CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-14 11:07:10 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'master' into compilade/mamba2

Browse Source
This commit is contained in:
Francis Couture-Harpin
2025-06-10 19:22:15 -04:00
parent e94f3932f2 dad5c44398
commit 9864bfcd01
581 changed files with 159374 additions and 37809 deletions
Download Patch File Download Diff File Expand all files Collapse all files

141
gguf-py/gguf/constants.py
View File

@@ -178,6 +178,9 @@ class Keys:
EMBEDDING_LENGTH = "{arch}.convnext.embedding_length"
BLOCK_COUNT = "{arch}.convnext.block_count"
class Classifier:
OUTPUT_LABELS = "{arch}.classifier.output_labels"
class Tokenizer:
MODEL = "tokenizer.ggml.model"
PRE = "tokenizer.ggml.pre"
@@ -220,10 +223,13 @@ class Keys:
TYPE = "adapter.type"
LORA_ALPHA = "adapter.lora.alpha"
class ClipVision:
class Clip:
PROJECTOR_TYPE = "clip.projector_type"
HAS_VISION_ENCODER = "clip.has_vision_encoder"
HAS_AUDIO_ENCODER = "clip.has_audio_encoder"
HAS_LLAVA_PROJECTOR = "clip.has_llava_projector"
class ClipVision:
IMAGE_SIZE = "clip.vision.image_size"
PATCH_SIZE = "clip.vision.patch_size"
EMBEDDING_LENGTH = "clip.vision.embedding_length"
@@ -235,6 +241,7 @@ class Keys:
SPATIAL_MERGE_SIZE = "clip.vision.spatial_merge_size"
USE_GELU = "clip.use_gelu"
USE_SILU = "clip.use_silu"
N_WA_PATTERN = "clip.vision.n_wa_pattern" # used by qwen2.5vl
class Attention:
HEAD_COUNT = "clip.vision.attention.head_count"
@@ -243,19 +250,33 @@ class Keys:
class Projector:
SCALE_FACTOR = "clip.vision.projector.scale_factor"
class ClipAudio:
NUM_MEL_BINS = "clip.audio.num_mel_bins"
EMBEDDING_LENGTH = "clip.audio.embedding_length"
FEED_FORWARD_LENGTH = "clip.audio.feed_forward_length"
PROJECTION_DIM = "clip.audio.projection_dim"
BLOCK_COUNT = "clip.audio.block_count"
class Attention:
HEAD_COUNT = "clip.audio.attention.head_count"
LAYERNORM_EPS = "clip.audio.attention.layer_norm_epsilon"
class Projector:
STACK_FACTOR = "clip.audio.projector.stack_factor"
#
# recommended mapping of model tensor names for storage in gguf
#
class GGUFType:
MODEL = "model"
ADAPTER = "adapter"
CLIP_VISION = "clip-vision"
MODEL = "model"
ADAPTER = "adapter"
MMPROJ = "mmproj" # dummy, unused for now
class MODEL_ARCH(IntEnum):
CLIP_VISION = auto() # dummy arch for clip.cpp
MMPROJ = auto() # dummy arch for clip.cpp
LLAMA = auto()
LLAMA4 = auto()
DECI = auto()
@@ -484,15 +505,20 @@ class MODEL_TENSOR(IntEnum):
V_ENC_EMBD_CLS = auto()
V_ENC_EMBD_PATCH = auto()
V_ENC_EMBD_POS = auto()
V_ENC_ATTN_Q = auto()
V_ENC_ATTN_K = auto()
V_ENC_ATTN_V = auto()
V_ENC_INPUT_NORM = auto()
V_ENC_OUTPUT = auto()
V_ENC_OUTPUT_NORM = auto()
V_ENC_ATTN_Q = auto()
V_ENC_ATTN_Q_NORM = auto()
V_ENC_ATTN_K = auto()
V_ENC_ATTN_K_NORM = auto()
V_ENC_ATTN_V = auto()
V_ENC_ATTN_O = auto()
V_ENC_ATTN_O_NORM = auto()
V_ENC_POST_ATTN_NORM = auto()
V_ENC_FFN_UP = auto()
