[MLX] Low speed on Qwen3.5-27B-4bit: 13.28 tok/s (M4 Pro 48GB)

[heykb]

python test_mlx.py

....

   import random
   # 在 partition 前加一句：
   arr[random.randint(low, high)], arr[high] = arr[high], arr[random.randint(low, high)]
2.

==================================================
Throughput: 13.28 tok/s
Download complete: : 0.00B [02:34, ?B/s]
a1-6@192 dflash % 

#!/usr/bin/env python3
"""DFlash MLX 测试脚本"""
from dflash.model_mlx import load, load_draft, stream_generate

# 加载主模型和草稿模型
print("Loading target model: mlx-community/Qwen3.6-27B-4bit...")
model, tokenizer = load("mlx-community/Qwen3.6-27B-4bit")

print("Loading draft model: z-lab/Qwen3.6-27B-DFlash...")
draft = load_draft("z-lab/Qwen3.6-27B-DFlash")

# 准备 prompt
messages = [{"role": "user", "content": "写个 Python 快排"}]
prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True
)

print(f"\nPrompt: {messages[0]['content']}")
print("=" * 50)

# 生成
tps = 0.0
print("\n生成结果:\n")
for r in stream_generate(
    model,
    draft,
    tokenizer,
    prompt,
    block_size=16,
    max_tokens=2048,
    temperature=0.6
):
    print(r.text, end="", flush=True)
    tps = r.generation_tps

print(f"\n\n{'=' * 50}")
print(f"Throughput: {tps:.2f} tok/s")

same Qwen3.6-27