llama.cpp/convert-hf-to-gguf.py

2297 lines
98 KiB
Python
Raw Normal View History

#!/usr/bin/env python3
from __future__ import annotations
import argparse
import contextlib
import json
import os
import re
import sys
from abc import ABC, abstractmethod
from enum import IntEnum
from pathlib import Path
from typing import TYPE_CHECKING, Any, Callable, ContextManager, Iterator, Sequence, TypeVar, cast
import numpy as np
import torch
if TYPE_CHECKING:
from torch import Tensor
if 'NO_LOCAL_GGUF' not in os.environ:
sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
import gguf
from convert import LlamaHfVocab, permute
###### MODEL DEFINITIONS ######
class SentencePieceTokenTypes(IntEnum):
NORMAL = 1
UNKNOWN = 2
CONTROL = 3
USER_DEFINED = 4
UNUSED = 5
BYTE = 6
2024-03-04 19:50:50 +00:00
AnyModel = TypeVar("AnyModel", bound="type[Model]")
2024-03-04 19:50:50 +00:00
class Model(ABC):
_model_classes: dict[str, type[Model]] = {}
def __init__(self, dir_model: Path, ftype: int, fname_out: Path, is_big_endian: bool):
self.dir_model = dir_model
self.ftype = ftype
self.fname_out = fname_out
self.is_big_endian = is_big_endian
self.endianess = gguf.GGUFEndian.BIG if is_big_endian else gguf.GGUFEndian.LITTLE
self.is_safetensors = self._is_model_safetensors()
self.num_parts = Model.count_model_parts(self.dir_model, ".safetensors" if self.is_safetensors else ".bin")
self.part_names = self._get_part_names()
self.hparams = Model.load_hparams(self.dir_model)
self.gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=False)
self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer"])
@property
@abstractmethod
def model_arch(self) -> gguf.MODEL_ARCH:
pass
def find_hparam(self, keys: Sequence[str], optional: bool = False) -> Any:
key = next((k for k in keys if k in self.hparams), None)
if key is not None:
return self.hparams[key]
if optional:
return None
raise KeyError(f"could not find any of: {keys}")
def set_vocab(self):
self._set_vocab_gpt2()
def get_tensors(self) -> Iterator[tuple[str, Tensor]]:
for part_name in self.part_names:
print(f"gguf: loading model part '{part_name}'")
ctx: ContextManager[Any]
if self.is_safetensors:
from safetensors import safe_open
ctx = cast(ContextManager[Any], safe_open(self.dir_model / part_name, framework="pt", device="cpu"))
else:
ctx = contextlib.nullcontext(torch.load(str(self.dir_model / part_name), map_location="cpu", mmap=True, weights_only=True))
with ctx as model_part:
for name in model_part.keys():
data = model_part.get_tensor(name) if self.is_safetensors else model_part[name]
yield name, data
def set_gguf_parameters(self):
self.gguf_writer.add_name(self.dir_model.name)
self.gguf_writer.add_block_count(self.block_count)
if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx"], optional=True)) is not None:
self.gguf_writer.add_context_length(n_ctx)
print(f"gguf: context length = {n_ctx}")
n_embd = self.find_hparam(["hidden_size", "n_embd"])
self.gguf_writer.add_embedding_length(n_embd)
print(f"gguf: embedding length = {n_embd}")
if (n_ff := self.find_hparam(["intermediate_size", "n_inner"], optional=True)) is not None:
self.gguf_writer.add_feed_forward_length(n_ff)
print(f"gguf: feed forward length = {n_ff}")
n_head = self.find_hparam(["num_attention_heads", "n_head"])
self.gguf_writer.add_head_count(n_head)
print(f"gguf: head count = {n_head}")
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 12:04:25 +00:00
if (n_head_kv := self.hparams.get("num_key_value_heads")) is not None:
self.gguf_writer.add_head_count_kv(n_head_kv)
print(f"gguf: key-value head count = {n_head_kv}")
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 12:04:25 +00:00
if (rope_theta := self.hparams.get("rope_theta")) is not None:
self.gguf_writer.add_rope_freq_base(rope_theta)
print(f"gguf: rope theta = {rope_theta}")
if (f_rms_eps := self.hparams.get("rms_norm_eps")) is not None:
self.gguf_writer.add_layer_norm_rms_eps(f_rms_eps)
print(f"gguf: rms norm epsilon = {f_rms_eps}")
if (f_norm_eps := self.find_hparam(["layer_norm_eps", "layer_norm_epsilon", "norm_epsilon"], optional=True)) is not None:
self.gguf_writer.add_layer_norm_eps(f_norm_eps)
print(f"gguf: layer norm epsilon = {f_norm_eps}")
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 12:04:25 +00:00
if (n_experts := self.hparams.get("num_local_experts")) is not None:
self.gguf_writer.add_expert_count(n_experts)
print(f"gguf: expert count = {n_experts}")
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 12:04:25 +00:00
if (n_experts_used := self.hparams.get("num_experts_per_tok")) is not None:
self.gguf_writer.add_expert_used_count(n_experts_used)
print(f"gguf: experts used count = {n_experts_used}")
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 12:04:25 +00:00
self.gguf_writer.add_file_type(self.ftype)
print(f"gguf: file type = {self.ftype}")
def write_tensors(self):
block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
for name, data_torch in self.get_tensors():
# we don't need these
if name.endswith((".attention.masked_bias", ".attention.bias", ".attention.rotary_emb.inv_freq")):
continue
old_dtype = data_torch.dtype
# convert any unsupported data types to float32
if data_torch.dtype not in (torch.float16, torch.float32):
data_torch = data_torch.to(torch.float32)
data = data_torch.squeeze().numpy()
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
n_dims = len(data.shape)
data_dtype = data.dtype
# if f32 desired, convert any float16 to float32
if self.ftype == 0 and data_dtype == np.float16:
data = data.astype(np.float32)
# TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
data = data.astype(np.float32)
# if f16 desired, convert any float32 2-dim weight tensors to float16
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
def write(self):
self.write_tensors()
self.gguf_writer.write_header_to_file()
self.gguf_writer.write_kv_data_to_file()
self.gguf_writer.write_tensors_to_file()
self.gguf_writer.close()
def write_vocab(self):
self.gguf_writer.write_header_to_file()
self.gguf_writer.write_kv_data_to_file()
self.gguf_writer.close()
@staticmethod
def count_model_parts(dir_model: Path, prefix: str) -> int:
num_parts = 0
for filename in os.listdir(dir_model):
if filename.endswith(prefix):
num_parts += 1
return num_parts
@staticmethod
def load_hparams(dir_model):
with open(dir_model / "config.json", "r", encoding="utf-8") as f:
return json.load(f)
@classmethod
def register(cls, *names: str) -> Callable[[AnyModel], AnyModel]:
assert names
2024-03-04 19:50:50 +00:00
def func(modelcls: type[Model]):
for name in names:
cls._model_classes[name] = modelcls
return modelcls
return func
@classmethod
def from_model_architecture(cls, arch):
try:
return cls._model_classes[arch]
except KeyError:
raise NotImplementedError(f'Architecture {arch!r} not supported!') from None
def _is_model_safetensors(self) -> bool:
return Model.count_model_parts(self.dir_model, ".safetensors") > 0
def _get_part_names(self):
if self.is_safetensors:
if self.num_parts == 1: # there's only one .safetensors file
return ("model.safetensors",)
return (f"model-{n:05}-of-{self.num_parts:05}.safetensors" for n in range(1, self.num_parts + 1))
if self.num_parts == 1: # there's only one .bin file
return ("pytorch_model.bin",)
return (f"pytorch_model-{n:05}-of-{self.num_parts:05}.bin" for n in range(1, self.num_parts + 1))
def _set_vocab_gpt2(self):
dir_model = self.dir_model
hparams = self.hparams
tokens: list[str] = []
toktypes: list[int] = []
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(dir_model)
vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
assert max(tokenizer.vocab.values()) < vocab_size
reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()}
added_vocab = tokenizer.get_added_vocab()
for i in range(vocab_size):
if i not in reverse_vocab:
tokens.append(f"[PAD{i}]")
toktypes.append(gguf.TokenType.USER_DEFINED)
elif reverse_vocab[i] in added_vocab:
tokens.append(reverse_vocab[i])
if tokenizer.added_tokens_decoder[i].special:
toktypes.append(gguf.TokenType.CONTROL)
else:
toktypes.append(gguf.TokenType.USER_DEFINED)
else:
tokens.append(reverse_vocab[i])
toktypes.append(gguf.TokenType.NORMAL)
self.gguf_writer.add_tokenizer_model("gpt2")
self.gguf_writer.add_token_list(tokens)
self.gguf_writer.add_token_types(toktypes)
special_vocab = gguf.SpecialVocab(dir_model, load_merges=True)
special_vocab.add_to_gguf(self.gguf_writer)
def _set_vocab_qwen(self):
dir_model = self.dir_model
hparams = self.hparams
tokens: list[str] = []
toktypes: list[int] = []
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(dir_model, trust_remote_code=True)
vocab_size = hparams["vocab_size"]
assert max(tokenizer.get_vocab().values()) < vocab_size
merges = []
vocab = {}
mergeable_ranks = tokenizer.mergeable_ranks
for token, rank in mergeable_ranks.items():
vocab[QwenModel.token_bytes_to_string(token)] = rank
if len(token) == 1:
continue
merged = QwenModel.bpe(mergeable_ranks, token, max_rank=rank)
assert len(merged) == 2
merges.append(' '.join(map(QwenModel.token_bytes_to_string, merged)))
# for this kind of tokenizer, added_vocab is not a subset of vocab, so they need to be combined
added_vocab = tokenizer.special_tokens
reverse_vocab = {id_ : encoded_tok for encoded_tok, id_ in (vocab | added_vocab).items()}
for i in range(vocab_size):
if i not in reverse_vocab:
tokens.append(f"[PAD{i}]")
toktypes.append(gguf.TokenType.USER_DEFINED)
elif reverse_vocab[i] in added_vocab:
tokens.append(reverse_vocab[i])
toktypes.append(gguf.TokenType.CONTROL)
else:
tokens.append(reverse_vocab[i])
toktypes.append(gguf.TokenType.NORMAL)
self.gguf_writer.add_tokenizer_model("gpt2")
self.gguf_writer.add_token_list(tokens)
self.gguf_writer.add_token_types(toktypes)
special_vocab = gguf.SpecialVocab(dir_model, load_merges=False)
special_vocab.merges = merges
# only add special tokens when they were not already loaded from config.json
if len(special_vocab.special_token_ids) == 0:
special_vocab._set_special_token("bos", tokenizer.special_tokens["<|endoftext|>"])
special_vocab._set_special_token("eos", tokenizer.special_tokens["<|endoftext|>"])
# this one is usually not in config.json anyway
special_vocab._set_special_token("unk", tokenizer.special_tokens["<|endoftext|>"])
special_vocab.add_to_gguf(self.gguf_writer)
def _set_vocab_sentencepiece(self):
from sentencepiece import SentencePieceProcessor
tokenizer_path = self.dir_model / 'tokenizer.model'
tokens: list[bytes] = []
scores: list[float] = []
toktypes: list[int] = []
if not tokenizer_path.is_file():
print(f'Error: Missing {tokenizer_path}', file=sys.stderr)
sys.exit(1)
tokenizer = SentencePieceProcessor(str(tokenizer_path))
vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
for token_id in range(tokenizer.vocab_size()):
piece = tokenizer.id_to_piece(token_id)
text = piece.encode("utf-8")
score = tokenizer.get_score(token_id)
toktype = SentencePieceTokenTypes.NORMAL