V_ENC_FFN_GATE = auto()
V_ENC_FFN_DOWN = auto()
V_LAYER_SCALE_1 = auto()
V_LAYER_SCALE_2 = auto()
V_PRE_NORM = auto()
V_POST_NORM = auto()
V_MM_INP_NORM = auto()
@@ -511,10 +537,28 @@ class MODEL_TENSOR(IntEnum):
V_RESMPL_QUERY = auto() # minicpmv
V_TOK_EMBD_IMG_BREAK = auto() # pixtral
V_MM_PATCH_MERGER = auto() # mistral small 3.1
# audio (mtmd)
A_ENC_EMBD_POS = auto()
A_ENC_CONV1D = auto()
A_PRE_NORM = auto()
A_POST_NORM = auto()
A_ENC_ATTN_Q = auto()
A_ENC_ATTN_K = auto()
A_ENC_ATTN_V = auto()
A_ENC_INPUT_NORM = auto()
A_ENC_OUTPUT = auto()
A_ENC_OUTPUT_NORM = auto()
A_ENC_FFN_UP = auto()
A_ENC_FFN_GATE = auto()
A_ENC_FFN_DOWN = auto()
A_MMPROJ = auto()
A_MMPROJ_FC = auto()
A_MM_NORM_PRE = auto()
A_MM_NORM_MID = auto()
MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
MODEL_ARCH.CLIP_VISION: "clip", # dummy arch for clip.cpp
MODEL_ARCH.MMPROJ: "clip", # dummy arch for clip.cpp
MODEL_ARCH.LLAMA: "llama",
MODEL_ARCH.LLAMA4: "llama4",
MODEL_ARCH.DECI: "deci",
@@ -744,14 +788,19 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
MODEL_TENSOR.V_ENC_EMBD_PATCH: "v.patch_embd",
MODEL_TENSOR.V_ENC_EMBD_POS: "v.position_embd",
MODEL_TENSOR.V_ENC_ATTN_Q: "v.blk.{bid}.attn_q",
MODEL_TENSOR.V_ENC_ATTN_Q_NORM: "v.blk.{bid}.attn_q_norm",
MODEL_TENSOR.V_ENC_ATTN_K: "v.blk.{bid}.attn_k",
MODEL_TENSOR.V_ENC_ATTN_K_NORM: "v.blk.{bid}.attn_k_norm",
MODEL_TENSOR.V_ENC_ATTN_V: "v.blk.{bid}.attn_v",
MODEL_TENSOR.V_ENC_INPUT_NORM: "v.blk.{bid}.ln1",
MODEL_TENSOR.V_ENC_OUTPUT: "v.blk.{bid}.attn_out",
MODEL_TENSOR.V_ENC_OUTPUT_NORM: "v.blk.{bid}.ln2",
MODEL_TENSOR.V_ENC_ATTN_O: "v.blk.{bid}.attn_out",
MODEL_TENSOR.V_ENC_ATTN_O_NORM: "v.blk.{bid}.attn_out_norm",
MODEL_TENSOR.V_ENC_POST_ATTN_NORM: "v.blk.{bid}.ln2",
MODEL_TENSOR.V_ENC_FFN_UP: "v.blk.{bid}.ffn_up",
MODEL_TENSOR.V_ENC_FFN_GATE: "v.blk.{bid}.ffn_gate",
MODEL_TENSOR.V_ENC_FFN_DOWN: "v.blk.{bid}.ffn_down",
MODEL_TENSOR.V_LAYER_SCALE_1: "v.blk.{bid}.ls1",
MODEL_TENSOR.V_LAYER_SCALE_2: "v.blk.{bid}.ls2",
MODEL_TENSOR.V_PRE_NORM: "v.pre_ln",
MODEL_TENSOR.V_POST_NORM: "v.post_ln",
MODEL_TENSOR.V_MM_INP_PROJ: "mm.input_projection",
@@ -770,10 +819,28 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
MODEL_TENSOR.V_RESMPL_QUERY: "resampler.query",
MODEL_TENSOR.V_TOK_EMBD_IMG_BREAK: "v.token_embd.img_break", # pixtral
MODEL_TENSOR.V_MM_PATCH_MERGER: "mm.patch_merger", # mistral small 3.1
# audio (mtmd)
MODEL_TENSOR.A_ENC_EMBD_POS: "a.position_embd",
MODEL_TENSOR.A_ENC_CONV1D: "a.conv1d.{bid}",
MODEL_TENSOR.A_PRE_NORM: "a.pre_ln",
MODEL_TENSOR.A_POST_NORM: "a.post_ln",
MODEL_TENSOR.A_ENC_ATTN_Q: "a.blk.{bid}.attn_q",
MODEL_TENSOR.A_ENC_ATTN_K: "a.blk.{bid}.attn_k",
MODEL_TENSOR.A_ENC_ATTN_V: "a.blk.{bid}.attn_v",
MODEL_TENSOR.A_ENC_INPUT_NORM: "a.blk.{bid}.ln1",
MODEL_TENSOR.A_ENC_OUTPUT: "a.blk.{bid}.attn_out",
MODEL_TENSOR.A_ENC_OUTPUT_NORM: "a.blk.{bid}.ln2",
MODEL_TENSOR.A_ENC_FFN_UP: "a.blk.{bid}.ffn_up",
