: n_shards=0
print(f" {C.yel}config.json is missing{C.r}: coli picks the engine from it, so nothing can run here yet.")
- print(f" Copy the checkpoint's config.json (with tokenizer.json and model.safetensors.index.json)")
+ print(" Copy the checkpoint's config.json (with tokenizer.json and model.safetensors.index.json)")
print(f" from the model repo next to the {n_shards} shard(s) found here, then run coli info again.")
try:
mi=open('/proc/meminfo').read()
@@ -1511,9 +1511,7 @@ def chat_commands_help():
def install_chat_completer():
"""TAB completa i comandi. Senza readline (Windows) si perde solo il TAB."""
- try:
- import readline
- except ImportError:
+ if readline is None:
return
names = [p + n for n in CHAT_COMMANDS for p in ("/", ":")]
def complete(text, state):
diff --git a/c/family_registry.py b/c/family_registry.py
index 7b3c2e36..ddd8ed31 100644
--- a/c/family_registry.py
+++ b/c/family_registry.py
@@ -702,7 +702,7 @@ def _dsv4_geometry(config, context, _model_dir):
"""
layers = _required_int(config, "num_hidden_layers", "deepseek_v4")
experts = _required_int(config, "n_routed_experts", "deepseek_v4")
- hidden = _required_int(config, "hidden_size", "deepseek_v4")
+ _required_int(config, "hidden_size", "deepseek_v4")
heads = _required_int(config, "num_attention_heads", "deepseek_v4")
head_dim = _required_int(config, "head_dim", "deepseek_v4")
q_rank = _required_int(config, "q_lora_rank", "deepseek_v4")
diff --git a/c/tools/convert_qwen36.py b/c/tools/convert_qwen36.py
index 7ec413f9..0cf3e07a 100644
--- a/c/tools/convert_qwen36.py
+++ b/c/tools/convert_qwen36.py
@@ -39,7 +39,7 @@ Usage (local tiny model for end-to-end testing):
python tools/convert_qwen36.py --model ../qwen36_tiny --out ../qwen36_tiny_i4 --ebits 8
"""
-import argparse, json, math, os, struct, sys
+import argparse, json, math, os, struct, sys, tempfile
from pathlib import Path
# Windows: force UTF-8 output
@@ -54,7 +54,7 @@ try:
import torch
from safetensors.numpy import safe_open as safe_open_np
from safetensors.torch import safe_open as safe_open_pt
- from safetensors.torch import save_file
+ from safetensors.torch import save_file, load_file
except ImportError as exc:
sys.exit(f"Missing dependencies: {exc}. Install: pip install torch safetensors")
@@ -149,7 +149,6 @@ def _unpack_int4(packed: "torch.Tensor") -> "torch.Tensor":
def _selftest():
"""Round-trip check for the true-int4 packing used by --ebits<=4."""
- import random
print("=== int4 pack/unpack selftest ===")
ok = True
for ebits in (2, 3, 4):
@@ -160,9 +159,7 @@ def _selftest():
d = torch.randn(H, inter) * 3
mw, qs = make_merged(g, u, d, ebits)
# serialize to safetensors + reload (exercises the real dtype path)
- import tempfile, os
td = tempfile.mkdtemp()
- from safetensors.torch import save_file, load_file
save_file({"merged_weight": mw, "qs": qs}, os.path.join(td, "e0.safetensors"))
back = load_file(os.path.join(td, "e0.safetensors"))
mw_r = back["merged_weight"]