if tokenizer.is_unknown(token_id):
toktype = SentencePieceTokenTypes.UNKNOWN
elif tokenizer.is_control(token_id):
toktype = SentencePieceTokenTypes.CONTROL
elif tokenizer.is_unused(token_id):
toktype = SentencePieceTokenTypes.UNUSED
elif tokenizer.is_byte(token_id):
toktype = SentencePieceTokenTypes.BYTE
tokens.append(text)
scores.append(score)
toktypes.append(toktype)
added_tokens_file = self.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:
key = key.encode("utf-8")
if key not in tokens:
tokens.append(key)
scores.append(-1000.0)
toktypes.append(SentencePieceTokenTypes.USER_DEFINED)
assert len(tokens) == vocab_size
self.gguf_writer.add_tokenizer_model("llama")
self.gguf_writer.add_token_list(tokens)
self.gguf_writer.add_token_scores(scores)
self.gguf_writer.add_token_types(toktypes)
special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
special_vocab.add_to_gguf(self.gguf_writer)
def _set_vocab_llama_hf(self):
vocab = LlamaHfVocab(self.dir_model)
tokens = []
scores = []
toktypes = []
for text, score, toktype in vocab.all_tokens():
tokens.append(text)
scores.append(score)
toktypes.append(toktype)
assert len(tokens) == vocab.vocab_size
self.gguf_writer.add_tokenizer_model("llama")
self.gguf_writer.add_token_list(tokens)
self.gguf_writer.add_token_scores(scores)
self.gguf_writer.add_token_types(toktypes)
special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
special_vocab.add_to_gguf(self.gguf_writer)
@Model.register("GPTNeoXForCausalLM")
class GPTNeoXModel(Model):
model_arch = gguf.MODEL_ARCH.GPTNEOX
def set_gguf_parameters(self):
block_count = self.hparams["num_hidden_layers"]
self.gguf_writer.add_name(self.dir_model.name)
self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"])
self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
self.gguf_writer.add_block_count(block_count)
self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
self.gguf_writer.add_rope_dimension_count(
int(self.hparams["rotary_pct"] * (self.hparams["hidden_size"] // self.hparams["num_attention_heads"])),
)
self.gguf_writer.add_head_count(self.hparams["num_attention_heads"])
self.gguf_writer.add_parallel_residual(self.hparams.get("use_parallel_residual", True))
self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_eps"])
@Model.register("BloomForCausalLM")
class BloomModel(Model):
model_arch = gguf.MODEL_ARCH.BLOOM
def set_gguf_parameters(self):
self.gguf_writer.add_name("Bloom")
n_embed = self.hparams.get("hidden_size", self.hparams.get("n_embed"))
n_head = self.hparams.get("n_head", self.hparams.get("num_attention_heads"))
self.gguf_writer.add_context_length(self.hparams.get("seq_length", n_embed))
self.gguf_writer.add_embedding_length(n_embed)
self.gguf_writer.add_feed_forward_length(4 * n_embed)
self.gguf_writer.add_block_count(self.hparams["n_layer"])
self.gguf_writer.add_head_count(n_head)
self.gguf_writer.add_head_count_kv(n_head)
self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
self.gguf_writer.add_file_type(self.ftype)
def write_tensors(self):
block_count = self.hparams["n_layer"]
tensors = dict(self.get_tensors())
tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
has_lm_head = True
n_head = self.hparams.get("n_head", self.hparams.get("num_attention_heads"))
n_embed = self.hparams.get("hidden_size", self.hparams.get("n_embed"))
for name, data_torch in tensors.items():
if "lm_head.weight" not in tensors.keys() and "output.weight" not in tensors.keys():
has_lm_head = False
name = re.sub(r'transformer\.', '', name)
old_dtype = data_torch.dtype
# convert any unsupported data types to float32
if data_torch.dtype not in (torch.float16, torch.float32):
data_torch = data_torch.to(torch.float32)
data = data_torch.squeeze().numpy()
if re.match(r"h\.\d+\.self_attention\.query_key_value\.weight", name):
# Map bloom-style qkv_linear to gpt-style qkv_linear
# bloom: https://github.com/huggingface/transformers/blob/main/src/transformers/models/bloom/modeling_bloom.py#L238-L252 # noqa
# gpt-2: https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt2/modeling_gpt2.py#L312 # noqa
qkv_weights = data.reshape((n_head, 3, n_embed // n_head, n_embed))
data = np.concatenate(
(
qkv_weights[:, 0, :, :].reshape((-1, n_embed)),
qkv_weights[:, 1, :, :].reshape((-1, n_embed)),
qkv_weights[:, 2, :, :].reshape((-1, n_embed)),
),
axis=0,
)
print("re-format attention.linear_qkv.weight")
elif re.match(r"h\.\d+\.self_attention\.query_key_value\.bias", name):
qkv_bias = data.reshape((n_head, 3, n_embed // n_head))
data = np.concatenate(
(
qkv_bias[:, 0, :].reshape((n_embed,)),
qkv_bias[:, 1, :].reshape((n_embed,)),
qkv_bias[:, 2, :].reshape((n_embed,)),
),
axis=0,
)
print("re-format attention.linear_qkv.bias")
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
n_dims = len(data.shape)
data_dtype = data.dtype
# if f32 desired, convert any float16 to float32
if self.ftype == 0 and data_dtype == np.float16:
data = data.astype(np.float32)
# TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
data = data.astype(np.float32)
# if f16 desired, convert any float32 2-dim weight tensors to float16
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"=> {new_name}, shape = {data.shape}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
if not has_lm_head and name == "word_embeddings.weight":
self.gguf_writer.add_tensor("output.weight", data)
print(name, f"=> output.weight, shape = {data.shape}, {old_dtype} --> {data.dtype}")
@Model.register("MPTForCausalLM")
class MPTModel(Model):
model_arch = gguf.MODEL_ARCH.MPT
def set_gguf_parameters(self):
block_count = self.hparams["n_layers"]
self.gguf_writer.add_name(self.dir_model.name)
self.gguf_writer.add_context_length(self.hparams["max_seq_len"])
self.gguf_writer.add_embedding_length(self.hparams["d_model"])
self.gguf_writer.add_block_count(block_count)
self.gguf_writer.add_feed_forward_length(4 * self.hparams["d_model"])
self.gguf_writer.add_head_count(self.hparams["n_heads"])
if kv_n_heads := self.hparams["attn_config"].get("kv_n_heads"):
self.gguf_writer.add_head_count_kv(kv_n_heads)
self.gguf_writer.add_layer_norm_eps(1e-5)
if self.hparams["attn_config"]["clip_qkv"] is not None:
self.gguf_writer.add_clamp_kqv(self.hparams["attn_config"]["clip_qkv"])
self.gguf_writer.add_max_alibi_bias(self.hparams["attn_config"]["alibi_bias_max"])
def write_tensors(self):
block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers"))
tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
for name, data_torch in self.get_tensors():
# we don't need these
if name.endswith((".attention.masked_bias", ".attention.bias", ".attention.rotary_emb.inv_freq")):
continue
old_dtype = data_torch.dtype
# convert any unsupported data types to float32
if data_torch.dtype not in (torch.float16, torch.float32):
data_torch = data_torch.to(torch.float32)
data = data_torch.squeeze().numpy()
# map tensor names
if "scales" in name:
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias", ".scales"))
if new_name is not None:
new_name = new_name.replace("scales", "act.scales")
else:
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
n_dims = len(data.shape)
data_dtype = data.dtype
# if f32 desired, convert any float16 to float32
if self.ftype == 0 and data_dtype == np.float16:
data = data.astype(np.float32)
# TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
data = data.astype(np.float32)
# if f16 desired, convert any float32 2-dim weight tensors to float16
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
@Model.register("OrionForCausalLM")
class OrionModel(Model):
model_arch = gguf.MODEL_ARCH.ORION
def set_vocab(self):
self._set_vocab_sentencepiece()
def set_gguf_parameters(self):
block_count = self.hparams["num_hidden_layers"]
head_count = self.hparams["num_attention_heads"]
head_count_kv = self.hparams.get("num_key_value_heads", head_count)
hf_repo = self.hparams.get("_name_or_path", "")
ctx_length = 0
if "max_sequence_length" in self.hparams:
ctx_length = self.hparams["max_sequence_length"]
elif "max_position_embeddings" in self.hparams:
ctx_length = self.hparams["max_position_embeddings"]
elif "model_max_length" in self.hparams:
ctx_length = self.hparams["model_max_length"]
else:
print("gguf: can not find ctx length parameter.")
sys.exit()
self.gguf_writer.add_file_type(self.ftype)
self.gguf_writer.add_name(self.dir_model.name)
self.gguf_writer.add_source_hf_repo(hf_repo)
self.gguf_writer.add_tensor_data_layout("Meta AI original pth")
self.gguf_writer.add_context_length(ctx_length)
self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
self.gguf_writer.add_block_count(block_count)
self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
self.gguf_writer.add_head_count(head_count)
self.gguf_writer.add_head_count_kv(head_count_kv)
2024-02-22 18:13:25 +00:00
# note: config provides rms norm but it is actually layer norm
# ref: https://huggingface.co/OrionStarAI/Orion-14B-Chat/blob/276a17221ce42beb45f66fac657a41540e71f4f5/modeling_orion.py#L570-L571
self.gguf_writer.add_layer_norm_eps(self.hparams["rms_norm_eps"])
def write_tensors(self):
# Collect tensors from generator object
model_kv = dict(self.get_tensors())
block_count = self.hparams["num_hidden_layers"]
tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
for name, data_torch in model_kv.items():
# we don't need these
if name.endswith(".rotary_emb.inv_freq"):
continue
old_dtype = data_torch.dtype
# convert any unsupported data types to float32
if data_torch.dtype not in (torch.float16, torch.float32):
data_torch = data_torch.to(torch.float32)
data = data_torch.squeeze().numpy()
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
n_dims = len(data.shape)
data_dtype = data.dtype
# if f32 desired, convert any float16 to float32
if self.ftype == 0 and data_dtype == np.float16:
data = data.astype(np.float32)
# TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
data = data.astype(np.float32)
# if f16 desired, convert any float32 2-dim weight tensors to float16
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{name} -> {new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
@Model.register("BaichuanForCausalLM", "BaiChuanForCausalLM")
class BaichuanModel(Model):
model_arch = gguf.MODEL_ARCH.BAICHUAN
def set_vocab(self):
self._set_vocab_sentencepiece()
def set_gguf_parameters(self):
block_count = self.hparams["num_hidden_layers"]
head_count = self.hparams["num_attention_heads"]
head_count_kv = self.hparams.get("num_key_value_heads", head_count)
hf_repo = self.hparams.get("_name_or_path", "")
ctx_length = 0
if "max_sequence_length" in self.hparams:
ctx_length = self.hparams["max_sequence_length"]
elif "max_position_embeddings" in self.hparams:
ctx_length = self.hparams["max_position_embeddings"]
elif "model_max_length" in self.hparams:
ctx_length = self.hparams["model_max_length"]
else:
print("gguf: can not find ctx length parameter.")
