Files
llama.cpp/convert_lora_to_gguf.py
Gabe Goodhart fd621880f3 aLoRA Support (#15327)
* feat: Add python-side constants and conversion for adapter.lora.invocation_string

Branch: gabe-l-hart/alora-support

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

* feat: Add c++ side constants for adapter.lora.invocation_string

Branch: gabe-l-hart/alora-support

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

* feat: Parse invocation string for adapters from GGUF

Branch: gabe-l-hart/alora-support

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

* fix(python): Update conversion to alora_invocation_tokens

This is the preferred method in PEFT which is the source of ground truth

https://github.com/huggingface/peft/pull/2609/files#diff-13380145401d203d5935c5189dd09879f990b81aa63e8e3aaff8ce9110333f0e

Branch: gabe-l-hart/alora-support

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

* fix(cpp): Update to alora_invocation_tokens on c++ side

Branch: gabe-l-hart/alora-support

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

* feat: Add C APIs to get alora invocation token array from lora

Branch: gabe-l-hart/alora-support

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

* feat: Initial implementation of alora cache logic in server

This does not yet do the part to identify the invocation tokens and only
apply the lora adapter afterwards, but it does seem to produce correct
results if the invocation tokens are the beginning of the uncached input.

Branch: gabe-l-hart/alora-support

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

* feat: Identify alora invocation sequences

This currently limits to a single enabled alora per slot. Multiple aloras
with different invocation sequences would be possible, but it would require
a more complex integration of the adapter toggling and is not really a well
studied case for alora since it's unclear if one alora can reuse cache from
previous prefill computed with a different alora.

Branch: gabe-l-hart/alora-support

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

* feat: Only reuse cache for tokens before the alora invocation start

This is a bit of an edge case, but theoretically a user could try the same
query with the alora disabled (just using the base model), then retry with
the alora. The cached tokens from the first pass should be invalid.

Branch: gabe-l-hart/alora-support

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

* feat: Handle un-cached tokens that come before the alora activation

The solution is to only fill up to the token before the invocation start in
the batch if there are any tokens to be prefilled between those pulled from
cache and the invocation start. When this is detected, the alora is
temporarily disabled with a scale of 0.0, then immediately re-enabled after
it has been initialized for the internal graph. Since the batch does not
complete the prompt tokens, the remaining prompt tokens are handled in the
next task, pulling all of the non-alora tokens from cache and proceeding
with prefill for the alora tokens.

Branch: gabe-l-hart/alora-support

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

* fix: Use || instead of 'or'

Too much python 🤦

Branch: gabe-l-hart/alora-support

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

* fix: Fix off-by-one for limiting cached tokens to before alora start

This was the cause of the inconsistent results from the dummy test script
with and without the turn that runs the prompt without the adapter before
running it with the adapter.

Branch: gabe-l-hart/alora-support

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

* fix: Support backwards-compatibility for "invocation_string" in adapter_config.json

While this has been replaced in the PEFT PR in favor of
alora_invocation_tokens, the existing adapters in the ibm-granite org on HF
use "invocation_string," so this will enable backwards compatibility and
enable testing now (before PEFT PR changes have percolated everywhere).

Branch: gabe-l-hart/alora-support

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

* fix: Remove duplicate logging

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

* feat: Report alora_invocation_string and alora_invocation_tokens from /lora-adapters

