c111b3a9b0
Phase 01 (LLM Tuning): - Gemma4 26B: 74.65 t/s (fast) - Qwen 35B: 61.62 t/s (balanced) - Gemma4 31B: 16.0 t/s (deep-coder) - Qwen 27B: 16.7 t/s (deep-logic) - Qwen 122B: 8.95 t/s (ultra, GPU 1 only) Phase 02 (API Engine): - FastAPI reverse proxy on port 8000 - /engine/switch hot-swap with 503 protection - config/engine_models.json as single source of truth - Replaced 4 individual .bat files with unified engine File cleanup: - scripts/ 85 files -> 9 + _archive/ - Root .bat files -> _archive/
142 lines
4.1 KiB
Python
142 lines
4.1 KiB
Python
"""
|
|
Qwen 3.5 35B-A3B Q4_K_M - Tensor Split 0.42/0.58 Speed Test
|
|
64 t/s 달성 설정 기반, 스플릿 비율만 변경
|
|
"""
|
|
import subprocess, time, json, urllib.request, sys, os
|
|
|
|
PYTHON = sys.executable
|
|
LLAMA = os.path.join(os.getcwd(), "llama_bin_run", "llama-server.exe")
|
|
MODEL = os.path.join(os.getcwd(), "models", "Qwen3.5-35B-A3B-Q4_K_M.gguf")
|
|
TS = "0.55,0.45"
|
|
|
|
# 1. Kill any existing server
|
|
print("[1/4] Killing existing llama-server...")
|
|
subprocess.run(["taskkill", "/F", "/IM", "llama-server.exe"],
|
|
stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
|
|
time.sleep(3)
|
|
|
|
# 2. Start server with 64t/s config + custom split
|
|
args = [
|
|
LLAMA, "--model", MODEL,
|
|
"-ngl", "999",
|
|
"-c", "262144",
|
|
"-np", "1",
|
|
"-fa", "on",
|
|
"--cache-type-k", "q4_0",
|
|
"--cache-type-v", "q4_0",
|
|
"-ub", "128",
|
|
"-b", "512",
|
|
"-t", "6",
|
|
"-tb", "6",
|
|
"--prio", "3",
|
|
"--mlock",
|
|
"--poll", "50",
|
|
"--port", "8000",
|
|
"--host", "0.0.0.0",
|
|
"-ts", TS,
|
|
]
|
|
print(f"[2/4] Starting server with -ts {TS}")
|
|
print(f" CMD: {' '.join(args[-6:])}")
|
|
server = subprocess.Popen(args, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
|
|
|
|
# 3. Wait for health
|
|
print("[3/4] Waiting for server to become healthy...")
|
|
t_boot = time.time()
|
|
healthy = False
|
|
for sec in range(180): # max 3 min
|
|
time.sleep(1)
|
|
# Check if process crashed
|
|
if server.poll() is not None:
|
|
print(f" !! Server process CRASHED (exit code {server.returncode})")
|
|
sys.exit(1)
|
|
try:
|
|
r = urllib.request.urlopen("http://127.0.0.1:8000/health", timeout=1)
|
|
body = json.loads(r.read())
|
|
if body.get("status") == "ok":
|
|
healthy = True
|
|
break
|
|
except Exception:
|
|
pass
|
|
if sec % 10 == 9:
|
|
print(f" ... {sec+1}s elapsed")
|
|
|
|
if not healthy:
|
|
print(f" FAIL: Server not healthy after 180 seconds")
|
|
server.kill()
|
|
sys.exit(1)
|
|
|
|
boot_secs = time.time() - t_boot
|
|
print(f" OK: Booted in {boot_secs:.1f}s")
|
|
|
|
# VRAM check
|
|
try:
|
|
v = subprocess.run(
|
|
["nvidia-smi", "--query-gpu=index,memory.used,memory.total",
|
|
"--format=csv,noheader,nounits"],
|
|
capture_output=True, text=True, timeout=5)
|
|
print(f" VRAM: {v.stdout.strip()}")
|
|
except:
|
|
pass
|
|
|
|
# 4. Benchmark
|
|
print("[4/4] Running token speed benchmark (200 tokens)...")
|
|
|
|
def do_bench(max_tok):
|
|
payload = json.dumps({
|
|
"messages": [{"role": "user", "content": "Count from 1 to 50, each number on a new line."}],
|
|
"max_tokens": max_tok,
|
|
"temperature": 0
|
|
}).encode("utf-8")
|
|
req = urllib.request.Request(
|
|
"http://127.0.0.1:8000/v1/chat/completions",
|
|
data=payload,
|
|
headers={"Content-Type": "application/json"})
|
|
t0 = time.time()
|
|
with urllib.request.urlopen(req, timeout=120) as resp:
|
|
result = json.loads(resp.read())
|
|
elapsed = time.time() - t0
|
|
ct = result["usage"]["completion_tokens"]
|
|
return ct / elapsed, ct, elapsed
|
|
|
|
# warmup
|
|
try:
|
|
do_bench(10)
|
|
except:
|
|
pass
|
|
|
|
# real runs - 5회
|
|
print("[4/4] Running 5x benchmark (200 tokens each)...")
|
|
results = []
|
|
for i in range(5):
|
|
tps, tokens, elapsed = do_bench(200)
|
|
results.append(tps)
|
|
# VRAM check per run
|
|
try:
|
|
v = subprocess.run(["nvidia-smi", "--query-gpu=index,memory.used,memory.total,memory.free", "--format=csv,noheader,nounits"], capture_output=True, text=True, timeout=5)
|
|
vram_info = v.stdout.strip()
|
|
except:
|
|
vram_info = "?"
|
|
print(f" Run {i+1}: {tps:.2f} t/s ({tokens} tok / {elapsed:.2f}s) | VRAM: {vram_info}")
|
|
|
|
avg = sum(results) / len(results)
|
|
best = max(results)
|
|
worst = min(results)
|
|
summary = f"""
|
|
==================================================
|
|
TS={TS} 5-Run Results (with --mlock --poll 50):
|
|
AVG: {avg:.2f} t/s
|
|
BEST: {best:.2f} t/s
|
|
MIN: {worst:.2f} t/s
|
|
Runs: {[f'{r:.2f}' for r in results]}
|
|
==================================================
|
|
"""
|
|
print(summary)
|
|
with open("scripts/split_test_result.txt", "w") as f:
|
|
f.write(summary)
|
|
|
|
# cleanup
|
|
server.kill()
|
|
subprocess.run(["taskkill", "/F", "/IM", "llama-server.exe"],
|
|
stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
|
|
|