MODEL_TENSOR.A_ENC_FFN_GATE: "a.blk.{bid}.ffn_gate",
MODEL_TENSOR.A_ENC_FFN_DOWN: "a.blk.{bid}.ffn_down",
MODEL_TENSOR.A_MMPROJ: "mm.a.mlp.{bid}",
MODEL_TENSOR.A_MMPROJ_FC: "mm.a.fc",
MODEL_TENSOR.A_MM_NORM_PRE: "mm.a.norm_pre",
MODEL_TENSOR.A_MM_NORM_MID: "mm.a.norm_mid",
}
MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_ARCH.CLIP_VISION: [
MODEL_ARCH.MMPROJ: [
MODEL_TENSOR.V_MMPROJ,
MODEL_TENSOR.V_MMPROJ_FC,
MODEL_TENSOR.V_MMPROJ_MLP,
@@ -781,15 +848,20 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.V_ENC_EMBD_CLS,
MODEL_TENSOR.V_ENC_EMBD_PATCH,
MODEL_TENSOR.V_ENC_EMBD_POS,
MODEL_TENSOR.V_ENC_ATTN_Q,
MODEL_TENSOR.V_ENC_ATTN_K,
MODEL_TENSOR.V_ENC_ATTN_V,
MODEL_TENSOR.V_ENC_INPUT_NORM,
MODEL_TENSOR.V_ENC_OUTPUT,
MODEL_TENSOR.V_ENC_OUTPUT_NORM,
MODEL_TENSOR.V_ENC_ATTN_Q,
MODEL_TENSOR.V_ENC_ATTN_Q_NORM,
MODEL_TENSOR.V_ENC_ATTN_K,
MODEL_TENSOR.V_ENC_ATTN_K_NORM,
MODEL_TENSOR.V_ENC_ATTN_V,
MODEL_TENSOR.V_ENC_ATTN_O,
MODEL_TENSOR.V_ENC_ATTN_O_NORM,
MODEL_TENSOR.V_ENC_POST_ATTN_NORM,
MODEL_TENSOR.V_ENC_FFN_UP,
MODEL_TENSOR.V_ENC_FFN_GATE,
MODEL_TENSOR.V_ENC_FFN_DOWN,
MODEL_TENSOR.V_LAYER_SCALE_1,
MODEL_TENSOR.V_LAYER_SCALE_2,
MODEL_TENSOR.V_PRE_NORM,
MODEL_TENSOR.V_POST_NORM,
MODEL_TENSOR.V_MM_INP_PROJ,
@@ -808,6 +880,24 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.V_RESMPL_QUERY,
MODEL_TENSOR.V_TOK_EMBD_IMG_BREAK,
MODEL_TENSOR.V_MM_PATCH_MERGER,
# audio
MODEL_TENSOR.A_ENC_EMBD_POS,
MODEL_TENSOR.A_ENC_CONV1D,
MODEL_TENSOR.A_PRE_NORM,
MODEL_TENSOR.A_POST_NORM,
MODEL_TENSOR.A_ENC_ATTN_Q,
MODEL_TENSOR.A_ENC_ATTN_K,
MODEL_TENSOR.A_ENC_ATTN_V,
MODEL_TENSOR.A_ENC_INPUT_NORM,
MODEL_TENSOR.A_ENC_OUTPUT,
MODEL_TENSOR.A_ENC_OUTPUT_NORM,
MODEL_TENSOR.A_ENC_FFN_UP,
MODEL_TENSOR.A_ENC_FFN_GATE,
MODEL_TENSOR.A_ENC_FFN_DOWN,
MODEL_TENSOR.A_MMPROJ,
MODEL_TENSOR.A_MMPROJ_FC,
MODEL_TENSOR.A_MM_NORM_PRE,
MODEL_TENSOR.A_MM_NORM_MID,
],
MODEL_ARCH.LLAMA: [
MODEL_TENSOR.TOKEN_EMBD,
@@ -951,6 +1041,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.POS_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.ATTN_OUT_NORM,
MODEL_TENSOR.ATTN_QKV,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_V,
@@ -1910,6 +2001,9 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_GATE_EXP,
MODEL_TENSOR.FFN_DOWN_EXP,
MODEL_TENSOR.FFN_UP_EXP,
MODEL_TENSOR.FFN_GATE_SHEXP,
MODEL_TENSOR.FFN_UP_SHEXP,
MODEL_TENSOR.FFN_DOWN_SHEXP,
],
MODEL_ARCH.CHAMELEON: [
MODEL_TENSOR.TOKEN_EMBD,
@@ -2050,6 +2144,8 @@ class PoolingType(IntEnum):
NONE = 0
MEAN = 1
CLS = 2
LAST = 3
RANK = 4
class GGMLQuantizationType(IntEnum):
@@ -2180,6 +2276,13 @@ class VisionProjectorType:
GEMMA3 = "gemma3"
IDEFICS3 = "idefics3"
PIXTRAL = "pixtral"
LLAMA4 = "llama4"
QWEN2VL = "qwen2vl_merger"
QWEN25VL = "qwen2.5vl_merger"
ULTRAVOX = "ultravox"
INTERNVL = "internvl"
QWEN2A = "qwen2a" # audio
QWEN25O = "qwen2.5o" # omni
# Items here are (block size, type size)

