Skip to content

Override Transformers defaults by GGUF defaults #42770

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
SunMarc merged 3 commits into from
Dec 10, 2025
Merged

Override Transformers defaults by GGUF defaults #42770

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
68d2175
Override Transformers defaults by GGUF defaults
a4lg Dec 10, 2025
21012b0
Apply suggestions from code review
a4lg Dec 10, 2025
ca3b023
Merge branch 'main' into gguf-config-override-defaults
MekkCyber Dec 10, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions src/transformers/integrations/__init__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -54,6 +54,7 @@
"finegrained_fp8": ["FP8Linear", "replace_with_fp8_linear"],
"fsdp": ["is_fsdp_enabled", "is_fsdp_managed_module"],
"ggml": [
"GGUF_CONFIG_DEFAULTS_MAPPING",
"GGUF_CONFIG_MAPPING",
"GGUF_TOKENIZER_MAPPING",
"_gguf_parse_value",
Expand Down Expand Up @@ -201,6 +202,7 @@
from .finegrained_fp8 import FP8Linear, replace_with_fp8_linear
from .fsdp import is_fsdp_enabled, is_fsdp_managed_module
from .ggml import (
GGUF_CONFIG_DEFAULTS_MAPPING,
GGUF_CONFIG_MAPPING,
GGUF_TOKENIZER_MAPPING,
_gguf_parse_value,
Expand Down
10 changes: 10 additions & 0 deletions src/transformers/integrations/ggml.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -313,6 +313,16 @@
},
}

# We only need to set here the parameters that default to different values between transformers and llamacpp.
GGUF_CONFIG_DEFAULTS_MAPPING = {
"qwen3_moe": {
# NOTE: Qwen3MoeConfig defaults to false but llama.cpp needs this to be true.
# See: https://github.com/ggml-org/llama.cpp/blob/17f7f4baad8b3a716ee139da7bb56ae984e8c0fa/src/models/qwen3moe.cpp#L85-L96
# (the parameter right after LLM_FFN_SILU corresponds to norm_topk_prob)
"norm_topk_prob": True,
},
}
Comment thread
a4lg marked this conversation as resolved.
Resolved
Comment thread
a4lg marked this conversation as resolved.
Resolved


def _gguf_parse_value(_value, data_type):
if not isinstance(data_type, list):
Expand Down
8 changes: 8 additions & 0 deletions src/transformers/modeling_gguf_pytorch_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,6 +20,7 @@
from tqdm.auto import tqdm

from .integrations import (
GGUF_CONFIG_DEFAULTS_MAPPING,
GGUF_CONFIG_MAPPING,
GGUF_TOKENIZER_MAPPING,
_gguf_parse_value,
Expand Down Expand Up @@ -437,6 +438,13 @@ def load_gguf_checkpoint(gguf_checkpoint_path, return_tensors=False, model_to_lo
all("output.weight" != tensor.name for tensor in reader.tensors) or architecture in exceptions
)

# Set GGUF-specific default values
config_defaults = GGUF_CONFIG_DEFAULTS_MAPPING.get(
updated_architecture, GGUF_CONFIG_DEFAULTS_MAPPING.get(architecture) or {}
)
for key, value in config_defaults.items():
parsed_parameters["config"].setdefault(key, value)

# List all key-value pairs in a columnized format
for gguf_key, field in reader.fields.items():
gguf_key = gguf_key.replace(architecture, updated_architecture)
Expand Down