create_paligemma_gguf.py

"""Copyright (c) 2024 Andrei Betlen

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."""

import os
import json
import typing
import pathlib
import argparse

import numpy as np
import numpy.typing as npt

import gguf

from safetensors import safe_open


class SafetensorsIndexFile(typing.TypedDict):
    weight_map: typing.Dict[str, str]


class SafetensorsIndex:
    def __init__(self, index_file_path: str):
        directory = os.path.dirname(index_file_path)
        self.index = typing.cast(SafetensorsIndexFile, json.load(open(index_file_path)))
        self.weight_map = self.index["weight_map"]
        files = set(self.weight_map.values())
        self.tensors = {file: safe_open(os.path.join(directory, file), framework="np") for file in files}

    def get_tensor(self, key: str) -> npt.NDArray[np.float32]:
        return typing.cast(npt.NDArray[np.float32], self.tensors[self.weight_map[key]].get_tensor(key)) # type: ignore


def k(raw_key: str, arch: str) -> str:
    return raw_key.format(arch=arch)

def does_token_look_special(token: typing.Union[str, bytes]) -> bool:
    if isinstance(token, (bytes, bytearray)):
        token_text = token.decode(encoding="utf-8")
    elif isinstance(token, memoryview):
        token_text = token.tobytes().decode(encoding="utf-8")
    else:
        token_text = token

    # Some models mark some added tokens which ought to be control tokens as not special.
    # (e.g. command-r, command-r-plus, deepseek-coder, gemma{,-2})
    seems_special = token_text in (
        "<pad>",  # deepseek-coder
        "<mask>", "<2mass>", "[@BOS@]",  # gemma{,-2}
    )

    seems_special = seems_special or (token_text.startswith("<|") and token_text.endswith("|>"))
    seems_special = seems_special or (token_text.startswith("<｜") and token_text.endswith("｜>"))  # deepseek-coder

    # TODO: should these be marked as UNUSED instead? (maybe not)
    seems_special = seems_special or (token_text.startswith("<unused") and token_text.endswith(">"))  # gemma{,-2}

    return seems_special


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "-d",
        "--dir-model",
        required=True,
        help="path to directory containing the tokenizer",
    )
    args = parser.parse_args()

    dir_model = pathlib.Path(args.dir_model)

    # set model name to folder name
    name = dir_model.name

    tensors = SafetensorsIndex((dir_model / "model.safetensors.index.json").as_posix())

    config = json.load(open(dir_model / "config.json"))
    text_config = {
        "max_position_embeddings": 8192,
        "rms_norm_eps": 1e-6,
        "head_dim": 256
    }
    text_config.update(config["text_config"])
    vision_config = config["vision_config"]

    preprocessor_config = json.load(open(dir_model / "preprocessor_config.json"))

    ### Vision model

    ftype = 1  # fp16

    fname_middle = "mmproj-"
    has_text_encoder = False
    has_llava_projector = True

    n_layers_clip = vision_config["num_hidden_layers"]

    fname_out = f"{name}-mmproj-f16.gguf"
    fout = gguf.GGUFWriter(fname_out, arch="clip")

    fout.add_bool("clip.has_text_encoder", False)
    fout.add_bool("clip.has_vision_encoder", True)
    fout.add_bool("clip.has_llava_projector", True)
    fout.add_file_type(ftype)  # fp16

    model_name = f"google/{name}"
    fout.add_name(model_name)
    fout.add_description("image encoder for " + model_name)
    fout.add_string("clip.projector_type", "mlp")

    image_size = vision_config.get("image_size", preprocessor_config["size"]["height"])

    # vision model hparams
    VISION = "clip.vision"
    fout.add_uint32("clip.vision.image_size", image_size)
    fout.add_uint32("clip.vision.patch_size", vision_config["patch_size"])
    fout.add_uint32(k(gguf.KEY_EMBEDDING_LENGTH, VISION), vision_config["hidden_size"])
    fout.add_uint32(k(gguf.KEY_FEED_FORWARD_LENGTH, VISION), vision_config["intermediate_size"])
    fout.add_uint32("clip.vision.projection_dim", vision_config["projection_dim"])
    fout.add_uint32(k(gguf.KEY_ATTENTION_HEAD_COUNT, VISION), vision_config["num_attention_heads"])
    fout.add_float32(k(gguf.KEY_ATTENTION_LAYERNORM_EPS, VISION), 1e-6)
    fout.add_uint32(k(gguf.KEY_BLOCK_COUNT, VISION), n_layers_clip + 1)

    fout.add_array("clip.vision.image_mean", preprocessor_config["image_mean"])
    fout.add_array("clip.vision.image_std", preprocessor_config["image_std"])
    fout.add_bool("clip.use_gelu", vision_config["projector_hidden_act"] == "gelu")
    fout.add_float32("clip.embeddings_scale", 1.0 / (config["projection_dim"]**0.5))