2
gguf-py/gguf/gguf_reader.py
View File

@@ -251,7 +251,7 @@ class GGUFReader:
offs += curr_size
return offs - orig_offs, aparts, data_idxs, types
# We can't deal with this one.
raise ValueError('Unknown/unhandled field type {gtype}')
raise ValueError(f'Unknown/unhandled field type {gtype}')
def _get_tensor_info_field(self, orig_offs: int) -> ReaderField:
offs = orig_offs

62
gguf-py/gguf/gguf_writer.py
View File

@@ -49,6 +49,7 @@ class TensorInfo:
class GGUFValue:
value: Any
type: GGUFValueType
sub_type: GGUFValueType | None = None
class WriterState(Enum):
@@ -238,7 +239,7 @@ class GGUFWriter:
for key, val in kv_data.items():
kv_bytes += self._pack_val(key, GGUFValueType.STRING, add_vtype=False)
kv_bytes += self._pack_val(val.value, val.type, add_vtype=True)
kv_bytes += self._pack_val(val.value, val.type, add_vtype=True, sub_type=val.sub_type)
fout.write(kv_bytes)
@@ -268,11 +269,11 @@ class GGUFWriter:
fout.flush()
self.state = WriterState.TI_DATA
def add_key_value(self, key: str, val: Any, vtype: GGUFValueType) -> None:
def add_key_value(self, key: str, val: Any, vtype: GGUFValueType, sub_type: GGUFValueType | None = None) -> None:
if any(key in kv_data for kv_data in self.kv_data):
raise ValueError(f'Duplicated key name {key!r}')
self.kv_data[0][key] = GGUFValue(value=val, type=vtype)
self.kv_data[0][key] = GGUFValue(value=val, type=vtype, sub_type=sub_type)
def add_uint8(self, key: str, val: int) -> None:
self.add_key_value(key,val, GGUFValueType.UINT8)
@@ -899,7 +900,7 @@ class GGUFWriter:
def add_remove_extra_whitespaces(self, value: bool) -> None:
self.add_bool(Keys.Tokenizer.REMOVE_EXTRA_WS, value)
def add_precompiled_charsmap(self, charsmap: Sequence[bytes]) -> None:
def add_precompiled_charsmap(self, charsmap: bytes) -> None:
self.add_array(Keys.Tokenizer.PRECOMPILED_CHARSMAP, charsmap)
def add_chat_template(self, value: str | Sequence[Mapping[str, str]]) -> None:
@@ -937,14 +938,23 @@ class GGUFWriter:
def add_eom_token_id(self, id: int) -> None:
self.add_uint32(Keys.Tokenizer.EOM_ID, id)
def add_classifier_output_labels(self, labels: Sequence[str]) -> None:
self.add_array(Keys.Classifier.OUTPUT_LABELS.format(arch=self.arch), labels)
# for vision models
def add_clip_has_vision_encoder(self, value: bool) -> None:
self.add_bool(Keys.Clip.HAS_VISION_ENCODER, value)
def add_clip_has_audio_encoder(self, value: bool) -> None:
self.add_bool(Keys.Clip.HAS_AUDIO_ENCODER, value)
def add_clip_projector_type(self, value: str) -> None:
self.add_string(Keys.Clip.PROJECTOR_TYPE, value)
def add_vision_projection_dim(self, value: int) -> None:
self.add_uint32(Keys.ClipVision.PROJECTION_DIM, value)
def add_vision_has_vision_encoder(self, value: bool) -> None:
self.add_bool(Keys.ClipVision.HAS_VISION_ENCODER, value)
def add_vision_patch_size(self, value: int) -> None:
self.add_uint32(Keys.ClipVision.PATCH_SIZE, value)
@@ -960,9 +970,6 @@ class GGUFWriter:
def add_vision_head_count(self, value: int) -> None:
self.add_uint32(Keys.ClipVision.Attention.HEAD_COUNT, value)