sys.exit()
self.gguf_writer.add_name(self.dir_model.name)
self.gguf_writer.add_source_hf_repo(hf_repo)
self.gguf_writer.add_tensor_data_layout("Meta AI original pth")
self.gguf_writer.add_context_length(ctx_length)
self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
self.gguf_writer.add_block_count(block_count)
self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
self.gguf_writer.add_rope_dimension_count(self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
self.gguf_writer.add_head_count(head_count)
self.gguf_writer.add_head_count_kv(head_count_kv)
self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
if self.hparams["rope_scaling"].get("type") == "linear":
self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
def write_tensors(self):
# Collect tensors from generator object
model_kv = dict(self.get_tensors())
block_count = self.hparams["num_hidden_layers"]
head_count = self.hparams["num_attention_heads"]
tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
head_count_kv = self.hparams.get("num_key_value_heads", head_count)
for i in range(block_count):
if (w := model_kv.get(f"model.layers.{i}.self_attn.W_pack.weight")) is not None:
print(f"Unpacking and permuting layer {i}")
model_kv[f"model.layers.{i}.self_attn.q_proj.weight"] = \
self._reverse_hf_permute_part(w, 0, head_count, head_count)
model_kv[f"model.layers.{i}.self_attn.k_proj.weight"] = \
self._reverse_hf_permute_part(w, 1, head_count, head_count_kv)
model_kv[f"model.layers.{i}.self_attn.v_proj.weight"] = \
self._reverse_hf_part(w, 2)
del model_kv[f"model.layers.{i}.self_attn.W_pack.weight"]
for name, data_torch in model_kv.items():
# we don't need these
if name.endswith(".rotary_emb.inv_freq"):
continue
old_dtype = data_torch.dtype
# convert any unsupported data types to float32
if data_torch.dtype not in (torch.float16, torch.float32):
data_torch = data_torch.to(torch.float32)
data = data_torch.squeeze().numpy()
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
n_dims = len(data.shape)
data_dtype = data.dtype
# if f32 desired, convert any float16 to float32
if self.ftype == 0 and data_dtype == np.float16:
data = data.astype(np.float32)
# TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
data = data.astype(np.float32)
# if f16 desired, convert any float32 2-dim weight tensors to float16
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{name} -> {new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
def _reverse_hf_permute(self, weights: Tensor, n_head: int, n_kv_head: int | None = None) -> Tensor:
if n_kv_head is not None and n_head != n_kv_head:
n_head //= n_kv_head
return (
weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
.swapaxes(1, 2)
.reshape(weights.shape)
)
def _reverse_hf_permute_part(
self, weights: Tensor, n_part: int, n_head: int, n_head_kv: int | None = None,
) -> Tensor:
r = weights.shape[0] // 3
return self._reverse_hf_permute(weights[r * n_part:r * n_part + r, ...], n_head, n_head_kv)
def _reverse_hf_part(self, weights: Tensor, n_part: int) -> Tensor:
r = weights.shape[0] // 3
return weights[r * n_part:r * n_part + r, ...]
@Model.register("XverseForCausalLM")
class XverseModel(Model):
model_arch = gguf.MODEL_ARCH.XVERSE
def set_vocab(self):
assert (self.dir_model / "tokenizer.json").is_file()
dir_model = self.dir_model
hparams = self.hparams
tokens: list[bytearray] = []
toktypes: list[int] = []
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(dir_model)
vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
assert max(tokenizer.vocab.values()) < vocab_size
reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()}
added_vocab = tokenizer.get_added_vocab()
for token_id in range(vocab_size):
token_text = reverse_vocab[token_id].encode('utf-8')
# replace "\x00" to string with length > 0
if token_text == b"\x00":
toktype = gguf.TokenType.BYTE # special
token_text = f"<{token_text}>".encode('utf-8')
elif re.fullmatch(br"<0x[0-9A-Fa-f]{2}>", token_text):
toktype = gguf.TokenType.BYTE # special
elif reverse_vocab[token_id] in added_vocab:
if tokenizer.added_tokens_decoder[token_id].special:
toktype = gguf.TokenType.CONTROL
else:
toktype = gguf.TokenType.USER_DEFINED
else:
toktype = gguf.TokenType.NORMAL
tokens.append(token_text)
toktypes.append(toktype)
self.gguf_writer.add_tokenizer_model("llama")
self.gguf_writer.add_token_list(tokens)
self.gguf_writer.add_token_types(toktypes)
special_vocab = gguf.SpecialVocab(dir_model, n_vocab=len(tokens))
special_vocab.add_to_gguf(self.gguf_writer)
def set_gguf_parameters(self):
block_count = self.hparams["num_hidden_layers"]
head_count = self.hparams["num_attention_heads"]
head_count_kv = self.hparams.get("num_key_value_heads", head_count)
hf_repo = self.hparams.get("_name_or_path", "")
ctx_length = 0
if "max_sequence_length" in self.hparams:
ctx_length = self.hparams["max_sequence_length"]
elif "max_position_embeddings" in self.hparams:
ctx_length = self.hparams["max_position_embeddings"]
elif "model_max_length" in self.hparams:
ctx_length = self.hparams["model_max_length"]
else:
print("gguf: can not find ctx length parameter.")
sys.exit()
self.gguf_writer.add_name(self.dir_model.name)
self.gguf_writer.add_source_hf_repo(hf_repo)
self.gguf_writer.add_tensor_data_layout("Meta AI original pth")
self.gguf_writer.add_context_length(ctx_length)
self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
self.gguf_writer.add_block_count(block_count)
self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
self.gguf_writer.add_rope_dimension_count(self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
self.gguf_writer.add_head_count(head_count)
self.gguf_writer.add_head_count_kv(head_count_kv)
self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
if self.hparams["rope_scaling"].get("type") == "linear":
self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
def write_tensors(self):
# Collect tensors from generator object
model_kv = dict(self.get_tensors())
block_count = self.hparams["num_hidden_layers"]
head_count = self.hparams["num_attention_heads"]
tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
head_count_kv = self.hparams.get("num_key_value_heads", head_count)
for name, data_torch in model_kv.items():
# we don't need these
if name.endswith(".rotary_emb.inv_freq"):
continue
old_dtype = data_torch.dtype
# convert any unsupported data types to float32
if data_torch.dtype not in (torch.float16, torch.float32):
data_torch = data_torch.to(torch.float32)
# HF models permute some of the tensors, so we need to undo that
if name.endswith(("q_proj.weight")):
data_torch = self._reverse_hf_permute(data_torch, head_count, head_count)
if name.endswith(("k_proj.weight")):
data_torch = self._reverse_hf_permute(data_torch, head_count, head_count_kv)
data = data_torch.squeeze().numpy()
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
n_dims = len(data.shape)
data_dtype = data.dtype
# if f32 desired, convert any float16 to float32
if self.ftype == 0 and data_dtype == np.float16:
data = data.astype(np.float32)
# TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
data = data.astype(np.float32)
# if f16 desired, convert any float32 2-dim weight tensors to float16
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{name} -> {new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
def _reverse_hf_permute(self, weights: Tensor, n_head: int, n_kv_head: int | None = None) -> Tensor:
if n_kv_head is not None and n_head != n_kv_head:
n_head //= n_kv_head
return (
weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
.swapaxes(1, 2)
.reshape(weights.shape)
)
@Model.register("FalconForCausalLM", "RWForCausalLM")
class FalconModel(Model):
model_arch = gguf.MODEL_ARCH.FALCON
def set_gguf_parameters(self):
block_count = self.hparams.get("num_hidden_layers")
if block_count is None:
block_count = self.hparams["n_layer"] # old name
n_head = self.hparams.get("num_attention_heads")
if n_head is None:
n_head = self.hparams["n_head"] # old name
n_head_kv = self.hparams.get("num_kv_heads")
if n_head_kv is None:
n_head_kv = self.hparams.get("n_head_kv", 1) # old name
self.gguf_writer.add_name("Falcon")
self.gguf_writer.add_context_length(2048) # not in config.json
self.gguf_writer.add_tensor_data_layout("jploski") # qkv tensor transform
self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
self.gguf_writer.add_feed_forward_length(4 * self.hparams["hidden_size"])
self.gguf_writer.add_block_count(block_count)
self.gguf_writer.add_head_count(n_head)
self.gguf_writer.add_head_count_kv(n_head_kv)
self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
self.gguf_writer.add_file_type(self.ftype)
def write_tensors(self):
block_count = self.hparams.get("num_hidden_layers")
if block_count is None:
block_count = self.hparams["n_layer"] # old name
n_head = self.hparams.get("num_attention_heads")
if n_head is None:
n_head = self.hparams["n_head"] # old name
n_head_kv = self.hparams.get("num_kv_heads")
if n_head_kv is None:
n_head_kv = self.hparams.get("n_head_kv", 1) # old name
head_dim = self.hparams["hidden_size"] // n_head
tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
for name, data_torch in self.get_tensors():
old_dtype = data_torch.dtype
# convert any unsupported data types to float32
if data_torch.dtype not in (torch.float16, torch.float32):
data_torch = data_torch.to(torch.float32)
