Upload 8 files

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

config.json

@@ -0,0 +1,41 @@
+{
+  "architectures": [
+    "NandiForCausalLM"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_nandi.NandiConfig",
+    "AutoModel": "modeling_nandi.NandiModel",
+    "AutoModelForCausalLM": "modeling_nandi.NandiForCausalLM"
+  },
+  "model_type": "nandi",
+  "vocab_size": 131072,
+  "hidden_size": 1248,
+  "intermediate_size": 3556,
+  "num_hidden_layers": 28,
+  "num_attention_heads": 16,
+  "num_key_value_heads": 8,
+  "head_dim": 78,
+  "hidden_act": "silu",
+  "max_position_embeddings": 2048,
+  "initializer_range": 0.008,
+  "rms_norm_eps": 1e-06,
+  "use_cache": true,
+  "pad_token_id": null,
+  "bos_token_id": 1,
+  "eos_token_id": 0,
+  "tie_word_embeddings": true,
+  "rope_parameters": {
+    "rope_theta": 1000000.0
+  },
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "mlp_bias": false,
+  "factorized_embedding": false,
+  "embedding_rank": 768,
+  "layer_sharing": false,
+  "layer_sharing_repeats": 2,
+  "qk_norm": true,
+  "shared_kv": true,
+  "kv_cache_mode": "shared",
+  "torch_dtype": "bfloat16"
+}

configuration_nandi.py

@@ -0,0 +1,147 @@
+# Copyright 2026 RTA AI Labs. All rights reserved.
+#
+# 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.
+
+from transformers.configuration_utils import PretrainedConfig
+
+
+class NandiConfig(PretrainedConfig):
+    r"""
+    Configuration class for the Nandi model.
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoConfig, AutoModelForCausalLM
+
+    >>> configuration = AutoConfig.from_pretrained("Rta-AILabs/Nandi-500M-remote", trust_remote_code=True)
+
+    >>> model = AutoModelForCausalLM.from_pretrained("Rta-AILabs/Nandi-500M-remote", trust_remote_code=True)
+
+    >>> configuration = model.config
+    ```
+    """
+
+    model_type = "nandi"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    base_model_tp_plan = {
+        "layers.*.self_attn.q_proj": "colwise",
+        "layers.*.self_attn.k_proj": "colwise",
+        "layers.*.self_attn.v_proj": "colwise",
+        "layers.*.self_attn.o_proj": "rowwise",
+        "layers.*.mlp.gate_proj": "colwise",
+        "layers.*.mlp.up_proj": "colwise",
+        "layers.*.mlp.down_proj": "rowwise",
+    }
+
+    def __init__(
+        self,
+        vocab_size=131072,
+        hidden_size=1248,
+        intermediate_size=3556,
+        num_hidden_layers=28,
+        num_attention_heads=16,
+        num_key_value_heads=8,
+        head_dim=None,
+        hidden_act="silu",
+        max_position_embeddings=2048,
+        initializer_range=0.008,
+        rms_norm_eps=1e-6,
+        use_cache=True,
+        pad_token_id=None,
+        bos_token_id=1,
+        eos_token_id=0,
+        pretraining_tp=1,
+        tie_word_embeddings=True,
+        rope_parameters=None,
+        attention_bias=False,
+        attention_dropout=0.0,
+        mlp_bias=False,
+        factorized_embedding=False,
+        embedding_rank=768,
+        layer_sharing=False,
+        layer_sharing_repeats=1,
+        qk_norm=True,
+        shared_kv=True,
+        kv_cache_mode="shared",
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.num_key_value_heads = num_key_value_heads if num_key_value_heads is not None else num_attention_heads
+        self.head_dim = head_dim if head_dim is not None else hidden_size // num_attention_heads
+        self.hidden_act = hidden_act
+        self.max_position_embeddings = max_position_embeddings
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.pretraining_tp = pretraining_tp
+        self.rope_parameters = rope_parameters if rope_parameters is not None else {"rope_theta": 1000000.0}
+        self.attention_bias = attention_bias
+        self.attention_dropout = attention_dropout
+        self.mlp_bias = mlp_bias
+        self.factorized_embedding = factorized_embedding
+        self.embedding_rank = embedding_rank
+        self.layer_sharing = layer_sharing
+        # Smoltron training loops over `layer_sharing_repeats` unconditionally
+        # (it does NOT check `layer_sharing`). Preserve the raw value here so
+        # the modeling code can honor it; the `layer_sharing` bool is now just
+        # metadata describing intent.
+        self.layer_sharing_repeats = max(1, int(layer_sharing_repeats or 1))
+        self.qk_norm = qk_norm
+        # `shared_kv` records that V was tied to K at pretraining time. In the
+        # HF model V is recomputed from `k_proj` at runtime (no `v_proj` module
+        # is materialised); see `NandiAttention.forward`.
+        self.shared_kv = shared_kv
+        # `kv_cache_mode` controls the inference-time K/V cache strategy when
+        # `shared_kv=True`. Both modes produce identical outputs (numerical
+        # round-off only); they trade memory for compute:
+        #   "shared"  -> cache ONLY raw K (single tensor per layer). Each
+        #                decode step re-applies k_norm + RoPE to the full
+        #                cached raw K. Halves KV-cache memory.