Branch: gabe-l-hart/alora-support

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

---------

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
2025-09-05 17:32:39 -06:00

488 lines
20 KiB
Python
Executable File

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
from __future__ import annotations
from dataclasses import dataclass
import logging
import argparse
import os
import sys
import json
from math import prod
from pathlib import Path
from typing import TYPE_CHECKING, Any, Callable, Iterable, Iterator, Sequence, SupportsIndex, cast
from transformers import AutoConfig, AutoTokenizer
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
# reuse model definitions from convert_hf_to_gguf.py
from convert_hf_to_gguf import LazyTorchTensor, ModelBase
from gguf.constants import GGUFValueType
logger = logging.getLogger("lora-to-gguf")
@dataclass
class PartialLoraTensor:
A: Tensor | None = None
B: Tensor | None = None
# magic to support tensor shape modifications and splitting
class LoraTorchTensor:
_lora_A: Tensor # (n_rank, row_size)
_lora_B: Tensor # (col_size, n_rank)
_rank: int
def __init__(self, A: Tensor, B: Tensor):
assert len(A.shape) == len(B.shape)
assert A.shape[-2] == B.shape[-1]
if A.dtype != B.dtype:
A = A.to(torch.float32)
B = B.to(torch.float32)
self._lora_A = A
self._lora_B = B
self._rank = B.shape[-1]
def get_lora_A_B(self) -> tuple[Tensor, Tensor]:
return (self._lora_A, self._lora_B)
def __getitem__(
self,
indices: (
SupportsIndex
| slice
| tuple[SupportsIndex | slice | Tensor, ...] # TODO: add ellipsis in the type signature
),
) -> LoraTorchTensor:
shape = self.shape
if isinstance(indices, SupportsIndex):
if len(shape) > 2:
return LoraTorchTensor(self._lora_A[indices], self._lora_B[indices])
else:
raise NotImplementedError # can't return a vector
elif isinstance(indices, slice):
if len(shape) > 2:
return LoraTorchTensor(self._lora_A[indices], self._lora_B[indices])
else:
return LoraTorchTensor(self._lora_A, self._lora_B[indices])
elif isinstance(indices, tuple):
assert len(indices) > 0
if indices[-1] is Ellipsis:
return self[indices[:-1]]
# expand ellipsis
indices = tuple(
u
for v in (
(
(slice(None, None) for _ in range(len(indices) - 1))
if i is Ellipsis
else (i,)
)
for i in indices
)
for u in v
)
if len(indices) < len(shape):
indices = (*indices, *(slice(None, None) for _ in range(len(indices), len(shape))))
# TODO: make sure this is correct
indices_A = (
*(
(
j.__index__() % self._lora_A.shape[i]
if isinstance(j, SupportsIndex)
else slice(None, None)
)
for i, j in enumerate(indices[:-2])
),
slice(None, None),
indices[-1],
)
indices_B = indices[:-1]
return LoraTorchTensor(self._lora_A[indices_A], self._lora_B[indices_B])
else:
raise NotImplementedError # unknown indice type
@property
def dtype(self) -> torch.dtype:
assert self._lora_A.dtype == self._lora_B.dtype
return self._lora_A.dtype
@property
def shape(self) -> tuple[int, ...]:
assert len(self._lora_A.shape) == len(self._lora_B.shape)
return (*self._lora_B.shape[:-1], self._lora_A.shape[-1])
def size(self, dim=None):
assert dim is None
return self.shape
def reshape(self, *shape: int | tuple[int, ...]) -> LoraTorchTensor:
if isinstance(shape[0], tuple):
new_shape: tuple[int, ...] = shape[0]
else:
new_shape = cast(tuple[int, ...], shape)
orig_shape = self.shape
if len(new_shape) < 2:
raise NotImplementedError # can't become a vector
# expand -1 in the shape
if any(dim == -1 for dim in new_shape):
n_elems = prod(orig_shape)
n_new_elems = prod(dim if dim != -1 else 1 for dim in new_shape)
assert n_elems % n_new_elems == 0
new_shape = (*(dim if dim != -1 else n_elems // n_new_elems for dim in new_shape),)
if new_shape[-1] != orig_shape[-1]:
raise NotImplementedError # can't reshape the row size trivially
shape_A = (*(1 for _ in new_shape[:-2]), self._rank, orig_shape[-1])
shape_B = (*new_shape[:-1], self._rank)
return LoraTorchTensor(
