Image-Text-to-Text
MLX
Safetensors
English
inkling2_loop
mlx-vlm
inkling2
Mixture of Experts
quantized
gptq
apple-silicon
tool-calling
agents
conversational
8-bit precision
Instructions to use VisualInference/Inkling-2-Mini-MLX-3bit with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use VisualInference/Inkling-2-Mini-MLX-3bit with MLX:
# Make sure mlx-vlm is installed # pip install --upgrade mlx-vlm from mlx_vlm import load, generate from mlx_vlm.prompt_utils import apply_chat_template from mlx_vlm.utils import load_config # Load the model model, processor = load("VisualInference/Inkling-2-Mini-MLX-3bit") config = load_config("VisualInference/Inkling-2-Mini-MLX-3bit") # Prepare input image = ["http://images.cocodataset.org/val2017/000000039769.jpg"] prompt = "Describe this image." # Apply chat template formatted_prompt = apply_chat_template( processor, config, prompt, num_images=1 ) # Generate output output = generate(model, processor, formatted_prompt, image) print(output) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- Pi
How to use VisualInference/Inkling-2-Mini-MLX-3bit with Pi:
Start the MLX server
# Install MLX LM: uv tool install mlx-lm # Start a local OpenAI-compatible server: mlx_lm.server --model "VisualInference/Inkling-2-Mini-MLX-3bit"
Configure the model in Pi
# Install Pi: npm install -g @earendil-works/pi-coding-agent # Add to ~/.pi/agent/models.json: { "providers": { "mlx-lm": { "baseUrl": "http://localhost:8080/v1", "api": "openai-completions", "apiKey": "none", "models": [ { "id": "VisualInference/Inkling-2-Mini-MLX-3bit" } ] } } }Run Pi
# Start Pi in your project directory: pi
- Hermes Agent
How to use VisualInference/Inkling-2-Mini-MLX-3bit with Hermes Agent:
Start the MLX server
# Install MLX LM: uv tool install mlx-lm # Start a local OpenAI-compatible server: mlx_lm.server --model "VisualInference/Inkling-2-Mini-MLX-3bit"
Configure Hermes
# Install Hermes: curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash hermes setup # Point Hermes at the local server: hermes config set model.provider custom hermes config set model.base_url http://127.0.0.1:8080/v1 hermes config set model.default VisualInference/Inkling-2-Mini-MLX-3bit
Run Hermes
hermes
- Atomic Chat
- OpenClaw
How to use VisualInference/Inkling-2-Mini-MLX-3bit with OpenClaw:
Start the MLX server
# Install MLX LM: uv tool install mlx-lm # Start a local OpenAI-compatible server: mlx_lm.server --model "VisualInference/Inkling-2-Mini-MLX-3bit"
Configure OpenClaw
# Install OpenClaw: npm install -g openclaw@latest # Register the local server and set it as the default model: openclaw onboard --non-interactive --mode local \ --auth-choice custom-api-key \ --custom-base-url http://127.0.0.1:8080/v1 \ --custom-model-id "VisualInference/Inkling-2-Mini-MLX-3bit" \ --custom-provider-id mlx-lm \ --custom-compatibility openai \ --custom-text-input \ --accept-risk \ --skip-health
Run OpenClaw
openclaw agent --local --agent main --message "Hello from Hugging Face"
File size: 5,799 Bytes
e0b70af | 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 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 | """Shared helpers for the Inkling 2 Mini quantization work (mlx-vlm layout)."""
import json
import math
import os
import mlx.core as mx
import mlx.nn as nn
import numpy as np
from mlx.utils import tree_map
from mlx_vlm import load
from mlx_vlm.models.switch_layers import QuantizedSwitchLinear
ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
NATIVE = os.path.join(ROOT, "upstream", "Inkling-2-Mini")
PROJECTIONS = ("gate_proj", "up_proj", "down_proj")
# ----------------------------------------------------------------------------- model access
def load_native(fp32=False, path=NATIVE):
model, processor = load(path)
if fp32:
to_fp32(model)
mx.eval(model.parameters())
return model, processor
def to_fp32(model):
model.update(tree_map(
lambda p: p.astype(mx.float32) if isinstance(p, mx.array) and p.dtype == mx.bfloat16 else p,
model.parameters()))
def layers(model):
return model.language_model.model.layers
def switch(model, layer):
return layers(model)[layer].mlp.switch_mlp
def dense_expert(proj):
"""Exact float32 weights (experts, out, in) of a native or affine expert projection."""
