Skip to content

Pretrained models

Loaders that import pretrained weights and fuse the model into one ANE program. This covers the vision, encoder, reranker, CLIP, GPT-2, and Whisper loaders in aneforge.models. Decoder LLMs load via af.load_llm (see the LLM guide) and ONNX models via af.load_onnx (see ONNX import).

models

Pretrained-model loaders (load and CrossEncoder for BERT/RoBERTa encoders and rerankers, load_resnet/load_vit image classifiers, load_gpt2 text generation) and trainable-graph builders (group_norm_train, conv_block, cifar_cnn). See docs/developer/models.md.

Vision

Source code in aneforge/models.py 
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
class Vision:
  def __init__(self, name_or_depth: int | str = 18, int8: bool = False, compress: str | None = None,
               compress_atol: float = 0.05, build_dir: str | None = None,
               weights: str = "IMAGENET1K_V1") -> None:
    if isinstance(name_or_depth, str) and "/" in name_or_depth:   # a Hugging Face repo id, not a depth
      self.depth = None
      self.sd, self.block, self.stages = _load_hf_resnet(name_or_depth)
    else:
      import torchvision  # lazy
      self.depth = _resnet_depth(name_or_depth)
      self.block, self.stages = _RESNETS[self.depth]
      m = getattr(torchvision.models, f"resnet{self.depth}")(weights=weights).eval()
      self.sd = {k: v.detach().numpy().astype(np.float32) for k, v in m.state_dict().items()}
    self.int8 = int8
    self.compress = compress
    self.compress_atol = compress_atol
    self.build_dir = build_dir
    self._model = self._build()

  def _fold(self, conv_key: str, bn: str):
    """Fold BatchNorm(`bn`) into conv(`conv_key`) -> (weight, bias)."""
    W = self.sd[conv_key + ".weight"]
    g, b = self.sd[bn + ".weight"], self.sd[bn + ".bias"]
    mu, var = self.sd[bn + ".running_mean"], self.sd[bn + ".running_var"]
    sc = g / np.sqrt(var + 1e-5)
    return (W * sc[:, None, None, None]).astype(np.float32), (b - mu * sc).astype(np.float32)

  def _shortcut(self, x: Tensor, prefix: str, stride: int) -> Tensor:
    """The residual branch: a projection when the block carries one, otherwise the identity.

    Presence is read from the weights rather than derived from a rule. torchvision projects when
    `stride != 1 or inplanes != planes * expansion`, so with Bottleneck even stage 1 is projected
    (64 -> 256 at stride 1) while with BasicBlock it is not. A "stage 1 never downsamples" rule holds
    for BasicBlock only, and would silently drop layer1's projection on ResNet-50/101."""
    if prefix + ".downsample.0.weight" not in self.sd:
      return x
    wd, bd = self._fold(prefix + ".downsample.0", prefix + ".downsample.1")
    return conv(x, wd, stride=stride, pad=0, bias=bd)

  def _block(self, x: Tensor, prefix: str, stride: int) -> Tensor:
    """BasicBlock: 3x3 -> 3x3, stride on the first conv."""
    w1, b1 = self._fold(prefix + ".conv1", prefix + ".bn1")
    w2, b2 = self._fold(prefix + ".conv2", prefix + ".bn2")
    out = conv(x, w1, stride=stride, pad=1, bias=b1).relu()
    out = conv(out, w2, stride=1, pad=1, bias=b2)
    return (out + self._shortcut(x, prefix, stride)).relu()

  def _bottleneck(self, x: Tensor, prefix: str, stride: int) -> Tensor:
    """Bottleneck: 1x1 -> 3x3 -> 1x1 with a 4x expansion. The stride sits on the 3x3, not on the
    first 1x1: that is torchvision's ResNet V1.5 variant, and putting it on conv1 instead changes
    which pixels survive."""
    w1, b1 = self._fold(prefix + ".conv1", prefix + ".bn1")
    w2, b2 = self._fold(prefix + ".conv2", prefix + ".bn2")
    w3, b3 = self._fold(prefix + ".conv3", prefix + ".bn3")
    out = conv(x, w1, stride=1, pad=0, bias=b1).relu()
    out = conv(out, w2, stride=stride, pad=1, bias=b2).relu()
    out = conv(out, w3, stride=1, pad=0, bias=b3)
    return (out + self._shortcut(x, prefix, stride)).relu()

  def _build(self) -> Model | SegmentedModel:
    x = input((1, 3, 224, 224))
    w, b = self._fold("conv1", "bn1")
    h = conv(x, w, stride=2, pad=3, bias=b).relu().max_pool(3, stride=2, pad=1)
    block = self._bottleneck if self.block == "bottleneck" else self._block
    for name, stride, n in zip(("layer1", "layer2", "layer3", "layer4"), (1, 2, 2, 2), self.stages):
      for i in range(n):
        h = block(h, f"{name}.{i}", stride if i == 0 else 1)
    feat = self.sd["fc.weight"].shape[1]          # 512 for BasicBlock, 2048 once expanded 4x
    h = h.mean((2, 3)).reshape(1, feat)
    out = h.linear(self.sd["fc.weight"], self.sd["fc.bias"])
    return compile(out, int8=self.int8, compress=self.compress,
                   compress_atol=self.compress_atol, build_dir=self.build_dir)

  def release(self) -> None: self._model.release()

  def __call__(self, image: np.ndarray) -> np.ndarray:
    image = np.asarray(image, dtype=np.float32)
    if image.ndim == 3: image = image[None]
    return self._model(image)            # [1, 1000] logits

  @property
  def n_ops(self) -> int: return self._model.n_ops

Encoder

Source code in aneforge/models.py 
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
class Encoder:
  _POOL = ("mean", "cls", "max")

  def __init__(self, name: str, int8: bool = False, pooling: str = "mean") -> None:
    if pooling not in self._POOL:
      raise ValueError(f"pooling must be one of {self._POOL}, got {pooling!r}")
    self.pooling = pooling
    from transformers import AutoConfig, AutoModel, AutoTokenizer  # lazy
    cfg = AutoConfig.from_pretrained(name)
    self.tok = AutoTokenizer.from_pretrained(name)
    sd = AutoModel.from_pretrained(name).state_dict()
    g = lambda k: sd[k].detach().numpy().astype(np.float32)
    self.D, self.H = cfg.hidden_size, cfg.num_attention_heads
    self.L, self.eps, self.int8 = cfg.num_hidden_layers, cfg.layer_norm_eps, int8
    mpnet = cfg.model_type == "mpnet"
    keys = _MPNET_KEYS if mpnet else _BERT_KEYS
    self.word = g("embeddings.word_embeddings.weight")
    self.pos = g("embeddings.position_embeddings.weight")
    self.typ = g("embeddings.token_type_embeddings.weight") if "embeddings.token_type_embeddings.weight" in sd else None
    self.eln_w, self.eln_b = g("embeddings.LayerNorm.weight"), g("embeddings.LayerNorm.bias")
    self.layers = [{k: g(f"encoder.layer.{i}." + v) for k, v in keys.items()}
                   for i in range(self.L)]
    # MPNet adds a T5-style relative-position bias to the attention scores; MPNet and the RoBERTa family
    # count positions from pad_id+1 (HF create_position_ids_from_input_ids), unlike BERT's 0-based.
    self.rel_bias = g("encoder.relative_attention_bias.weight") if mpnet else None
    self.n_buckets = int(getattr(cfg, "relative_attention_num_buckets", 32))
    offset_pos = cfg.model_type in ("mpnet", "roberta", "xlm-roberta", "camembert")
    self.pos_offset = int(cfg.pad_token_id) + 1 if offset_pos else 0
    self._bias_cache: dict[int, np.ndarray] = {}
    self._cache: dict[int, Model | SegmentedModel] = {}

  def _build(self, S: int) -> Model | SegmentedModel:
    h = input((S, self.D))
    mh = self.H if self.rel_bias is not None else 1   # MPNet folds a per-head relative bias into the mask
    m = input((mh, S, S))                    # additive key-padding mask: 0 on real keys, -1e4 on padded ones
    for w in self.layers:
      attn = mha(h, w["Wq"], w["bq"], w["Wk"], w["bk"], w["Wv"], w["bv"], w["Wo"], w["bo"], self.H, mask=m)
      h = (h + attn).layer_norm(w["ln1w"], w["ln1b"], self.eps)
      ff = h.linear(w["Wi"], w["bi"]).gelu().linear(w["Wd"], w["bd"])
      h = (h + ff).layer_norm(w["ln2w"], w["ln2b"], self.eps)
    return compile(h, int8=self.int8)

