nanbeige-mlx 0.2.0__py3-none-any.whl

nanbeige_mlx/__init__.py

@@ -0,0 +1,22 @@
+"""nanbeige-mlx: an MLX port of the Nanbeige4.2-3B Looped Transformer.
+
+The model definition lives in :mod:`nanbeige_mlx.model` and is the single
+source of truth — it is also copied verbatim into each converted weight
+directory as the ``model_file`` mlx-lm loads. Conversion (HF -> MLX quant) is
+in :mod:`nanbeige_mlx.convert`; publishing helpers in
+:mod:`nanbeige_mlx.upload` and :mod:`nanbeige_mlx.pull`.
+"""
+
+from __future__ import annotations
+
+__version__ = "0.2.0"
+
+
+def __getattr__(name: str):  # pragma: no cover - thin convenience export
+    # Lazy re-export so ``from nanbeige_mlx import pull`` works without importing
+    # huggingface_hub at package-import time.
+    if name == "pull":
+        from .pull import pull
+
+        return pull
+    raise AttributeError(f"module 'nanbeige_mlx' has no attribute {name!r}")

nanbeige_mlx/convert.py

@@ -0,0 +1,235 @@
+"""Convert Nanbeige4.2-3B HF weights into MLX quants.
+
+Uses mlx-lm's ``model_file`` hook: the MLX port (``model.py``) is dropped into a
+*staged copy* of the HF repo and referenced from that copy's ``config.json``, so
+the standard ``mlx_lm.convert`` pipeline (load -> quantize -> save) works without
+a registry entry. ``auto_map`` is stripped from the *output* config because the
+MLX side loads via ``model_file``; it is left intact in the source so the HF
+parity reference can still load the remote modeling code.
+
+Two correctness properties (P3.3, P3.4 in docs/investigation-log.md):
+
+* **Never mutate the source.** ``prepare_source`` stages a conversion-ready copy
+  in a temp dir, symlinks the (large) safetensors shards instead of duplicating
+  them, and writes the ``model_file`` config into the *copy*. Pointing ``--src``
+  at an HF cache snapshot is therefore safe — the snapshot's hashes are
+  untouched. ``to_mlx`` calls ``prepare_source`` itself and cleans up the stage,
+  so ``convert`` is one step and idempotent from a pristine HF download.
+* **Preserve tokenizer files and verify them.** mlx-lm's convert writes a
+  trimmed ``tokenizer_config.json`` and drops ``added_tokens.json`` /
+  ``special_tokens_map.json`` / ``tokenizer.model``. We copy those through,
+  drop transient init keys, restore ``model_max_length``, and assert the
+  tokenizer round-trips the EOS token before declaring success.
+"""
+
+from __future__ import annotations
+
+import json
+import shutil
+import tempfile
+from pathlib import Path
+
+from mlx_lm.convert import convert
+
+# The single source of truth for the port. Copied verbatim into each staged
+# (and therefore each converted) weight repo as the shipped ``model_file``.
+PORT_FILE = Path(__file__).resolve().parent / "model.py"
+
+# Tokenizer / template files mlx-lm's convert trims that we must carry through.
+CARRY = [
+    "added_tokens.json",
+    "special_tokens_map.json",
+    "tokenizer.model",
+    "chat_template.jinja",
+]
+
+# Transient keys mlx-lm leaks into tokenizer_config.json from
+# tokenizer.init_kwargs; they don't belong in a published config.
+_TRANSIENT_TOKENIZER_KEYS = ("is_local", "local_files_only", "backend")
+
+# HF ``trust_remote_code`` files that the source repo carries for the PyTorch
+# reference path. They must NOT ship in an MLX repo (which loads via
+# ``model_file``), so remove them from the output after convert.
+_STRIP_FROM_OUTPUT = {
+    "configuration_nanbeige.py",
+    "modeling_nanbeige.py",
+}
+
+# Nanbeige4.2-3B supports 256K context; mlx-lm's convert writes the int64
+# sentinel instead of this concrete value. Restore it.
+MODEL_MAX_LENGTH = 262144
+
+# EOS token the model must be able to emit to ever stop generating.
+EOS_TOKEN = "<|im_end|>"
+EOS_TOKEN_ID = 166101
+
+
+def prepare_source(
+    src_dir: str | Path,
+    model_file: str | Path = PORT_FILE,
+    workdir: str | Path | None = None,
+) -> Path:
+    """Stage a conversion-ready copy of ``src_dir``. **Never writes into ``src_dir``**.
+
+    Non-safetensor files are copied; ``*.safetensors`` shards are *symlinked*
+    (not duplicated — the BF16 checkpoint is ~8 GB). The port is copied in as
+    ``nanbeige.py`` and ``config.json["model_file"]`` is set on the *copy*.
+
+    Returns the stage directory. The caller is responsible for cleanup if they
+    don't go through :func:`to_mlx` (which handles it).
+    """
+    src = Path(src_dir)
+    if not src.is_dir():
+        raise FileNotFoundError(f"source repo not found: {src}")
+
+    stage = Path(workdir) if workdir else Path(
+        tempfile.mkdtemp(prefix="nanbeige-stage-")
+    )
+    stage.mkdir(parents=True, exist_ok=True)
+
+    # Copy small files verbatim.
+    for f in src.iterdir():
+        if f.is_file() and f.suffix != ".safetensors":
+            shutil.copy2(f, stage / f.name)
+
+    # Symlink the weight shards — don't duplicate ~8 GB.
+    for f in src.glob("*.safetensors"):
+        (stage / f.name).symlink_to(f.resolve())
+