# QKV tensor transform
# The original query_key_value tensor contains n_head_kv "kv groups",
# each consisting of n_head/n_head_kv query weights followed by one key
# and one value weight (shared by all query heads in the kv group).
# This layout makes it a big pain to work with in GGML.
# So we rearrange them here,, so that we have n_head query weights
# followed by n_head_kv key weights followed by n_head_kv value weights,
# in contiguous fashion.
# ref: https://github.com/jploski/ggml/blob/falcon40b/examples/falcon/convert-hf-to-ggml.py
if "query_key_value" in name:
qkv = data_torch.view(n_head_kv, n_head // n_head_kv + 2, head_dim, head_dim * n_head)
q = qkv[:, :-2].reshape(n_head * head_dim, head_dim * n_head)
k = qkv[:, [-2]].reshape(n_head_kv * head_dim, head_dim * n_head)
v = qkv[:, [-1]].reshape(n_head_kv * head_dim, head_dim * n_head)
data_torch = torch.cat((q, k, v)).reshape_as(data_torch)
data = data_torch.squeeze().numpy()
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
n_dims = len(data.shape)
data_dtype = data.dtype
# if f32 desired, convert any float16 to float32
if self.ftype == 0 and data_dtype == np.float16:
data = data.astype(np.float32)
# TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
data = data.astype(np.float32)
# if f16 desired, convert any float32 2-dim weight tensors to float16
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
@Model.register("GPTBigCodeForCausalLM")
class StarCoderModel(Model):
model_arch = gguf.MODEL_ARCH.STARCODER
def set_gguf_parameters(self):
block_count = self.hparams["n_layer"]
self.gguf_writer.add_name("StarCoder")
self.gguf_writer.add_context_length(self.hparams["n_positions"])
self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
self.gguf_writer.add_feed_forward_length(4 * self.hparams["n_embd"])
self.gguf_writer.add_block_count(block_count)
self.gguf_writer.add_head_count(self.hparams["n_head"])
self.gguf_writer.add_head_count_kv(1)
self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
self.gguf_writer.add_file_type(self.ftype)
@Model.register("GPTRefactForCausalLM")
class RefactModel(Model):
model_arch = gguf.MODEL_ARCH.REFACT
def set_gguf_parameters(self):
hidden_dim = self.hparams["n_embd"]
inner_dim = 4 * hidden_dim
hidden_dim = int(2 * inner_dim / 3)
multiple_of = 256
ff_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
block_count = self.hparams["n_layer"]
self.gguf_writer.add_name("Refact")
# refact uses Alibi. So this is from config.json which might be used by training.
self.gguf_writer.add_context_length(self.hparams["n_positions"])
self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
self.gguf_writer.add_feed_forward_length(ff_dim)
self.gguf_writer.add_block_count(block_count)
self.gguf_writer.add_head_count(self.hparams["n_head"])
self.gguf_writer.add_head_count_kv(1)
self.gguf_writer.add_layer_norm_rms_eps(self.hparams["layer_norm_epsilon"])
self.gguf_writer.add_file_type(self.ftype)
def write_tensors(self):
hidden_dim = self.hparams["n_embd"]
inner_dim = 4 * hidden_dim
hidden_dim = int(2 * inner_dim / 3)
multiple_of = 256
ff_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
n_head = self.hparams["n_head"]
n_head_kv = 1
head_dim = self.hparams["n_embd"] // n_head
block_count = self.hparams["n_layer"]
tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
tensors = dict(self.get_tensors())
for i in range(block_count):
if (w := tensors.get(f"transformer.h.{i}.attn.kv.weight")) is not None:
tensors[f"model.layers.{i}.self_attn.k_proj.weight"] = w[:n_head_kv * head_dim]
tensors[f"model.layers.{i}.self_attn.v_proj.weight"] = w[n_head_kv * head_dim:]
del tensors[f"transformer.h.{i}.attn.kv.weight"]
if (w := tensors.get(f"transformer.h.{i}.attn.q.weight")) is not None:
tensors[f"model.layers.{i}.self_attn.q_proj.weight"] = w
del tensors[f"transformer.h.{i}.attn.q.weight"]
if (w := tensors.get(f"transformer.h.{i}.mlp.gate_up_proj.weight")) is not None:
tensors[f"model.layers.{i}.mlp.gate_proj.weight"] = w[:ff_dim]
tensors[f"model.layers.{i}.mlp.up_proj.weight"] = w[ff_dim:]
del tensors[f"transformer.h.{i}.mlp.gate_up_proj.weight"]
for name, data_torch in tensors.items():
old_dtype = data_torch.dtype
# convert any unsupported data types to float32
if data_torch.dtype not in (torch.float16, torch.float32):
data_torch = data_torch.to(torch.float32)
data = data_torch.squeeze().numpy()
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight",))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
n_dims = len(data.shape)
data_dtype = data.dtype
# if f32 desired, convert any float16 to float32
if self.ftype == 0 and data_dtype == np.float16:
data = data.astype(np.float32)
# TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
data = data.astype(np.float32)
# if f16 desired, convert any float32 2-dim weight tensors to float16
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
@Model.register("PersimmonForCausalLM")
class PersimmonModel(Model):
model_arch = gguf.MODEL_ARCH.PERSIMMON
def set_gguf_parameters(self):
block_count = self.hparams.get("num_layers", self.hparams.get("num_hidden_layers"))
head_count = self.hparams["num_attention_heads"]
head_count_kv = head_count
hidden_size = self.hparams["hidden_size"]
self.gguf_writer.add_name('persimmon-8b-chat')
self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"])
self.gguf_writer.add_embedding_length(hidden_size)
self.gguf_writer.add_block_count(block_count)
self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
# NOTE: not sure about this change - why does the model not have a rope dimension count when it is smaller
# than the head size?
# ref: https://github.com/ggerganov/llama.cpp/pull/4889
2024-01-12 11:03:38 +00:00
# self.gguf_writer.add_rope_dimension_count(hidden_size // head_count)
self.gguf_writer.add_rope_dimension_count(hidden_size // head_count // 2)
self.gguf_writer.add_head_count(head_count)
self.gguf_writer.add_head_count_kv(head_count_kv)
self.gguf_writer.add_rope_freq_base(self.hparams["rope_theta"])
self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_eps"])
def set_vocab(self):
self._set_vocab_sentencepiece()
# self.gguf_writer.add_bos_token_id(71013)
# self.gguf_writer.add_eos_token_id(71013)
def write_tensors(self):
block_count = self.hparams.get("num_layers", self.hparams.get("num_hidden_layers"))
tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
for name, data_torch in self.get_tensors():
if name.endswith(".self_attention.rotary_emb.inv_freq"):
continue
old_dtype = data_torch.dtype
# TODO: FP16 conversion produces garbage outputs. (Q8_0 does not, so..?)
data = data_torch.to(torch.float32).squeeze().numpy()
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
n_dims = len(data.shape)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
@Model.register("StableLmForCausalLM", "StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM")
class StableLMModel(Model):
model_arch = gguf.MODEL_ARCH.STABLELM
def set_vocab(self):
if (self.dir_model / "tokenizer.json").is_file():
self._set_vocab_gpt2()
else:
# StableLM 2 1.6B uses a vocab in a similar format to Qwen's vocab
self._set_vocab_qwen()
def set_gguf_parameters(self):
hparams = self.hparams
block_count = hparams["num_hidden_layers"]
self.gguf_writer.add_name(self.dir_model.name)
self.gguf_writer.add_context_length(hparams["max_position_embeddings"])
self.gguf_writer.add_embedding_length(hparams["hidden_size"])
self.gguf_writer.add_block_count(block_count)
self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
rotary_factor = self.find_hparam(["partial_rotary_factor", "rope_pct"])
self.gguf_writer.add_rope_dimension_count(int(rotary_factor * (hparams["hidden_size"] // hparams["num_attention_heads"])))
self.gguf_writer.add_head_count(hparams["num_attention_heads"])
self.gguf_writer.add_parallel_residual(hparams["use_parallel_residual"] if "use_parallel_residual" in hparams else True)
self.gguf_writer.add_layer_norm_eps(self.find_hparam(["layer_norm_eps", "norm_eps"]))
@Model.register("LlamaForCausalLM", "MistralForCausalLM", "MixtralForCausalLM")
class LlamaModel(Model):
model_arch = gguf.MODEL_ARCH.LLAMA
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 12:04:25 +00:00
def set_vocab(self):
try:
self. _set_vocab_sentencepiece()
except FileNotFoundError:
self._set_vocab_llama_hf()
def set_gguf_parameters(self):
super().set_gguf_parameters()
hparams = self.hparams
self.gguf_writer.add_vocab_size(hparams["vocab_size"])
self.gguf_writer.add_rope_dimension_count(hparams["hidden_size"] // hparams["num_attention_heads"])
# Same as super class, but permuting q_proj, k_proj
def write_tensors(self):
block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
n_head = self.hparams.get("num_attention_heads")
n_kv_head = self.hparams.get("num_key_value_heads")
for name, data_torch in self.get_tensors():
# we don't need these
if name.endswith((".attention.masked_bias", ".attention.bias", ".attention.rotary_emb.inv_freq")):
continue
old_dtype = data_torch.dtype
# convert any unsupported data types to float32
if data_torch.dtype not in (torch.float16, torch.float32):
data_torch = data_torch.to(torch.float32)
data = data_torch.numpy()
if name.endswith("q_proj.weight"):
data = permute(data, n_head, n_head)
if name.endswith("k_proj.weight"):
data = permute(data, n_head, n_kv_head)
data = data.squeeze()
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
n_dims = len(data.shape)
data_dtype = data.dtype
# if f32 desired, convert any float16 to float32
if self.ftype == 0 and data_dtype == np.float16:
data = data.astype(np.float32)
# TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
data = data.astype(np.float32)
# if f16 desired, convert any float32 2-dim weight tensors to float16
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
llama : add Mixtral support (#4406) * convert : support Mixtral as LLAMA arch * convert : fix n_ff typo * llama : model loading * ggml : sync latest ggml_mul_mat_id * llama : update graph to support MoE * llama : fix cur -> cur_expert * llama : first working version * llama : fix expert weighting in the FFN * ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only) * ggml : add n_as argument to ggml_mul_mat_id * ggml : fix ggml_get_rows to take into account ne02 / ne11 * metal : add more general support for ggml_get_rows + tests * llama : add basic support for offloading moe with CUDA * metal : add/mul/div use general kernel when src1 not cont * metal : reduce the kernel launches for ggml_mul_mat_id * ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D * ggml : update get_rows f16 and q * cuda : support non-contiguous src1 in get_rows * llama : offload missing ffn_moe_silu * metal : fix ggml_get_rows to work with non-cont src1 * metal : add indirect mat-vec kernels for all quantization types * llama : do not quantize expert gating tensors * llama : add n_expert and n_expert_used to hparams + change quants * test-backend-ops : add moe test * cuda : fix get_rows when ncols is odd * convert : determine n_ctx correctly * metal : fix ggml_mul_mat_id for F32 * test-backend-ops : make experts more evenly probable (test_moe) * test-backend-ops : cleanup, add moe test for batches * test-backend-ops : add cpy from f32 -> all types test * test-backend-ops : fix dequantize block offset * llama : fix hard-coded number of experts * test-backend-ops : simplify and disable slow tests to avoid CI timeout * test-backend-ops : disable MOE test with thread sanitizer * cuda : fix mul_mat_id with multi gpu * convert : use 1e6 rope_freq_base for mixtral * convert : fix style * convert : support safetensors format * gguf-py : bump version * metal : add cpy f16 -> f32 kernel * metal : fix binary ops for ne10 % 4 != 0 * test-backend-ops : add one more sum_rows test * ggml : do not use BLAS with ggml_mul_mat_id * convert-hf : support for mixtral-instruct (#4428) * convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct * convert : use sentencepiece tokenizer for Mixtral-instruct * convert : make flake8 happy * metal : fix soft_max kernels ref: https://github.com/ggerganov/ggml/pull/621/commits/1914017863d2f9ab8ecc0281cc2a56d683668b92 * metal : limit kernels to not use more than the allowed threads --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Radek Pilar <github@mrkva.eu>
2023-12-13 12:04:25 +00:00
@Model.register("GrokForCausalLM")
class GrokModel(Model):
model_arch = gguf.MODEL_ARCH.GROK
def set_vocab(self):
self._set_vocab_sentencepiece()
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