  def _embed(self, ids: np.ndarray) -> np.ndarray:
    """Host-side token + position + type embedding lookup, then LayerNorm."""  # gather is not an ANE op
    e = self.word[ids] + self.pos[self.pos_offset + np.arange(len(ids))]
    if self.typ is not None: e = e + self.typ[0]         # MPNet has no token-type embedding
    m = e.mean(-1, keepdims=True)
    v = ((e - m) ** 2).mean(-1, keepdims=True)
    return ((e - m) / np.sqrt(v + self.eps) * self.eln_w + self.eln_b).astype(np.float32)

  def __call__(self, texts, normalize: bool = True) -> np.ndarray:
    if isinstance(texts, str): texts = [texts]
    enc = [np.asarray(self.tok(t)["input_ids"], dtype=np.int64) for t in texts]
    # Pad the whole batch to its longest sequence and compile ONE program for that length -- the padded
    # keys are masked out per text, so a corpus of varied lengths shares one program instead of one each.
    S = max((len(ids) for ids in enc), default=1)
    net = self._cache.get(S) or self._cache.setdefault(S, self._build(S))
    pad_id = self.tok.pad_token_id or 0
    vecs = []
    for ids in enc:
      n = len(ids)
      padded = np.full(S, pad_id, dtype=np.int64); padded[:n] = ids
      mask = np.zeros((1, S, S), dtype=np.float32); mask[0, :, n:] = -1e4   # mask the padded key columns
      if self.rel_bias is not None:                       # MPNet: add the per-head relative-position bias
        if S not in self._bias_cache:
          self._bias_cache[S] = _mpnet_position_bias(self.rel_bias, S, self.n_buckets)
        mask = mask + self._bias_cache[S]
      states = net(self._embed(padded), mask)[:n]  # real-token states on the ANE (pads masked + dropped)
      if self.pooling == "cls":
        v = states[0]
      elif self.pooling == "max":
        v = states.max(0)
      else:
        v = states.mean(0)
      if normalize:
        v = _l2_normalizer(self.D)(v.reshape(1, self.D))[0]
      vecs.append(v)
    return np.asarray(vecs, dtype=np.float32)

CrossEncoder

Reranker: scores (query, passage) pairs with a BERT-, RoBERTa-, or DistilBERT-family sequence-classification model, the transformer running on the ANE. Mirrors sentence_transformers.CrossEncoder: CrossEncoder(name).predict([(query, passage), ...]) returns one relevance score per pair (raw logits -- order is what a reranker needs). Higher is more relevant.

BERT and RoBERTa share the encoder graph and _BERT_KEYS; DistilBERT uses the same graph operations with a family-specific key map. The families differ only in weight selection and host-side head plumbing, rather than requiring separate graph code paths:

  • Head. See _seqcls_head: BERT/RoBERTa use tanh; DistilBERT uses ReLU.
  • Position ids. See _position_ids: RoBERTa counts from padding_idx + 1; BERT and DistilBERT count from 0.
Source code in aneforge/models.py 
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
class CrossEncoder:
  """Reranker: scores (query, passage) pairs with a BERT-, RoBERTa-, or DistilBERT-family sequence-classification
  model, the transformer running on the ANE. Mirrors `sentence_transformers.CrossEncoder`:
  `CrossEncoder(name).predict([(query, passage), ...])` returns one relevance score per pair
  (raw logits -- order is what a reranker needs). Higher is more relevant.

  BERT and RoBERTa share the encoder graph and `_BERT_KEYS`; DistilBERT uses the same graph
  operations with a family-specific key map. The families differ only in weight selection and
  host-side head plumbing, rather than requiring separate graph code paths:

  - **Head.** See `_seqcls_head`: BERT/RoBERTa use tanh; DistilBERT uses ReLU.
  - **Position ids.** See `_position_ids`: RoBERTa counts from `padding_idx + 1`; BERT and DistilBERT count from 0."""

  def __init__(self, name: str, int8: bool = False) -> None:
    from transformers import AutoConfig, AutoModelForSequenceClassification, AutoTokenizer  # lazy
    cfg = AutoConfig.from_pretrained(name)
    self.tok = AutoTokenizer.from_pretrained(name)
    sd = AutoModelForSequenceClassification.from_pretrained(name).state_dict()
    g = lambda k: sd[k].detach().numpy().astype(np.float32)
    # The base transformer sits under a model-specific prefix (bert / roberta / ...).
    marker = ".embeddings.word_embeddings.weight"
    pref = next((k[: -len(marker)] for k in sd if k.endswith(marker)), None)
    head = _seqcls_head(sd, pref, cfg.model_type)
    if pref is None or head is None:
      raise ValueError(
        f"CrossEncoder: {name!r} is not a supported reranker -- it needs either a BERT-family head "
        "(a pooler plus `classifier`) or a RoBERTa-family one (`classifier.dense` plus "
        "`classifier.out_proj`), or a DistilBERT-family head (`pre_classifier` plus `classifier`).")
    roberta = head[0].startswith("classifier.")
    layer_prefix, layer_keys = _encoder_layer_spec(cfg.model_type)
    self.D, self.H = cfg.hidden_size, cfg.num_attention_heads
    # DistilBertConfig omits layer_norm_eps; its reference implementation uses 1e-12.
    self.L, self.eps, self.int8 = cfg.num_hidden_layers, getattr(cfg, "layer_norm_eps", 1e-12), int8
    self.pad_id = int(cfg.pad_token_id) if roberta else None   # None -> positions count from 0
    self.word = g(f"{pref}.embeddings.word_embeddings.weight")
    self.pos = g(f"{pref}.embeddings.position_embeddings.weight")
    typ_key = f"{pref}.embeddings.token_type_embeddings.weight"
    self._has_typ = typ_key in sd
    self.typ = g(typ_key) if self._has_typ else np.zeros((1, self.D), np.float32)
    self.eln_w, self.eln_b = g(f"{pref}.embeddings.LayerNorm.weight"), g(f"{pref}.embeddings.LayerNorm.bias")
    self.layers = [{k: g(f"{pref}.{layer_prefix.format(i=i)}" + v) for k, v in layer_keys.items()}
                   for i in range(self.L)]
    self.pool_w, self.pool_b, self.cls_w, self.cls_b = (g(k) for k in head)
    self._model_type = cfg.model_type
    self._cache: dict[int, Model | SegmentedModel] = {}

  def _build(self, S: int) -> Model | SegmentedModel:
    h = input((S, self.D))
    m = input((1, S, S))                     # additive key-padding mask: 0 on real keys, -1e4 on padded ones
    for w in self.layers:
      attn = mha(h, w["Wq"], w["bq"], w["Wk"], w["bk"], w["Wv"], w["bv"], w["Wo"], w["bo"], self.H, mask=m)
      h = (h + attn).layer_norm(w["ln1w"], w["ln1b"], self.eps)
      ff = h.linear(w["Wi"], w["bi"]).gelu().linear(w["Wd"], w["bd"])
      h = (h + ff).layer_norm(w["ln2w"], w["ln2b"], self.eps)
    return compile(h, int8=self.int8)

  def _embed(self, ids: np.ndarray, typ_ids: np.ndarray) -> np.ndarray:
    """Host-side token + position + segment embedding lookup, then LayerNorm."""
    e = self.word[ids] + self.pos[_position_ids(ids, self.pad_id)] + self.typ[typ_ids]
    m = e.mean(-1, keepdims=True)
    v = ((e - m) ** 2).mean(-1, keepdims=True)
    return ((e - m) / np.sqrt(v + self.eps) * self.eln_w + self.eln_b).astype(np.float32)

