File size: 14,837 Bytes
0e41b61 | 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 | #!/usr/bin/env python3
"""
Vision-aware Gemma 4 sniper for single-machine inference.
Combines:
- mlx-vlm vision_tower + embed_vision (~1.2 GB, no SSD streaming)
- Our own MoESniperEngineGemma4 (~5.8 GB pinned + experts streamed from SSD)
Total RAM: ~7 GB. Verified working on M4 Mac Mini 16 GB.
Requirements:
pip install "mlx>=0.31" "mlx-vlm>=0.4" "mlx-lm>=0.31"
(these need Python 3.11+, install in a venv)
"""
import os
import sys
import json
import time
import glob
import numpy as np
import mlx.core as mx
import mlx.nn as nn
# Gemma 4 special token IDs
IMAGE_TOKEN_ID = 262144 # <image_soft_token> — replaced by vision features
BOI_TOKEN_ID = 255999 # <start_of_image>
EOI_TOKEN_ID = 256000 # <end_of_image>
class VisionGemma4Sniper:
"""Single-machine vision-aware Gemma 4 sniper.
The vision encoder (~1.2 GB, bfloat16) is loaded from a source 4-bit
Gemma 4 model directory. The LLM (text + experts) is loaded via our
existing MoESniperEngineGemma4 from a pre-split streaming directory.
"""
def __init__(self, stream_dir, source_dir, cache_size=4000):
"""
Args:
stream_dir: directory with pinned.safetensors + bin/layer_XX.bin
(created by split_gemma4.py)
source_dir: directory with the original mlx-community/gemma-4-26b-a4b-it-4bit
safetensors files (used to extract vision weights)
cache_size: LRU cache size for the expert reader
"""
self.stream_dir = os.path.expanduser(stream_dir)
self.source_dir = os.path.expanduser(source_dir)
self.cache_size = cache_size
self.sniper = None
self.vision_tower = None
self.embed_vision = None
self.image_processor = None
def load(self):
"""Load all three components."""
# Make sure the existing single-machine sniper code is importable
SNIPER_PATHS = [
os.path.expanduser("~/expert-sniper-mlx"),
os.path.expanduser("~/cli-agent/src"),
]
for p in SNIPER_PATHS:
if os.path.isdir(p) and p not in sys.path:
sys.path.insert(0, p)
try:
from moe_agent_gemma4 import MoESniperEngineGemma4
except ImportError as e:
raise ImportError(
"Could not import moe_agent_gemma4. The vision engine wraps "
"the existing single-machine sniper. Make sure these directories "
"exist and contain the sniper code:\n"
f" ~/expert-sniper-mlx/moe_agent_gemma4.py\n"
f" ~/cli-agent/src/mlx_expert_sniper/models/gemma4.py\n"
f"Original error: {e}"
)
from mlx_vlm.models.gemma4 import VisionModel, VisionConfig, Gemma4ImageProcessor
from mlx_vlm.models.gemma4.gemma4 import MultimodalEmbedder
print("Loading Vision Gemma 4 Sniper...")
# Load configs
with open(os.path.join(self.stream_dir, "config.json")) as f:
text_config = json.load(f)
with open(os.path.join(self.source_dir, "config.json")) as f:
source_config = json.load(f)
vision_config_dict = source_config.get("vision_config", {})
if not vision_config_dict:
raise RuntimeError(f"No vision_config in {self.source_dir}/config.json")
# 1. Sniper LLM (proven 4.15 tok/s on M4)
print(" [1/4] Sniper LLM...")
self.sniper = MoESniperEngineGemma4(
model_dir=self.stream_dir, cache_size=self.cache_size
)
self.sniper.load()
# 2. Vision tower (bfloat16, no quantization)
print(" [2/4] Vision tower...")
vc = VisionConfig.from_dict(vision_config_dict)
self.vision_tower = VisionModel(vc)
# Find safetensors files — accept both the original sharded format
# and the extracted standalone vision.safetensors
safetensors_files = sorted(glob.glob(os.path.join(self.source_dir, "model-*.safetensors")))
if not safetensors_files:
standalone = os.path.join(self.source_dir, "vision.safetensors")
if os.path.exists(standalone):
safetensors_files = [standalone]
if not safetensors_files:
raise FileNotFoundError(
f"No safetensors files found in {self.source_dir}. "
f"Expected either model-*.safetensors or vision.safetensors"
)
vision_weights = {}
for sf in safetensors_files:
w = mx.load(sf)
for k, v in w.items():
if k.startswith("vision_tower."):
vision_weights[k[len("vision_tower."):]] = v
del w
self.vision_tower.load_weights(list(vision_weights.items()), strict=False)
mx.eval(self.vision_tower.parameters())
# 3. embed_vision projector (1 quantized layer)
print(" [3/4] embed_vision...")