    # vision projection
    fout.add_tensor(
        "mm.0.weight",
        tensors.get_tensor("multi_modal_projector.linear.weight").astype(np.float16),
    )
    fout.add_tensor(
        "mm.0.bias",
        tensors.get_tensor("multi_modal_projector.linear.bias").astype(np.float32),
    )

    # encoder (siglip)
    fout.add_tensor(
        "v.position_embd.weight",
        tensors.get_tensor("vision_tower.vision_model.embeddings.position_embedding.weight").astype(np.float16),
    )
    fout.add_tensor(
        "v.patch_embd.weight",
        tensors.get_tensor("vision_tower.vision_model.embeddings.patch_embedding.weight")
            .reshape(vision_config["hidden_size"], 3, vision_config["patch_size"], vision_config["patch_size"])
            .astype(np.float16),
    )
    fout.add_tensor(
        "v.patch_embd.bias",
        tensors.get_tensor("vision_tower.vision_model.embeddings.patch_embedding.bias").astype(np.float32),
    )

    fout.add_tensor(
        "v.post_ln.weight",
        tensors.get_tensor("vision_tower.vision_model.post_layernorm.weight").astype(np.float32),
    )
    fout.add_tensor(
        "v.post_ln.bias",
        tensors.get_tensor("vision_tower.vision_model.post_layernorm.bias").astype(np.float32),
    )

    def blk_tensor(i: int, name: str):
        return tensors.get_tensor(
            rf"vision_tower.vision_model.encoder.layers.{i}.{name}"
        )

    def add_tensor(blk_id: int, gguf_id: typing.Optional[int] = None):
        if gguf_id is None:
            gguf_id = blk_id

        q_w = blk_tensor(blk_id, "self_attn.q_proj.weight")
        k_w = blk_tensor(blk_id, "self_attn.k_proj.weight")
        v_w = blk_tensor(blk_id, "self_attn.v_proj.weight")
        q_b = blk_tensor(blk_id, "self_attn.q_proj.bias")
        k_b = blk_tensor(blk_id, "self_attn.k_proj.bias")
        v_b = blk_tensor(blk_id, "self_attn.v_proj.bias")

        fout.add_tensor(f"v.blk.{gguf_id}.attn_q.weight", q_w.astype(np.float16))
        fout.add_tensor(f"v.blk.{gguf_id}.attn_q.bias", q_b.astype(np.float32))
        fout.add_tensor(f"v.blk.{gguf_id}.attn_k.weight", k_w.astype(np.float16))
        fout.add_tensor(f"v.blk.{gguf_id}.attn_k.bias", k_b.astype(np.float32))
        fout.add_tensor(f"v.blk.{gguf_id}.attn_v.weight", v_w.astype(np.float16))
        fout.add_tensor(f"v.blk.{gguf_id}.attn_v.bias", v_b.astype(np.float32))
        fout.add_tensor(
            f"v.blk.{gguf_id}.attn_out.weight",
            blk_tensor(blk_id, "self_attn.out_proj.weight").astype(np.float16),
        )
        fout.add_tensor(
            f"v.blk.{gguf_id}.attn_out.bias",
            blk_tensor(blk_id, "self_attn.out_proj.bias").astype(np.float32),
        )

        fout.add_tensor(
            f"v.blk.{gguf_id}.ln1.weight",
            blk_tensor(blk_id, "layer_norm1.weight").astype(np.float32),
        )
        fout.add_tensor(
            f"v.blk.{gguf_id}.ln1.bias",
            blk_tensor(blk_id, "layer_norm1.bias").astype(np.float32),
        )

        fout.add_tensor(
            f"v.blk.{gguf_id}.ffn_down.weight",
            blk_tensor(blk_id, "mlp.fc1.weight").astype(np.float16),
        )
        fout.add_tensor(
            f"v.blk.{gguf_id}.ffn_down.bias",
            blk_tensor(blk_id, "mlp.fc1.bias").astype(np.float32),
        )
        fout.add_tensor(
            f"v.blk.{gguf_id}.ffn_up.weight",
            blk_tensor(blk_id, "mlp.fc2.weight").astype(np.float16),
        )
        fout.add_tensor(
            f"v.blk.{gguf_id}.ffn_up.bias",
            blk_tensor(blk_id, "mlp.fc2.bias").astype(np.float32),
        )