+    # Drop in the port and wire up model_file on the staged config copy.
+    shutil.copy2(model_file, stage / "nanbeige.py")
+    cfg_path = stage / "config.json"
+    cfg = json.loads(cfg_path.read_text(encoding="utf-8"))
+    cfg["model_file"] = "nanbeige.py"
+    cfg_path.write_text(json.dumps(cfg, indent=2, ensure_ascii=False), encoding="utf-8")
+
+    return stage
+
+
+def _finalize(src: Path, out: Path) -> None:
+    """Carry tokenizer files through, scrub transient keys, restore max length.
+
+    Idempotent: only copies a CARRY file when the output is missing it.
+    """
+    for name in CARRY:
+        if (src / name).exists() and not (out / name).exists():
+            shutil.copy2(src / name, out / name)
+
+    tc = out / "tokenizer_config.json"
+    if tc.exists():
+        cfg = json.loads(tc.read_text(encoding="utf-8"))
+        for k in _TRANSIENT_TOKENIZER_KEYS:
+            cfg.pop(k, None)
+        cfg["model_max_length"] = MODEL_MAX_LENGTH
+        tc.write_text(json.dumps(cfg, indent=2, ensure_ascii=False), encoding="utf-8")
+
+    # Strip HF trust_remote_code files the convert step may have copied through;
+    # the MLX repo loads via model_file and shouldn't ship the PyTorch reference.
+    for name in _STRIP_FROM_OUTPUT:
+        p = out / name
+        if p.exists():
+            p.unlink()
+
+
+def verify(out_dir: str | Path) -> None:
+    """Assert the converted repo's tokenizer round-trips the EOS token.
+
+    A conversion that silently loses ``<|im_end|>`` produces a model that never
+    stops generating — cheap to assert, expensive to debug.
+    """
+    from transformers import AutoTokenizer  # type: ignore
+
+    out = Path(out_dir)
+    tok = AutoTokenizer.from_pretrained(str(out), trust_remote_code=False)
+    got = tok.convert_tokens_to_ids(EOS_TOKEN)
+    assert got == EOS_TOKEN_ID, (
+        f"EOS token lost in conversion: {EOS_TOKEN!r} -> {got} (expected {EOS_TOKEN_ID})"
+    )
+    # ``eos_token_id`` may be stored as a string in some tokenizer configs.
+    assert int(tok.eos_token_id) == EOS_TOKEN_ID, (
+        f"tokenizer.eos_token_id={tok.eos_token_id} != {EOS_TOKEN_ID}"
+    )
+    rendered = tok.apply_chat_template(
+        [{"role": "user", "content": "hi"}],
+        tokenize=False,
+        add_generation_prompt=True,
+    )
+    assert "<|im_start|>assistant" in rendered, "chat template missing assistant marker"
+    assert "<think>" in rendered, "chat template missing the reasoning (<think>) block"
+
+
+def to_mlx(
+    src_dir: str | Path,
+    out_dir: str | Path,
+    bits: int,
+    group_size: int = 64,
+    *,
+    skip_verify: bool = False,
+) -> Path:
+    """Convert ``src_dir`` to a quantized MLX repo at ``out_dir`` (``bits`` per weight).
+
+    Stages a non-mutating copy (P3.3), runs mlx-lm's convert, strips
+    ``auto_map`` from the output config, carries tokenizer files through and
+    verifies the round-trip (P3.4). The stage dir is always cleaned up.
+    """
+    src = Path(src_dir)
+    out = Path(out_dir)
+
+    stage = prepare_source(src)
+    try:
+        convert(
+            hf_path=str(stage),
+            mlx_path=str(out),
+            quantize=True,
+            q_group_size=group_size,
+            q_bits=bits,
+        )
+    finally:
+        shutil.rmtree(stage, ignore_errors=True)
+
+    # Strip auto_map from the output config: MLX loads via model_file, and
+    # leaving auto_map triggers an irrelevant trust_remote_code prompt on load.
+    cfg_path = out / "config.json"
+    cfg = json.loads(cfg_path.read_text(encoding="utf-8"))
+    cfg.pop("auto_map", None)
+    cfg_path.write_text(json.dumps(cfg, indent=2, ensure_ascii=False), encoding="utf-8")
+
+    _finalize(src, out)
+
+    if not skip_verify:
+        try:
+            verify(out)
+            print(
+                f"PASS: tokenizer round-trips {EOS_TOKEN!r} -> {EOS_TOKEN_ID}; "
+                f"chat template intact."
+            )
+        except Exception as exc:  # pragma: no cover - surfaces as a hard error
+            print(f"VERIFY FAILED: {exc}")
+            raise
+
+    return out
+
+
+def main(argv: list[str] | None = None) -> None:  # pragma: no cover
+    import argparse
+
+    ap = argparse.ArgumentParser(
+        prog="nanbeige-mlx-convert",
+        description="Convert Nanbeige4.2-3B HF weights to a quantized MLX repo.",
+    )
+    ap.add_argument("--src", required=True, help="local Nanbeige HF repo (left untouched)")
+    ap.add_argument("--out", required=True, help="output MLX repo path")
+    ap.add_argument("--bits", type=int, required=True, help="quantization bits (4, 6, 8)")
+    ap.add_argument("--group-size", type=int, default=64)
+    ap.add_argument(
+        "--skip-verify",
+        action="store_true",
+        help="skip the post-convert tokenizer round-trip assertion",
+    )
+    a = ap.parse_args(argv)
+    to_mlx(a.src, a.out, a.bits, a.group_size, skip_verify=a.skip_verify)
+    print(a.out)
+
+
+if __name__ == "__main__":  # pragma: no cover
+    main()