def set_gguf_parameters(self):
super().set_gguf_parameters()
self.gguf_writer.add_name("Grok")
@Model.register("MiniCPMForCausalLM")
class MiniCPMModel(Model):
model_arch = gguf.MODEL_ARCH.MINICPM
def set_gguf_parameters(self):
block_count = self.hparams["num_hidden_layers"]
self.gguf_writer.add_name("MiniCPM")
self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"])
self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
self.gguf_writer.add_block_count(block_count)
self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
self.gguf_writer.add_rope_dimension_count(self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
self.gguf_writer.add_head_count(self.hparams["num_attention_heads"])
self.gguf_writer.add_head_count_kv(self.hparams["num_key_value_heads"])
self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
self.gguf_writer.add_file_type(self.ftype)
def set_vocab(self):
self._set_vocab_llama_hf()
def _reverse_hf_permute(self, weights: Tensor, n_head: int, n_kv_head: int | None = None) -> Tensor:
if n_kv_head is not None and n_head != n_kv_head:
n_head //= n_kv_head
return (
weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
.swapaxes(1, 2)
.reshape(weights.shape)
)
def write_tensors(self):
block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
n_head = self.hparams.get("num_attention_heads")
n_kv_head = self.hparams.get("num_key_value_heads")
for name, data_torch in self.get_tensors():
# we don't need these
if name.endswith((".attention.masked_bias", ".attention.bias", ".attention.rotary_emb.inv_freq")):
continue
old_dtype = data_torch.dtype
# convert any unsupported data types to float32
if data_torch.dtype not in (torch.float16, torch.float32):
data_torch = data_torch.to(torch.float32)
# HF models permute some of the tensors, so we need to undo that
if name.endswith(("q_proj.weight")):
data_torch = self._reverse_hf_permute(data_torch, n_head, n_head)
if name.endswith(("k_proj.weight")):
data_torch = self._reverse_hf_permute(data_torch, n_head, n_kv_head)
data = data_torch.squeeze().numpy()
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
n_dims = len(data.shape)
data_dtype = data.dtype
# if f32 desired, convert any float16 to float32
if self.ftype == 0 and data_dtype == np.float16:
data = data.astype(np.float32)
# TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
data = data.astype(np.float32)
# if f16 desired, convert any float32 2-dim weight tensors to float16
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
@Model.register("QWenLMHeadModel")
class QwenModel(Model):
model_arch = gguf.MODEL_ARCH.QWEN
@staticmethod
def token_bytes_to_string(b):
from transformers.models.gpt2.tokenization_gpt2 import bytes_to_unicode
byte_encoder = bytes_to_unicode()
return ''.join([byte_encoder[ord(char)] for char in b.decode('latin-1')])
@staticmethod
def bpe(mergeable_ranks: dict[bytes, int], token: bytes, max_rank: int | None = None) -> list[bytes]:
parts = [bytes([b]) for b in token]
while True:
min_idx = None
min_rank = None
for i, pair in enumerate(zip(parts[:-1], parts[1:])):
rank = mergeable_ranks.get(pair[0] + pair[1])
if rank is not None and (min_rank is None or rank < min_rank):
min_idx = i
min_rank = rank
if min_rank is None or (max_rank is not None and min_rank >= max_rank):
break
assert min_idx is not None
parts = parts[:min_idx] + [parts[min_idx] + parts[min_idx + 1]] + parts[min_idx + 2:]
return parts
def set_vocab(self):
self._set_vocab_qwen()
def set_gguf_parameters(self):
self.gguf_writer.add_name("Qwen")
self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"])
self.gguf_writer.add_block_count(self.hparams["num_hidden_layers"])
self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
self.gguf_writer.add_rope_freq_base(self.hparams["rotary_emb_base"])
self.gguf_writer.add_rope_dimension_count(self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
self.gguf_writer.add_head_count(self.hparams["num_attention_heads"])
self.gguf_writer.add_layer_norm_rms_eps(self.hparams["layer_norm_epsilon"])
def write_tensors(self):
block_count = self.hparams["num_hidden_layers"]
model_kv = dict(self.get_tensors())
tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
for name, data_torch in model_kv.items():
# we don't need these
if name.endswith(".rotary_emb.inv_freq"):
continue
old_dtype = data_torch.dtype
# convert any unsupported data types to float32
if data_torch.dtype not in (torch.float16, torch.float32):
data_torch = data_torch.to(torch.float32)
data = data_torch.squeeze().numpy()
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
n_dims = len(data.shape)
data_dtype = data.dtype
# if f32 desired, convert any float16 to float32
if self.ftype == 0 and data_dtype == np.float16:
data = data.astype(np.float32)
# TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
data = data.astype(np.float32)
# if f16 desired, convert any float32 2-dim weight tensors to float16
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
@Model.register("Qwen2ForCausalLM")
class Qwen2Model(Model):
model_arch = gguf.MODEL_ARCH.QWEN2
@Model.register("GPT2LMHeadModel")
class GPT2Model(Model):
model_arch = gguf.MODEL_ARCH.GPT2
def set_gguf_parameters(self):
self.gguf_writer.add_name(self.dir_model.name)
self.gguf_writer.add_block_count(self.hparams["n_layer"])
self.gguf_writer.add_context_length(self.hparams["n_ctx"])
self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
self.gguf_writer.add_feed_forward_length(4 * self.hparams["n_embd"])
self.gguf_writer.add_head_count(self.hparams["n_head"])
self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
self.gguf_writer.add_file_type(self.ftype)
def write_tensors(self):
block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
for name, data_torch in self.get_tensors():
# we don't need these
if name.endswith((".attention.masked_bias", ".attention.bias", ".attention.rotary_emb.inv_freq", ".attn.bias", ".attn.masked_bias")):
continue
if name.endswith((".c_attn.weight", ".c_proj.weight", ".c_fc.weight", ".c_proj.weight")):
data_torch = data_torch.transpose(1, 0)
old_dtype = data_torch.dtype
# convert any unsupported data types to float32
if data_torch.dtype not in (torch.float16, torch.float32):
data_torch = data_torch.to(torch.float32)
data = data_torch.squeeze().numpy()
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
n_dims = len(data.shape)
data_dtype = data.dtype
# if f32 desired, convert any float16 to float32
if self.ftype == 0 and data_dtype == np.float16:
data = data.astype(np.float32)
# TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
data = data.astype(np.float32)
# if f16 desired, convert any float32 2-dim weight tensors to float16
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
# note: GPT2 output is tied to (same as) wte in original model
if new_name == "token_embd.weight":
print(f"output.weight, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor("output.weight", data)
@Model.register("PhiForCausalLM")
class Phi2Model(Model):
model_arch = gguf.MODEL_ARCH.PHI2
def set_gguf_parameters(self):
block_count = self.find_hparam(["num_hidden_layers", "n_layer"])
rot_pct = self.find_hparam(["partial_rotary_factor"])
n_embd = self.find_hparam(["hidden_size", "n_embd"])
n_head = self.find_hparam(["num_attention_heads", "n_head"])
self.gguf_writer.add_name("Phi2")
self.gguf_writer.add_context_length(self.find_hparam(["n_positions", "max_position_embeddings"]))
self.gguf_writer.add_embedding_length(n_embd)
self.gguf_writer.add_feed_forward_length(4 * n_embd)
self.gguf_writer.add_block_count(block_count)
self.gguf_writer.add_head_count(n_head)
self.gguf_writer.add_head_count_kv(n_head)
self.gguf_writer.add_layer_norm_eps(self.find_hparam(["layer_norm_epsilon", "layer_norm_eps"]))
self.gguf_writer.add_rope_dimension_count(int(rot_pct * n_embd) // n_head)
self.gguf_writer.add_file_type(self.ftype)
self.gguf_writer.add_add_bos_token(False)
@Model.register("PlamoForCausalLM")
class PlamoModel(Model):
model_arch = gguf.MODEL_ARCH.PLAMO
def set_vocab(self):
self._set_vocab_sentencepiece()
def set_gguf_parameters(self):
hparams = self.hparams
block_count = hparams["num_hidden_layers"]
self.gguf_writer.add_name("PLaMo")
self.gguf_writer.add_context_length(4096) # not in config.json
self.gguf_writer.add_embedding_length(hparams["hidden_size"])
self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
self.gguf_writer.add_block_count(block_count)
self.gguf_writer.add_head_count(hparams["num_attention_heads"])
self.gguf_writer.add_head_count_kv(5) # hparams["num_key_value_heads"]) is wrong
self.gguf_writer.add_layer_norm_rms_eps(hparams["rms_norm_eps"])
def shuffle_attn_q_weight(self, data_torch):
assert data_torch.size() == (5120, 5120)
data_torch = data_torch.reshape(8, 5, 128, 5120)
data_torch = torch.permute(data_torch, (1, 0, 2, 3))
data_torch = torch.reshape(data_torch, (5120, 5120))
return data_torch
def shuffle_attn_output_weight(self, data_torch):
assert data_torch.size() == (5120, 5120)
data_torch = data_torch.reshape(5120, 8, 5, 128)
data_torch = torch.permute(data_torch, (0, 2, 1, 3))
data_torch = torch.reshape(data_torch, (5120, 5120))
return data_torch
def write_tensors(self):
block_count = self.hparams.get("num_layers", self.hparams.get("num_hidden_layers"))
tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
for name, data_torch in self.get_tensors():
if "self_attn.rotary_emb.inv_freq" in name:
continue
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
# shuffle for broadcasting of gqa in ggml_mul_mat
if new_name.endswith("attn_q.weight"):
data_torch = self.shuffle_attn_q_weight(data_torch)
elif new_name.endswith("attn_output.weight"):
data_torch = self.shuffle_attn_output_weight(data_torch)
old_dtype = data_torch.dtype
# convert any unsupported data types to float32
if data_torch.dtype not in (torch.float16, torch.float32):
data_torch = data_torch.to(torch.float32)
data = data_torch.squeeze().numpy()
n_dims = len(data.shape)
data_dtype = data.dtype
# if f32 desired, convert any float16 to float32
if self.ftype == 0 and data_dtype == np.float16:
data = data.astype(np.float32)
# TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
data = data.astype(np.float32)
# if f16 desired, convert any float32 2-dim weight tensors to float16
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
2024-01-19 11:52:22 +00:00
@Model.register("CodeShellForCausalLM")
class CodeShellModel(Model):
model_arch = gguf.MODEL_ARCH.CODESHELL
def set_gguf_parameters(self):
block_count = self.hparams["n_layer"]
self.gguf_writer.add_name("CodeShell")
self.gguf_writer.add_context_length(self.hparams["n_positions"])
self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
self.gguf_writer.add_feed_forward_length(4 * self.hparams["n_embd"])
self.gguf_writer.add_block_count(block_count)
self.gguf_writer.add_head_count(self.hparams["n_head"])
self.gguf_writer.add_head_count_kv(self.hparams["num_query_groups"])
self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
self.gguf_writer.add_file_type(self.ftype)
self.gguf_writer.add_rope_freq_base(10000.0)
self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
self.gguf_writer.add_rope_scaling_factor(1.0)
def write_tensors(self):
block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
tensors = dict(self.get_tensors())
has_lm_head = "lm_head.weight" in tensors.keys() or "output.weight" in tensors.keys()
for name, data_torch in tensors.items():
# we don't need these
if name.endswith((".attn.rotary_emb.inv_freq")):
continue
old_dtype = data_torch.dtype
# convert any unsupported data types to float32
if data_torch.dtype not in (torch.float16, torch.float32):
data_torch = data_torch.to(torch.float32)
data = data_torch.squeeze().numpy()
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
n_dims = len(data.shape)
data_dtype = data.dtype
# if f32 desired, convert any float16 to float32
if self.ftype == 0 and data_dtype == np.float16:
data = data.astype(np.float32)
# TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
data = data.astype(np.float32)
# if f16 desired, convert any float32 2-dim weight tensors to float16
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
if not has_lm_head and name == "transformer.wte.weight":
self.gguf_writer.add_tensor("output.weight", data)
print(name, f"=> output.weight, shape = {data.shape}, {old_dtype} --> {data.dtype}")
@Model.register("InternLM2ForCausalLM")
class InternLM2Model(Model):
model_arch = gguf.MODEL_ARCH.INTERNLM2
def set_vocab(self):
