Text Generation
MLX
Safetensors
modilify_mk2
diffusion
mixture-of-experts
custom-code
modilify-mk2
conversational
Instructions to use modilify/Modilify-Mk2-mlx with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use modilify/Modilify-Mk2-mlx with MLX:
# Make sure mlx-lm is installed # pip install --upgrade mlx-lm # Generate text with mlx-lm from mlx_lm import load, generate model, tokenizer = load("modilify/Modilify-Mk2-mlx") prompt = "Write a story about Einstein" messages = [{"role": "user", "content": prompt}] prompt = tokenizer.apply_chat_template( messages, add_generation_prompt=True ) text = generate(model, tokenizer, prompt=prompt, verbose=True) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- Pi
How to use modilify/Modilify-Mk2-mlx with Pi:
Start the MLX server
# Install MLX LM: uv tool install mlx-lm # Start a local OpenAI-compatible server: mlx_lm.server --model "modilify/Modilify-Mk2-mlx"
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": "modilify/Modilify-Mk2-mlx" } ] } } }Run Pi
# Start Pi in your project directory: pi
- MLX LM
How to use modilify/Modilify-Mk2-mlx with MLX LM:
Generate or start a chat session
# Install MLX LM uv tool install mlx-lm # Interactive chat REPL mlx_lm.chat --model "modilify/Modilify-Mk2-mlx"
Run an OpenAI-compatible server
# Install MLX LM uv tool install mlx-lm # Start the server mlx_lm.server --model "modilify/Modilify-Mk2-mlx" # Calling the OpenAI-compatible server with curl curl -X POST "http://localhost:8080/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "modilify/Modilify-Mk2-mlx", "messages": [ {"role": "user", "content": "Hello"} ] }' - Hermes Agent
How to use modilify/Modilify-Mk2-mlx 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 "modilify/Modilify-Mk2-mlx"
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 modilify/Modilify-Mk2-mlx
Run Hermes
hermes
- Atomic Chat
- OpenClaw
How to use modilify/Modilify-Mk2-mlx with OpenClaw:
Start the MLX server
# Install MLX LM: uv tool install mlx-lm # Start a local OpenAI-compatible server: mlx_lm.server --model "modilify/Modilify-Mk2-mlx"
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 "modilify/Modilify-Mk2-mlx" \ --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: 17,962 Bytes
1ac2be1 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 | """Text generation and commit policy for the schema25 MLX release."""
from __future__ import annotations
from collections.abc import Callable, Sequence
from dataclasses import dataclass, replace
import math
from pathlib import Path
from types import SimpleNamespace
import mlx.core as mx
from tokenizers.decoders import DecodeStream
from transformers import GemmaTokenizer
from .model import GDN2TrajectoryState, MLXModilifyMk2, _read_json, load_model
JUMP_FAILURE_BUDGET = 2.0
FUSED_EPS = 1.0e-6
def commit_target_confidence_bias(
target_confidence: float | None,
failure_budget: float = 0.2,
budget_safety_ratio: float = 0.85,
) -> float:
if target_confidence is None or target_confidence <= 0.0 or target_confidence >= 1.0:
return 0.0
target_failure = min(budget_safety_ratio * failure_budget, 1.0 - target_confidence)
target_failure = max(target_failure, 1.0e-4)
target_conf = 1.0 - target_failure
logit_c = math.log(target_conf / (1.0 - target_conf))
logit_p = math.log(target_confidence / (1.0 - target_confidence))
return max(logit_c - logit_p, 0.0)
def fused_commit_confidence(
proposal_confidence: mx.array,
token_entropy: mx.array,
*,
eps: float = FUSED_EPS,
entropy_weight: float = 1.0,
confidence_power: float = 2.0,
top_k: int | None = None,
min_p: float | None = None,
target_confidence: float | None = None,
failure_budget: float = 0.2,
) -> mx.array:
"""Match the schema25 excess-entropy sigmoid and checkpoint power."""