+        #   "vanilla" -> cache post-norm post-RoPE K AND raw V (two tensors
+        #                per layer). k_norm + RoPE are applied only to the
+        #                current step's tokens. Standard HF behavior.
+        # Ignored when `shared_kv=False`. Defaults to "shared".
+        if kv_cache_mode not in ("shared", "vanilla"):
+            raise ValueError(
+                f"`kv_cache_mode` must be 'shared' or 'vanilla', got {kv_cache_mode!r}."
+            )
+        self.kv_cache_mode = kv_cache_mode
+
+        if self.factorized_embedding and self.embedding_rank <= 0:
+            raise ValueError(
+                f"`embedding_rank` must be positive when `factorized_embedding=True`, got {self.embedding_rank}."
+            )
+        if self.hidden_size % self.num_attention_heads != 0:
+            raise ValueError(
+                f"`hidden_size` ({self.hidden_size}) must be divisible by "
+                f"`num_attention_heads` ({self.num_attention_heads})."
+            )
+        if self.layer_sharing_repeats < 1:
+            raise ValueError(f"`layer_sharing_repeats` must be >= 1, got {self.layer_sharing_repeats}.")
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )
+
+
+__all__ = ["NandiConfig"]

convert_smoltron_to_hf.py

@@ -0,0 +1,320 @@
+"""Convert a Smoltron (nanotron LlamaForTraining) on-disk checkpoint into an HF
+Nandi (`NandiForCausalLM`) checkpoint with custom code support.
+
+This script reads the per-parameter safetensors shards that nanotron writes
+to disk and re-bundles them into a single ``model.safetensors`` + ``config.json``
+that matches ``NandiConfig`` / ``NandiForCausalLM`` (see the sibling
+``configuration_nandi.py`` / ``modeling_nandi.py`` files).
+
+Why this script (and not ``Smoltron/examples/llama/convert_nanotron_to_hf.py``):
+    - The official converter has to instantiate a full nanotron model on CUDA
+      to call ``nanotron.serialize.load_weights``. Here we only need the tensors,
+      not the parallel runtime, so we read the safetensors files straight from
+      the layout that ``nanotron.serialize.save_weights`` produced.
+    - The 500M run uses ``shared_kv=True`` (V tied to K) and ``qk_norm=True``,
+      which the existing converter only knows how to express in a separate
+      ``transformers``-5.x environment (see ``build_nandi_hf_from_dump.py``).
+      This script collapses both phases into a single dependency-light step
+      (``torch`` + ``safetensors``) so we can run it locally.
+
+Layout we expect under ``--checkpoint_path``:
+
+    model_config.json
+    model/
+        token_position_embeddings/pp_block/token_embedding/model_weight_*.safetensors
+        final_layer_norm/pp_block/model_weight.safetensors
+        decoder/{i}/pp_block/attn/qkv_proj/model_weight_*.safetensors
+        decoder/{i}/pp_block/attn/o_proj/model_weight_*.safetensors
+        decoder/{i}/pp_block/attn/q_norm/model_weight.safetensors       # if qk_norm
+        decoder/{i}/pp_block/attn/k_norm/model_weight.safetensors       # if qk_norm
+        decoder/{i}/pp_block/mlp/gate_up_proj/model_weight_*.safetensors
+        decoder/{i}/pp_block/mlp/down_proj/model_weight_*.safetensors
+        decoder/{i}/pp_block/input_layernorm/model_weight.safetensors
+        decoder/{i}/pp_block/post_attention_layernorm/model_weight.safetensors
+        (optionally) lm_head/pp_block/model_weight_*.safetensors        # if not tied
+        (optionally) token_position_embeddings/pp_block/embedding_proj/...   # if factorized
+
+Each per-parameter file stores a single tensor under the key ``"data"``.
+
+Usage::
+
+    python convert_smoltron_to_hf.py \
+        --checkpoint_path /path/to/checkpoint_45000 \
+        --save_path      /path/to/Nandi-Mini-500M \
+        --dtype          bfloat16
+
+When ``--save_path`` already contains the Nandi custom code files (this is the
+intended layout for the Nandi-Mini-500M HF repo), they are left untouched and
+only ``config.json`` + ``model.safetensors`` are (over)written.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+from pathlib import Path
+
+import torch
+from safetensors.torch import load_file, save_file
+
+
+DTYPE_MAP = {"float32": torch.float32, "float16": torch.float16, "bfloat16": torch.bfloat16}
+
+
+def _resolve(path: Path) -> Path:
+    """Return the safetensors file inside ``path``.
+
+    Nanotron writes either ``model_weight.safetensors`` (replicated tensors:
+    norms, embeddings under TP=1) or ``model_weight_pp-rank-*-of-*_tp-rank-*-of-*.safetensors``
+    (sharded tensors). With pp=tp=1 there's exactly one file in either case.