# (TODO): Is there a better way?
# Copy from _set_vocab_sentencepiece, The only difference is that we will treat the character
# \x00 specially and convert it into an emoji character to prevent it from being mistakenly
# recognized as an empty string in C++.
from sentencepiece import SentencePieceProcessor
from sentencepiece import sentencepiece_model_pb2 as model
tokenizer_path = self.dir_model / 'tokenizer.model'
tokens: list[bytes] = []
scores: list[float] = []
toktypes: list[int] = []
if not tokenizer_path.is_file():
print(f'Error: Missing {tokenizer_path}', file=sys.stderr)
sys.exit(1)
sentencepiece_model = model.ModelProto()
sentencepiece_model.ParseFromString(open(tokenizer_path, "rb").read())
add_prefix = sentencepiece_model.normalizer_spec.add_dummy_prefix
tokenizer = SentencePieceProcessor(str(tokenizer_path))
vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
for token_id in range(vocab_size):
piece = tokenizer.id_to_piece(token_id)
text = piece.encode("utf-8")
score = tokenizer.get_score(token_id)
if text == b"\x00":
# (TODO): fixme
# Hack here and replace the \x00 characters.
print(f"InternLM2 convert token '{text}' to '🐉'!")
text = "🐉"
toktype = SentencePieceTokenTypes.NORMAL
if tokenizer.is_unknown(token_id):
toktype = SentencePieceTokenTypes.UNKNOWN
elif tokenizer.is_control(token_id):
toktype = SentencePieceTokenTypes.CONTROL
elif tokenizer.is_unused(token_id):
toktype = SentencePieceTokenTypes.UNUSED
elif tokenizer.is_byte(token_id):
toktype = SentencePieceTokenTypes.BYTE
tokens.append(text)
scores.append(score)
toktypes.append(toktype)
added_tokens_file = self.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:
tokens.append(key.encode("utf-8"))
scores.append(-1000.0)
toktypes.append(SentencePieceTokenTypes.USER_DEFINED)
self.gguf_writer.add_tokenizer_model("llama")
self.gguf_writer.add_token_list(tokens)
self.gguf_writer.add_token_scores(scores)
self.gguf_writer.add_token_types(toktypes)
self.gguf_writer.add_add_space_prefix(add_prefix)
special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
old_eos = special_vocab.special_token_ids["eos"]
if "chat" in os.path.basename(self.dir_model.absolute()):
# For the chat model, we replace the eos with '<|im_end|>'.
special_vocab.special_token_ids["eos"] = self._try_get_sft_eos(tokenizer)
print(f"Replace eos:{old_eos} with a special token:{special_vocab.special_token_ids['eos']} \
in chat mode so that the conversation can end normally.")
special_vocab.add_to_gguf(self.gguf_writer)
def _try_get_sft_eos(self, tokenizer):
unused_145_list = tokenizer.encode('[UNUSED_TOKEN_145]')
im_end_list = tokenizer.encode('<|im_end|>')
assert (len(unused_145_list) == 1) ^ (len(im_end_list) == 1)
if len(unused_145_list) == 1:
eos_token = unused_145_list[0]
if len(im_end_list) == 1:
eos_token = im_end_list[0]
return eos_token
def _hf_permute_qk(self, weights, n_head: int, n_head_kv: int):
if n_head_kv is not None and n_head != n_head_kv:
n_head = n_head_kv
return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
.swapaxes(1, 2)
.reshape(weights.shape))
def set_gguf_parameters(self):
self.gguf_writer.add_name("InternLM2")
self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"])
self.gguf_writer.add_block_count(self.hparams["num_hidden_layers"])
self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
self.gguf_writer.add_rope_freq_base(self.hparams["rope_theta"])
self.gguf_writer.add_head_count(self.hparams["num_attention_heads"])
self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
self.gguf_writer.add_head_count_kv(self.hparams["num_key_value_heads"])
def post_write_tensors(self, tensor_map, name, data_torch):
old_dtype = data_torch.dtype
# convert any unsupported data types to float32
if data_torch.dtype not in (torch.float16, torch.float32):
data_torch = data_torch.to(torch.float32)
data = data_torch.squeeze().numpy()
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
n_dims = len(data.shape)
data_dtype = data.dtype
# if f32 desired, convert any float16 to float32
if self.ftype == 0 and data_dtype == np.float16:
data = data.astype(np.float32)
# TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
data = data.astype(np.float32)
# if f16 desired, convert any float32 2-dim weight tensors to float16
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
def write_tensors(self):
from einops import rearrange
num_heads = self.hparams.get("num_attention_heads")
num_kv_heads = self.hparams.get("num_key_value_heads")
hidden_size = self.hparams.get("hidden_size")
q_per_kv = num_heads // num_kv_heads
head_dim = hidden_size // num_heads
num_groups = num_heads // q_per_kv
block_count = self.hparams["num_hidden_layers"]
model_kv = dict(self.get_tensors())
tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
qkv_pattern = r"model\.layers\.(\d+)\.attention\.wqkv"
for name, data_torch in model_kv.items():
# we don't need these
if name.endswith(".rotary_emb.inv_freq"):
continue
if re.match(qkv_pattern, name):
bid = re.findall(qkv_pattern, name)[0]
qkv = data_torch
qkv = rearrange(qkv.T, " o (g n i) ->o g n i", g=num_groups, n=q_per_kv + 2, i=head_dim)
q, k, v = qkv[..., : q_per_kv, :], qkv[..., q_per_kv: q_per_kv + 1, :], qkv[..., q_per_kv + 1: q_per_kv + 2, :]
# The model weights of q and k equire additional reshape.
q = self._hf_permute_qk(rearrange(q, " o g n i -> o (g n i)").T, num_heads, num_heads)
k = self._hf_permute_qk(rearrange(k, " o g n i -> o (g n i)").T, num_heads, num_kv_heads)
v = rearrange(v, " o g n i -> o (g n i)").T
self.post_write_tensors(tensor_map, f"model.layers.{bid}.attention.wq.weight", q)
self.post_write_tensors(tensor_map, f"model.layers.{bid}.attention.wk.weight", k)
self.post_write_tensors(tensor_map, f"model.layers.{bid}.attention.wv.weight", v)
else:
self.post_write_tensors(tensor_map, name, data_torch)
@Model.register("BertModel", "CamembertModel")
class BertModel(Model):
model_arch = gguf.MODEL_ARCH.BERT
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.vocab_size = None
def set_gguf_parameters(self):
super().set_gguf_parameters()
self.gguf_writer.add_causal_attention(False)
# get pooling path
pooling_path = None
module_path = self.dir_model / "modules.json"
if module_path.is_file():
with open(module_path, encoding="utf-8") as f:
modules = json.load(f)
for mod in modules:
if mod["type"] == "sentence_transformers.models.Pooling":
pooling_path = mod["path"]
break
# get pooling type
if pooling_path is not None:
with open(self.dir_model / pooling_path / "config.json", encoding="utf-8") as f:
pooling = json.load(f)
if pooling["pooling_mode_mean_tokens"]:
pooling_type = gguf.PoolingType.MEAN
elif pooling["pooling_mode_cls_token"]:
pooling_type = gguf.PoolingType.CLS
else:
raise NotImplementedError("Only MEAN and CLS pooling types supported")
self.gguf_writer.add_pooling_type(pooling_type)
def set_vocab(self):
# use huggingface vocab to get all tokens
vocab = LlamaHfVocab(self.dir_model, ignore_nonllama=True)
tokens, scores, toktypes = zip(*vocab.all_tokens())
assert len(tokens) == vocab.vocab_size
self.vocab_size = vocab.vocab_size
# we need this to validate the size of the token_type embeddings
# though currently we are passing all zeros to the token_type embeddings
n_token_types = len(set(toktypes))
self.gguf_writer.add_token_type_count(n_token_types)
# convert to phantom space vocab
def phantom(tok, typ):
if tok.startswith(b"[") and tok.endswith(b"]"):
return tok
if tok.startswith(b"##"):
return tok[2:]
return b"\xe2\x96\x81" + tok
tokens = tuple(phantom(t, y) for t, y in zip(tokens, toktypes))
# set up bos and eos tokens (cls and sep)
self.gguf_writer.add_bos_token_id(vocab.tokenizer.cls_token_id)
self.gguf_writer.add_eos_token_id(vocab.tokenizer.sep_token_id)
# add vocab to gguf
self.gguf_writer.add_tokenizer_model("bert")
self.gguf_writer.add_token_list(tokens)
self.gguf_writer.add_token_scores(scores)
self.gguf_writer.add_token_types(toktypes)
# handle special tokens
special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
special_vocab.add_to_gguf(self.gguf_writer)
def write_tensors(self):
tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
tensors = dict(self.get_tensors())
for name, data_torch in tensors.items():
# we are only using BERT for embeddings so we don't need the pooling layer
if name in ("embeddings.position_ids", "pooler.dense.weight", "pooler.dense.bias"):
continue # we don't need these
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
data = data_torch.squeeze().numpy()
n_dims = len(data.shape)
new_dtype: type[np.floating[Any]]
if (
self.ftype == 1 and name.endswith(".weight") and n_dims == 2
and name != "embeddings.token_type_embeddings.weight" # not used with get_rows, must be F32
):
# if f16 desired, convert any float32 2-dim weight tensors to float16
new_dtype = np.float16
else:
# if f32 desired, convert any float16 to float32
new_dtype = np.float32
print(f"{new_name}, n_dims = {n_dims}, {data_torch.dtype} --> {new_dtype}")
if data.dtype != new_dtype:
data = data.astype(new_dtype)
self.gguf_writer.add_tensor(new_name, data)
@Model.register("NomicBertModel")
class NomicBertModel(BertModel):
model_arch = gguf.MODEL_ARCH.NOMIC_BERT
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
# the HF config claims n_ctx=8192, but it uses RoPE scaling
self.hparams["n_ctx"] = 2048
# SwigLU activation
assert self.hparams["activation_function"] == "swiglu"
# this doesn't do anything in the HF version
assert self.hparams["causal"] is False
# no bias tensors
assert self.hparams["qkv_proj_bias"] is False
assert self.hparams["mlp_fc1_bias"] is False
assert self.hparams["mlp_fc2_bias"] is False
# norm at end of layer
assert self.hparams["prenorm"] is False
# standard RoPE
assert self.hparams["rotary_emb_fraction"] == 1.0
assert self.hparams["rotary_emb_interleaved"] is False
assert self.hparams["rotary_emb_scale_base"] is None
def set_gguf_parameters(self):
super().set_gguf_parameters()
self.gguf_writer.add_rope_freq_base(self.hparams["rotary_emb_base"])
def get_tensors(self):
assert self.vocab_size is not None
for name, data in super().get_tensors():