p = mx.clip(mx.nan_to_num(proposal_confidence.astype(mx.float32), nan=0.5), eps, 1.0 - eps)
entropy = mx.maximum(mx.nan_to_num(token_entropy.astype(mx.float32), nan=0.0), 0.0)
binary_entropy = -p * mx.log(p) - (1.0 - p) * mx.log1p(-p)
excess = mx.maximum(entropy - binary_entropy, 0.0)
k_eff = None
if top_k is not None and top_k > 0:
k_eff = mx.full(p.shape, float(top_k), dtype=mx.float32)
if min_p is not None and min_p > 0:
thresh = mx.maximum(float(min_p) * p, 1.0e-6)
k_min_p = mx.maximum((1.0 - p) / thresh, 1.0)
k_eff = k_min_p if k_eff is None else mx.minimum(k_eff, k_min_p)
if k_eff is not None:
max_excess = (1.0 - p) * mx.log(k_eff)
excess = mx.minimum(excess, max_excess)
bias = commit_target_confidence_bias(target_confidence, failure_budget=failure_budget)
fused_logit = mx.log(p) - mx.log1p(-p) + bias - entropy_weight * excess
return mx.clip(mx.sigmoid(fused_logit) ** confidence_power, eps, 1.0 - eps)
def fused_commit_failure_rate(
proposal_confidence: mx.array, token_entropy: mx.array, **kwargs: object
) -> mx.array:
return 1.0 - fused_commit_confidence(proposal_confidence, token_entropy, **kwargs)
def _terminal_mask(tokens: mx.array, stop_token_id: int | Sequence[int]) -> mx.array:
ids = (stop_token_id,) if isinstance(stop_token_id, int) else tuple(dict.fromkeys(stop_token_id))
if not ids:
raise ValueError("At least one stop token ID is required.")
matches = tokens == int(ids[0])
for value in ids[1:]:
matches = matches | (tokens == int(value))
return matches
def _prefix_length(risk: mx.array, valid: mx.array, budget: float) -> mx.array:
# Nonnegative risk and contiguous validity make the allowed prefix monotonic.
allowed = (mx.cumsum(mx.clip(risk, 0.0, 1.0) * valid, axis=-1) < budget) & valid
return mx.sum(allowed, axis=-1).astype(mx.int32)
def _commit(
sampled, risk, previous_risk, greedy, greedy_risk, *,
remaining, stops, budget, ponder, stagnation, generation,
):
positions = mx.arange(sampled.shape[-1])[None, :]
valid = positions < remaining
def bounded(ids, rates, limit):
length = _prefix_length(rates, valid, limit)
terminal = _terminal_mask(ids, stops) & (positions < length[:, None])
first = mx.min(mx.where(terminal, positions + 1, sampled.shape[-1]), axis=-1)
return mx.minimum(length, first)
normal = bounded(sampled, risk, budget)
previous = _prefix_length(previous_risk, valid, budget)
frontier = mx.minimum(mx.maximum(previous, normal) + 1, mx.sum(valid, axis=-1))
progress = mx.sum((previous_risk - risk) * (positions < frontier[:, None]), axis=-1)
progress = progress / mx.maximum(frontier, 1)
next_ponder = mx.where(normal > 0, 0, ponder + 1)
next_stagnation = mx.where(normal > 0, 0, stagnation + 1)
jump = (normal == 0) & (
((next_stagnation >= generation.jump_on_no_progress_after)
& (progress <= generation.min_trajectory_progress))
| (next_ponder >= generation.max_ponder_steps)
)
lengths = mx.where(jump, bounded(greedy, greedy_risk, JUMP_FAILURE_BUDGET), normal)
return (lengths, mx.where(jump[:, None], greedy, sampled),
mx.where(lengths > 0, 0, next_ponder), mx.where(lengths > 0, 0, next_stagnation))
def seed_key(seed: int, salt: int) -> mx.array:
return mx.random.key((seed + salt) % (2**32))
def proposal_statistics(
hidden: mx.array, weight: mx.array, *, seed: int, step: int,
softcap: float, temperature: float, top_k: int, min_p: float,
repetition_penalty: float, repetition_seen: set[int], chunk_size: int = 32768,
) -> tuple[mx.array, ...]:
canvas = hidden.shape[1]
flat = hidden[0]
log_z = mx.full((canvas,), -mx.inf, mx.float32)
greedy_score = log_z
greedy = mx.zeros((canvas,), mx.int32)
moment_sum = mx.zeros_like(log_z)
moment_weighted = mx.zeros_like(log_z)
candidates = []
candidate_ids = []
if repetition_penalty != 1:
seen = mx.zeros((weight.shape[0],), mx.bool_)
if repetition_seen:
seen[mx.array(sorted(repetition_seen), mx.int32)] = True
for start in range(0, weight.shape[0], chunk_size):
stop = min(start + chunk_size, weight.shape[0])
score = mx.tanh((flat @ weight[start:stop].T).astype(mx.float32) / softcap) * softcap
if repetition_penalty != 1:
penalized = mx.where(score < 0, score * repetition_penalty, score / repetition_penalty)
score = mx.where(seen[start:stop], penalized, score)
score = score / temperature
log_z = mx.logaddexp(log_z, mx.logsumexp(score, axis=-1))
local_max = mx.max(score, axis=-1)
greedy = mx.where(local_max > greedy_score, mx.argmax(score, axis=-1) + start, greedy)
next_max = mx.maximum(greedy_score, local_max)
previous_scale = mx.exp(greedy_score - next_max)
local_weights = mx.exp(score - next_max[:, None])
moment_sum = moment_sum * previous_scale + mx.sum(local_weights, axis=-1)
moment_weighted = moment_weighted * previous_scale + mx.sum(local_weights * score, axis=-1)
greedy_score = next_max
k = min(top_k, stop - start)
indices = mx.argpartition(-score, kth=k - 1, axis=-1)[:, :k]
candidates.append(mx.take_along_axis(score, indices, axis=-1))
candidate_ids.append(indices + start)
mx.async_eval(log_z, greedy, greedy_score, moment_sum, moment_weighted,
candidates[-1], candidate_ids[-1])
all_scores = mx.concatenate(candidates, axis=-1)
all_ids = mx.concatenate(candidate_ids, axis=-1)
k = min(top_k, all_scores.shape[-1])
indices = mx.argpartition(-all_scores, kth=k - 1, axis=-1)[:, :k]
scores = mx.take_along_axis(all_scores, indices, axis=-1)
ids = mx.take_along_axis(all_ids, indices, axis=-1)
eligible = mx.where(scores >= greedy_score[:, None] + math.log(min_p), scores, -mx.inf)
uniform = mx.random.uniform(shape=(canvas, k), key=seed_key(seed, 0xC09A + step * 1000003))
uniform = mx.clip(uniform, 1.17549435e-38, 1.0 - 1.19209290e-7)
chosen = mx.argmax(eligible - mx.log(-mx.log(uniform)), axis=-1, keepdims=True)
token_ids = mx.take_along_axis(ids, chosen, axis=-1)[:, 0]
chosen_score = mx.take_along_axis(scores, chosen, axis=-1)[:, 0]
confidence = mx.clip(mx.exp(chosen_score - log_z), 0.0, 1.0)
greedy_confidence = mx.clip(mx.exp(greedy_score - log_z), 0.0, 1.0)
entropy = log_z - moment_weighted / moment_sum
return tuple(value[None] for value in (token_ids, confidence, entropy, greedy, greedy_confidence))
@dataclass(frozen=True)
class GenerationResult:
text: str
token_ids: tuple[int, ...]
denoise_steps: int
stop_reason: str
def _positive_int(value: int, name: str) -> None:
if type(value) is not int or value < 1:
raise ValueError(f"{name} must be a positive integer.")
class Runtime:
def __init__(self, model: MLXModilifyMk2, tokenizer, generation: dict):
self.model = model
self.tokenizer = tokenizer
self.generation = SimpleNamespace(**generation)
def generate(
self, messages: Sequence[dict[str, str]], *, enable_thinking: bool = True,
seed: int = 42, max_new_tokens: int = 256, max_denoising_steps: int = 256,
on_text: Callable[[str], None] | None = None,
) -> GenerationResult:
"""Generate one text response and optionally emit committed text increments."""