  def predict(self, pairs) -> np.ndarray:
    """`pairs` is a (query, passage) tuple or a list of them; returns a relevance score each."""
    if pairs and isinstance(pairs[0], str):     # a single (query, passage) pair
      pairs = [pairs]
    enc = [self.tok(q, p, truncation=True) for q, p in pairs]
    ids = [np.asarray(e["input_ids"], dtype=np.int64) for e in enc]
    # Pad the batch to its longest pair and compile ONE program; padded keys are masked per pair, so a
    # query's candidates (and the next query) share one program instead of one per pair length.
    S = max((len(i) for i in ids), default=1)
    net = self._cache.get(S) or self._cache.setdefault(S, self._build(S))
    pad_id = self.pad_id or 0
    scores = []
    for e, tid in zip(enc, ids):
      n = len(tid)
      typ = _token_type_ids(e, n, self._has_typ)
      padded = np.full(S, pad_id, dtype=np.int64); padded[:n] = tid
      typ_p = np.zeros(S, dtype=np.int64); typ_p[:n] = typ
      mask = np.zeros((1, S, S), dtype=np.float32); mask[0, :, n:] = -1e4   # mask the padded key columns
      cls = net(self._embed(padded, typ_p), mask)[0]                   # [CLS] / <s> state (row 0), on the ANE
      pooled = _seqcls_activation(cls @ self.pool_w.T + self.pool_b, self._model_type)
      logits = pooled @ self.cls_w.T + self.cls_b                      # [num_labels]
      scores.append(float(logits[0]))
    return np.asarray(scores, dtype=np.float32)

predict 

predict(pairs) -> np.ndarray

pairs is a (query, passage) tuple or a list of them; returns a relevance score each.

Source code in aneforge/models.py 
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
def predict(self, pairs) -> np.ndarray:
  """`pairs` is a (query, passage) tuple or a list of them; returns a relevance score each."""
  if pairs and isinstance(pairs[0], str):     # a single (query, passage) pair
    pairs = [pairs]
  enc = [self.tok(q, p, truncation=True) for q, p in pairs]
  ids = [np.asarray(e["input_ids"], dtype=np.int64) for e in enc]
  # Pad the batch to its longest pair and compile ONE program; padded keys are masked per pair, so a
  # query's candidates (and the next query) share one program instead of one per pair length.
  S = max((len(i) for i in ids), default=1)
  net = self._cache.get(S) or self._cache.setdefault(S, self._build(S))
  pad_id = self.pad_id or 0
  scores = []
  for e, tid in zip(enc, ids):
    n = len(tid)
    typ = _token_type_ids(e, n, self._has_typ)
    padded = np.full(S, pad_id, dtype=np.int64); padded[:n] = tid
    typ_p = np.zeros(S, dtype=np.int64); typ_p[:n] = typ
    mask = np.zeros((1, S, S), dtype=np.float32); mask[0, :, n:] = -1e4   # mask the padded key columns
    cls = net(self._embed(padded, typ_p), mask)[0]                   # [CLS] / <s> state (row 0), on the ANE
    pooled = _seqcls_activation(cls @ self.pool_w.T + self.pool_b, self._model_type)
    logits = pooled @ self.cls_w.T + self.cls_b                      # [num_labels]
    scores.append(float(logits[0]))
  return np.asarray(scores, dtype=np.float32)

ViT

Vision Transformer image classifier from Hugging Face, running on the ANE. ViT(name)(image) returns logits [1, num_labels]; .classify(image) returns top labels. Scope: ViT-family classifiers with a CLS token and a pre-norm encoder (ViTForImageClassification and compatible DeiT/BEiT-style models).

Source code in aneforge/models.py 
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
class ViT:
  """Vision Transformer image classifier from Hugging Face, running on the ANE.
  `ViT(name)(image)` returns logits [1, num_labels]; `.classify(image)` returns top labels.
  Scope: ViT-family classifiers with a CLS token and a pre-norm encoder (ViTForImageClassification
  and compatible DeiT/BEiT-style models)."""

  def __init__(self, name: str, int8: bool = False) -> None:
    from transformers import AutoConfig, AutoImageProcessor, AutoModelForImageClassification  # lazy
    cfg = AutoConfig.from_pretrained(name)
    self.proc = AutoImageProcessor.from_pretrained(name)
    sd = {k: v.detach().numpy().astype(np.float32)
          for k, v in AutoModelForImageClassification.from_pretrained(name).state_dict().items()}
    marker = ".embeddings.cls_token"
    pref = next((k[: -len(marker)] for k in sd if k.endswith(marker)), None)
    if pref is None or "classifier.weight" not in sd:
      raise ValueError(f"load_vit: {name!r} is not a ViT-family image classifier with a CLS token "
                       "(needs <base>.embeddings.cls_token and a classifier head).")
    self.int8 = int8
    self.D, self.H, self.L = cfg.hidden_size, cfg.num_attention_heads, cfg.num_hidden_layers
    self.P, self.img, self.eps = cfg.patch_size, cfg.image_size, cfg.layer_norm_eps
    self.id2label = getattr(cfg, "id2label", {}) or {}
    self._pref, self._g = pref, (lambda k: sd[k])
    self._lfmt, self._km = _vit_layer_keys(sd, pref)
    n = (self.img // self.P) ** 2                                      # number of patches
    self._cls = sd[f"{pref}.embeddings.cls_token"].reshape(1, self.D)
    self._pos = sd[f"{pref}.embeddings.position_embeddings"].reshape(n + 1, self.D)
    self._model = self._build(n)

  def _build(self, n: int) -> Model | SegmentedModel:
    g, pref, D, eps, H = self._g, self._pref, self.D, self.eps, self.H
    x = input((1, 3, self.img, self.img)); cls = input((1, D)); pos = input((n + 1, D))
    # a strided PxP patch conv is walled on the ANE -> space_to_depth(P) + a 1x1 conv
    w_pe = np.ascontiguousarray(
      g(f"{pref}.embeddings.patch_embeddings.projection.weight").transpose(0, 2, 3, 1)).reshape(D, -1, 1, 1)
    h = conv(space_to_depth(x, self.P), w_pe,
             bias=g(f"{pref}.embeddings.patch_embeddings.projection.bias"))
    patches = h.reshape(1, D, n).transpose([0, 2, 1]).reshape(n, D)
    seq = concat([cls, patches], axis=0) + pos
    km = self._km
    for i in range(self.L):
      p = self._lfmt.format(i=i)
      xn = seq.layer_norm(g(p + km["ln1"] + ".weight"), g(p + km["ln1"] + ".bias"), eps)
      attn = mha(xn,
                 g(p + km["q"] + ".weight"), g(p + km["q"] + ".bias"),
                 g(p + km["k"] + ".weight"), g(p + km["k"] + ".bias"),
                 g(p + km["v"] + ".weight"), g(p + km["v"] + ".bias"),
                 g(p + km["o"] + ".weight"), g(p + km["o"] + ".bias"), H)
      seq = seq + attn
      yn = seq.layer_norm(g(p + km["ln2"] + ".weight"), g(p + km["ln2"] + ".bias"), eps)
      y = yn.linear(g(p + km["fc1"] + ".weight"), g(p + km["fc1"] + ".bias")).gelu()
      seq = seq + y.linear(g(p + km["fc2"] + ".weight"), g(p + km["fc2"] + ".bias"))
    seq = seq.layer_norm(g(f"{pref}.layernorm.weight"), g(f"{pref}.layernorm.bias"), eps)
    return compile(_row0(seq).linear(g("classifier.weight"), g("classifier.bias")), int8=self.int8)

  def _pixels(self, image) -> np.ndarray:
    if isinstance(image, np.ndarray) and image.ndim == 4:
      return image.astype(np.float32)
    return np.asarray(self.proc(image, return_tensors="np")["pixel_values"], np.float32)

  def __call__(self, image) -> np.ndarray:
    return np.asarray(self._model(self._pixels(image), self._cls, self._pos), np.float32)

  def classify(self, image, top_k: int = 5):
    """Top-k (label, logit) for an image (PIL, path, or preprocessed pixel array)."""
    logits = self(image)[0]
    return [(self.id2label.get(int(i), str(int(i))), float(logits[i]))
            for i in np.argsort(-logits)[:top_k]]

  def release(self) -> None: self._model.release()

  @property
  def n_ops(self) -> int: return self._model.n_ops

classify 

classify(image, top_k: int = 5)

Top-k (label, logit) for an image (PIL, path, or preprocessed pixel array).