self.embed_vision = MultimodalEmbedder(
embedding_dim=vision_config_dict["hidden_size"],
text_hidden_size=text_config["hidden_size"],
eps=vision_config_dict.get("rms_norm_eps", 1e-6),
)
nn.quantize(self.embed_vision, group_size=64, bits=4)
ev_weights = {}
for sf in safetensors_files: # reuse the same files
w = mx.load(sf)
for k, v in w.items():
if k.startswith("embed_vision."):
ev_weights[k[len("embed_vision."):]] = v
del w
self.embed_vision.load_weights(list(ev_weights.items()), strict=False)
mx.eval(self.embed_vision.parameters())
# 4. Image processor
print(" [4/4] Image processor...")
self.image_processor = Gemma4ImageProcessor.from_pretrained(self.source_dir)
total_gb = mx.get_active_memory() / 1e9
print(f" Total active memory: {total_gb:.2f} GB")
print("Ready!")
def encode_image(self, image_path):
"""Encode an image into projected embeddings.
Returns:
(image_features, n_tokens) where image_features is
mx.array [1, n_tokens, text_hidden_size]
"""
from PIL import Image
img = Image.open(image_path).convert("RGB")
processed = self.image_processor(images=[img], return_tensors="mlx")
# Unpack tuple (dict, list)
if isinstance(processed, tuple):
pv_obj = processed[0]
n_tokens_list = processed[1]
else:
pv_obj = processed
n_tokens_list = None
if isinstance(pv_obj, dict):
pixel_values = pv_obj["pixel_values"]
else:
pixel_values = pv_obj
if not isinstance(pixel_values, mx.array):
pixel_values = mx.array(np.array(pixel_values))
# Run vision tower → embed_vision (the projector)
image_features = self.vision_tower(pixel_values)
image_features = self.embed_vision(image_features)
mx.eval(image_features)
n_tokens = n_tokens_list[0] if n_tokens_list else image_features.shape[1]
return image_features, n_tokens
def encode_chat(self, prompt, image_path=None):
"""Build the input token sequence with optional image placeholders.
Returns: (token_ids, image_features_or_None, n_image_tokens)
"""
NL = chr(10)
prompt_toks = self.sniper.tokenizer.encode(prompt).ids
user_toks = self.sniper.tokenizer.encode("user" + NL).ids
model_toks = self.sniper.tokenizer.encode("model" + NL).ids
image_features = None
n_image_tokens = 0
if image_path:
image_features, n_image_tokens = self.encode_image(image_path)
tokens = (
[2, 105] # <bos> <|turn>
+ user_toks # user\n
+ [BOI_TOKEN_ID] # <start_of_image>
+ [IMAGE_TOKEN_ID] * n_image_tokens # image placeholders
+ [EOI_TOKEN_ID] # <end_of_image>
+ prompt_toks # text question
+ [106, 107, 105] # <turn|> \n <|turn>
+ model_toks # model\n
)
else:
tokens = (
[2, 105] + user_toks + prompt_toks
+ [106, 107, 105] + model_toks
)
return tokens, image_features, n_image_tokens
def generate(self, prompt, image_path=None, max_tokens=200, temperature=0.7,
on_chunk=None):
"""Generate a response, optionally conditioned on an image.
on_chunk: optional callback that receives streaming text chunks.
"""
tokens, image_features, n_image_tokens = self.encode_chat(prompt, image_path)
# Run the prefill (with image embedding injection)
next_token = self._prefill(tokens, image_features)
# Stream subsequent tokens
generated = [next_token]
eos_set = {1, 106} # <eos>, <turn|>
from collections import deque
recent = deque(maxlen=64)
recent.append(next_token)
printed = ""
for step in range(max_tokens - 1):
# Decode buffer for streaming
full = self.sniper.tokenizer.decode(generated)
new_chunk = full[len(printed):]
if new_chunk and on_chunk:
on_chunk(new_chunk)
printed = full
if next_token in eos_set:
break
input_ids = mx.array([[next_token]])
logits = self.sniper.forward(input_ids)
mx.eval(logits)
# Apply repetition penalty (mild)
if recent:
last_logits = logits[0, -1]
last_np = np.array(last_logits.astype(mx.float32))
for tid in set(recent):
if last_np[tid] > 0:
last_np[tid] /= 1.1
else:
last_np[tid] *= 1.1
last_logits = mx.array(last_np)
else:
last_logits = logits[0, -1]
if temperature <= 0:
next_token = int(mx.argmax(last_logits).item())
else:
probs = mx.softmax(last_logits / temperature, axis=-1)
next_token = int(mx.random.categorical(mx.log(probs + 1e-10)).item())
generated.append(next_token)
recent.append(next_token)
# Final flush
full = self.sniper.tokenizer.decode(generated)
new_chunk = full[len(printed):]
if new_chunk and on_chunk:
on_chunk(new_chunk)
return full
def _prefill(self, input_token_ids, image_features):
"""Run the first forward pass with image embeddings injected."""