+    """
+    if not path.is_dir():
+        raise FileNotFoundError(f"Expected directory: {path}")
+    candidates = sorted(p for p in path.iterdir() if p.suffix == ".safetensors")
+    if not candidates:
+        raise FileNotFoundError(f"No safetensors files in {path}")
+    if len(candidates) > 1:
+        # Multiple shards (pp>1 or tp>1) — would need to concat along the right dim,
+        # which depends on the parameter. We don't support that here.
+        raise NotImplementedError(
+            f"Multiple shards under {path}; this script only supports pp=tp=1 checkpoints."
+        )
+    return candidates[0]
+
+
+def _load_tensor(path: Path) -> torch.Tensor:
+    f = _resolve(path)
+    sd = load_file(str(f))
+    if "data" not in sd:
+        raise KeyError(f"Expected key 'data' in {f}, got {list(sd)}")
+    return sd["data"]
+
+
+def build_hf_state_dict(
+    checkpoint_path: Path,
+    nt_cfg: dict,
+    dtype: torch.dtype,
+) -> dict[str, torch.Tensor]:
+    """Read the nanotron on-disk layout and produce an HF Nandi state_dict."""
+    n_layers = nt_cfg["num_hidden_layers"]
+    n_q = nt_cfg["num_attention_heads"]
+    n_kv = nt_cfg["num_key_value_heads"]
+    hidden = nt_cfg["hidden_size"]
+    head_dim = hidden // n_q
+    inter = nt_cfg["intermediate_size"]
+    hq = n_q * head_dim
+    hk = n_kv * head_dim
+
+    shared_kv = bool(nt_cfg.get("shared_kv", False))
+    qk_norm = bool(nt_cfg.get("qk_norm", False))
+    factorized_embedding = bool(nt_cfg.get("factorized_embedding", False))
+    tied = bool(nt_cfg.get("tie_word_embeddings", True))
+
+    # Nanotron writes parameters under ``<root_folder>/model/<param_path>``.
+    # Smoltron historically called ``serialize.save_weights(root_folder=<ckpt>/model)``,
+    # producing a ``<ckpt>/model/model/...`` layout, which is what we see in the wild.
+    # Accept either nesting depth so the script also works on freshly produced ckpts.
+    candidate_roots = [
+        checkpoint_path / "model" / "model",
+        checkpoint_path / "model",
+    ]
+    for c in candidate_roots:
+        if (c / "token_position_embeddings").is_dir():
+            root = c
+            break
+    else:
+        raise FileNotFoundError(
+            f"Could not locate the nanotron weight tree under {checkpoint_path}; "
+            f"looked at {[str(c) for c in candidate_roots]}"
+        )
+    print(f"  resolved weight root -> {root}")
+    sd: dict[str, torch.Tensor] = {}
+
+    embed = _load_tensor(root / "token_position_embeddings" / "pp_block" / "token_embedding")
+    sd["model.embed_tokens.weight"] = embed
+    sd["model.norm.weight"] = _load_tensor(root / "final_layer_norm" / "pp_block")
+
+    if factorized_embedding:
+        # Either side may live in one of a few possible directories depending on
+        # the Smoltron version that wrote the checkpoint.
+        candidates_in = [
+            root / "token_position_embeddings" / "pp_block" / "embedding_proj",
+            root / "token_position_embeddings" / "pp_block" / "token_embedding_projection",
+            root / "embedding_proj" / "pp_block",
+        ]
+        for c in candidates_in:
+            if c.is_dir():
+                sd["model.embedding_proj.weight"] = _load_tensor(c)
+                break
+        else:
+            raise FileNotFoundError(
+                f"factorized_embedding=True but no embedding_proj dir under {root}"
+            )
+
+        candidates_out = [
+            root / "lm_head_proj" / "pp_block",
+            root / "lm_head" / "pp_block" / "lm_head_proj",
+        ]
+        for c in candidates_out:
+            if c.is_dir():
+                sd["lm_head_proj.weight"] = _load_tensor(c)
+                break
+        else:
+            raise FileNotFoundError(
+                f"factorized_embedding=True but no lm_head_proj dir under {root}"
+            )
+
+    if not tied:
+        sd["lm_head.weight"] = _load_tensor(root / "lm_head" / "pp_block")
+
+    for i in range(n_layers):
+        layer_root = root / "decoder" / str(i) / "pp_block"
+        hf_p = f"model.layers.{i}"
+
+        # Combined qkv -> q, k, v.
+        qkv = _load_tensor(layer_root / "attn" / "qkv_proj")
+        expected_first = hq + hk if shared_kv else hq + 2 * hk
+        if qkv.shape[0] != expected_first:
+            raise ValueError(
+                f"Layer {i}: qkv_proj first dim {qkv.shape[0]} != expected {expected_first} "
+                f"(hq={hq}, hk={hk}, shared_kv={shared_kv})"
+            )
+        sd[f"{hf_p}.self_attn.q_proj.weight"] = qkv[:hq].contiguous()
+        sd[f"{hf_p}.self_attn.k_proj.weight"] = qkv[hq : hq + hk].contiguous()
+        if not shared_kv:
+            sd[f"{hf_p}.self_attn.v_proj.weight"] = qkv[hq + hk : hq + 2 * hk].contiguous()
+        # else: `shared_kv=True` -> the HF `NandiAttention` has no v_proj module;
+        # V is recomputed from `k_proj` inside attention. Skip writing v_proj.weight.
