loader.py

# (c) City96 || Apache-2.0 (apache.org/licenses/LICENSE-2.0)
import torch
import gguf

from .ops import GGMLTensor
from .dequant import is_quantized, dequantize_tensor

IMG_ARCH_LIST = {"flux", "sd1", "sdxl", "sd3", "aura", "ltxv", "hyvid", "wan"}
TXT_ARCH_LIST = {"t5", "t5encoder", "llama"}

def get_orig_shape(reader, tensor_name):
    field_key = f"comfy.gguf.orig_shape.{tensor_name}"
    field = reader.get_field(field_key)
    if field is None:
        return None
    # Has original shape metadata, so we try to decode it.
    if len(field.types) != 2 or field.types[0] != gguf.GGUFValueType.ARRAY or field.types[1] != gguf.GGUFValueType.INT32:
        raise TypeError(f"Bad original shape metadata for {field_key}: Expected ARRAY of INT32, got {field.types}")
    return torch.Size(tuple(int(field.parts[part_idx][0]) for part_idx in field.data))

def get_field(reader, field_name, field_type):
    field = reader.get_field(field_name)
    if field is None:
        return None
    elif field_type == str:
        # extra check here as this is used for checking arch string
        if len(field.types) != 1 or field.types[0] != gguf.GGUFValueType.STRING:
            raise TypeError(f"Bad type for GGUF {field_name} key: expected string, got {field.types!r}")
        return str(field.parts[field.data[-1]], encoding="utf-8")
    elif field_type in [int, float, bool]:
        return field_type(field.parts[field.data[-1]])
    else:
        raise TypeError(f"Unknown field type {field_type}")

def get_list_field(reader, field_name, field_type):
    field = reader.get_field(field_name)
    if field is None:
        return None
    elif field_type == str:
        return tuple(str(field.parts[part_idx], encoding="utf-8") for part_idx in field.data)
    elif field_type in [int, float, bool]:
        return tuple(field_type(field.parts[part_idx][0]) for part_idx in field.data)
    else:
        raise TypeError(f"Unknown field type {field_type}")

def gguf_sd_loader(path, handle_prefix="model.diffusion_model.", return_arch=False):
    """
    Read state dict as fake tensors
    """
    reader = gguf.GGUFReader(path)

    # filter and strip prefix
    has_prefix = False
    if handle_prefix is not None:
        prefix_len = len(handle_prefix)
        tensor_names = set(tensor.name for tensor in reader.tensors)
        has_prefix = any(s.startswith(handle_prefix) for s in tensor_names)

    tensors = []
    for tensor in reader.tensors:
        sd_key = tensor_name = tensor.name
        if has_prefix:
            if not tensor_name.startswith(handle_prefix):
                continue
            sd_key = tensor_name[prefix_len:]
        tensors.append((sd_key, tensor))

    # detect and verify architecture
    compat = None
    arch_str = get_field(reader, "general.architecture", str)
    if arch_str is None: # stable-diffusion.cpp
        # import here to avoid changes to convert.py breaking regular models
        from .tools.convert import detect_arch
        arch_str = detect_arch(set(val[0] for val in tensors)).arch
        compat = "sd.cpp"
    elif arch_str in ["pig"]:
        from .tools.convert import detect_arch
        arch_str = detect_arch(set(val[0] for val in tensors)).arch
        compat = "pig"
    elif arch_str not in IMG_ARCH_LIST and arch_str not in TXT_ARCH_LIST:
        raise ValueError(f"Unexpected architecture type in GGUF file: {arch_str!r}")

    if compat:
        print(f"Warning: This model file is loaded in compatibility mode '{compat}' [arch:{arch_str}]")

    # main loading loop
    state_dict = {}
    qtype_dict = {}
    for sd_key, tensor in tensors:
        tensor_name = tensor.name
        torch_tensor = torch.from_numpy(tensor.data) # mmap

        shape = get_orig_shape(reader, tensor_name)
        if shape is None:
            shape = torch.Size(tuple(int(v) for v in reversed(tensor.shape)))
            # Workaround for stable-diffusion.cpp SDXL detection.
            if compat == "sd.cpp" and arch_str == "sdxl":
                if any([tensor_name.endswith(x) for x in (".proj_in.weight", ".proj_out.weight")]):
                    while len(shape) > 2 and shape[-1] == 1:
                        shape = shape[:-1]

        # add to state dict
        if tensor.tensor_type in {gguf.GGMLQuantizationType.F32, gguf.GGMLQuantizationType.F16}:
            torch_tensor = torch_tensor.view(*shape)
        state_dict[sd_key] = GGMLTensor(torch_tensor, tensor_type=tensor.tensor_type, tensor_shape=shape)

        # keep track of loaded tensor types
        tensor_type_str = getattr(tensor.tensor_type, "name", repr(tensor.tensor_type))
        qtype_dict[tensor_type_str] = qtype_dict.get(tensor_type_str, 0) + 1

    # print loaded tensor type counts
    print("gguf qtypes: " + ", ".join(f"{k} ({v})" for k, v in qtype_dict.items()))

    # mark largest tensor for vram estimation
    qsd = {k:v for k,v in state_dict.items() if is_quantized(v)}
    if len(qsd) > 0:
        max_key = max(qsd.keys(), key=lambda k: qsd[k].numel())
        state_dict[max_key].is_largest_weight = True

    if return_arch:
        return (state_dict, arch_str)
    return state_dict

# for remapping llama.cpp -> original key names
T5_SD_MAP = {
    "enc.": "encoder.",
    ".blk.": ".block.",
    "token_embd": "shared",
    "output_norm": "final_layer_norm",
    "attn_q": "layer.0.SelfAttention.q",
    "attn_k": "layer.0.SelfAttention.k",
    "attn_v": "layer.0.SelfAttention.v",
    "attn_o": "layer.0.SelfAttention.o",
    "attn_norm": "layer.0.layer_norm",
    "attn_rel_b": "layer.0.SelfAttention.relative_attention_bias",
    "ffn_up": "layer.1.DenseReluDense.wi_1",
    "ffn_down": "layer.1.DenseReluDense.wo",
    "ffn_gate": "layer.1.DenseReluDense.wi_0",
    "ffn_norm": "layer.1.layer_norm",
}