Source code in aneforge/models.py 
553
554
555
556
557
def classify(self, image, top_k: int = 5):
  """Top-k (label, logit) for an image (PIL, path, or preprocessed pixel array)."""
  logits = self(image)[0]
  return [(self.id2label.get(int(i), str(int(i))), float(logits[i]))
          for i in np.argsort(-logits)[:top_k]]

GPT2

GPT-2 causal LM from Hugging Face, running on the ANE via the unified LLM runner with resident KV-cache decode, LayerNorm, and learned positional embeddings. Activations fp16; int8=True quantizes weights per-channel int8. GPT2(name)(ids) -> logits [S, vocab]; .generate(prompt, K) autoregressively generates K tokens using the resident KV cache.

Source code in aneforge/models.py 
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
class GPT2:
  """GPT-2 causal LM from Hugging Face, running on the ANE via the unified LLM runner with
  resident KV-cache decode, LayerNorm, and learned positional embeddings. Activations fp16;
  `int8=True` quantizes weights per-channel int8. `GPT2(name)(ids)` -> logits [S, vocab];
  `.generate(prompt, K)` autoregressively generates K tokens using the resident KV cache."""

  def __init__(self, name: str, int8: bool = False, max_layers: int | None = None) -> None:
    from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer  # lazy
    from .llm import _gpt2_adapter, LlamaPrefill, ModelType
    cfg_hf = AutoConfig.from_pretrained(name)
    if cfg_hf.model_type != ModelType.GPT2:
      raise ValueError(f"load_gpt2: {name!r} is not a GPT-2 model (model_type={cfg_hf.model_type!r})")
    self.tok = AutoTokenizer.from_pretrained(name)
    sd = {k: v.detach().numpy().astype(np.float32)
          for k, v in AutoModelForCausalLM.from_pretrained(name).state_dict().items()}
    cfg, weights = _gpt2_adapter(cfg_hf, sd)
    if max_layers is not None:
      if not 1 <= max_layers <= cfg.n_layers:
        raise ValueError(f"load_gpt2: max_layers must be in [1, {cfg.n_layers}], got {max_layers}")
      cfg.n_layers = max_layers
      cfg.layers = cfg.layers[:max_layers]
      weights["layers"] = weights["layers"][:max_layers]
    self.cfg = cfg
    self.w = weights
    self.D, self.H, self.L = cfg.dim, cfg.n_heads, cfg.n_layers
    self.dh = cfg.dh
    self.Dff = cfg.ffn_dim
    self.eps = cfg.norm_eps
    self.int8 = int8
    self.wte = weights["embed"]
    self.wpe = weights["wpe"]
    self.runner = LlamaPrefill(cfg, weights, compress="int8" if int8 else None)

  def __call__(self, ids) -> np.ndarray:
    """1-D token ids -> logits [S, vocab]."""
    ids = np.asarray(ids, dtype=np.int64)
    if ids.ndim != 1:
      raise ValueError(f"GPT2.__call__ expects 1-D token ids, got shape {ids.shape}")
    hidden = self.runner._hidden(ids)
    return hidden @ np.asarray(self.w["lm_head"]).T

  def generate(self, prompt: str, max_new_tokens: int = 16, **kwargs) -> list[int]:
    """Autoregressive generation with resident KV-cache."""
    base = np.asarray(self.tok.encode(prompt), dtype=np.int64)
    return self.runner.generate(base, max_new_tokens=max_new_tokens, **kwargs)

  def generate_text(self, prompt: str, max_new_tokens: int = 16) -> str:
    """Greedy-decode and return the newly generated tokens as text."""
    return self.tok.decode(self.generate(prompt, max_new_tokens))

  def release(self) -> None:
    self.runner.release()

generate 

generate(prompt: str, max_new_tokens: int = 16, **kwargs) -> list[int]

Autoregressive generation with resident KV-cache.

Source code in aneforge/models.py 
650
651
652
653
def generate(self, prompt: str, max_new_tokens: int = 16, **kwargs) -> list[int]:
  """Autoregressive generation with resident KV-cache."""
  base = np.asarray(self.tok.encode(prompt), dtype=np.int64)
  return self.runner.generate(base, max_new_tokens=max_new_tokens, **kwargs)

generate_text 

generate_text(prompt: str, max_new_tokens: int = 16) -> str

Greedy-decode and return the newly generated tokens as text.

Source code in aneforge/models.py 
655
656
657
def generate_text(self, prompt: str, max_new_tokens: int = 16) -> str:
  """Greedy-decode and return the newly generated tokens as text."""
  return self.tok.decode(self.generate(prompt, max_new_tokens))

CLIP

CLIP dual-encoder model from Hugging Face (CLIPModel), running both vision and text encoders on the ANE. .encode_image(image) -> [1, proj_dim] normalized image embedding. .encode_text(texts) -> [N, proj_dim] normalized text embeddings. .classify(image, candidate_labels) -> sorted list of (label, probability).

Source code in aneforge/models.py 
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
class CLIP:
  """CLIP dual-encoder model from Hugging Face (CLIPModel), running both vision and text encoders on the ANE.
  `.encode_image(image)` -> [1, proj_dim] normalized image embedding.
  `.encode_text(texts)` -> [N, proj_dim] normalized text embeddings.
  `.classify(image, candidate_labels)` -> sorted list of (label, probability)."""

  def __init__(self, name: str = "openai/clip-vit-base-patch32", int8: bool = False) -> None:
    from transformers import AutoConfig, AutoImageProcessor, AutoTokenizer, CLIPModel  # lazy
    cfg = AutoConfig.from_pretrained(name)
    if getattr(cfg, "model_type", None) != "clip":
      raise ValueError(f"load_clip: {name!r} is not a CLIP model (model_type={getattr(cfg, 'model_type', None)!r})")
    self.tok = AutoTokenizer.from_pretrained(name)
    try:
      self.proc = AutoImageProcessor.from_pretrained(name)
    except Exception:
      self.proc = None
    sd = {k: v.detach().numpy().astype(np.float32)
          for k, v in CLIPModel.from_pretrained(name).state_dict().items()}
    self.int8 = int8
    v_cfg, t_cfg = cfg.vision_config, cfg.text_config
    self.Dv, self.Hv, self.Lv = v_cfg.hidden_size, v_cfg.num_attention_heads, v_cfg.num_hidden_layers
    self.Pv, self.img, self.v_eps = v_cfg.patch_size, v_cfg.image_size, v_cfg.layer_norm_eps
    self.Dt, self.Ht, self.Lt = t_cfg.hidden_size, t_cfg.num_attention_heads, t_cfg.num_hidden_layers
    self.t_eps = t_cfg.layer_norm_eps
    self.St = int(getattr(t_cfg, "max_position_embeddings", 77))
    self.proj_dim = getattr(cfg, "projection_dim", 512)
    self.logit_scale = float(np.exp(sd.get("logit_scale", np.log(100.0))))

    # Vision weights & embeddings
    self.nv = (self.img // self.Pv) ** 2
    self._cls_v = sd["vision_model.embeddings.class_embedding"].reshape(1, self.Dv)
    self._pos_v = sd["vision_model.embeddings.position_embedding.weight"].reshape(self.nv + 1, self.Dv)
    self._w_pe_v = np.ascontiguousarray(
      sd["vision_model.embeddings.patch_embedding.weight"].transpose(0, 2, 3, 1)).reshape(self.Dv, -1, 1, 1)