# Nomic Embed's token embeddings tensor is padded, but llama.cpp wants tensor sizes to match exactly.
if name == 'embeddings.word_embeddings.weight' and data.shape[1] != self.vocab_size:
rounded_vocab_size = (self.vocab_size + 63) // 64 * 64
assert data.shape == (rounded_vocab_size, self.hparams["n_embd"])
data = data[:self.vocab_size, :]
yield name, data
@Model.register("GemmaForCausalLM")
class GemmaModel(Model):
model_arch = gguf.MODEL_ARCH.GEMMA
def set_vocab(self):
self._set_vocab_sentencepiece()
def set_gguf_parameters(self):
hparams = self.hparams
block_count = hparams["num_hidden_layers"]
self.gguf_writer.add_name(self.dir_model.name)
self.gguf_writer.add_context_length(hparams["max_position_embeddings"])
self.gguf_writer.add_embedding_length(hparams["hidden_size"])
self.gguf_writer.add_block_count(block_count)
self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
self.gguf_writer.add_head_count(hparams["num_attention_heads"])
self.gguf_writer.add_head_count_kv(self.hparams["num_key_value_heads"] if "num_key_value_heads" in hparams else hparams["num_attention_heads"])
self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
self.gguf_writer.add_key_length(hparams["head_dim"])
self.gguf_writer.add_value_length(hparams["head_dim"])
self.gguf_writer.add_file_type(self.ftype)
def write_tensors(self):
block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
for name, data_torch in self.get_tensors():
old_dtype = data_torch.dtype
# convert any unsupported data types to float32
if data_torch.dtype not in (torch.float16, torch.float32):
data_torch = data_torch.to(torch.float32)
# ref: https://github.com/huggingface/transformers/blob/fc37f38915372c15992b540dfcbbe00a916d4fc6/src/transformers/models/gemma/modeling_gemma.py#L89
if name.endswith("norm.weight"):
data_torch = data_torch + 1
data = data_torch.squeeze().numpy()
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
n_dims = len(data.shape)
data_dtype = data.dtype
data = data.astype(np.float32)
# if f16 desired, convert any float32 2-dim weight tensors to float16
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
@Model.register("Starcoder2ForCausalLM")
class StarCoder2Model(Model):
model_arch = gguf.MODEL_ARCH.STARCODER2
llama : support Mamba Selective State Space Models (#5328) * mamba : begin working on support for Mamba SSM * mamba : begin figuring out how to (ab)use the kv cache for Mamba * mamba : recurrent inference almost works, but incoherent * mamba : recurrent inference WORKS!!! * convert : optionally use d_conv and d_state from config.json for Mamba * mamba : refactor recurrent conv, resulting in 20% perf increase It's still slower than I'd like, but I did not really optimize `ggml_exp` yet. I also refactored `ggml_exp` to work with tensors with more than 2 dimensions. * ggml : parallelize ggml_exp This results in 8% faster token generation for Mamba-130M. * mamba : simplify the conv step with a self-overlapping view Turns out the conv_state can be made smaller by one column. Note that this breaks existing GGUFs of Mamba, because the key_value_length field is tied to the conv_state size. Convolution with a self-overlapping view is cool! And it's much simpler than what I initially thought would be necessary to make the convolution step work with more than 1 token at a time. Next step is to make the SSM step work on batches of tokens too, and thus I need to figure out a way to make a parallel selective scan which will keep the ssm_state small and won't make it bigger by a factor of (n_layer * batch_size). * llama : fix Mamba KV self size wrongly displaying as f16 instead of f32 Relatedly, I also tried to see if other types than f32 worked for the states, but they don't, because of the operators used. It's probably better anyway to keep lots of precision there, since the states are small anyway. * mamba : fix self-overlapping view depth stride * mamba : handle batches of more than 1 token This means running Mamba no longer crashes when using the default settings! And probably also slightly faster prompt processing. Both batched and non-batched processing yield the same output. Previously, the state was not cleared when starting a sequence. Next step is to make the KV cache API work as expected for Mamba models. * ggml: add ggml_ssm_scan to help with parallel selective scan If the selective scan was implemented without a custom operator, there would be waaay too many nodes in the graph. For example, for Mamba-130M, with a batch size of 512 (the default), a naive selective scan could add at least 24*512=12288 nodes, which is more than LLAMA_MAX_NODES (8192), and that's only for the smallest Mamba model. So it's much cleaner with a custom operator. Not sure about the name, though. * ggml : in ggml_ssm_scan, merge multiple rows in the same vec operation This will help with performance on CPU if ggml_vec_mul_f32 and ggml_vec_add_f32 are ever optimized with SIMD. * mamba : very basic quantization support Mostly works, but there is currently no difference between the variants of a k-quant (e.g. Q4_K_S and Q4_K_M are the same). Most of the SSM-specific weights can be kept in f32 without affecting the size that much, since they are relatively small. (the linear projection weights are responsible for most of Mamba's size) Too much quantization seems to make the state degrade quite fast, and the model begins to output gibberish. It seems to affect bigger models to a lesser extent than small models, but I'm not sure by how much. Experimentation will be needed to figure out which weights are more important for the _M (and _L?) variants of k-quants for Mamba. * convert : fix wrong name for layer norm weight of offical Mamba models I was using Q-bert/Mamba-* models before, which have a slighlty different naming scheme for the weights. (they start with "model.layers" instead of "backbone.layers") * mamba : fuse more steps of the SSM scan in the ggml_ssm_scan operator This increases performance on CPU by around 30% for prompt processing, and by around 20% for text generation. However, it also makes the ggml_exp and ggml_soft_plus operators unused. Whether or not they should be kept will be decided later. * convert : for Mamba, also consider the "MambaLMHeadModel" arch name It's the name of the class of the official implementation, though they don't use it (yet) in the "architectures" field of config.json * mamba : fix vocab size problems with official models The perplexity was waaaay to high for models with a non-round vocab size. Not sure why, but it needed to be fixed in the metadata. Note that this breaks existing GGUF-converted Mamba models, but **only if** the vocab size was not already rounded. * ggml : remove ggml_exp and ggml_soft_plus They did not exist anyway outside of this branch, and since ggml_ssm_scan fused operations together, they are unused. It's always possible to bring them back if needed. * mamba : remove some useless comments No code change. * convert : fix flake8 linter errors * mamba : apply suggestions from code review * mamba : remove unecessary branch for row-wise ssm_state and C multiplication It was previously done to avoid permuting when only one token is processed at a time (like when generating text), but permuting is cheap, and dynamically changing the compute graph is not future-proof. * ggml : in ggml_ssm_scan, use more appropriate asserts * ggml : rename the destination pointer in ggml_compute_forward_ssm_scan_f32 * mamba : multiple sequences, but one at a time This is a step towards making this Mamba implementation usable with the server example (the way the system prompt is kept when clearing the client slots will need to be changed before this can work, though). The KV cache size for this kind of model is tied to the maximum number of sequences kept at any single time. For now, this number is obtained from n_parallel (plus one, to have an extra sequence to dedicate to the system prompt), but there might be a better way to do this which won't also make the main example use 2 cells even if only 1 is really used. (for this specific case, --parallel 0 helps) Simultaneous sequence processing will probably require changes to ggml_ssm_scan, and possibly a new operator for the conv step. * mamba : support llama_kv_cache_seq_cp This (mis)uses the logic around K shifts, because tokens in a state can't be shifted anyway, and because inp_K_shift has the right shape and type. Using ggml_get_rows is a nice way to do copies, but copy chains can't work. Fortunately, copy chains don't really seem to be used in the examples. Each KV cell is dedicated to the sequence ID corresponding to its own index. * mamba : use a state mask It's cleaner than the previous heuristic of checking for the pos of the first token in the batch. inp_KQ_mask could not be re-used for this, because it has the wrong shape and because it seems more suited to the next step of simultaneous sequence processing (helping with the problem of remembering which token belongs to which sequence(s)/state(s)). * llama : replace the usage of n_ctx with kv_self.size in many places * mamba : use n_tokens directly instead of n_tok * mamba : in comments, properly refer to KV cells instead of slots * mamba : reduce memory usage of ggml_ssm_scan From 290.37 MiB to 140.68 MiB of CPU compute buffer size with Mamba 3B with a batch size of 512. The result tensor of ggml_ssm_scan was previously a big part of the CPU compute buffer size. To make it smaller, it does not contain the intermediate ssm states anymore. Both y and the last ssm state are combined in the result tensor, because it seems only a single tensor can be returned by an operator with the way the graph is built. * mamba : simultaneous sequence processing A batch can now contain tokens from multiple sequences. This is necessary for at least the parallel example, the server example, and the HellaSwag test in the perplexity example. However, for this to be useful, uses of llama_kv_cache_seq_rm/cp will need to be changed to work on whole sequences. * ggml : add ggml_ssm_conv as a new operator for the conv step of Mamba This operator makes it possible to use and update the correct states for each token of the batch in the same way as ggml_ssm_scan. Other solutions which use existing operators would need loops which would add too many nodes to the graph (at least the ones I thought of). Using this operator further reduces the size of the CPU compute buffer from 140.68 MiB to 103.20 MiB with Mamba 3B with a batch size of 512. And (at least on CPU), it's a bit faster than before. Note that "ggml_ssm_conv" is probably not the most appropriate name, and it could be changed if a better one is found. * llama : add inp_s_seq as a new input tensor The most convenient implementation to select the correct state (for Mamba) for each token is to directly get the correct index from a tensor. This is why inp_s_seq is storing int32_t and not floats. The other, less convenient way to select the correct state would be to have inp_KQ_mask contain 1.0f for each state used by a token and 0.0f otherwise. This complicates quickly fetching the first used state of a token, and is also less efficient because a whole row of the mask would always need to be read for each token. Using indexes makes it easy to stop searching when there are no more sequences for a token, and the first sequence assigned is always very quickly available (it's the first element of each row). * mamba : support llama_kv_cache_seq_cp copy chains * mamba : support shifting and dividing the kv cache pos * mamba : make the server and parallel examples work with whole sequences A seq_id is dedicated to the system prompt in both cases. * llama : make llama_kv_cache_seq_rm return whether it succeeded or not * mamba : dedicate an input tensor for state copy indices This is cleaner and makes it easier to adapt when/if token positions (and by extension, inp_K_shift) are no longer integers. * mamba : adapt perplexity, batched, and batched-bench examples * perplexity : limit the max number of sequences This adapts to what the loaded model can provide. * llama : add llama_n_max_seq to get the upper limit for seq_ids Used by the perplexity example. * batched : pass n_parallel to the model's context params This should have been there already, but it wasn't. * batched-bench : reserve sequences to support Mamba * batched-bench : fix tokens being put in wrong sequences Generation quality isn't what's measured in there anyway, but at least using the correct sequences avoids using non-consecutive token positions. * mamba : stop abusing attention metadata This breaks existing converted-to-GGUF Mamba models, but will allow supporting mixed architectures like MambaFormer without needing to break Mamba models. This will also allow changing the size of Mamba's states without having to reconvert models in the future. (e.g. using something else than d_conv - 1 columns for the conv_states will not require breaking existing converted Mamba models again) * gguf-py : add new KV metadata key-value pairs for Mamba * llama : add new metadata key-value pairs for Mamba * llama : guard against divisions by zero when n_head is 0 * mamba : rename "unlimited" KV cache property to "recurrent" * mamba : more correctly update the "used" field of the KV cache * ggml : in ggml_ssm_scan, use a threshold for soft_plus This is how the official Mamba implementation does it, and it's also what torch.nn.Softplus does. * convert : for Mamba, fallback to internal NeoX tokenizer The resulting models are exactly the same as if the tokenizer.json and tokenizer_config.json of GPT-NeoX were there. * mamba : support state saving and restoring * ggml : implicitly pass src tensors through dst for Mamba-related ops * mamba : clarify some comments * server : fix cache_tokens not getting correctly resized Otherwise, when the "we have to evaluate at least 1 token" special case was triggered, an extra token was kept in cache_tokens even if it was removed from the KV cache. For Mamba, this caused useless prompt reprocessing when the previous request triggered the above case. * convert-hf : support new metadata keys for Mamba For the models available at https://huggingface.co/collections/state-spaces/transformers-compatible-mamba-65e7b40ab87e5297e45ae406 * mamba : rename metadata to be more similar to transformers library This breaks existing converted-to-GGUF models, but the metadata names are more "standard". * mamba : support mamba-*-hf models These models share their token_embd.weight with their output.weight * mamba : add missing spaces This is purely a formatting change. * convert-hf : omit output.weight when identical with token_embd.weight Only for Mamba for now, but it might be relevant for other models eventually. Most Mamba models actually share these two tensors, albeit implicitly. * readme : add Mamba to supported models, and add recent API changes * mamba : move state_seq and state_mask views outside layer loop A few tensors were also missing `struct` in front of `ggml_tensor`.