nanbeige_mlx/model.py

@@ -0,0 +1,355 @@
+# Copyright © 2026 Jishnu Venugopal
+#
+# MLX port of the Nanbeige *Looped Transformer* (Nanbeige4.2-3B).
+#
+# This file is the single source of truth for the port. It is dual-purpose:
+#   1. Importable as ``nanbeige_mlx.model`` (the package this file ships in).
+#   2. A self-contained ``model_file`` that mlx-lm can import directly from a
+#      converted weight directory (see ``config.json`` -> ``"model_file"``), so
+#      anyone can ``mlx_lm.load`` a converted Nanbeige repo without a registry
+#      entry. The conversion step copies this file verbatim into each quant dir.
+#
+# Dependency surface of the shipped ``model_file``:
+#   The trivial helpers (SwiGLU, the unscaled RoPE branch, BaseModelArgs
+#   field-filtering) are inlined below so the shipped file has no dependency on
+#   mlx-lm's private modules for behavior it can express directly. The helpers
+#   that carry real behavior (``scaled_dot_product_attention`` for quantized-KV
+#   dispatch, ``KVCache``, ``create_attention_mask``) are imported from
+#   ``mlx_lm.models.*`` behind a single guarded import that raises an actionable
+#   error. This file is therefore **pinned to mlx-lm >= 0.31 internals** (tested
+#   0.31.3); an upstream mlx-lm refactor of those modules would require a new
+#   model_file. That's an honest trade for keeping quantized-KV support.
+#
+# Only absolute ``mlx`` / ``mlx_lm`` imports are used so the file works
+# standalone when dropped into a weight repo.
+#
+# Architecture for the Nanbeige4.2-3B checkpoint (see its ``config.json``):
+#   - Standard Llama-style decoder blocks: RMSNorm -> GQA attention -> residual,
+#     RMSNorm -> SwiGLU MLP -> residual.
+#   - ``num_loops = 2``: the 22 physical layers are executed twice per forward
+#     pass (weight-shared recurrence). The final RMSNorm is applied at *each*
+#     loop boundary (``skip_loop_final_norm = False``), i.e. after the 22nd and
+#     after the 44th effective layer.
+#   - GQA: 48 query heads, 8 key/value heads.
+#   - ``head_dim = 128`` while ``hidden_size / num_heads = 3072 / 48 = 64``.
+#     The projection widths therefore use the explicit head_dim, giving
+#     non-square q/o projections (q: 3072 -> 6144, o: 6144 -> 3072).
+#   - RoPE is the rotate-half convention (MLX ``rope_traditional=False``) with an
+#     unusually large base ``rope_theta = 70_000_000`` for the 256K context.
+#   - The richer Nanbeige config class also supports n-gram embeddings, hyper /
+#     mini-hyper connection and depth attention, but this checkpoint leaves all
+#     of those disabled, so the port implements the plain looped variant.
+
+import inspect
+from dataclasses import dataclass
+from typing import Any, Dict, Optional, Union
+
+import mlx.core as mx
+import mlx.nn as nn
+
+# --- mlx-lm internals this model_file depends on (pinned to >= 0.31) ---------
+# These carry real behavior (quantized-KV dispatch in particular) that we want
+# to keep, so they are guarded rather than inlined. An actionable error beats a
+# raw ModuleNotFoundError in a stranger's traceback years from now.
+try:
+    from mlx_lm.models.base import create_attention_mask, scaled_dot_product_attention
+    from mlx_lm.models.cache import KVCache
+except ImportError as exc:  # pragma: no cover - exercised only on bad installs
+    raise ImportError(
+        "nanbeige_mlx requires mlx-lm >= 0.31 (tested 0.31.3) and mlx >= 0.32. "
+        "The internal helpers this model_file uses "
+        "(mlx_lm.models.base.{scaled_dot_product_attention,create_attention_mask}, "
+        "mlx_lm.models.cache.KVCache) moved or were removed in your version. "
+        "Upgrade: pip install -U 'mlx-lm>=0.31' 'mlx>=0.32'."
+    ) from exc
+
+
+def _swiglu(gate: mx.array, x: mx.array) -> mx.array:
+    """SwiGLU activation: ``silu(gate) * x`` (reference: ``act_fn(gate) * up``).
+
+    Inlined from ``mlx_lm.models.activations.swiglu``; the only difference is
+    the upstream copy is wrapped in ``mx.compile`` (a speed optimization, not a
+    behavior change), which we deliberately do not bake into a shipped
+    ``model_file`` — mlx-lm applies its own compilation at generate time.
+    """
+    return nn.silu(gate) * x
+
+
+def _make_rope(
+    dims: int,
+    base: float,
+    traditional: bool,
+    scaling_config: Optional[Dict[str, Union[float, str]]],
+    max_position_embeddings: Optional[int],
+) -> nn.RoPE:
+    """Construct the RoPE module.
+
+    The unscaled branch (``scaling_config is None``) is inlined as
+    ``nn.RoPE(dims, traditional, base)`` so the shipped file does not depend on
+    ``mlx_lm.models.rope_utils`` for the common case. Scaled RoPE variants
+    (Llama-3 / dynamic / Yarn / ...) are deferred to ``initialize_rope`` behind a
+    guarded import — this checkpoint uses no scaling, but the branch is kept so
+    the file stays correct if a future Nanbeige checkpoint enables one.
+    """
+    if scaling_config is None:
+        return nn.RoPE(dims, traditional=traditional, base=base)
+    try:
+        from mlx_lm.models.rope_utils import initialize_rope
+    except ImportError as exc:  # pragma: no cover
+        raise ImportError(
+            "scaled RoPE requires mlx-lm >= 0.31 (mlx_lm.models.rope_utils)."
+        ) from exc
+    return initialize_rope(
+        dims,
+        base=base,
+        traditional=traditional,
+        scaling_config=scaling_config,
+        max_position_embeddings=max_position_embeddings,
+    )
+
+
+@dataclass
+class ModelArgs:
+    """Configuration for the Nanbeige looped transformer.
+
+    Field-filtering on construction (``from_dict``) is inlined from
+    ``mlx_lm.models.base.BaseModelArgs`` so this dataclass has no mlx-lm base
+    dependency in the shipped file.
+    """
+
+    model_type: str
+    hidden_size: int
+    num_hidden_layers: int
+    intermediate_size: int
+    num_attention_heads: int
+    num_key_value_heads: int
+    head_dim: int
+    rms_norm_eps: float
+    vocab_size: int
+    rope_theta: float = 70_000_000.0
+    num_loops: int = 1
+    max_position_embeddings: int = 262144
+    rope_traditional: bool = False
+    rope_scaling: Optional[Dict[str, Union[float, str]]] = None
+    tie_word_embeddings: bool = False
+    attention_bias: bool = False
+    mlp_bias: bool = False
+
+    @classmethod
+    def from_dict(cls, params: Dict[str, Any]) -> "ModelArgs":
+        """Build from a config dict, ignoring keys this dataclass doesn't define."""
+        return cls(
+            **{
+                k: v
+                for k, v in params.items()
+                if k in inspect.signature(cls).parameters
+            }
+        )
+
+
+class Attention(nn.Module):
+    """Grouped-query attention with an explicit (non-default) head dimension."""
+
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        dim = args.hidden_size
+        self.n_heads = n_heads = args.num_attention_heads
+        self.n_kv_heads = n_kv_heads = args.num_key_value_heads
+        # NOTE: head_dim is taken from the config (128), not hidden_size // n_heads.
+        self.head_dim = head_dim = args.head_dim
+        self.scale = head_dim ** -0.5
+
+        self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=args.attention_bias)
+        self.k_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=args.attention_bias)
+        self.v_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=args.attention_bias)