+
+        sd[f"{hf_p}.self_attn.o_proj.weight"] = _load_tensor(layer_root / "attn" / "o_proj")
+
+        if qk_norm:
+            sd[f"{hf_p}.self_attn.q_norm.weight"] = _load_tensor(layer_root / "attn" / "q_norm")
+            sd[f"{hf_p}.self_attn.k_norm.weight"] = _load_tensor(layer_root / "attn" / "k_norm")
+
+        # Combined gate_up -> gate, up.
+        gate_up = _load_tensor(layer_root / "mlp" / "gate_up_proj")
+        if gate_up.shape[0] != 2 * inter:
+            raise ValueError(
+                f"Layer {i}: gate_up_proj first dim {gate_up.shape[0]} != 2*inter={2 * inter}"
+            )
+        sd[f"{hf_p}.mlp.gate_proj.weight"] = gate_up[:inter].contiguous()
+        sd[f"{hf_p}.mlp.up_proj.weight"] = gate_up[inter:].contiguous()
+        sd[f"{hf_p}.mlp.down_proj.weight"] = _load_tensor(layer_root / "mlp" / "down_proj")
+
+        sd[f"{hf_p}.input_layernorm.weight"] = _load_tensor(layer_root / "input_layernorm")
+        sd[f"{hf_p}.post_attention_layernorm.weight"] = _load_tensor(layer_root / "post_attention_layernorm")
+
+    # Cast every tensor to the requested dtype (after slicing so we slice the
+    # original storage, not a copy).
+    for k, v in list(sd.items()):
+        v = v.detach().to(dtype)
+        if not v.is_contiguous():
+            v = v.contiguous()
+        sd[k] = v
+
+    return sd
+
+
+def build_hf_config(nt_cfg: dict) -> dict:
+    """Build the dict that ends up as ``config.json``."""
+    hidden = nt_cfg["hidden_size"]
+    n_heads = nt_cfg["num_attention_heads"]
+    head_dim = hidden // n_heads
+
+    return {
+        "architectures": ["NandiForCausalLM"],
+        "auto_map": {
+            "AutoConfig": "configuration_nandi.NandiConfig",
+            "AutoModel": "modeling_nandi.NandiModel",
+            "AutoModelForCausalLM": "modeling_nandi.NandiForCausalLM",
+        },
+        "model_type": "nandi",
+        "vocab_size": nt_cfg["vocab_size"],
+        "hidden_size": hidden,
+        "intermediate_size": nt_cfg["intermediate_size"],
+        "num_hidden_layers": nt_cfg["num_hidden_layers"],
+        "num_attention_heads": n_heads,
+        "num_key_value_heads": nt_cfg["num_key_value_heads"],
+        "head_dim": head_dim,
+        "hidden_act": nt_cfg["hidden_act"],
+        "max_position_embeddings": nt_cfg["max_position_embeddings"],
+        "initializer_range": nt_cfg.get("initializer_range", 0.008),
+        "rms_norm_eps": nt_cfg["rms_norm_eps"],
+        "use_cache": True,
+        "pad_token_id": nt_cfg.get("pad_token_id", None),
+        "bos_token_id": nt_cfg["bos_token_id"],
+        "eos_token_id": nt_cfg["eos_token_id"],
+        "tie_word_embeddings": bool(nt_cfg.get("tie_word_embeddings", True)),
+        "rope_parameters": {"rope_theta": float(nt_cfg["rope_theta"])},
+        "attention_bias": bool(nt_cfg.get("attention_bias", False)),
+        "attention_dropout": 0.0,
+        "mlp_bias": bool(nt_cfg.get("mlp_bias", False)),
+        "factorized_embedding": bool(nt_cfg.get("factorized_embedding", False)),
+        "embedding_rank": nt_cfg.get("embedding_rank", hidden),
+        "layer_sharing": bool(nt_cfg.get("layer_sharing", False)),
+        "layer_sharing_repeats": nt_cfg.get("layer_sharing_repeats", 1),
+        "qk_norm": bool(nt_cfg.get("qk_norm", False)),
+        "shared_kv": bool(nt_cfg.get("shared_kv", False)),
+        # Default inference-time cache strategy for shared_kv layers. "shared"
+        # caches raw K only (memory-saver, recomputes k_norm+RoPE per step);
+        # "vanilla" caches post-norm post-RoPE K and raw V separately (standard
+        # HF layout, faster decode). User can flip at runtime by setting
+        # `model.config.kv_cache_mode = "vanilla"` before generate.