self._lora_A.reshape(shape_A),
self._lora_B.reshape(shape_B),
)
def reshape_as(self, other: Tensor) -> LoraTorchTensor:
return self.reshape(*other.shape)
def view(self, *size: int) -> LoraTorchTensor:
return self.reshape(*size)
def permute(self, *dims: int) -> LoraTorchTensor:
shape = self.shape
dims = tuple(dim - len(shape) if dim >= 0 else dim for dim in dims)
if dims[-1] == -1:
# TODO: support higher dimensional A shapes bigger than 1
assert all(dim == 1 for dim in self._lora_A.shape[:-2])
return LoraTorchTensor(self._lora_A, self._lora_B.permute(*dims))
if len(shape) == 2 and dims[-1] == -2 and dims[-2] == -1:
return LoraTorchTensor(self._lora_B.permute(*dims), self._lora_A.permute(*dims))
else:
# TODO: compose the above two
raise NotImplementedError
def transpose(self, dim0: int, dim1: int) -> LoraTorchTensor:
shape = self.shape
dims = [i for i in range(len(shape))]
dims[dim0], dims[dim1] = dims[dim1], dims[dim0]
return self.permute(*dims)
def swapaxes(self, axis0: int, axis1: int) -> LoraTorchTensor:
return self.transpose(axis0, axis1)
def to(self, *args, **kwargs):
return LoraTorchTensor(self._lora_A.to(*args, **kwargs), self._lora_B.to(*args, **kwargs))
@classmethod
def __torch_function__(cls, func: Callable, types, args=(), kwargs=None):
del types # unused
if kwargs is None:
kwargs = {}
if func is torch.permute:
return type(args[0]).permute(*args, **kwargs)
elif func is torch.reshape:
return type(args[0]).reshape(*args, **kwargs)
elif func is torch.stack:
assert isinstance(args[0], Sequence)
dim = kwargs.get("dim", 0)
assert dim == 0
return LoraTorchTensor(
torch.stack([a._lora_A for a in args[0]], dim),
torch.stack([b._lora_B for b in args[0]], dim),
)
elif func is torch.cat:
assert isinstance(args[0], Sequence)
dim = kwargs.get("dim", 0)
assert dim == 0
if len(args[0][0].shape) > 2:
return LoraTorchTensor(
torch.cat([a._lora_A for a in args[0]], dim),
torch.cat([b._lora_B for b in args[0]], dim),
)
elif all(torch.equal(args[0][0]._lora_A, t._lora_A) for t in args[0][1:]):
return LoraTorchTensor(
args[0][0]._lora_A,
torch.cat([b._lora_B for b in args[0]], dim),
)
else:
raise NotImplementedError
else:
raise NotImplementedError
def get_base_tensor_name(lora_tensor_name: str) -> str:
base_name = lora_tensor_name.replace("base_model.model.", "")
base_name = base_name.replace(".lora_A.weight", ".weight")
base_name = base_name.replace(".lora_B.weight", ".weight")
# models produced by mergekit-extract-lora have token embeddings in the adapter
base_name = base_name.replace(".lora_embedding_A", ".weight")
base_name = base_name.replace(".lora_embedding_B", ".weight")
return base_name
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser(
description="Convert a Hugging Face PEFT LoRA adapter to a GGUF file")
parser.add_argument(
"--outfile", type=Path,
help="path to write to; default: based on input. {ftype} will be replaced by the outtype.",
)
parser.add_argument(
"--outtype", type=str, choices=["f32", "f16", "bf16", "q8_0", "auto"], default="f16",
help="output format - use f32 for float32, f16 for float16, bf16 for bfloat16, q8_0 for Q8_0, auto for the highest-fidelity 16-bit float type depending on the first loaded tensor type",
)
parser.add_argument(
"--bigendian", action="store_true",
help="model is executed on big endian machine",
)
parser.add_argument(
"--no-lazy", action="store_true",
help="use more RAM by computing all outputs before writing (use in case lazy evaluation is broken)",
)
parser.add_argument(
"--verbose", action="store_true",
help="increase output verbosity",
)
parser.add_argument(
"--dry-run", action="store_true",
help="only print out what will be done, without writing any new files",
)
parser.add_argument(
"--base", type=Path,
help="directory containing Hugging Face model config files (config.json, tokenizer.json) for the base model that the adapter is based on - only config is needed, actual model weights are not required. If base model is unspecified, it will be loaded from Hugging Face hub based on the adapter config",