    # Text weights & embeddings needed at runtime
    self.token_embed = sd["text_model.embeddings.token_embedding.weight"]
    self.pos_embed = sd["text_model.embeddings.position_embedding.weight"]
    self.text_proj = sd["text_projection.weight"]

    self._v_model = self._build_vision(sd)
    self._t_model = self._build_text(sd, self.St)

  def _build_vision(self, sd: dict) -> Model | SegmentedModel:
    Dv, Hv, Lv, Pv, img, eps = self.Dv, self.Hv, self.Lv, self.Pv, self.img, self.v_eps
    nv = self.nv
    x_img = input((1, 3, img, img)); cls_in = input((1, Dv)); pos_in = input((nv + 1, Dv))
    h = conv(space_to_depth(x_img, Pv), self._w_pe_v, bias=None)
    patches = h.reshape(1, Dv, nv).transpose([0, 2, 1]).reshape(nv, Dv)
    seq = concat([cls_in, patches], axis=0) + pos_in
    seq = seq.layer_norm(sd["vision_model.pre_layrnorm.weight"], sd["vision_model.pre_layrnorm.bias"], eps)
    for i in range(Lv):
      p = f"vision_model.encoder.layers.{i}."
      xn = seq.layer_norm(sd[p + "layer_norm1.weight"], sd[p + "layer_norm1.bias"], eps)
      attn = mha(xn,
                 sd[p + "self_attn.q_proj.weight"], sd[p + "self_attn.q_proj.bias"],
                 sd[p + "self_attn.k_proj.weight"], sd[p + "self_attn.k_proj.bias"],
                 sd[p + "self_attn.v_proj.weight"], sd[p + "self_attn.v_proj.bias"],
                 sd[p + "self_attn.out_proj.weight"], sd[p + "self_attn.out_proj.bias"], Hv)
      seq = seq + attn
      yn = seq.layer_norm(sd[p + "layer_norm2.weight"], sd[p + "layer_norm2.bias"], eps)
      mlp = _quick_gelu(yn.linear(sd[p + "mlp.fc1.weight"], sd[p + "mlp.fc1.bias"])).linear(
        sd[p + "mlp.fc2.weight"], sd[p + "mlp.fc2.bias"])
      seq = seq + mlp
    cls_out = _row0(seq).layer_norm(sd["vision_model.post_layernorm.weight"], sd["vision_model.post_layernorm.bias"], eps)
    v_proj = cls_out.linear(sd["visual_projection.weight"])
    return compile(v_proj.l2_norm(axis=-1), int8=self.int8)

  def _build_text(self, sd: dict, S: int) -> Model | SegmentedModel:
    Dt, Ht, Lt, eps = self.Dt, self.Ht, self.Lt, self.t_eps
    dht = Dt // Ht
    x_txt = input((S, Dt))
    seq = x_txt
    for i in range(Lt):
      p = f"text_model.encoder.layers.{i}."
      xn = seq.layer_norm(sd[p + "layer_norm1.weight"], sd[p + "layer_norm1.bias"], eps)
      q = xn.linear(sd[p + "self_attn.q_proj.weight"], sd[p + "self_attn.q_proj.bias"]).reshape(1, S, Ht, dht).transpose([0, 2, 1, 3])
      k = xn.linear(sd[p + "self_attn.k_proj.weight"], sd[p + "self_attn.k_proj.bias"]).reshape(1, S, Ht, dht).transpose([0, 2, 1, 3])
      v = xn.linear(sd[p + "self_attn.v_proj.weight"], sd[p + "self_attn.v_proj.bias"]).reshape(1, S, Ht, dht).transpose([0, 2, 1, 3])
      attn = _causal_attn(q, k, v).transpose([0, 2, 1, 3]).reshape(S, Dt)
      seq = seq + attn.linear(sd[p + "self_attn.out_proj.weight"], sd[p + "self_attn.out_proj.bias"])
      yn = seq.layer_norm(sd[p + "layer_norm2.weight"], sd[p + "layer_norm2.bias"], eps)
      mlp = _quick_gelu(yn.linear(sd[p + "mlp.fc1.weight"], sd[p + "mlp.fc1.bias"])).linear(
        sd[p + "mlp.fc2.weight"], sd[p + "mlp.fc2.bias"])
      seq = seq + mlp
    seq = seq.layer_norm(sd["text_model.final_layer_norm.weight"], sd["text_model.final_layer_norm.bias"], eps)
    return compile(seq, int8=self.int8)

  def _pixels(self, image) -> np.ndarray:
    if isinstance(image, np.ndarray) and image.ndim == 4 and image.shape[1:] == (3, self.img, self.img):
      return image.astype(np.float32)
    if self.proc is not None:
      return np.asarray(self.proc(images=image, return_tensors="np")["pixel_values"], np.float32)
    # Fallback normalization: PIL if present, else pure-numpy resampling
    try:
      import PIL.Image
      if not isinstance(image, PIL.Image.Image):
        arr = np.asarray(image)
        if arr.ndim == 3 and arr.shape[0] == 3: arr = arr.transpose(1, 2, 0)
        u8 = np.asarray(arr if arr.max() > 1.0 else arr * 255.0, dtype=np.uint8)
        image = PIL.Image.fromarray(u8)
      image = image.resize((self.img, self.img), PIL.Image.Resampling.BICUBIC)
      arr = np.asarray(image, dtype=np.float32).transpose(2, 0, 1) / 255.0
    except Exception:
      arr = np.asarray(image, dtype=np.float32)
      if arr.ndim == 3 and arr.shape[-1] == 3: arr = arr.transpose(2, 0, 1)
      if arr.ndim == 2: arr = np.repeat(arr[None], 3, axis=0)
      if arr.shape[1:] != (self.img, self.img):
        H, W = arr.shape[1], arr.shape[2]
        r = np.linspace(0, H - 1, self.img).astype(int)
        c = np.linspace(0, W - 1, self.img).astype(int)
        arr = arr[:, r[:, None], c[None, :]]
      if arr.max() > 1.0: arr /= 255.0
    mean = np.array([0.48145466, 0.4578275, 0.40821073], dtype=np.float32).reshape(3, 1, 1)
    std = np.array([0.26862954, 0.26130258, 0.27577711], dtype=np.float32).reshape(3, 1, 1)
    return ((arr - mean) / std)[None].astype(np.float32)

  def encode_image(self, image) -> np.ndarray:
    """Encode an image -> L2-normalized embedding vector [1, proj_dim]."""
    px = self._pixels(image)
    return np.asarray(self._v_model(px, self._cls_v, self._pos_v), np.float32)

  def encode_text(self, texts: str | list[str]) -> np.ndarray:
    """Encode text string(s) -> L2-normalized embedding vectors [N, proj_dim]."""
    if isinstance(texts, str): texts = [texts]
    ids = np.asarray(self.tok(texts, padding="max_length", max_length=self.St, truncation=True, return_tensors="np")["input_ids"], dtype=np.int64)
    vecs = []
    for row in ids:
      eot_pos = int(row.argmax())
      emb = self.token_embed[row] + self.pos_embed[:self.St]
      hidden = np.asarray(self._t_model(emb), np.float32)
      eot = hidden[eot_pos:eot_pos + 1]
      proj = eot @ self.text_proj.T
      norm_proj = proj / np.linalg.norm(proj, axis=-1, keepdims=True)
      vecs.append(norm_proj[0])
    return np.asarray(vecs, dtype=np.float32)

  def classify(self, image, candidate_labels: list[str]) -> list[tuple[str, float]]:
    """Zero-shot image classification: score image against candidate labels, returning sorted (label, prob) pairs."""
    img_feat = self.encode_image(image)
    txt_feat = self.encode_text(candidate_labels)
    logits = (img_feat @ txt_feat.T * self.logit_scale)[0]
    exp = np.exp(logits - np.max(logits))
    probs = exp / np.sum(exp)
    ranked = sorted(zip(probs, candidate_labels), key=lambda p: p[0], reverse=True)
    return [(label, float(prob)) for prob, label in ranked]

  def release(self) -> None:
    self._v_model.release()
    self._t_model.release()

encode_image 

encode_image(image) -> np.ndarray

Encode an image -> L2-normalized embedding vector [1, proj_dim].

Source code in aneforge/models.py 
798
799
800
801
def encode_image(self, image) -> np.ndarray:
  """Encode an image -> L2-normalized embedding vector [1, proj_dim]."""
  px = self._pixels(image)
  return np.asarray(self._v_model(px, self._cls_v, self._pos_v), np.float32)

encode_text 

encode_text(texts: str | list[str]) -> np.ndarray

Encode text string(s) -> L2-normalized embedding vectors [N, proj_dim].

Source code in aneforge/models.py 
803
804
805
806
807
808
809
810
811
812
813
814
815
816
def encode_text(self, texts: str | list[str]) -> np.ndarray:
  """Encode text string(s) -> L2-normalized embedding vectors [N, proj_dim]."""
  if isinstance(texts, str): texts = [texts]
  ids = np.asarray(self.tok(texts, padding="max_length", max_length=self.St, truncation=True, return_tensors="np")["input_ids"], dtype=np.int64)
  vecs = []
  for row in ids:
    eot_pos = int(row.argmax())
    emb = self.token_embed[row] + self.pos_embed[:self.St]
    hidden = np.asarray(self._t_model(emb), np.float32)
    eot = hidden[eot_pos:eot_pos + 1]
    proj = eot @ self.text_proj.T
    norm_proj = proj / np.linalg.norm(proj, axis=-1, keepdims=True)
    vecs.append(norm_proj[0])
  return np.asarray(vecs, dtype=np.float32)

classify 

classify(image, candidate_labels: list[str]) -> list[tuple[str, float]]

Zero-shot image classification: score image against candidate labels, returning sorted (label, prob) pairs.