+        "kv_cache_mode": "shared",
+        "torch_dtype": "bfloat16",
+    }
+
+
+def main() -> None:
+    p = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
+    p.add_argument("--checkpoint_path", type=Path, required=True, help="Smoltron checkpoint directory")
+    p.add_argument("--save_path", type=Path, required=True, help="Output HF directory")
+    p.add_argument("--dtype", choices=list(DTYPE_MAP), default="bfloat16")
+    args = p.parse_args()
+
+    dtype = DTYPE_MAP[args.dtype]
+    cfg_path = args.checkpoint_path / "model_config.json"
+    if not cfg_path.is_file():
+        raise FileNotFoundError(f"Missing nanotron config: {cfg_path}")
+    with open(cfg_path) as f:
+        nt_cfg = json.load(f)
+
+    print(
+        f"Source: {args.checkpoint_path}\n"
+        f"  layers={nt_cfg['num_hidden_layers']} "
+        f"hidden={nt_cfg['hidden_size']} "
+        f"heads={nt_cfg['num_attention_heads']} "
+        f"kv_heads={nt_cfg['num_key_value_heads']} "
+        f"shared_kv={nt_cfg.get('shared_kv', False)} "
+        f"qk_norm={nt_cfg.get('qk_norm', False)} "
+        f"factorized_embedding={nt_cfg.get('factorized_embedding', False)} "
+        f"tied_embeddings={nt_cfg.get('tie_word_embeddings', True)}"
+    )
+
+    print("Reading & remapping tensors...")
+    sd = build_hf_state_dict(args.checkpoint_path, nt_cfg, dtype)
+    print(f"  built {len(sd)} HF tensors")
+
+    args.save_path.mkdir(parents=True, exist_ok=True)
+
+    cfg = build_hf_config(nt_cfg)
+    cfg_out = args.save_path / "config.json"
+    with open(cfg_out, "w") as f:
+        json.dump(cfg, f, indent=2)
+    print(f"Wrote {cfg_out}")
+
+    weights_out = args.save_path / "model.safetensors"
+    save_file(sd, str(weights_out), metadata={"format": "pt"})
+    n_params = sum(v.numel() for v in sd.values())
+    print(
+        f"Wrote {weights_out} "
+        f"({os.path.getsize(weights_out) / 1e6:.1f} MB, ~{n_params / 1e6:.1f}M tensors elements)"
+    )
+
+    print("Done.")
+
+
+if __name__ == "__main__":
+    main()

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d6a1debc23b05874a5bc3a0d7f2f399714f7e6cef3ba1795542ee00d192b400d
+size 1290954432

modeling_nandi.py

@@ -0,0 +1,533 @@
+from collections.abc import Callable
+
+import torch
+import torch.nn as nn
+
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache, DynamicLayer
+from transformers.generation import GenerationMixin
+from transformers.integrations import use_kernel_forward_from_hub
+from transformers.masking_utils import create_causal_mask
+from transformers.modeling_layers import GradientCheckpointingLayer
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
+from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
+from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from transformers.processing_utils import Unpack
+from transformers.utils import TransformersKwargs, auto_docstring
+from transformers.utils.deprecation import deprecate_kwarg
+from transformers.utils.generic import can_return_tuple, merge_with_config_defaults
+from transformers.utils.output_capturing import capture_outputs
+from .configuration_nandi import NandiConfig
+
+
+@use_kernel_forward_from_hub("RMSNorm")
+class NandiRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+    def extra_repr(self):
+        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
+
+
+class NandiRotaryEmbedding(nn.Module):
+    inv_freq: torch.Tensor
+
+    def __init__(self, config: NandiConfig, device=None):
+        super().__init__()
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+        self.rope_type = self.config.rope_parameters.get("rope_type", "default")
+        rope_init_fn: Callable = self.compute_default_rope_parameters
+        if self.rope_type != "default":
+            rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        inv_freq, self.attention_scaling = rope_init_fn(self.config, device)
+
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)
+
+    @staticmethod
+    def compute_default_rope_parameters(
+        config: NandiConfig | None = None,
+        device: torch.device | None = None,
+        seq_len: int | None = None,
+    ) -> tuple[torch.Tensor, float]:
+        del seq_len
+        base = config.rope_parameters["rope_theta"]
+        dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads
+        attention_factor = 1.0
+        inv_freq = 1.0 / (
+            base ** (torch.arange(0, dim, 2, dtype=torch.int64).to(device=device, dtype=torch.float) / dim)
+        )
+        return inv_freq, attention_factor
+
+    @torch.no_grad()
+    @dynamic_rope_update
+    def forward(self, x, position_ids):
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
+        position_ids_expanded = position_ids[:, None, :].float()
+
+        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos() * self.attention_scaling
+            sin = emb.sin() * self.attention_scaling
+
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    del position_ids
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+def eager_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: torch.Tensor | None,
+    scaling: float,
+    dropout: float = 0.0,
+    **kwargs: Unpack[TransformersKwargs],
+):
+    del kwargs
+    key_states = repeat_kv(key, module.num_key_value_groups)
+    value_states = repeat_kv(value, module.num_key_value_groups)
+
+    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
+    if attention_mask is not None:
+        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+        attn_weights = attn_weights + causal_mask
+
+    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
+    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
+    attn_output = torch.matmul(attn_weights, value_states)
+    attn_output = attn_output.transpose(1, 2).contiguous()
+
+    return attn_output, attn_weights
+
+
+class NandiAttention(nn.Module):
+    def __init__(self, config: NandiConfig, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.head_dim = config.head_dim