2024-03-08 22:31:00 +00:00
@Model.register("MambaForCausalLM", "MambaLMHeadModel")
class MambaModel(Model):
model_arch = gguf.MODEL_ARCH.MAMBA
def set_vocab(self):
vocab_size = self.hparams["vocab_size"]
# Round vocab size to next multiple of 8
pad_vocab = self.hparams.get("pad_vocab_size_multiple", 8)
# pad using ceiling division
# ref: https://stackoverflow.com/a/17511341/22827863
vocab_size = -(vocab_size // -pad_vocab) * pad_vocab
self.hparams["vocab_size"] = vocab_size
if (self.dir_model / "tokenizer.json").is_file():
self._set_vocab_gpt2()
else:
# Use the GPT-NeoX tokenizer when no tokenizer files are present
tokenizer_path = Path(sys.path[0]) / "models" / "ggml-vocab-gpt-neox.gguf"
print(f"Using tokenizer from '{os.path.relpath(tokenizer_path, os.getcwd())}'")
neox_reader = gguf.GGUFReader(tokenizer_path, "r")
field = neox_reader.get_field(gguf.Keys.Tokenizer.MODEL)
self.gguf_writer.add_tokenizer_model(bytes(field.parts[-1]))
field = neox_reader.get_field(gguf.Keys.Tokenizer.LIST)
self.gguf_writer.add_token_list([bytes(field.parts[i]) for i in field.data][:vocab_size])
field = neox_reader.get_field(gguf.Keys.Tokenizer.TOKEN_TYPE)
self.gguf_writer.add_token_types([field.parts[i].tolist()[0] for i in field.data][:vocab_size])
field = neox_reader.get_field(gguf.Keys.Tokenizer.MERGES)
self.gguf_writer.add_token_merges([bytes(field.parts[i]) for i in field.data])
field = neox_reader.get_field(gguf.Keys.Tokenizer.BOS_ID)
self.gguf_writer.add_bos_token_id(field.parts[-1].tolist()[0])
field = neox_reader.get_field(gguf.Keys.Tokenizer.EOS_ID)
self.gguf_writer.add_eos_token_id(field.parts[-1].tolist()[0])
field = neox_reader.get_field(gguf.Keys.Tokenizer.UNK_ID)
self.gguf_writer.add_unk_token_id(field.parts[-1].tolist()[0])
def set_gguf_parameters(self):
d_model = self.find_hparam(["hidden_size", "d_model"])
d_conv = self.find_hparam(["conv_kernel", "d_conv"], optional=True) or 4
d_inner = self.find_hparam(["intermediate_size", "d_inner"], optional=True) or 2 * d_model
d_state = self.find_hparam(["state_size", "d_state"], optional=True) or 16
# ceiling division
# ref: https://stackoverflow.com/a/17511341/22827863
# ref: https://github.com/state-spaces/mamba/blob/ce59daea3a090d011d6476c6e5b97f6d58ddad8b/mamba_ssm/modules/mamba_simple.py#L58
dt_rank = self.find_hparam(["time_step_rank", "dt_rank"], optional=True) or -(d_model // -16)
rms_norm_eps = self.find_hparam(["layer_norm_epsilon", "rms_norm_eps"], optional=True) or 1e-5
# Fail early for models which don't have a block expansion factor of 2
assert d_inner == 2 * d_model
self.gguf_writer.add_name(self.dir_model.name)
self.gguf_writer.add_context_length(2**20) # arbitrary value; for those who use the default
self.gguf_writer.add_embedding_length(d_model)
self.gguf_writer.add_feed_forward_length(0) # unused, but seemingly required when loading
self.gguf_writer.add_head_count(0) # unused, but seemingly required when loading
self.gguf_writer.add_block_count(self.hparams["n_layer"])
self.gguf_writer.add_ssm_conv_kernel(d_conv)
self.gguf_writer.add_ssm_inner_size(d_inner)
self.gguf_writer.add_ssm_state_size(d_state)
self.gguf_writer.add_ssm_time_step_rank(dt_rank)
self.gguf_writer.add_layer_norm_rms_eps(rms_norm_eps)
self.gguf_writer.add_file_type(self.ftype)
def write_tensors(self):
block_count = self.hparams["n_layer"]
tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
tok_embd = None
tok_embd_name = gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.TOKEN_EMBD] + ".weight"
output_name = gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.OUTPUT] + ".weight"
for name, data_torch in self.get_tensors():
old_dtype = data_torch.dtype
# convert any unsupported data types to float32
if data_torch.dtype not in (torch.float16, torch.float32):
data_torch = data_torch.to(torch.float32)
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
if name.endswith(".A_log"):
print("A_log --> A ==> " + new_name)
data_torch = -torch.exp(data_torch)
# assuming token_embd.weight is seen before output.weight
if tok_embd is not None and new_name == output_name:
if torch.equal(tok_embd, data_torch):
print(f"{output_name} is equivalent to {tok_embd_name}, omitting")
continue
if new_name == tok_embd_name:
tok_embd = data_torch
data = data_torch.squeeze().numpy()
n_dims = len(data.shape)
data_dtype = data.dtype
# if f32 desired, convert any float16 to float32
if self.ftype == 0 and data_dtype == np.float16:
data = data.astype(np.float32)
# TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
data = data.astype(np.float32)
# if f16 desired, convert big float32 2-dim weight tensors to float16
if self.ftype == 1 and data_dtype == np.float32 and new_name.removesuffix(".weight").endswith((".ssm_in", ".ssm_out", "token_embd", "output")) and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
@Model.register("CohereForCausalLM")
class CommandR2Model(Model):
model_arch = gguf.MODEL_ARCH.COMMAND_R
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
# max_position_embeddings = 8192 in config.json but model was actually
# trained on 128k context length
self.hparams["max_position_embeddings"] = self.hparams["model_max_length"]
def set_gguf_parameters(self):
super().set_gguf_parameters()
self.gguf_writer.add_logit_scale(self.hparams["logit_scale"])
self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
###### CONVERSION LOGIC ######
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser(
description="Convert a huggingface model to a GGML compatible file")
parser.add_argument(
"--vocab-only", action="store_true",
help="extract only the vocab",
)
parser.add_argument(
"--awq-path", type=Path, default=None,
help="Path to scale awq cache file")
parser.add_argument(
"--outfile", type=Path,
help="path to write to; default: based on input",
)
parser.add_argument(
"--outtype", type=str, choices=["f32", "f16"], default="f16",
help="output format - use f32 for float32, f16 for float16",
)
parser.add_argument("--bigendian", action="store_true", help="model is executed on big endian machine")
parser.add_argument(
"model", type=Path,
help="directory containing model file",
)
return parser.parse_args()
def main() -> None:
args = parse_args()
dir_model = args.model
if args.awq_path:
sys.path.insert(1, str(Path(__file__).parent / 'awq-py'))
from awq.apply_awq import add_scale_weights # type: ignore[import-not-found]
tmp_model_path = args.model / "weighted_model"
dir_model = tmp_model_path
if tmp_model_path.is_dir():
print(f"{tmp_model_path} exists as a weighted model.")
else:
tmp_model_path.mkdir(parents=True, exist_ok=True)
print("Saving new weighted model ...")
add_scale_weights(str(args.model), str(args.awq_path), str(tmp_model_path))
print(f"Saved weighted model at {tmp_model_path}.")
if not dir_model.is_dir():
print(f'Error: {args.model} is not a directory', file=sys.stderr)
sys.exit(1)
ftype_map = {
"f32": gguf.GGMLQuantizationType.F32,
"f16": gguf.GGMLQuantizationType.F16,
}
if args.outfile is not None:
fname_out = args.outfile
else:
# output in the same directory as the model by default
fname_out = dir_model / f'ggml-model-{args.outtype}.gguf'
print(f"Loading model: {dir_model.name}")
hparams = Model.load_hparams(dir_model)
with torch.inference_mode():
model_class = Model.from_model_architecture(hparams["architectures"][0])
model_instance = model_class(dir_model, ftype_map[args.outtype], fname_out, args.bigendian)
print("Set model parameters")
model_instance.set_gguf_parameters()
print("Set model tokenizer")
model_instance.set_vocab()
if args.vocab_only:
print(f"Exporting model vocab to '{fname_out}'")
model_instance.write_vocab()
else:
print(f"Exporting model to '{fname_out}'")
model_instance.write()
print(f"Model successfully exported to '{fname_out}'")
if __name__ == '__main__':
main()