+        # bias follows config.attention_bias (harmless: this checkpoint is False).
+        self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=args.attention_bias)
+
+        self.rope = _make_rope(
+            head_dim,
+            base=args.rope_theta,
+            traditional=args.rope_traditional,
+            scaling_config=args.rope_scaling,
+            max_position_embeddings=args.max_position_embeddings,
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        B, L, _ = x.shape
+
+        queries = self.q_proj(x)
+        keys = self.k_proj(x)
+        values = self.v_proj(x)
+
+        queries = queries.reshape(B, L, self.n_heads, self.head_dim).transpose(0, 2, 1, 3)
+        keys = keys.reshape(B, L, self.n_kv_heads, self.head_dim).transpose(0, 2, 1, 3)
+        values = values.reshape(B, L, self.n_kv_heads, self.head_dim).transpose(0, 2, 1, 3)
+
+        # RoPE precision: nothing to do here, and this is a measured result, not
+        # an assumption. `mx.fast.rope` computes its frequencies and cos/sin in
+        # fp32 inside the kernel regardless of the input dtype, so wrapping these
+        # calls in an fp32 upcast/downcast is arithmetically inert. Measured: an
+        # fp32 round-trip around the two calls below left the end-to-end parity
+        # cosine *bit-identical* (0.846566 both ways, 6 prompts, 166k logits).
+        # Bit-identity is only possible if the kernel was already fp32-internal.
+        #
+        # Contrast the HF reference, which is vulnerable here: it holds `inv_freq`
+        # as a non-persistent fp32 buffer, so a blanket `.to(bfloat16)` on the
+        # module silently degrades it (see docs/investigation-log.md and the
+        # comment in nanbeige_mlx_eval/bisect.py::_real_layer_stages). The port
+        # has no such buffer to clobber -- nn.RoPE stores four Python scalars.
+        if cache is not None:
+            queries = self.rope(queries, offset=cache.offset)
+            keys = self.rope(keys, offset=cache.offset)
+            keys, values = cache.update_and_fetch(keys, values)
+        else:
+            queries = self.rope(queries)
+            keys = self.rope(keys)
+
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache=cache, scale=self.scale, mask=mask
+        )
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output)
+
+
+class MLP(nn.Module):
+    """SwiGLU feed-forward block."""
+
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.gate_proj = nn.Linear(args.hidden_size, args.intermediate_size, bias=args.mlp_bias)
+        self.up_proj = nn.Linear(args.hidden_size, args.intermediate_size, bias=args.mlp_bias)
+        self.down_proj = nn.Linear(args.intermediate_size, args.hidden_size, bias=args.mlp_bias)
+
+    def __call__(self, x: mx.array) -> mx.array:
+        return self.down_proj(_swiglu(self.gate_proj(x), self.up_proj(x)))
+
+
+class TransformerBlock(nn.Module):
+    """A standard pre-norm Llama-style decoder layer."""
+
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.self_attn = Attention(args)
+        self.mlp = MLP(args)
+        self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.args = args
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        r = self.self_attn(self.input_layernorm(x), mask, cache)
+        h = x + r
+        r = self.mlp(self.post_attention_layernorm(h))
+        return h + r
+
+
+class NanbeigeModel(nn.Module):
+    """The looped transformer trunk.
+
+    ``num_loops`` passes are made over the same ``num_hidden_layers`` physical
+    blocks (weight sharing). The final RMSNorm is applied at the end of every
+    pass, matching ``skip_loop_final_norm = False`` in the reference checkpoint.
+
+    Because each (loop, layer) pair attends over a distinct key/value stream,
+    the cache needs ``num_loops * num_hidden_layers`` virtual slots even though
+    only ``num_hidden_layers`` blocks of weights exist.
+    """
+
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.num_hidden_layers = args.num_hidden_layers
+        self.num_loops = args.num_loops
+        self.vocab_size = args.vocab_size
+
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [TransformerBlock(args) for _ in range(args.num_hidden_layers)]
+        self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+
+    def __call__(self, inputs: mx.array, cache=None) -> mx.array:
+        h = self.embed_tokens(inputs)
+
+        if cache is None:
+            cache = [None] * (self.num_loops * self.num_hidden_layers)
+
+        # All virtual cache slots stay in lockstep (same offset) within a single
+        # call, so a single mask derived from the first slot is valid for every
+        # layer in both loops.
+        mask = create_attention_mask(h, cache[0])
+
+        for loop in range(self.num_loops):
+            for i, layer in enumerate(self.layers):
+                # Virtual index: each (loop, layer) pair owns its own KV slot so
+                # the two passes over the shared weights do not collide.
+                h = layer(h, mask, cache[loop * self.num_hidden_layers + i])
+            h = self.norm(h)
+
+        return h
+
+
+class Model(nn.Module):
+    """Top-level causal LM wrapper, conforming to the mlx-lm model contract."""
+
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.model = NanbeigeModel(args)
+        if not args.tie_word_embeddings:
+            self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+
+    def __call__(self, inputs: mx.array, cache=None):
+        out = self.model(inputs, cache)
+        if self.args.tie_word_embeddings:
+            return self.model.embed_tokens.as_linear(out)
+        return self.lm_head(out)
+
+    def make_cache(self):
+        """Allocate the loop-aware KV cache (``num_loops * num_hidden_layers`` slots).
+
+        ``mlx_lm``'s :func:`make_prompt_cache` defers to this method when present,
+        which is what lets a model with weight-shared looped layers expose the
+        larger effective depth the cache requires.
+
+        Note on the ``--max-kv-size`` knob: ``make_prompt_cache`` ignores
+        ``max_kv_size`` whenever a model supplies ``make_cache``, so the
+        rotating-cache / max-length knob is **inert** for this model.
+        ``--kv-bits`` still works (``KVCache.to_quantized``). The practical
+        ceiling is real: at full context the looped design needs
+        ``num_loops * num_hidden_layers`` KV slots, i.e. for this checkpoint
+        44 x 8 KV-heads x 128 dim x 2 (K+V) x 262144 positions x 2 bytes
+        ~= 47 GB — unreachable on a 16 GB machine. This is a genuine,
+        quantifiable cost of the looped design, not a bug.
+        """
+        n = self.args.num_loops * self.args.num_hidden_layers
+        return [KVCache() for _ in range(n)]
+
+    @property
+    def layers(self):
+        return self.model.layers
+
+    @property
+    def head_dim(self):
+        return self.args.head_dim
+
+    @property
+    def n_kv_heads(self):
+        return self.args.num_key_value_heads
+
+    def sanitize(self, weights):
+        """Weight keys map 1:1 to the module tree for this checkpoint.
+
+        Drop any non-persistent buffers a checkpoint might carry (e.g. a cached
+        rotary ``inv_freq``) so a strict load succeeds.
+        """
+        return {k: v for k, v in weights.items() if "rotary_emb.inv_freq" not in k}