+        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
+        self.scaling = self.head_dim**-0.5
+        self.attention_dropout = config.attention_dropout
+        self.is_causal = True
+
+        self.shared_kv = getattr(config, "shared_kv", False)
+
+        self.q_proj = nn.Linear(
+            config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias
+        )
+        self.k_proj = nn.Linear(
+            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
+        )
+        if self.shared_kv:
+            self.v_proj = None
+        else:
+            self.v_proj = nn.Linear(
+                config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
+            )
+        self.o_proj = nn.Linear(
+            config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
+        )
+
+        self.qk_norm = getattr(config, "qk_norm", False)
+        if self.qk_norm:
+            self.q_norm = NandiRMSNorm(self.head_dim, eps=config.rms_norm_eps)
+            self.k_norm = NandiRMSNorm(self.head_dim, eps=config.rms_norm_eps)
+        else:
+            self.q_norm = None
+            self.k_norm = None
+
+    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor],
+        attention_mask: torch.Tensor | None,
+        past_key_values: Cache | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        input_shape = hidden_states.shape[:-1]
+        hidden_shape = (*input_shape, -1, self.head_dim)
+
+        query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+        k_raw = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+
+        if self.shared_kv:
+            kv_cache_mode = getattr(self.config, "kv_cache_mode", "shared")
+
+            if self.qk_norm:
+                query_states = self.q_norm(query_states)
+
+            if kv_cache_mode == "shared":
+                if past_key_values is not None:
+                    empty_v = torch.empty(
+                        k_raw.shape[0],
+                        k_raw.shape[1],
+                        0,
+                        k_raw.shape[3],
+                        device=k_raw.device,
+                        dtype=k_raw.dtype,
+                    )
+                    k_raw_full, _ = past_key_values.update(k_raw, empty_v, self.layer_idx)
+                else:
+                    k_raw_full = k_raw
+
+                value_states = k_raw_full
+                key_states = self.k_norm(k_raw_full) if self.qk_norm else k_raw_full
+
+                cos, sin = position_embeddings
+                q_len = query_states.shape[-2]
+                cos_q = cos[..., -q_len:, :]
+                sin_q = sin[..., -q_len:, :]
+                query_states, _ = apply_rotary_pos_emb(query_states, query_states, cos_q, sin_q)
+                _, key_states = apply_rotary_pos_emb(key_states, key_states, cos, sin)
+
+            else:  
+                key_states = self.k_norm(k_raw) if self.qk_norm else k_raw
+                value_states = k_raw  
+
+                cos, sin = position_embeddings  
+                query_states, key_states = apply_rotary_pos_emb(
+                    query_states, key_states, cos, sin
+                )
+
+                if past_key_values is not None:
+                    key_states, value_states = past_key_values.update(
+                        key_states, value_states, self.layer_idx
+                    )
+
+        else:
+            value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+            key_states = k_raw
+
+            if self.qk_norm:
+                query_states = self.q_norm(query_states)
+                key_states = self.k_norm(key_states)
+
+            cos, sin = position_embeddings
+            query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+            if past_key_values is not None:
+                key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx)
+
+        attention_interface: Callable = eager_attention_forward
+        if self.config._attn_implementation != "eager":
+            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+
+        attn_output, attn_weights = attention_interface(
+            self,
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            dropout=0.0 if not self.training else self.attention_dropout,
+            scaling=self.scaling,
+            **kwargs,
+        )
+
+        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+        return attn_output, attn_weights
+
+
+class NandiMLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.gate_proj = nn.Linear(config.hidden_size, config.intermediate_size, bias=config.mlp_bias)
+        self.up_proj = nn.Linear(config.hidden_size, config.intermediate_size, bias=config.mlp_bias)
+        self.down_proj = nn.Linear(config.intermediate_size, config.hidden_size, bias=config.mlp_bias)
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, x):
+        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+
+
+class NandiDecoderLayer(GradientCheckpointingLayer):
+    def __init__(self, config: NandiConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.self_attn = NandiAttention(config=config, layer_idx=layer_idx)
+        self.mlp = NandiMLP(config)
+        self.input_layernorm = NandiRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = NandiRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: Cache | None = None,
+        use_cache: bool | None = False,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> torch.Tensor:
+        residual = hidden_states
+        hidden_states = self.input_layernorm(hidden_states)
+
+        hidden_states, _ = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            position_embeddings=position_embeddings,
+            **kwargs,
+        )
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+        return hidden_states
+
+
+class _VirtualLayerCache:
+    """Proxy that shifts cache layer indices by `offset` to give each repeat its own virtual slots."""