)
parser.add_argument(
"--base-model-id", type=str,
help="the model ID of the base model, if it is not available locally or in the adapter config. If specified, it will ignore --base and load the base model config from the Hugging Face hub (Example: 'meta-llama/Llama-3.2-1B-Instruct')",
)
parser.add_argument(
"lora_path", type=Path,
help="directory containing Hugging Face PEFT LoRA config (adapter_model.json) and weights (adapter_model.safetensors or adapter_model.bin)",
)
return parser.parse_args()
def load_hparams_from_hf(hf_model_id: str) -> dict[str, Any]:
# normally, adapter does not come with base model config, we need to load it from AutoConfig
config = AutoConfig.from_pretrained(hf_model_id)
return config.to_dict()
if __name__ == '__main__':
args = parse_args()
logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
ftype_map: dict[str, gguf.LlamaFileType] = {
"f32": gguf.LlamaFileType.ALL_F32,
"f16": gguf.LlamaFileType.MOSTLY_F16,
"bf16": gguf.LlamaFileType.MOSTLY_BF16,
"q8_0": gguf.LlamaFileType.MOSTLY_Q8_0,
"auto": gguf.LlamaFileType.GUESSED,
}
ftype = ftype_map[args.outtype]
dir_base_model: Path | None = args.base
dir_lora: Path = args.lora_path
base_model_id: str | None = args.base_model_id
lora_config = dir_lora / "adapter_config.json"
input_model = dir_lora / "adapter_model.safetensors"
if args.outfile is not None:
fname_out = args.outfile
else:
# output in the same directory as the model by default
fname_out = dir_lora
if os.path.exists(input_model):
# lazy import load_file only if lora is in safetensors format.
from safetensors.torch import load_file
lora_model = load_file(input_model, device="cpu")
else:
input_model = os.path.join(dir_lora, "adapter_model.bin")
lora_model = torch.load(input_model, map_location="cpu", weights_only=True)
# load LoRA config
with open(lora_config, "r") as f:
lparams: dict[str, Any] = json.load(f)
# load base model
if base_model_id is not None:
logger.info(f"Loading base model from Hugging Face: {base_model_id}")
hparams = load_hparams_from_hf(base_model_id)
elif dir_base_model is None:
if "base_model_name_or_path" in lparams:
model_id = lparams["base_model_name_or_path"]
logger.info(f"Loading base model from Hugging Face: {model_id}")
try:
hparams = load_hparams_from_hf(model_id)
except OSError as e:
logger.error(f"Failed to load base model config: {e}")
logger.error("Please try downloading the base model and add its path to --base")
sys.exit(1)
else:
logger.error("'base_model_name_or_path' is not found in adapter_config.json")
logger.error("Base model config is required. Please download the base model and add its path to --base")
sys.exit(1)
else:
logger.info(f"Loading base model: {dir_base_model.name}")
hparams = ModelBase.load_hparams(dir_base_model, False)
with torch.inference_mode():
try:
model_class = ModelBase.from_model_architecture(hparams["architectures"][0])
except NotImplementedError:
logger.error(f"Model {hparams['architectures'][0]} is not supported")
sys.exit(1)
class LoraModel(model_class):
model_arch = model_class.model_arch
lora_alpha: float
def __init__(self, *args, dir_lora_model: Path, lora_alpha: float, **kwargs):
super().__init__(*args, **kwargs)
self.dir_model_card = dir_lora_model
self.lora_alpha = float(lora_alpha)
def set_vocab(self):
pass
def set_type(self):
self.gguf_writer.add_type(gguf.GGUFType.ADAPTER)
self.gguf_writer.add_string(gguf.Keys.Adapter.TYPE, "lora")
def set_gguf_parameters(self):
logger.debug("GGUF KV: %s = %d", gguf.Keys.Adapter.LORA_ALPHA, self.lora_alpha)
self.gguf_writer.add_float32(gguf.Keys.Adapter.LORA_ALPHA, self.lora_alpha)
alora_invocation_tokens = lparams.get("alora_invocation_tokens")
invocation_string = lparams.get("invocation_string")
if invocation_string and not alora_invocation_tokens:
logger.debug("Tokenizing invocation_string -> alora_invocation_tokens")
base_model_path_or_id = hparams.get("_name_or_path")
try:
tokenizer = AutoTokenizer.from_pretrained(base_model_path_or_id)
except ValueError:
logger.error("Unable to load tokenizer from %s", base_model_path_or_id)
raise
# NOTE: There's an off-by-one with the older aLoRAs where
# the invocation string includes the "<|start_of_turn|>"
# token, but the adapters themselves were trained to
# activate _after_ that first token, so we drop it here.
alora_invocation_tokens = tokenizer(invocation_string)["input_ids"][1:]
if alora_invocation_tokens:
logger.debug("GGUF KV: %s = %s", gguf.Keys.Adapter.ALORA_INVOCATION_TOKENS, alora_invocation_tokens)
self.gguf_writer.add_key_value(
gguf.Keys.Adapter.ALORA_INVOCATION_TOKENS,
alora_invocation_tokens,
GGUFValueType.ARRAY,
GGUFValueType.UINT32,
)
def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
# Never add extra tensors (e.g. rope_freqs) for LoRA adapters
return ()
def get_tensors(self) -> Iterator[tuple[str, Tensor]]:
tensor_map: dict[str, PartialLoraTensor] = {}
for name, tensor in lora_model.items():
if self.lazy:
tensor = LazyTorchTensor.from_eager(tensor)
base_name = get_base_tensor_name(name)
# note: mergekit-extract-lora also adds token embeddings to the adapter
is_lora_a = ".lora_A.weight" in name or ".lora_embedding_A" in name
is_lora_b = ".lora_B.weight" in name or ".lora_embedding_B" in name
if not is_lora_a and not is_lora_b:
if ".base_layer.weight" in name:
continue
# mergekit-extract-lora add these layernorm to the adapter, we need to keep them
if "_layernorm" in name or ".norm" in name:
yield (base_name, tensor)
continue
logger.error(f"Unexpected name '{name}': Not a lora_A or lora_B tensor")
if ".embed_tokens.weight" in name or ".lm_head.weight" in name:
logger.error("Embeddings is present in the adapter. This can be due to new tokens added during fine tuning")
logger.error("Please refer to https://github.com/ggml-org/llama.cpp/pull/9948")
sys.exit(1)
if base_name in tensor_map:
if is_lora_a:
tensor_map[base_name].A = tensor
else:
tensor_map[base_name].B = tensor
else:
if is_lora_a:
tensor_map[base_name] = PartialLoraTensor(A=tensor)
else:
tensor_map[base_name] = PartialLoraTensor(B=tensor)
for name, tensor in tensor_map.items():
assert tensor.A is not None
assert tensor.B is not None
yield (name, cast(torch.Tensor, LoraTorchTensor(tensor.A, tensor.B)))
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
dest = list(super().modify_tensors(data_torch, name, bid))
# some archs may have the same tensor for lm_head and output (tie word embeddings)
# in this case, adapters targeting lm_head will fail when using llama-export-lora
# therefore, we ignore them for now
# see: https://github.com/ggml-org/llama.cpp/issues/9065
if name == "lm_head.weight" and len(dest) == 0:
raise ValueError("lm_head is present in adapter, but is ignored in base model")
for dest_name, dest_data in dest:
# mergekit-extract-lora add these layernorm to the adapter
if "_norm" in dest_name:
assert dest_data.dim() == 1
yield (dest_name, dest_data)
continue
# otherwise, we must get the lora_A and lora_B tensors
assert isinstance(dest_data, LoraTorchTensor)
lora_a, lora_b = dest_data.get_lora_A_B()
# note: mergekit-extract-lora flip and transpose A and B
# here we only need to transpose token_embd.lora_a, see llm_build_inp_embd()
if "token_embd.weight" in dest_name:
lora_a = lora_a.T
yield (dest_name + ".lora_a", lora_a)
yield (dest_name + ".lora_b", lora_b)
alpha: float = lparams["lora_alpha"]
model_instance = LoraModel(
dir_base_model,
ftype,
fname_out,
is_big_endian=args.bigendian,
use_temp_file=False,
eager=args.no_lazy,
dry_run=args.dry_run,
dir_lora_model=dir_lora,
lora_alpha=alpha,
hparams=hparams,
)
logger.info("Exporting model...")
model_instance.write()
logger.info(f"Model successfully exported to {model_instance.fname_out}")