_positive_int(max_new_tokens, "max_new_tokens")
_positive_int(max_denoising_steps, "max_denoising_steps")
if type(seed) is not int:
raise ValueError("seed must be an integer.")
if not messages or any(
not isinstance(message, dict)
or message.get("role") not in ("system", "user", "assistant")
or not isinstance(message.get("content"), str)
for message in messages
):
raise ValueError("Expected a non-empty sequence of system, user or assistant text messages.")
prompt_ids = self.tokenizer.apply_chat_template(
list(messages), tokenize=True, add_generation_prompt=True,
enable_thinking=enable_thinking, return_dict=True,
)["input_ids"]
max_positions = self.model.config.text_config.max_position_embeddings
if not prompt_ids or len(prompt_ids) + max_new_tokens > max_positions:
raise ValueError("Prompt and response budget must fit within the model position limit.")
with mx.stream(mx.gpu):
return self._generate(prompt_ids, seed, max_new_tokens, max_denoising_steps, on_text)
def _generate(
self, prompt_ids: list[int], seed: int, max_new_tokens: int,
max_denoising_steps: int, on_text: Callable[[str], None] | None,
) -> GenerationResult:
config = self.model.config
decoder = self.model.model.decoder
encoder = self.model.model.encoder
cache = encoder.make_cache()
for start in range(0, len(prompt_ids), 256):
_, cache = encoder(mx.array([prompt_ids[start:start + 256]], mx.int32), cache=cache)
mx.eval(*(value for layer in cache for value in (layer.keys, layer.values)
if isinstance(value, mx.array)))
generation = self.generation
vocab = config.text_config.vocab_size
canvas = config.canvas_length
positions = mx.arange(canvas)[None, :]
unknown_entropy = math.log(vocab)
memory = GDN2TrajectoryState(
mx.zeros((1, canvas, 16, 64, 64), mx.float32),
mx.zeros((1, 16, 64, 64), mx.float32),
mx.zeros((1, 16, 128, 128), mx.float32),
mx.zeros((1, canvas), mx.bool_),
)
confidence = mx.zeros((1, canvas), mx.float32)
entropy = mx.full((1, canvas), unknown_entropy, mx.float32)
clocks = mx.zeros((1,), mx.int32)
ponder = stagnation = head = clocks
def noise(step: int) -> mx.array:
return mx.random.randint(0, vocab, (1, canvas), key=seed_key(seed, 0x51A7 + step * 1000003))
tokens = mx.where(positions < max_new_tokens, noise(0), generation.pad_token_id)
turn_end = generation.turn_end_token_id
eos = generation.eos_token_id
stops = tuple(dict.fromkeys((turn_end, *(eos if isinstance(eos, list) else [eos]))))
excluded = set(generation.repetition_penalty_exclude_token_ids)
excluded.update((generation.pad_token_id, generation.bos_token_id, *stops))
seen = set(prompt_ids) - excluded
generated = []
stream = DecodeStream(skip_special_tokens=True) if on_text is not None else None
emitted = ""
watchdog = max_new_tokens * generation.max_ponder_steps
def failure(confidence: mx.array, entropy: mx.array) -> mx.array:
return fused_commit_failure_rate(
confidence, entropy, entropy_weight=config.commit_entropy_weight,
confidence_power=config.commit_confidence_power, top_k=config.commit_top_k,
min_p=config.commit_min_p, target_confidence=config.commit_target_confidence,
failure_budget=config.commit_failure_budget,
)
limit = min(max_denoising_steps, watchdog)
for step in range(1, limit + 1):
output = self.model.forward_canvas(
tokens, prefix_cache=cache, state=memory, canvas_head=head,
first_step=step == 1, has_commits=bool(generated),
)
sampled, next_confidence, next_entropy, greedy, greedy_confidence = proposal_statistics(
output.heavy_hidden, decoder.embed_tokens.weight, seed=seed, step=step - 1,