+
+    def __init__(self, cache: Cache, offset: int):
+        self._cache = cache
+        self._offset = offset
+
+    def __getattr__(self, name):
+        return getattr(self._cache, name)
+
+    def update(self, key_states, value_states, layer_idx, cache_kwargs=None):
+        virtual_idx = layer_idx + self._offset
+        # grow the backing cache if generate() pre-allocated fewer slots than needed
+        while len(self._cache.layers) <= virtual_idx:
+            self._cache.layers.append(DynamicLayer())
+        return self._cache.update(key_states, value_states, virtual_idx, cache_kwargs)
+
+    def get_seq_length(self, layer_idx: int = 0) -> int:
+        return self._cache.get_seq_length(layer_idx + self._offset)
+
+
+@auto_docstring
+class NandiPreTrainedModel(PreTrainedModel):
+    config: NandiConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["NandiDecoderLayer"]
+    _skip_keys_device_placement = ["past_key_values"]
+    _supports_flash_attn = True
+    _supports_sdpa = True
+    _supports_flex_attn = True
+    _can_compile_fullgraph = True
+    _supports_attention_backend = True
+    _can_record_outputs = {
+        "hidden_states": NandiDecoderLayer,
+        "attentions": NandiAttention,
+    }
+
+    def __init__(self, config: NandiConfig):
+        super().__init__(config)
+
+
+@auto_docstring
+class NandiModel(NandiPreTrainedModel):
+    def __init__(self, config: NandiConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+        embedding_dim = config.embedding_rank if config.factorized_embedding else config.hidden_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, embedding_dim, self.padding_idx)
+        self.embedding_proj = (
+            nn.Linear(config.embedding_rank, config.hidden_size, bias=False) if config.factorized_embedding else None
+        )
+        self.layers = nn.ModuleList(
+            [NandiDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.norm = NandiRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = NandiRotaryEmbedding(config=config)
+        self.gradient_checkpointing = False
+
+        self.post_init()
+
+    @merge_with_config_defaults
+    @capture_outputs
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: torch.LongTensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: Cache | None = None,
+        inputs_embeds: torch.FloatTensor | None = None,
+        use_cache: bool | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> BaseModelOutputWithPast:
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if self.embedding_proj is not None:
+            inputs_embeds = self.embedding_proj(inputs_embeds)
+        repeats = max(1, int(getattr(self.config, "layer_sharing_repeats", 1) or 1))
+
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache()
+
+        if position_ids is None:
+            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+            position_ids = torch.arange(inputs_embeds.shape[1], device=inputs_embeds.device) + past_seen_tokens
+            position_ids = position_ids.unsqueeze(0)
+
+        causal_mask = create_causal_mask(
+            config=self.config,
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            position_ids=position_ids,
+        )
+
+        hidden_states = inputs_embeds
+        kv_cache_mode = getattr(self.config, "kv_cache_mode", "shared")
+        if (
+            getattr(self.config, "shared_kv", False)
+            and kv_cache_mode == "shared"
+            and past_key_values is not None
+        ):
+            past_len = past_key_values.get_seq_length(0)
+            cur_len = inputs_embeds.shape[1]
+            full_position_ids = torch.arange(
+                past_len + cur_len, device=inputs_embeds.device
+            ).unsqueeze(0)
+            position_embeddings = self.rotary_emb(hidden_states, position_ids=full_position_ids)
+        else:
+            position_embeddings = self.rotary_emb(hidden_states, position_ids=position_ids)
+
+        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
+            for repeat_idx in range(repeats):
+
+                repeat_cache = (
+                    _VirtualLayerCache(past_key_values, repeat_idx * self.config.num_hidden_layers)
+                    if (past_key_values is not None and repeat_idx > 0)
+                    else past_key_values
+                )
+                hidden_states = decoder_layer(
+                    hidden_states,
+                    attention_mask=causal_mask,
+                    position_embeddings=position_embeddings,
+                    position_ids=position_ids,
+                    past_key_values=repeat_cache,
+                    use_cache=use_cache,
+                    **kwargs,
+                )
+
+        hidden_states = self.norm(hidden_states)
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values,
+        )
+
+
+@auto_docstring
+class NandiForCausalLM(NandiPreTrainedModel, GenerationMixin):
+    _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
+    _tp_plan = {"lm_head": "colwise_gather_output"}
+    _pp_plan = {
+        "lm_head_proj": (["hidden_states"], ["hidden_states"]),
+        "lm_head": (["hidden_states"], ["logits"]),
+    }
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = NandiModel(config)
+        self.vocab_size = config.vocab_size
+
+        lm_head_in_features = config.embedding_rank if config.factorized_embedding else config.hidden_size
+        self.lm_head_proj = (
+            nn.Linear(config.hidden_size, config.embedding_rank, bias=False) if config.factorized_embedding else None
+        )
+        self.lm_head = nn.Linear(lm_head_in_features, config.vocab_size, bias=False)
+
+        self.post_init()
+
+    @can_return_tuple
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: torch.LongTensor | None = None,
+        attention_mask: torch.Tensor | None = None,
+        position_ids: torch.LongTensor | None = None,
+        past_key_values: Cache | None = None,
+        inputs_embeds: torch.FloatTensor | None = None,
+        labels: torch.LongTensor | None = None,
+        use_cache: bool | None = None,
+        logits_to_keep: int | torch.Tensor = 0,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> CausalLMOutputWithPast:
+        outputs: BaseModelOutputWithPast = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            **kwargs,
+        )
+
+        hidden_states = outputs.last_hidden_state
+        if self.lm_head_proj is not None:
+            hidden_states = self.lm_head_proj(hidden_states)
+
+        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+__all__ = ["NandiPreTrainedModel", "NandiModel", "NandiForCausalLM"]

tokenization_nandi.py

@@ -0,0 +1,143 @@
+# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
+#
+# 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.