mode = proj.mode
if mode == "nvfp4":
w = mx.dequantize(proj.weight, proj.scales, None, group_size=16, bits=4, mode="nvfp4", dtype=mx.float32)
scale = proj.global_scale.astype(mx.float32).reshape(-1, 1, 1)
return w * scale if proj._scale_name == "weight_scale_2" else w / scale
if mode == "mxfp8":
return mx.dequantize(proj.weight, proj.scales, None, group_size=32, bits=8, mode="mxfp8", dtype=mx.float32)
return mx.dequantize(proj.weight, proj.scales, proj.get("biases"), group_size=proj.group_size, bits=proj.bits,
mode=mode, dtype=mx.float32)
# ----------------------------------------------------------------------------- affine packing
def pack_codes(codes, bits):
"""Pack integer codes (..., n) into MLX's uint32 layout: a little-endian bit stream along the last axis."""
n = codes.shape[-1]
per = math.lcm(bits, 32) // bits # codes per repeating unit
words = math.lcm(bits, 32) // 32 # uint32 words per unit
assert n % per == 0, (n, per)
c = codes.astype(mx.uint32).reshape(*codes.shape[:-1], n // per, per)
out = [None] * words
for j in range(per):
off = bits * j
k, sh = off // 32, off % 32
v = c[..., j]
lo = mx.left_shift(v, mx.array(sh, mx.uint32)) if sh else v
out[k] = lo if out[k] is None else mx.bitwise_or(out[k], lo)
if sh + bits > 32:
hi = mx.right_shift(v, mx.array(32 - sh, mx.uint32))
out[k + 1] = hi if out[k + 1] is None else mx.bitwise_or(out[k + 1], hi)
return mx.stack(out, axis=-1).reshape(*codes.shape[:-1], n // per * words)
def affine_switch(packed, scales, biases, bits, group_size):
q = QuantizedSwitchLinear.__new__(QuantizedSwitchLinear)
nn.Module.__init__(q)
q.weight, q.scales, q.biases = packed, scales, biases
q.group_size, q.bits, q.mode = group_size, bits, "affine"
q.freeze()
return q
def rtn_affine(w, bits, group_size, scale_dtype=mx.float16):
"""MLX's own round-to-nearest affine quantizer, with scales stored in `scale_dtype`."""
wq, s, b = mx.quantize(w, group_size=group_size, bits=bits, mode="affine")
return wq, s.astype(scale_dtype), b.astype(scale_dtype)
# ----------------------------------------------------------------------------- data
def render_ids(tok, rec):
"""Token ids of a recorded conversation (scenario record from gen_chats.py)."""
import sys
sys.path.insert(0, os.path.join(ROOT, "scripts"))
from scenarios import TOOLS
tools = [TOOLS[t] for t in rec.get("tools", [])] or None
out = tok.apply_chat_template(rec["messages"], tools=tools, reasoning_effort=rec.get("effort"), tokenize=True)
return list(out["input_ids"]) if hasattr(out, "keys") else list(out)
def load_sequences(path, split=None, max_len=1024):
"""Token sequences from a prepared JSON file: [{"name", "split", "kind", "ids"}]."""
seqs = json.load(open(path))
out = []
for s in seqs:
if split and s["split"] != split:
continue
ids = s["ids"]
for start in range(0, len(ids), max_len):
chunk = ids[start:start + max_len]
if len(chunk) >= 16:
out.append(dict(name=f"{s['name']}@{start}", kind=s["kind"], ids=chunk))
return out
def logits_for(model, ids):
out = model(mx.array([ids])).logits[0].astype(mx.float32)
mx.eval(out)
return out
def log_softmax(x):
return x - mx.logsumexp(x, axis=-1, keepdims=True)
def compare(teacher_logits, student_logits, ids):
"""Per-sequence statistics; teacher and student are (T, V) float32 logits."""
lt, ls = log_softmax(teacher_logits), log_softmax(student_logits)
kl = (mx.exp(lt) * (lt - ls)).sum(-1)
nxt = mx.array(ids[1:])
nll_t = -mx.take_along_axis(lt[:-1], nxt[:, None], axis=-1)[:, 0]
nll_s = -mx.take_along_axis(ls[:-1], nxt[:, None], axis=-1)[:, 0]
agree = (mx.argmax(lt, -1) == mx.argmax(ls, -1)).astype(mx.float32)
vals = [kl.sum(), kl.max(), agree.sum(), nll_t.sum(), nll_s.sum()]
mx.eval(vals)
return dict(n=len(ids), kl_sum=vals[0].item(), kl_max=vals[1].item(), agree=vals[2].item(),
nll_t=vals[3].item(), nll_s=vals[4].item())
def summarize(rows):
n = sum(r["n"] for r in rows)
m = sum(r["n"] - 1 for r in rows)
return dict(tokens=n, kl=sum(r["kl_sum"] for r in rows) / n, kl_max=max(r["kl_max"] for r in rows),
top1=sum(r["agree"] for r in rows) / n, nll_teacher=sum(r["nll_t"] for r in rows) / m,
nll_student=sum(r["nll_s"] for r in rows) / m)
|