Source code in aneforge/models.py 
818
819
820
821
822
823
824
825
826
def classify(self, image, candidate_labels: list[str]) -> list[tuple[str, float]]:
  """Zero-shot image classification: score image against candidate labels, returning sorted (label, prob) pairs."""
  img_feat = self.encode_image(image)
  txt_feat = self.encode_text(candidate_labels)
  logits = (img_feat @ txt_feat.T * self.logit_scale)[0]
  exp = np.exp(logits - np.max(logits))
  probs = exp / np.sum(exp)
  ranked = sorted(zip(probs, candidate_labels), key=lambda p: p[0], reverse=True)
  return [(label, float(prob)) for prob, label in ranked]

Whisper

OpenAI Whisper (encoder-decoder ASR) on the ANE: one fused encoder program per clip, and a single-token decoder program per greedy step against a resident KV cache. Greedy matches HF generate. Host-side log-mel and tokenization only. Default whisper-base.en. See docs/developer/models.md. .transcribe(audio) -> str; .encode(audio) -> audio features [1500, 512].

Source code in aneforge/models.py 
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
class Whisper:
  """OpenAI Whisper (encoder-decoder ASR) on the ANE: one fused encoder program per clip, and a single-token
  decoder program per greedy step against a resident KV cache. Greedy matches HF `generate`. Host-side log-mel
  and tokenization only. Default `whisper-base.en`. See docs/developer/models.md.
  `.transcribe(audio)` -> str; `.encode(audio)` -> audio features [1500, 512]."""

  def __init__(self, name: str = "openai/whisper-base.en") -> None:
    from transformers import AutoConfig, GenerationConfig, WhisperForConditionalGeneration, WhisperProcessor  # lazy
    cfg = AutoConfig.from_pretrained(name)
    if getattr(cfg, "model_type", None) != "whisper":
      raise ValueError(f"load_whisper: {name!r} is not a Whisper model (model_type={cfg.model_type!r})")
    self.proc = WhisperProcessor.from_pretrained(name)
    gen = GenerationConfig.from_pretrained(name)
    sd = {k: v.detach().numpy().astype(np.float16)
          for k, v in WhisperForConditionalGeneration.from_pretrained(name).state_dict().items()}
    self.D, self.H = cfg.d_model, cfg.encoder_attention_heads
    self.dec_heads = cfg.decoder_attention_heads
    self.n_mels, self.vocab = cfg.num_mel_bins, cfg.vocab_size
    self.max_dec = cfg.max_target_positions                   # resident KV-cache length (== HF's max decode length)
    # decode prompt + logit rules from the generation config, so a non-.en checkpoint gets the right ids
    forced = getattr(gen, "forced_decoder_ids", None) or []
    self.sot = [cfg.decoder_start_token_id] + [int(t) for _, t in sorted(forced)]      # e.g. [sot, notimestamps]
    self.suppress = np.asarray(gen.suppress_tokens or [], dtype=np.int64)              # forced to -inf every step
    self.begin_suppress = np.asarray(gen.begin_suppress_tokens or [], dtype=np.int64)  # ... only the first token
    self.eot = cfg.eos_token_id                               # <|endoftext|>
    g = lambda k: sd[k]
    self._enc_conv1 = g("model.encoder.conv1.weight").reshape(self.D, self.n_mels, 1, 3)
    self._enc_conv1_b = g("model.encoder.conv1.bias")
    self._enc_conv2 = g("model.encoder.conv2.weight").reshape(self.D, self.D, 1, 3)
    self._enc_conv2_b = g("model.encoder.conv2.bias")
    self._enc_pos = g("model.encoder.embed_positions.weight")           # [1500, D] sinusoidal
    self._enc_lnw, self._enc_lnb = g("model.encoder.layer_norm.weight"), g("model.encoder.layer_norm.bias")
    self._enc_layers = _whisper_layers(sd, "encoder", cfg.encoder_layers)
    self._dec_tok = g("model.decoder.embed_tokens.weight")              # [vocab, D], tied to proj_out
    self._dec_pos = g("model.decoder.embed_positions.weight")          # [max_target_positions, D] learned
    self._dec_lnw, self._dec_lnb = g("model.decoder.layer_norm.weight"), g("model.decoder.layer_norm.bias")
    self._dec_layers = _whisper_layers(sd, "decoder", cfg.decoder_layers)
    self._encoder = self._build_encoder()
    self._decoder = None                                       # decode program built on first transcribe

  def _build_encoder(self) -> Model | SegmentedModel:
    mel = input((1, self.n_mels, 1, 3000))                    # log-mel as a 1xT "image" for the 1D convs
    h = conv(mel, self._enc_conv1, pad=(0, 0, 1, 1), bias=self._enc_conv1_b).gelu()      # [1, D, 1, 3000]
    h = conv(h, self._enc_conv2, stride=2, pad=(0, 0, 1, 1), bias=self._enc_conv2_b).gelu()  # [1, D, 1, 1500]
    h = h.reshape(self.D, 1500).transpose([1, 0]) + self._enc_pos    # [1500, D] + sinusoidal positions
    for w in self._enc_layers:
      a = mha(h.layer_norm(w["ln1w"], w["ln1b"]), w["Wq"], w["bq"], w["Wk"], None,
              w["Wv"], w["bv"], w["Wo"], w["bo"], self.H)
      h = h + a
      f = h.layer_norm(w["ln2w"], w["ln2b"]).linear(w["Wi"], w["bi"]).gelu().linear(w["Wd"], w["bd"])
      h = h + f
    return compile(h.layer_norm(self._enc_lnw, self._enc_lnb))

  def _features(self, audio: np.ndarray) -> np.ndarray:
    """Host log-mel [n_mels, 3000] via the Whisper feature extractor (no ANE)."""
    audio = np.asarray(audio, dtype=np.float32)
    mel = self.proc.feature_extractor(audio, sampling_rate=16000, return_tensors="np")["input_features"]
    return np.asarray(mel, np.float32).reshape(self.n_mels, 3000)

  def encode(self, audio: np.ndarray) -> np.ndarray:
    """Audio features [1500, D] on the ANE for a 16 kHz mono clip (truncated/padded to 30 s)."""
    mel = self._features(audio).reshape(1, self.n_mels, 1, 3000)
    return np.asarray(self._encoder(mel), np.float32)