+"""Tokenization classes for the Nandi family."""
+
+from tokenizers import Regex, Tokenizer, decoders, normalizers, pre_tokenizers
+from tokenizers.models import BPE
+
+from transformers.tokenization_utils_tokenizers import TokenizersBackend
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+PRETOKENIZE_REGEX = r"""(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?(?:\p{L}\p{M}*)+|\p{N}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+"""
+
+
+class NandiTokenizer(TokenizersBackend):
+    model_input_names = ["input_ids", "attention_mask"]
+    model = BPE
+
+    def __init__(
+        self,
+        vocab: str | dict[str, int] | None = None,
+        merges: str | list[str] | None = None,
+        vocab_file=None,
+        merges_file=None,
+        unk_token: str = "<|endoftext|>",
+        bos_token: str = "<|im_start|>",
+        eos_token: str = "<|endoftext|>",
+        pad_token: str = "<|pad|>",
+        add_prefix_space: bool | None = None,
+        **kwargs,
+    ):
+        self._vocab = (
+            vocab
+            if vocab is not None
+            else {
+                "<|endoftext|>": 0,
+            }
+        )
+        self._merges = merges or []
+
+        self._tokenizer = Tokenizer(
+            BPE(
+                vocab=self._vocab,
+                merges=self._merges,
+                dropout=None,
+                unk_token=None,
+                continuing_subword_prefix="",
+                end_of_word_suffix="",
+                fuse_unk=False,
+                byte_fallback=False,
+            )
+        )
+        self._tokenizer.decoder = decoders.ByteLevel()
+        self._tokenizer.normalizer = normalizers.NFC()
+        self._tokenizer.pre_tokenizer = pre_tokenizers.Sequence(
+            [
+                pre_tokenizers.Split(
+                    Regex(PRETOKENIZE_REGEX),
+                    behavior="isolated",
+                    invert=False,
+                ),
+                pre_tokenizers.ByteLevel(
+                    add_prefix_space=False,
+                    trim_offsets=True,
+                    use_regex=False
+                ),
+            ]
+        )
+
+        super().__init__(
+            vocab_file=vocab_file,
+            merges_file=merges_file,
+            unk_token=unk_token,
+            bos_token=bos_token,
+            eos_token=eos_token,
+            pad_token=pad_token,
+            add_prefix_space=add_prefix_space,
+            **kwargs,
+        )
+
+    def __call__(self, text, *args, **kwargs):
+        add_special_tokens = kwargs.get("add_special_tokens", False)
+    
+        def add_prefix(t):
+            if isinstance(t, str):
+                return "<|im_start|> " + t
+            return t
+    
+        # Only inject when special tokens are disabled
+        if not add_special_tokens:
+            if isinstance(text, list):
+                text = [add_prefix(t) for t in text]
+            else:
+                text = add_prefix(text)
+    
+        return super().__call__(text, *args, **kwargs)
+
+
+
+    # def encode(
+    #     self,
+    #     text,
+    #     text_pair=None,
+    #     add_special_tokens: bool = True,
+    #     padding=False,
+    #     truncation=None,
+    #     max_length=None,
+    #     stride: int = 0,
+    #     padding_side=None,
+    #     return_tensors=None,
+    #     **kwargs,
+    # ):
+    #     if isinstance(text, str):
+    #         # This is a temporary fix to match the behaviour of the training pipeline
+    #         text = "<|im_start|>" + " " + text
+    #     return super().encode(
+    #         text,
+    #         text_pair=text_pair,
+    #         add_special_tokens=add_special_tokens,
+    #         padding=padding,
+    #         truncation=truncation,
+    #         max_length=max_length,
+    #         stride=stride,
+    #         padding_side=padding_side,
+    #         return_tensors=return_tensors,
+    #         **kwargs,
+    #     )
+
+
+__all__ = ["NandiTokenizer"]

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dc858b896e5a8c86097420a23a90b8a0ad4d2ad23250e1fae99b89f525120c90
+size 12460256

tokenizer_config.json

@@ -0,0 +1,15 @@
+{
+  "backend": "tokenizers",
+  "bos_token": "<|im_start|>",
+  "bos_token_id": 1,
+  "eos_token": "<|endoftext|>",
+  "eos_token_id": 0,
+  "pad_token": "<|pad|>",
+  "pad_token_id": 3,
+  "model_max_length": 2048,
+  "tokenizer_class": "NandiTokenizer",
+  "unk_token": "<|endoftext|>",
+  "auto_map": {
+    "AutoTokenizer": ["tokenization_nandi.NandiTokenizer", null]
+  }
+}