Text Generation
MLX
English
German
next-concept-modeling
apple-silicon
diffusion-lm
research-code
negative-results
Instructions to use Red-fuchs/ncm-modular with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use Red-fuchs/ncm-modular with MLX:
# Make sure mlx-lm is installed # pip install --upgrade mlx-lm # if on a CUDA device, also pip install mlx[cuda] # Generate text with mlx-lm from mlx_lm import load, generate model, tokenizer = load("Red-fuchs/ncm-modular") prompt = "Once upon a time in" text = generate(model, tokenizer, prompt=prompt, verbose=True) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- MLX LM
How to use Red-fuchs/ncm-modular with MLX LM:
Generate or start a chat session
# Install MLX LM uv tool install mlx-lm # Generate some text mlx_lm.generate --model "Red-fuchs/ncm-modular" --prompt "Once upon a time"
- Atomic Chat
File size: 1,588 Bytes
b5731a6 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 | """Compare MLX codec vs torch dump."""
import sys, os, numpy as np, mlx.core as mx
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
from paths import WORK
import codec_mlx
mx.set_default_device(mx.cpu)
task = sys.argv[1]
R = np.load(f"{WORK}/ref_{task}.npz")
for dt, nm in [(mx.float32, "fp32"), (mx.float64, "fp64")]:
c = codec_mlx.load(f"{WORK}/mlx/{task}", dt)
x = mx.array(R["ids"])
z = c.embed(x); zq = z
codes = c.encode(x); zh = c.cumulative(codes)
mx.eval(codes, zh)
# pre-quant latent
print(f"--- {task} {nm}: pre-quant z max|diff| {np.abs(np.array(z) - R['z']).max():.2e}")
off = 0
for k, cd in enumerate(codes):
n = c.levels[k]
ref = R["idx"][:, off:off + n]; off += n
eq = (np.array(cd) == ref)
print(f" level {k:2d} (len {n:2d}) codes equal {eq.mean()*100:6.2f}% zhat max|diff| {np.abs(np.array(zh[k]) - R['zh'][k]).max():.2e}")
dec = np.stack([np.array(mx.argmax(c.decode_logits(h), -1)) for h in zh])
print(f" decoded ids per level equal: " + " ".join(f"{(dec[k]==R['dec'][k]).mean()*100:.1f}" for k in range(len(zh))))
print(f" final decoded ids equal: {(dec[-1]==R['dec_full']).mean()*100:.2f}%")
# decode-only parity: feed torch codes into MLX decode
cs, off = [], 0
for n in c.levels: cs.append(mx.array(R["idx"][:, off:off + n])); off += n
dd = np.array(c.decode_codes(cs))
zz = c.cumulative(cs)[-1]
print(f" decode-from-torch-codes: ids equal {(dd==R['dec_full']).mean()*100:.2f}% zhat max|diff| {np.abs(np.array(zz)-R['zh'][-1]).max():.2e}")
|