nanbeige_mlx/pull.py

@@ -0,0 +1,41 @@
+"""Download a published Nanbeige4.2-3B MLX quant by short name.
+
+Wraps :func:`huggingface_hub.snapshot_download` with the repo ids baked in, so
+the user experience is a one-liner with no conversion step::
+
+    from nanbeige_mlx import pull
+    import mlx_lm
+    model, tok = mlx_lm.load(pull("4bit"))
+
+Weights arrive on first use, cached in ``~/.cache/huggingface``, resumable and
+revision-pinnable — indistinguishable from bundling, and exactly how
+``mlx-community`` ships every model.
+"""
+
+from __future__ import annotations
+
+QUANTS: dict[str, str] = {
+    "4bit": "jishnuvenugopal/Nanbeige4.2-3B-mlx-4bit",
+    "6bit": "jishnuvenugopal/Nanbeige4.2-3B-mlx-6bit",
+    "8bit": "jishnuvenugopal/Nanbeige4.2-3B-mlx-8bit",
+}
+
+
+def pull(quant: str = "4bit", revision: str | None = None) -> str:
+    """Download a published quant; returns a local path for ``mlx_lm.load``.
+
+    Parameters
+    ----------
+    quant:
+        One of ``"4bit"``, ``"6bit"``, ``"8bit"`` (see :data:`QUANTS`).
+    revision:
+        Optional pinned HF revision (commit SHA or tag). Pinning is recommended
+        for reproducibility — see the eval README, which records exact SHAs.
+    """
+    if quant not in QUANTS:
+        raise ValueError(
+            f"unknown quant {quant!r}; choose one of {sorted(QUANTS)}"
+        )
+    from huggingface_hub import snapshot_download  # type: ignore
+
+    return snapshot_download(repo_id=QUANTS[quant], revision=revision)