softcap=config.text_config.final_logit_softcapping, temperature=generation.t_min,
top_k=config.commit_top_k, min_p=config.commit_min_p,
repetition_penalty=generation.repetition_penalty, repetition_seen=seen,
)
remaining = max_new_tokens - len(generated)
logical_order = (head[:, None] + positions) % canvas
physical_offsets = (positions - head[:, None]) % canvas
def logical(values: mx.array) -> mx.array:
return mx.take_along_axis(values, logical_order, axis=1)
lengths, selected, ponder, stagnation = _commit(
logical(sampled), logical(failure(next_confidence, next_entropy)),
logical(failure(confidence, entropy)),
logical(greedy), logical(failure(greedy_confidence, next_entropy)),
ponder=ponder, stagnation=stagnation, remaining=remaining,
budget=config.commit_failure_budget, stops=stops, generation=generation,
)
mx.eval(lengths, selected)
count = int(lengths.item())
block = selected[0, :count].tolist()
generated.extend(block)
seen.update(set(block) - excluded)
if stream is not None and block:
delta = stream.step(self.tokenizer.backend_tokenizer, block)
if delta:
on_text(delta)
emitted += delta
if turn_end in block:
reason = "turn_end"
break
if any(token in stops for token in block):
reason = "eos"
break
if len(generated) >= max_new_tokens:
reason = "max_new_tokens"
break
if step == limit:
continue
memory = self.model.latent_deliberation.trajectory.observe(
memory, output.heavy_hidden + output.working_state, physical_offsets < remaining, head,
)
tokens, confidence, entropy = sampled, next_confidence, next_entropy
if count:
committed_ids = selected[:, :count]
persistent = self.model.latent_deliberation.commit_write(
memory=memory.persistent, working_state=output.working_state,
heavy_hidden=output.heavy_hidden,
committed_token_embeddings=decoder.embed_tokens(committed_ids) * decoder.embed_scale,
canvas_head=head,
)
memory = replace(memory.clear_refill(lengths, head), persistent=persistent)
_, cache = encoder(committed_ids, cache=cache)
head = (head + count) % canvas
refill = physical_offsets < count
tail = mx.take_along_axis(noise(step), physical_offsets, axis=1)
pad = ((positions - head[:, None]) % canvas) >= remaining - count
tokens = mx.where(refill, mx.where(pad, generation.pad_token_id, tail), tokens)
confidence = mx.where(refill, 0.0, confidence)
entropy = mx.where(refill, unknown_entropy, entropy)
mx.async_eval(tokens, confidence, entropy, memory.cells, memory.row, memory.persistent, memory.seen,
*(value for layer in cache for value in (layer.keys, layer.values)
if isinstance(value, mx.array)))
else:
reason = "max_denoising_steps" if max_denoising_steps <= watchdog else "episode_watchdog"
text = self.tokenizer.backend_tokenizer.decode(generated, skip_special_tokens=True)
if on_text is not None and emitted != text:
if not text.startswith(emitted):
raise RuntimeError("Tokenizer streaming output diverged from complete decode.")
on_text(text[len(emitted):])
return GenerationResult(text, tuple(generated), step, reason)
def load(model_dir: str | Path | None = None) -> Runtime:
"""Open a local release directory with model computation on Metal."""
root = Path(model_dir).expanduser().resolve() if model_dir is not None else Path(__file__).resolve().parents[1]
generation = _read_json(root / "generation_config.json")
# Keep the released Gemma tokenizer regex as stored in tokenizer.json.
tokenizer = GemmaTokenizer.from_pretrained(
root, local_files_only=True, fix_mistral_regex=False,
)
with mx.stream(mx.gpu):
model = load_model(root)
return Runtime(model, tokenizer, generation)
|