LLAMA_SD_MAP = {
    "blk.": "model.layers.",
    "attn_norm": "input_layernorm",
    "attn_q": "self_attn.q_proj",
    "attn_k": "self_attn.k_proj",
    "attn_v": "self_attn.v_proj",
    "attn_output": "self_attn.o_proj",
    "ffn_up": "mlp.up_proj",
    "ffn_down": "mlp.down_proj",
    "ffn_gate": "mlp.gate_proj",
    "ffn_norm": "post_attention_layernorm",
    "token_embd": "model.embed_tokens",
    "output_norm": "model.norm",
    "output.weight": "lm_head.weight",
}

def sd_map_replace(raw_sd, key_map):
    sd = {}
    for k,v in raw_sd.items():
        for s,d in key_map.items():
            k = k.replace(s,d)
        sd[k] = v
    return sd

def llama_permute(raw_sd, n_head, n_head_kv):
    # Reverse version of LlamaModel.permute in llama.cpp convert script
    sd = {}
    permute = lambda x,h: x.reshape(h, x.shape[0] // h // 2, 2, *x.shape[1:]).swapaxes(1, 2).reshape(x.shape)
    for k,v in raw_sd.items():
        if k.endswith(("q_proj.weight", "q_proj.bias")):
            v.data = permute(v.data, n_head)
        if k.endswith(("k_proj.weight", "k_proj.bias")):
            v.data = permute(v.data, n_head_kv)
        sd[k] = v
    return sd

def gguf_tokenizer_loader(path, temb_shape):
    # convert gguf tokenizer to spiece
    print(f"Attempting to recreate sentencepiece tokenizer from GGUF file metadata...")
    try:
        from sentencepiece import sentencepiece_model_pb2 as model
    except ImportError:
        raise ImportError("Please make sure sentencepiece and protobuf are installed.\npip install sentencepiece protobuf")
    spm = model.ModelProto()

    reader = gguf.GGUFReader(path)

    if get_field(reader, "tokenizer.ggml.model", str) == "t5":
        if temb_shape == (256384, 4096): # probably UMT5
            spm.trainer_spec.model_type == 1 # Unigram (do we have a T5 w/ BPE?)
        else:
            raise NotImplementedError(f"Unknown model, can't set tokenizer!")
    else:
        raise NotImplementedError(f"Unknown model, can't set tokenizer!")

    spm.normalizer_spec.add_dummy_prefix = get_field(reader, "tokenizer.ggml.add_space_prefix", bool)
    spm.normalizer_spec.remove_extra_whitespaces = get_field(reader, "tokenizer.ggml.remove_extra_whitespaces", bool)

    tokens = get_list_field(reader, "tokenizer.ggml.tokens", str)
    scores = get_list_field(reader, "tokenizer.ggml.scores", float)
    toktypes = get_list_field(reader, "tokenizer.ggml.token_type", int)

    for idx, (token, score, toktype) in enumerate(zip(tokens, scores, toktypes)):
        # # These aren't present in the original?
        # if toktype == 5 and idx >= temb_shape[0]%1000):
        #     continue

        piece = spm.SentencePiece()
        piece.piece = token
        piece.score = score
        piece.type = toktype
        spm.pieces.append(piece)

    # unsure if any of these are correct
    spm.trainer_spec.byte_fallback = True
    spm.trainer_spec.vocab_size = len(tokens) # split off unused?
    spm.trainer_spec.max_sentence_length = 4096
    spm.trainer_spec.eos_id = get_field(reader, "tokenizer.ggml.eos_token_id", int)
    spm.trainer_spec.pad_id = get_field(reader, "tokenizer.ggml.padding_token_id", int)

    print(f"Created tokenizer with vocab size of {len(spm.pieces)}")
    del reader
    return torch.ByteTensor(list(spm.SerializeToString()))

def gguf_clip_loader(path):
    sd, arch = gguf_sd_loader(path, return_arch=True)
    if arch in {"t5", "t5encoder"}:
        temb_key = "token_embd.weight"
        if temb_key in sd and sd[temb_key].shape == (256384, 4096):
            # non-standard Comfy-Org tokenizer
            sd["spiece_model"] = gguf_tokenizer_loader(path, sd[temb_key].shape)
            # TODO: dequantizing token embed here is janky but otherwise we OOM due to tensor being massive.
            print(f"Dequantizing {temb_key} to prevent runtime OOM.")
            sd[temb_key] = dequantize_tensor(sd[temb_key], dtype=torch.float16)
        sd = sd_map_replace(sd, T5_SD_MAP)
    elif arch in {"llama"}:
        temb_key = "token_embd.weight"
        if temb_key in sd and sd[temb_key].shape != (128320, 4096):
            # This still works. Raise error?
            print("Warning! token_embd shape may be incorrect for llama 3 model!")
        sd = sd_map_replace(sd, LLAMA_SD_MAP)
        sd = llama_permute(sd, 32, 8) # L3
    else:
        pass
    return sd