nanbeige_mlx/upload.py

@@ -0,0 +1,230 @@
+"""Publish a converted Nanbeige MLX repo to the HuggingFace Hub.
+
+Nanbeige4.2-3B is Apache-2.0, which permits redistribution of derivative works
+(quantization is a modification), so §4(d) applies: retain the upstream
+LICENSE, retain notices, and *state that you changed the files*. This module
+writes proper model-card frontmatter, copies the upstream LICENSE through, and
+emits a NOTICE describing the modification, then (unless ``--dry-run``) uploads.
+
+Run with ``--dry-run`` first: it renders the card and lists the files that would
+be uploaded without touching the network. The actual upload needs
+``huggingface-cli login`` and is the one outward-facing step — by default this
+module does NOT push, it prepares everything up to it.
+"""
+
+from __future__ import annotations
+
+import argparse
+import fnmatch
+from pathlib import Path
+
+
+UPLOAD_IGNORE_PATTERNS = [
+    "__pycache__/**",
+    "**/__pycache__/**",
+    "*.pyc",
+    "**/*.pyc",
+    ".DS_Store",
+    "**/.DS_Store",
+    "*.egg-info/**",
+    "**/*.egg-info/**",
+]
+
+# Apache-2.0 §4(d): state what changed.
+NOTICE_TEMPLATE = """\
+Nanbeige4.2-3B MLX quantization
+================================
+
+This repository contains a derivative of `{base_model}` (Apache-2.0).
+
+Modifications by Jishnu Venugopal ({year}):
+  * Converted from the original PyTorch checkpoint to Apple's MLX format.
+  * Quantized to {bits}-bit (group size {group_size}) for Apple Silicon.
+
+The original model and its Apache-2.0 license are retained (see LICENSE).
+Upstream notices and the full license text are unchanged. This derivative is
+redistributed under the same Apache-2.0 license.
+
+The MLX model definition (`nanbeige.py`) is MIT-licensed source code from the
+`nanbeige-mlx` project; the weights themselves remain Apache-2.0.
+"""
+
+
+def _card_frontmatter(bits: int, group_size: int) -> str:
+    # Proper frontmatter for a bilingual quantized weights repo (C7).
+    return (
+        "---\n"
+        "license: apache-2.0\n"
+        "base_model: Nanbeige/Nanbeige4.2-3B\n"
+        "language: [en, zh]\n"
+        "library_name: mlx\n"
+        "pipeline_tag: text-generation\n"
+        f"tags: [mlx, nanbeige, looped-transformer, apple-silicon, {bits}-bit]\n"
+        f"quantized_from: Nanbeige/Nanbeige4.2-3B\n"
+        f"quantization:\n  bits: {bits}\n  group_size: {group_size}\n"
+        "---\n"
+    )
+
+
+def _card_body(bits: int, group_size: int) -> str:
+    return f"""
+# Nanbeige4.2-3B MLX ({bits}-bit)
+
+An MLX conversion and {bits}-bit quantization (group size {group_size}) of
+[`Nanbeige/Nanbeige4.2-3B`](https://huggingface.co/Nanbeige/Nanbeige4.2-3B) — a
+3-billion-parameter **Looped Transformer** (`num_loops=2`, weight-shared over 22
+layers for an effective depth of 44). Produced by the independent
+[`nanbeige-mlx`](https://github.com/jishnuvenugopal/nanbeige-mlx) project;
+not affiliated with the Nanbeige team.
+
+## Load (no conversion step)
+
+```python
+from nanbeige_mlx import pull
+import mlx_lm
+model, tok = mlx_lm.load(pull("{bits}bit"))
+```
+
+Weights arrive on first use and are cached in `~/.cache/huggingface`.
+
+## KV-cache ceiling (a real cost of the looped design)
+
+The looped architecture needs `num_loops * num_hidden_layers = 44` KV slots.
+At full context that is 44 x 8 KV-heads x 128 dim x 2 (K+V) x 262144 positions
+x 2 bytes ~= **47 GB** — unreachable on a 16 GB machine. Because this model
+supplies `make_cache`, mlx-lm's `--max-kv-size` knob is inert; use `--kv-bits`
+to reduce KV precision instead.
+
+## Fidelity of the port
+
+This quant is produced by a from-scratch MLX port of the looped architecture,
+checked against the HuggingFace reference. The headline finding carries over
+verbatim from the port's documentation, and must read identically wherever it
+appears (repo README, model card, investigation log):
+
+> Bottom line: **the source of the logit gap is documented-open, not blocking.**
+> Six candidate causes are ruled out by measurement (the seventh, "RoPE runs in
+> bf16," was killed by a bit-identical upcast experiment — `mx.fast.rope` is
+> fp32-internal regardless of input dtype). The `[1,1,5,6]` mask shape is a red
+> herring: the 6th column is HF's `past_seen_tokens + sequence_length + 1`
+> boilerplate for StaticCache sizing, present in every Llama-family model, and
+> eager attention slices it to `[:,:,:,:L]` before use — nothing to do with