  def _build_decode_step(self) -> dict:
    """Single-token decode program with a resident KV cache (built once): self-attn one-hot cache write
    (`Kout = Kin*inv + k*oh`), cross-attn to per-clip audio K/V, caches aliased resident via `share_buffer`."""
    M, D, H, dh = self.max_dec, self.D, self.dec_heads, self.D // self.dec_heads
    scale = 1.0 / dh ** 0.5
    x = input((1, D))                                          # this step's token+position embedding
    oh = input((1, M, 1)); inv = input((1, M, 1))             # one-hot write to the current position (inv = 1 - oh)
    mask = input((1, 1, M))                                    # causal: 0 on keys <= pos, -1e4 beyond
    h = x
    pairs, cross = [], []
    for w in self._dec_layers:
      Kin = input((H, M, dh)); Vin = input((H, M, dh))         # resident self-attn cache
      CKin = input((1500, D)); CVin = input((1500, D))         # resident cross-attn K/V (set once per clip)
      xn = h.layer_norm(w["ln1w"], w["ln1b"])
      q = xn.linear(w["Wq"], w["bq"]).reshape(H, 1, dh)
      k = xn.linear(w["Wk"], None).reshape(H, 1, dh)           # k_proj has no bias
      v = xn.linear(w["Wv"], w["bv"]).reshape(H, 1, dh)
      Kout = Kin * inv + k * oh; Vout = Vin * inv + v * oh      # positional write into the resident cache
      sc = ((q @ Kout.transpose([0, 2, 1])) * scale + mask).softmax(-1)   # [H, 1, M]
      a = (sc @ Vout).transpose([1, 0, 2]).reshape(1, D)                   # [H,1,dh] -> [1, D]
      h = h + a.linear(w["Wo"], w["bo"])
      pairs += [(Kout, Kin), (Vout, Vin)]
      cn = h.layer_norm(w["cln_w"], w["cln_b"])
      cq = cn.linear(w["CWq"], w["Cbq"]).reshape(H, 1, dh)
      ck = CKin.reshape(1500, H, dh).transpose([1, 0, 2])       # [H, 1500, dh]
      cv = CVin.reshape(1500, H, dh).transpose([1, 0, 2])
      cs = ((cq @ ck.transpose([0, 2, 1])) * scale).softmax(-1)             # [H, 1, 1500]
      ca = (cs @ cv).transpose([1, 0, 2]).reshape(1, D)
      h = h + ca.linear(w["CWo"], w["Cbo"])
      cross.append((CKin, CVin))
      f = h.layer_norm(w["ln2w"], w["ln2b"]).linear(w["Wi"], w["bi"]).gelu().linear(w["Wd"], w["bd"])
      h = h + f
    h = h.layer_norm(self._dec_lnw, self._dec_lnb)
    tiles = _lm_head_tiles(h, self._dec_tok)                   # tied head -> logits [1, vocab] (tiled by family)
    net = compile_multi(tiles + [o for o, _ in pairs])
    inm = {id(t): n for t, n in net.input_ports}; om = dict(net.output_ports)
    for o, i in pairs:                                         # alias each cache output back to its input (resident)
      net.prog.share_buffer(0, om[o], 0, inm[id(i)])
    return {"net": net, "x": inm[id(x)], "oh": inm[id(oh)], "inv": inm[id(inv)], "mask": inm[id(mask)],
            "logits": [om[t] for t in tiles], "states": {inm[id(i)]: i.shape for _, i in pairs},
            "cross": [(inm[id(ck)], inm[id(cv)]) for ck, cv in cross]}

  def transcribe(self, audio: np.ndarray) -> str:
    """Greedy transcript of a 16 kHz mono clip, both towers on the ANE. Matches HF `generate` up to `max_dec` tokens."""
    if self._decoder is None:
      self._decoder = self._build_decode_step()
    d = self._decoder; prog = d["net"].prog
    M = self.max_dec
    feats = self.encode(audio)                                # [1500, D] fp32 (fast BLAS for the projection below)
    for name, shape in d["states"].items():                   # reset the self-attn caches for this clip
      prog.set_input(name, np.zeros(shape, np.float16))
    for (ckn, cvn), w in zip(d["cross"], self._dec_layers):   # cross-attn K/V, computed once and held resident
      ck = feats @ w["CWk"].T.astype(np.float32)              # fp32 matmul (fp16 numpy has no BLAS path); k_proj no bias
      cv = feats @ w["CWv"].T.astype(np.float32) + w["Cbv"].astype(np.float32)
      prog.set_input(ckn, ck.astype(np.float16)); prog.set_input(cvn, cv.astype(np.float16))

    def step(tok: int, pos: int) -> np.ndarray:
      emb = (self._dec_tok[tok] + self._dec_pos[pos]).astype(np.float16)[None]   # [1, D]
      ohv = np.zeros((1, M, 1), np.float16); ohv[0, pos, 0] = 1.0
      invv = np.ones((1, M, 1), np.float16); invv[0, pos, 0] = 0.0
      mv = np.full((1, 1, M), -1e4, np.float16); mv[..., :pos + 1] = 0.0
      prog.set_input(d["x"], emb); prog.set_input(d["oh"], ohv)
      prog.set_input(d["inv"], invv); prog.set_input(d["mask"], mv)
      prog.execute()
      logits = np.concatenate([np.asarray(prog.read_output(n), np.float32) for n in d["logits"]], axis=1)[0]
      logits[self.suppress] = -np.inf                          # HF SuppressTokensLogitsProcessor (every step)
      return logits

    ids = list(self.sot)
    for p, tok in enumerate(ids[:-1]):                        # prime the cache with the start tokens
      step(tok, p)
    cur, pos = ids[-1], len(ids) - 1
    first = True
    while len(ids) < M:
      logits = step(cur, pos)
      if first:                                                # HF begin-suppress: no leading space / immediate EOT
        logits[self.begin_suppress] = -np.inf; first = False
      nxt = int(np.argmax(logits))
      ids.append(nxt)
      if nxt == self.eot:
        break
      cur, pos = nxt, pos + 1
    return self.proc.tokenizer.decode(ids[len(self.sot):], skip_special_tokens=True)

  def release(self) -> None:
    self._encoder.release()
    if self._decoder is not None:
      self._decoder["net"].release()

encode 

encode(audio: ndarray) -> np.ndarray

Audio features [1500, D] on the ANE for a 16 kHz mono clip (truncated/padded to 30 s).

Source code in aneforge/models.py 
898
899
900
901
def encode(self, audio: np.ndarray) -> np.ndarray:
  """Audio features [1500, D] on the ANE for a 16 kHz mono clip (truncated/padded to 30 s)."""
  mel = self._features(audio).reshape(1, self.n_mels, 1, 3000)
  return np.asarray(self._encoder(mel), np.float32)

transcribe 

transcribe(audio: ndarray) -> str

Greedy transcript of a 16 kHz mono clip, both towers on the ANE. Matches HF generate up to max_dec tokens.

Source code in aneforge/models.py 
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
def transcribe(self, audio: np.ndarray) -> str:
  """Greedy transcript of a 16 kHz mono clip, both towers on the ANE. Matches HF `generate` up to `max_dec` tokens."""
  if self._decoder is None:
    self._decoder = self._build_decode_step()
  d = self._decoder; prog = d["net"].prog
  M = self.max_dec
  feats = self.encode(audio)                                # [1500, D] fp32 (fast BLAS for the projection below)
  for name, shape in d["states"].items():                   # reset the self-attn caches for this clip
    prog.set_input(name, np.zeros(shape, np.float16))
  for (ckn, cvn), w in zip(d["cross"], self._dec_layers):   # cross-attn K/V, computed once and held resident
    ck = feats @ w["CWk"].T.astype(np.float32)              # fp32 matmul (fp16 numpy has no BLAS path); k_proj no bias
    cv = feats @ w["CWv"].T.astype(np.float32) + w["Cbv"].astype(np.float32)
    prog.set_input(ckn, ck.astype(np.float16)); prog.set_input(cvn, cv.astype(np.float16))

  def step(tok: int, pos: int) -> np.ndarray:
    emb = (self._dec_tok[tok] + self._dec_pos[pos]).astype(np.float16)[None]   # [1, D]
    ohv = np.zeros((1, M, 1), np.float16); ohv[0, pos, 0] = 1.0
    invv = np.ones((1, M, 1), np.float16); invv[0, pos, 0] = 0.0
    mv = np.full((1, 1, M), -1e4, np.float16); mv[..., :pos + 1] = 0.0
    prog.set_input(d["x"], emb); prog.set_input(d["oh"], ohv)
    prog.set_input(d["inv"], invv); prog.set_input(d["mask"], mv)
    prog.execute()
    logits = np.concatenate([np.asarray(prog.read_output(n), np.float32) for n in d["logits"]], axis=1)[0]
    logits[self.suppress] = -np.inf                          # HF SuppressTokensLogitsProcessor (every step)
    return logits

  ids = list(self.sot)
  for p, tok in enumerate(ids[:-1]):                        # prime the cache with the start tokens
    step(tok, p)
  cur, pos = ids[-1], len(ids) - 1
  first = True
  while len(ids) < M:
    logits = step(cur, pos)
    if first:                                                # HF begin-suppress: no leading space / immediate EOT
      logits[self.begin_suppress] = -np.inf; first = False
    nxt = int(np.argmax(logits))
    ids.append(nxt)
    if nxt == self.eot:
      break
    cur, pos = nxt, pos + 1
  return self.proc.tokenizer.decode(ids[len(self.sot):], skip_special_tokens=True)

load 

load(name: str, int8: bool = False, pooling: str = 'mean') -> 'Encoder'

Load a BERT-family sentence encoder from HF weights as an ANE embedder; pooling in mean/cls/max.