        fout.add_tensor(
            f"v.blk.{gguf_id}.ln2.weight",
            blk_tensor(blk_id, "layer_norm2.weight").astype(np.float32),
        )
        fout.add_tensor(
            f"v.blk.{gguf_id}.ln2.bias",
            blk_tensor(blk_id, "layer_norm2.bias").astype(np.float32),
        )

    for i in range(n_layers_clip):
        add_tensor(i)

    # Duplicate the last block (llava-cli skips over this)
    add_tensor(n_layers_clip - 1, n_layers_clip)

    fout.write_header_to_file()
    fout.write_kv_data_to_file()
    fout.write_tensors_to_file()
    fout.close()

    print(f"GGUF written to {fname_out}")

    ### Text model

    # general GGUF init
    fname_out = f"{name}-text-model-f16.gguf"
    fout = gguf.GGUFWriter(fname_out, arch="gemma")
    ftype = 1
    
    block_count = text_config["num_hidden_layers"]

    fout.add_name(name)
    fout.add_context_length(text_config["max_position_embeddings"])
    fout.add_embedding_length(text_config["hidden_size"])
    fout.add_block_count(block_count)
    fout.add_feed_forward_length(text_config["intermediate_size"])
    fout.add_head_count(text_config["num_attention_heads"])
    fout.add_head_count_kv(text_config.get("num_key_value_heads") or text_config["num_attention_heads"])
    fout.add_layer_norm_rms_eps(text_config["rms_norm_eps"])
    fout.add_key_length(text_config["head_dim"])
    fout.add_value_length(text_config["head_dim"])
    fout.add_file_type(ftype)
    # fout.add_add_bos_token(True)

    ### Tokenizer

    # Taken from _set_vocab_sentencepiece
    from enum import IntEnum
    class SentencePieceTokenTypes(IntEnum):
        NORMAL = 1
        UNKNOWN = 2
        CONTROL = 3
        USER_DEFINED = 4
        UNUSED = 5
        BYTE = 6

    from sentencepiece import SentencePieceProcessor
    tokenizer_path = dir_model / 'tokenizer.model'

    tokens: typing.List[bytes] = []
    scores: typing.List[float] = []
    toktypes: typing.List[int] = []

    if not tokenizer_path.is_file():
        raise FileNotFoundError(f"File not found: {tokenizer_path}")

    tokenizer = SentencePieceProcessor()
    tokenizer.LoadFromFile(str(tokenizer_path))

    vocab_size = config["vocab_size"]

    tokens: typing.List[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
    scores: typing.List[float] = [-10000.0] * vocab_size
    toktypes: typing.List[int] = [SentencePieceTokenTypes.UNKNOWN] * vocab_size

    for token_id in range(tokenizer.vocab_size()):
        piece = tokenizer.IdToPiece(token_id)
        text = piece.encode("utf-8")
        score = tokenizer.GetScore(token_id)

        toktype = SentencePieceTokenTypes.NORMAL
        if tokenizer.IsUnknown(token_id):
            toktype = SentencePieceTokenTypes.UNKNOWN
        elif tokenizer.IsControl(token_id):
            toktype = SentencePieceTokenTypes.CONTROL
        elif tokenizer.IsUnused(token_id):
            toktype = SentencePieceTokenTypes.UNUSED
        elif tokenizer.IsByte(token_id):
            toktype = SentencePieceTokenTypes.BYTE

        tokens[token_id] = text
        scores[token_id] = score
        toktypes[token_id] = toktype

    added_tokens_file = dir_model / 'added_tokens.json'
    if added_tokens_file.is_file():
        with open(added_tokens_file, "r", encoding="utf-8") as f:
            added_tokens_json = json.load(f)
            for key in added_tokens_json:
                token_id = added_tokens_json[key]
                if (token_id >= vocab_size):
                    print(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}')
                    continue

                tokens[token_id] = key.encode("utf-8")
                scores[token_id] = -1000.0
                toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED

    tokenizer_config_file = dir_model / 'tokenizer_config.json'
    if tokenizer_config_file.is_file():
        with open(tokenizer_config_file, "r", encoding="utf-8") as f:
            tokenizer_config_json = json.load(f)
            added_tokens_decoder = tokenizer_config_json.get("added_tokens_decoder", {})
            for token_id, token_data in added_tokens_decoder.items():
                token_id = int(token_id)
                token: str = token_data["content"]
                if toktypes[token_id] != SentencePieceTokenTypes.UNUSED:
                    if tokens[token_id] != token.encode("utf-8"):
                        logger.warning(f'replacing token {token_id}: {tokens[token_id].decode("utf-8")!r} -> {token!r}')
                if token_data.get("special") or does_token_look_special(token):
                    toktypes[token_id] = SentencePieceTokenTypes.CONTROL
                else:
                    token = token.replace(b"\xe2\x96\x81".decode("utf-8"), " ")  # pre-normalize user-defined spaces
                    toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED

                scores[token_id] = -1000.0
                tokens[token_id] = token.encode("utf-8")

    if vocab_size > len(tokens):
        pad_count = vocab_size - len(tokens)
        print(f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]")
        for i in range(1, pad_count + 1):
            tokens.append(bytes(f"[PAD{i}]", encoding="utf-8"))
            scores.append(-1000.0)
            toktypes.append(SentencePieceTokenTypes.UNUSED)

    fout.add_tokenizer_model("llama")
    fout.add_tokenizer_pre("default")
    fout.add_token_list(tokens)
    fout.add_token_scores(scores)
    fout.add_token_types(toktypes)

    special_vocab = gguf.SpecialVocab(dir_model, n_vocab=len(tokens))
    special_vocab.add_to_gguf(fout)
    fout.add_add_space_prefix(False)

    ### Text model

    fout.add_tensor(
        "token_embd.weight",
        tensors.get_tensor("language_model.model.embed_tokens.weight").astype(np.float16),
    )

    for i in range(text_config["num_hidden_layers"]):
        fout.add_tensor(
            f"blk.{i}.attn_norm.weight",
            tensors.get_tensor(f"language_model.model.layers.{i}.input_layernorm.weight").astype(
                np.float32
            # https://github.com/huggingface/transformers/blob/fc37f38915372c15992b540dfcbbe00a916d4fc6/src/transformers/models/gemma/modeling_gemma.py#L89
            ) + 1,
        )

        fout.add_tensor(
            f"blk.{i}.ffn_down.weight",
            tensors.get_tensor(f"language_model.model.layers.{i}.mlp.down_proj.weight").astype(
                np.float16
            ),
        )
        fout.add_tensor(
            f"blk.{i}.ffn_gate.weight",
            tensors.get_tensor(f"language_model.model.layers.{i}.mlp.gate_proj.weight").astype(
                np.float16
            ),
        )

        fout.add_tensor(
            f"blk.{i}.ffn_up.weight",
            tensors.get_tensor(f"language_model.model.layers.{i}.mlp.up_proj.weight").astype(
                np.float16
            ),
        )

        fout.add_tensor(
            f"blk.{i}.ffn_norm.weight",
            tensors.get_tensor(f"language_model.model.layers.{i}.post_attention_layernorm.weight").astype(
                np.float32
            ) + 1,
        )

        fout.add_tensor(
            f"blk.{i}.attn_k.weight",
            tensors.get_tensor(
                f"language_model.model.layers.{i}.self_attn.k_proj.weight"
            ).astype(np.float16),
        )
        fout.add_tensor(
            f"blk.{i}.attn_output.weight",
            tensors.get_tensor(
                f"language_model.model.layers.{i}.self_attn.o_proj.weight"
            ).astype(np.float16),
        )
        fout.add_tensor(
            f"blk.{i}.attn_q.weight",
            tensors.get_tensor(
                f"language_model.model.layers.{i}.self_attn.q_proj.weight"
            ).astype(np.float16),
        )
        fout.add_tensor(
            f"blk.{i}.attn_v.weight",
            tensors.get_tensor(
                f"language_model.model.layers.{i}.self_attn.v_proj.weight"
            ).astype(np.float16),
        )

    fout.add_tensor(
        "output_norm.weight",
        tensors.get_tensor("language_model.model.norm.weight").astype(np.float32) + 1,
    )

    # save gguf
    fout.write_header_to_file()
    fout.write_kv_data_to_file()
    fout.write_tensors_to_file()
    fout.close()

    print(f"GGUF written to {fname_out}")