+> `num_loops`. The gap does not affect behavior on the agentic suite. Bit-exact
+> parity is not claimed and not expected in bf16. Full record:
+> [`docs/investigation-log.md`](https://github.com/jishnuvenugopal/nanbeige-mlx-eval/blob/main/docs/investigation-log.md).
+
+## License
+
+Apache-2.0 (the upstream license; quantization is a modification under §4(d)).
+The `nanbeige.py` model definition is MIT-licensed source from `nanbeige-mlx`.
+See `LICENSE` and `NOTICE`.
+"""
+
+
+def write_card(model_dir: str | Path, bits: int, group_size: int, *, base_model: str = "Nanbeige/Nanbeige4.2-3B") -> Path:
+    """Write README.md, LICENSE (copied from upstream), and NOTICE into ``model_dir``."""
+    import datetime
+    import shutil
+
+    d = Path(model_dir)
+    card = d / "README.md"
+    card.write_text(_card_frontmatter(bits, group_size) + _card_body(bits, group_size), encoding="utf-8")
+
+    notice = d / "NOTICE"
+    notice.write_text(
+        NOTICE_TEMPLATE.format(
+            base_model=base_model,
+            bits=bits,
+            group_size=group_size,
+            year=datetime.datetime.now().year,
+        ),
+        encoding="utf-8",
+    )
+
+    # Always ship the full canonical Apache-2.0 LICENSE text. The weights are
+    # Apache-2.0 (not MIT — MIT is the code), so a 2-line pointer is not enough:
+    # §4(d) requires giving recipients a copy of the license. Source the text
+    # from the vendored LICENSE.apache (committed alongside this module).
+    license_path = d / "LICENSE"
+    candidates = [
+        Path(__file__).resolve().parent.parent / "LICENSE.apache",  # port repo root
+        d.parent / "nanbeige42-hf" / "LICENSE",                      # upstream checkout
+        d.parent / "LICENSE.apache",
+    ]
+    src_license = next((c for c in candidates if c.exists()), None)
+    if src_license is None:
+        raise FileNotFoundError(
+            "Full Apache-2.0 LICENSE text not found. Expected LICENSE.apache in the "
+            "port repo root (vendored), or nanbeige42-hf/LICENSE. Do not ship a "
+            "license pointer — the weights require the full license text under §4(d)."
+        )
+    shutil.copy2(src_license, license_path)
+    return card
+
+
+def upload(model_dir: str | Path, repo_id: str, *, dry_run: bool = True) -> None:
+    """Write the card/LICENSE/NOTICE and (unless ``dry_run``) upload to ``repo_id``."""
+    d = Path(model_dir)
+    cfg = __import__("json").loads((d / "config.json").read_text(encoding="utf-8"))
+    q = cfg.get("quantization") or cfg.get("quantization_config") or {}
+    bits = int(q.get("bits", 0))
+    group_size = int(q.get("group_size", 64))
+
+    write_card(d, bits, group_size)
+
+    files = sorted(
+        rel
+        for p in d.rglob("*")
+        if p.is_file()
+        and not any(
+            fnmatch.fnmatch((rel := p.relative_to(d).as_posix()), pattern)
+            for pattern in UPLOAD_IGNORE_PATTERNS
+        )
+    )
+    print(f"repo: {repo_id}")
+    print(f"model_dir: {d}")
+    print(f"quantization: {bits}-bit, group_size={group_size}")
+    print("files:")
+    for name in files:
+        print(f"  - {name}")
+
+    if dry_run:
+        print("\n[dry-run] no upload performed. Re-run without --dry-run to publish.")
+        return
+
+    from huggingface_hub import HfApi  # type: ignore
+
+    api = HfApi()
+    api.create_repo(repo_id=repo_id, repo_type="model", exist_ok=True)
+    api.upload_folder(
+        folder_path=str(d),
+        repo_id=repo_id,
+        repo_type="model",
+        ignore_patterns=UPLOAD_IGNORE_PATTERNS,
+    )
+    print(f"\nuploaded to {repo_id}")
+
+
+def main(argv: list[str] | None = None) -> None:  # pragma: no cover
+    ap = argparse.ArgumentParser(
+        prog="nanbeige-mlx-upload",
+        description="Write a proper model card + LICENSE + NOTICE and upload a Nanbeige MLX quant.",
+    )
+    ap.add_argument("--model-dir", required=True)
+    ap.add_argument("--repo-id", required=True)
+    ap.add_argument(
+        "--dry-run",
+        action="store_true",
+        help="render the card and list files without uploading (default behavior)",
+    )
+    ap.add_argument(
+        "--yes",
+        action="store_true",
+        help="actually upload (requires `huggingface-cli login`)",
+    )
+    a = ap.parse_args(argv)
+    upload(a.model_dir, a.repo_id, dry_run=a.dry_run or not a.yes)
+
+
+if __name__ == "__main__":  # pragma: no cover
+    main()