Source code in aneforge/models.py 
23
24
25
def load(name: str, int8: bool = False, pooling: str = "mean") -> "Encoder":
  """Load a BERT-family sentence encoder from HF weights as an ANE embedder; `pooling` in mean/cls/max."""
  return Encoder(name, int8=int8, pooling=pooling)

load_resnet 

load_resnet(name_or_depth: int | str = 18, int8: bool = False, compress: str | None = None, compress_atol: float = 0.05, build_dir: str | None = None, weights: str = 'IMAGENET1K_V1') -> 'Vision'

Load a ResNet as a fused ANE classifier; BatchNorm folded into the preceding conv at load. name_or_depth takes a torchvision depth (50, "50", "resnet50" -> 18/34/50/101 ImageNet weights) or a Hugging Face ResNet repo id (contains "/", e.g. "microsoft/resnet-50").

Source code in aneforge/models.py 
38
39
40
41
42
43
44
45
def load_resnet(name_or_depth: int | str = 18, int8: bool = False, compress: str | None = None,
                compress_atol: float = 0.05, build_dir: str | None = None,
                weights: str = "IMAGENET1K_V1") -> "Vision":
  """Load a ResNet as a fused ANE classifier; BatchNorm folded into the preceding conv at load.
  `name_or_depth` takes a torchvision depth (50, "50", "resnet50" -> 18/34/50/101 ImageNet weights)
  or a Hugging Face ResNet repo id (contains "/", e.g. "microsoft/resnet-50")."""
  return Vision(name_or_depth, int8=int8, compress=compress, compress_atol=compress_atol,
                build_dir=build_dir, weights=weights)

load_resnet18 

load_resnet18(int8: bool = False, compress: str | None = None, compress_atol: float = 0.05, build_dir: str | None = None) -> 'Vision'

Load torchvision ResNet-18 (ImageNet) as a fused ANE classifier; BatchNorm folded into the preceding conv at load.

Source code in aneforge/models.py 
48
49
50
51
def load_resnet18(int8: bool = False, compress: str | None = None,
                  compress_atol: float = 0.05, build_dir: str | None = None) -> "Vision":
  """Load torchvision ResNet-18 (ImageNet) as a fused ANE classifier; BatchNorm folded into the preceding conv at load."""
  return Vision(int8=int8, compress=compress, compress_atol=compress_atol, build_dir=build_dir)

load_vit 

load_vit(name: str, int8: bool = False) -> 'ViT'

Load a Hugging Face ViT image classifier (ViTForImageClassification) as a fused ANE program.

Source code in aneforge/models.py 
464
465
466
def load_vit(name: str, int8: bool = False) -> "ViT":
  """Load a Hugging Face ViT image classifier (ViTForImageClassification) as a fused ANE program."""
  return ViT(name, int8=int8)

load_gpt2 

load_gpt2(name: str, int8: bool = False, max_layers: int | None = None) -> 'GPT2'

Load a Hugging Face GPT-2 causal LM as fused ANE programs: a pre-norm transformer (native causal SDPA) plus the tied lm_head tiled along vocab. max_layers trims the stack (the compile-fallback knob examples use when the full model will not fit).

Source code in aneforge/models.py 
565
566
567
568
569
def load_gpt2(name: str, int8: bool = False, max_layers: int | None = None) -> "GPT2":
  """Load a Hugging Face GPT-2 causal LM as fused ANE programs: a pre-norm transformer
  (native causal SDPA) plus the tied lm_head tiled along vocab. `max_layers` trims the
  stack (the compile-fallback knob examples use when the full model will not fit)."""
  return GPT2(name, int8=int8, max_layers=max_layers)

load_clip 

load_clip(name: str = 'openai/clip-vit-base-patch32', int8: bool = False) -> 'CLIP'

Load a Hugging Face CLIP dual-encoder model (CLIPModel) for zero-shot image/text classification on the ANE.

Source code in aneforge/models.py 
676
677
678
def load_clip(name: str = "openai/clip-vit-base-patch32", int8: bool = False) -> "CLIP":
  """Load a Hugging Face CLIP dual-encoder model (CLIPModel) for zero-shot image/text classification on the ANE."""
  return CLIP(name, int8=int8)

load_whisper 

load_whisper(name: str = 'openai/whisper-base.en') -> 'Whisper'

Load a Hugging Face Whisper speech-to-text model with both towers (audio encoder + text decoder) running on the ANE. Whisper(name).transcribe(audio) -> greedy English transcript.

Source code in aneforge/models.py 
833
834
835
836
def load_whisper(name: str = "openai/whisper-base.en") -> "Whisper":
  """Load a Hugging Face Whisper speech-to-text model with both towers (audio encoder + text
  decoder) running on the ANE. `Whisper(name).transcribe(audio)` -> greedy English transcript."""
  return Whisper(name)

group_norm_train 

group_norm_train(x, gamma, beta, groups: int, eps: float = 1e-05)

Any-batch GroupNorm with trainable affine, built from VJP-bearing primitives; x is [N,C,H,W], gamma/beta [1,C,1,1].

Source code in aneforge/models.py 
 993
 994
 995
 996
 997
 998
 999
1000
1001
1002
1003
def group_norm_train(x, gamma, beta, groups: int, eps: float = 1e-5):
  """Any-batch GroupNorm with trainable affine, built from VJP-bearing primitives; `x` is [N,C,H,W], gamma/beta [1,C,1,1]."""
  N, C, H, W = x.shape
  if C % groups:
    raise ValueError(f"group_norm_train: channels {C} not divisible by groups {groups}")
  M = (C // groups) * H * W
  xg = x.reshape(N, groups, M)
  xc = xg - xg.mean((2,))
  var = xc.square().mean((2,))
  xn = (xc * var.adds(float(eps)).rsqrt()).reshape(N, C, H, W)
  return xn * gamma + beta

conv_block 

conv_block(x, conv_w, gamma, beta, groups: int, pool: int = 0)

conv2d(pad=1) -> GroupNorm(train) -> ReLU -> optional max_pool(pool); pool=0 skips pooling.

Source code in aneforge/models.py 
1006
1007
1008
1009
1010
def conv_block(x, conv_w, gamma, beta, groups: int, pool: int = 0):
  """conv2d(pad=1) -> GroupNorm(train) -> ReLU -> optional max_pool(pool); `pool=0` skips pooling."""
  h = conv2d(x, conv_w, pad=1)
  h = group_norm_train(h, gamma, beta, groups).relu()
  return h.max_pool(pool) if pool else h

cifar_cnn 

cifar_cnn(batch: int, widths=(32, 64, 128), groups: int = 8, classes: int = 10, seed: int = 0)

Build the CIFAR-10 CNN graph; returns (x_input, logits, params) with params in fixed trainable order.

Source code in aneforge/models.py 
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
def cifar_cnn(batch: int, widths=(32, 64, 128), groups: int = 8, classes: int = 10, seed: int = 0):
  """Build the CIFAR-10 CNN graph; returns (x_input, logits, params) with params in fixed trainable order."""
  rng = np.random.default_rng(seed)
  w0, w1, w2 = widths
  x = input((batch, 3, 32, 32))
  cW1 = conv_param(_he(rng, (w0, 3, 3, 3)))
  cW2 = conv_param(_he(rng, (w1, w0, 3, 3)))
  cW3 = conv_param(_he(rng, (w2, w1, 3, 3)))
  g1 = parameter(np.ones((1, w0, 1, 1), np.float32)); b1 = parameter(np.zeros((1, w0, 1, 1), np.float32))
  g2 = parameter(np.ones((1, w1, 1, 1), np.float32)); b2 = parameter(np.zeros((1, w1, 1, 1), np.float32))
  g3 = parameter(np.ones((1, w2, 1, 1), np.float32)); b3 = parameter(np.zeros((1, w2, 1, 1), np.float32))
  Wfc = parameter(_he(rng, (w2, classes))); bfc = parameter(np.zeros((1, classes), np.float32))

  h = conv_block(x, cW1, g1, b1, groups, pool=2)     # -> [B, w0, 16, 16]
  h = conv_block(h, cW2, g2, b2, groups, pool=2)     # -> [B, w1,  8,  8]
  h = conv_block(h, cW3, g3, b3, groups, pool=0)     # -> [B, w2,  8,  8]
  h = h.mean((2, 3)).reshape(batch, w2)              # global average pool -> [B, w2]
  logits = (h @ Wfc) + bfc                           # [B, classes]
  params = [cW1, g1, b1, cW2, g2, b2, cW3, g3, b3, Wfc, bfc]
  return x, logits, params