from mlx_lm.models.base import create_attention_mask
from mlx_vlm.models.gemma4.gemma4 import masked_scatter
from moe_agent_gemma4 import run_expert_ffn
self.sniper.reset_cache()
input_ids = mx.array([input_token_ids])
# Get text embeddings + Gemma sqrt(d) scaling
h = self.sniper.model.model.embed_tokens(input_ids)
h = h * (self.sniper.model.args.hidden_size ** 0.5)
# Inject image embeddings at IMAGE_TOKEN_ID positions
if image_features is not None:
image_mask = (input_ids == IMAGE_TOKEN_ID)
image_feats_flat = image_features.reshape(-1, image_features.shape[-1])
image_feats_flat = image_feats_flat.astype(h.dtype)
image_mask_expanded = mx.expand_dims(image_mask, -1)
image_mask_expanded = mx.broadcast_to(image_mask_expanded, h.shape)
h = masked_scatter(h, image_mask_expanded, image_feats_flat)
# Run the LLM forward pass — replicates moe_agent_gemma4 forward but
# starting from h (post-embedding) instead of input_ids
mask = create_attention_mask(h, self.sniper.cache[0] if self.sniper.cache else None)
for i in range(self.sniper.num_layers):
layer = self.sniper.model.model.layers[i]
cache_i = self.sniper.cache[i] if self.sniper.cache else None
residual = h
h_norm = layer.input_layernorm(h)
h_attn = layer.self_attn(h_norm, mask=mask, cache=cache_i)
h_attn = layer.post_attention_layernorm(h_attn)
h = residual + h_attn
mx.eval(h)
residual = h
h_ff = layer.pre_feedforward_layernorm(h)
h_ff = layer.mlp(h_ff)
if layer.enable_moe_block:
h_dense = layer.post_feedforward_layernorm_1(h_ff)
B, L, D = residual.shape
residual_flat = residual.reshape(-1, D)
router = layer.router
x_normed = router._inline_rms_norm(residual_flat)
x_normed = x_normed * router.scale * (router.hidden_size ** -0.5)
scores = router.proj(x_normed)
probs = mx.softmax(scores, axis=-1)
top_k_indices = mx.argpartition(-probs, kth=router.top_k - 1, axis=-1)[..., :router.top_k]
top_k_weights = mx.take_along_axis(probs, top_k_indices, axis=-1)
top_k_weights = top_k_weights / mx.sum(top_k_weights, axis=-1, keepdims=True)
expert_scales = router.per_expert_scale[top_k_indices]
top_k_weights = top_k_weights * expert_scales
moe_input = layer.pre_feedforward_layernorm_2(residual_flat)
mx.eval(moe_input, top_k_indices, top_k_weights)
top_k_indices_r = top_k_indices.reshape(B, L, -1)
top_k_weights_r = top_k_weights.reshape(B, L, -1)
active_ids = list(set(int(e) for e in np.array(top_k_indices_r).flatten()))
expert_data = self.sniper.reader.get_experts(i, active_ids)
moe_input_r = moe_input.reshape(B, L, D)
expert_out = run_expert_ffn(moe_input_r, expert_data, top_k_indices_r, top_k_weights_r)
h_moe = layer.post_feedforward_layernorm_2(expert_out)
h_ff = h_dense + h_moe
h_ff = layer.post_feedforward_layernorm(h_ff)
h = residual + h_ff
h = h * layer.layer_scalar
mx.eval(h)
mx.clear_cache()
h = self.sniper.model.model.norm(h)
if self.sniper.model.args.tie_word_embeddings:
logits = self.sniper.model.model.embed_tokens.as_linear(h)
else:
logits = self.sniper.model.lm_head(h)
mx.eval(logits)
# Sample first token
return int(mx.argmax(logits[0, -1]).item())
|