nanbeige_mlx-0.2.0.dist-info/METADATA

@@ -0,0 +1,104 @@
+Metadata-Version: 2.4
+Name: nanbeige-mlx
+Version: 0.2.0
+Summary: MLX port of the Nanbeige4.2-3B Looped Transformer for Apple Silicon — model definition, HF->MLX conversion, and publish helpers.
+Author: Jishnu Venugopal
+License: MIT
+Project-URL: Homepage, https://github.com/jishnuvenugopal/nanbeige-mlx
+Project-URL: Repository, https://github.com/jishnuvenugopal/nanbeige-mlx
+Project-URL: Issues, https://github.com/jishnuvenugopal/nanbeige-mlx/issues
+Project-URL: Changelog, https://github.com/jishnuvenugopal/nanbeige-mlx/blob/main/CHANGELOG.md
+Project-URL: Evaluation harness, https://github.com/jishnuvenugopal/nanbeige-mlx-eval
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: Intended Audience :: Science/Research
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Operating System :: MacOS
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+License-File: LICENSE.apache
+Requires-Dist: mlx<0.34,>=0.32
+Requires-Dist: mlx-lm<0.33,>=0.31
+Requires-Dist: huggingface_hub>=1.20
+Requires-Dist: transformers>=4.40
+Provides-Extra: dev
+Requires-Dist: pytest>=8; extra == "dev"
+Dynamic: license-file
+
+# nanbeige-mlx
+
+An MLX port of the **Nanbeige4.2-3B *Looped Transformer*** for Apple Silicon,
+plus the HF→MLX conversion and publishing helpers. The model definition in
+[`nanbeige_mlx/model.py`](nanbeige_mlx/model.py) is the **single source of truth**
+for the port — it is also copied verbatim into each converted weight directory
+as the `model_file` mlx-lm loads.
+
+This is an independent project; it is not affiliated with or endorsed by the
+Nanbeige team.
+
+## What's here
+
+| module | purpose |
+|---|---|
+| `nanbeige_mlx/model.py` | the port — also the shipped `model_file` |
+| `nanbeige_mlx/convert.py` | HF → MLX quant (non-mutating staging, tokenizer verify) |
+| `nanbeige_mlx/upload.py` | model-card + LICENSE + NOTICE + (opt-in) HF upload |
+| `nanbeige_mlx/pull.py` | `pull("4bit")` → local path for `mlx_lm.load` |
+
+## Load a published quant (one line)
+
+```python
+from nanbeige_mlx import pull
+import mlx_lm
+model, tok = mlx_lm.load(pull("4bit"))
+```
+
+## Convert from the BF16 checkpoint yourself
+
+```bash
+nanbeige-mlx-convert --src /path/to/Nanbeige4.2-3B --out ./nanbeige-mlx-4bit --bits 4
+```
+
+The source directory is never mutated; the tokenizer round-trip is asserted.
+
+## The 44-slot KV cache (a real cost of the looped design)
+
+The looped architecture needs `num_loops * num_hidden_layers = 44` KV slots.
+At full context that is 44 × 8 KV-heads × 128 dim × 2 (K+V) × 262 144 positions
+× 2 bytes ≈ **47 GB** — unreachable on a 16 GB machine. Because this model
+supplies `make_cache`, mlx-lm's `--max-kv-size` knob is inert; use `--kv-bits`
+to reduce KV precision instead. This is a quantifiable cost of the looped
+design, not a bug — see `make_cache()`'s docstring.
+
+## Dependencies
+
+Upper-bounded to `mlx>=0.32,<0.34` and `mlx-lm>=0.31,<0.33`. The bounds are
+deliberate: `model.py` uses `mlx_lm.models.base` / `.cache` internals and is
+copied verbatim into every published weight repo as the `model_file`, where it
+is frozen — a user who already downloaded a quant cannot receive a patch.
+Relaxing the bound is a decision to make after testing against a new mlx-lm.
+
+## Fidelity status
+
+This port is validated by behaviour, not bit-parity: per-layer arithmetic
+agrees with the reference's own `NanbeigeDecoderLayer` to fp32 precision, the
+44-slot loop-aware KV cache passes a prefill-vs-incremental-decode equality
+test, and the port's two code paths are bit-identical. End-to-end next-token
+logit cosine against the HF reference is 0.847 (top-1 agreement 83%), lower
+than a faithful port should give; six candidate causes have been eliminated by
+measurement. The full record — including every falsified hypothesis — is in
+the evaluation harness repo's
+[`docs/investigation-log.md`](https://github.com/jishnuvenugopal/nanbeige-mlx-eval/blob/main/docs/investigation-log.md).
+Behaviour on the bilingual agentic suite is unaffected (26–28/30 across
+4/6/8-bit).
+
+## License
+
+MIT for the code in this package. The Nanbeige model weights are governed by
+the upstream Apache-2.0 license; convert and redistribute them per that license.

nanbeige_mlx-0.2.0.dist-info/RECORD

@@ -0,0 +1,13 @@
+nanbeige_mlx/__init__.py,sha256=m8nnqRDABY8cmW5pnFquFESD1yvJygnp8ZZm9QEWdIU,828
+nanbeige_mlx/convert.py,sha256=FqowSx1lm0PPpg8aC8_1lTWS2nkdj5SvvaYaxnOgJFc,8823
+nanbeige_mlx/model.py,sha256=9kdRvMOL7-JvWxvTLMMpl1dG5UnY1x6_dWC-tR9zzWE,15129
+nanbeige_mlx/pull.py,sha256=s1-ivwytz-n1JGG8blJ2ZoI45nMBWGq5A5cCikJQLlk,1438
+nanbeige_mlx/py.typed,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+nanbeige_mlx/upload.py,sha256=J-6Ub8bRlh9HE1nS40GACZ64fYmUzSK2vq8miTVSUys,8691
+nanbeige_mlx-0.2.0.dist-info/licenses/LICENSE,sha256=MOWIyoKq-Q2hIHhINTeut6W1nzXeOol7E334yGUrN1k,1073
+nanbeige_mlx-0.2.0.dist-info/licenses/LICENSE.apache,sha256=z8d0m5b2O9McPEK1xHG_dWgUBT6EfBDz6wA0F7xSPTA,11358
+nanbeige_mlx-0.2.0.dist-info/METADATA,sha256=YPVvUfITgkrkNoJ_IJOCygl7Ksnv1c8RTrCTlkNqjRA,4540
+nanbeige_mlx-0.2.0.dist-info/WHEEL,sha256=K260EYznzXsJYBQGqmI8VTxEdiZYNvDZwW9cBh9-_MA,91
+nanbeige_mlx-0.2.0.dist-info/entry_points.txt,sha256=tmTzha-whpQsw8fosf67DUjOZvhe274GXj8rAjaMH1A,114
+nanbeige_mlx-0.2.0.dist-info/top_level.txt,sha256=g4XY571m4D6fG2Iv0Gcjadw6vr4QLkkKEjfSoo96Bx8,13
+nanbeige_mlx-0.2.0.dist-info/RECORD,,

nanbeige_mlx-0.2.0.dist-info/WHEEL

@@ -0,0 +1,5 @@
+Wheel-Version: 1.0
+Generator: setuptools (83.0.0)
+Root-Is-Purelib: true
+Tag: py3-none-any
+

nanbeige_mlx-0.2.0.dist-info/entry_points.txt

@@ -0,0 +1,3 @@
+[console_scripts]
+nanbeige-mlx-convert = nanbeige_mlx.convert:main
+nanbeige-mlx-upload = nanbeige_mlx.upload:main

nanbeige_mlx-0.2.0.dist-info/licenses/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Jishnu Venugopal
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

nanbeige_mlx-0.2.0.dist-info/licenses/LICENSE.apache

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

nanbeige_mlx-0.2.0.dist-info/top_level.txt

@@ -0,0 +1 @@
+nanbeige_mlx