def add_vision_projector_type(self, value: str) -> None:
self.add_string(Keys.ClipVision.PROJECTOR_TYPE, value)
def add_vision_attention_layernorm_eps(self, value: float) -> None:
self.add_float32(Keys.ClipVision.Attention.LAYERNORM_EPS, value)
@@ -987,13 +994,42 @@ class GGUFWriter:
def add_vision_projector_scale_factor(self, value: int) -> None:
self.add_uint32(Keys.ClipVision.Projector.SCALE_FACTOR, value)
def add_vision_n_wa_pattern(self, value: int) -> None:
self.add_uint32(Keys.ClipVision.N_WA_PATTERN, value)
# audio models
def add_audio_projection_dim(self, value: int) -> None:
self.add_uint32(Keys.ClipAudio.PROJECTION_DIM, value)
def add_audio_embedding_length(self, value: int) -> None:
self.add_uint32(Keys.ClipAudio.EMBEDDING_LENGTH, value)
def add_audio_feed_forward_length(self, value: int) -> None:
self.add_uint32(Keys.ClipAudio.FEED_FORWARD_LENGTH, value)
def add_audio_block_count(self, value: int) -> None:
self.add_uint32(Keys.ClipAudio.BLOCK_COUNT, value)
def add_audio_head_count(self, value: int) -> None:
self.add_uint32(Keys.ClipAudio.Attention.HEAD_COUNT, value)
def add_audio_attention_layernorm_eps(self, value: float) -> None:
self.add_float32(Keys.ClipAudio.Attention.LAYERNORM_EPS, value)
def add_audio_num_mel_bins(self, value: int) -> None:
self.add_uint32(Keys.ClipAudio.NUM_MEL_BINS, value)
def add_audio_stack_factor(self, value: int) -> None:
self.add_uint32(Keys.ClipAudio.Projector.STACK_FACTOR, value)
def _pack(self, fmt: str, value: Any, skip_pack_prefix: bool = False) -> bytes:
pack_prefix = ''
if not skip_pack_prefix:
pack_prefix = '<' if self.endianess == GGUFEndian.LITTLE else '>'
return struct.pack(f'{pack_prefix}{fmt}', value)
def _pack_val(self, val: Any, vtype: GGUFValueType, add_vtype: bool) -> bytes:
def _pack_val(self, val: Any, vtype: GGUFValueType, add_vtype: bool, sub_type: GGUFValueType | None = None) -> bytes:
kv_data = bytearray()
if add_vtype:
@@ -1014,7 +1050,9 @@ class GGUFWriter:
if len(val) == 0:
raise ValueError("Invalid GGUF metadata array. Empty array")
if isinstance(val, bytes):
if sub_type is not None:
ltype = sub_type
elif isinstance(val, bytes):
ltype = GGUFValueType.UINT8
else:
ltype = GGUFValueType.get_type(val[0])

7
gguf-py/gguf/scripts/__init__.py
View File

@@ -1,7 +0,0 @@
# pyright: reportUnusedImport=false
from .gguf_convert_endian import main as gguf_convert_endian_entrypoint
from .gguf_dump import main as gguf_dump_entrypoint
from .gguf_set_metadata import main as gguf_set_metadata_entrypoint
from .gguf_new_metadata import main as gguf_new_metadata_entrypoint
from .gguf_editor_gui import main as gguf_editor_gui_entrypoint

31
gguf-py/gguf/scripts/gguf_editor_gui.py
View File

@@ -823,6 +823,7 @@ class GGUFEditorWindow(QMainWindow):
self.modified = False
self.metadata_changes = {} # Store changes to apply when saving
self.metadata_to_remove = set() # Store keys to remove when saving
self.on_metadata_changed_is_connected = False
self.setup_ui()
@@ -941,9 +942,11 @@ class GGUFEditorWindow(QMainWindow):
return
# Disconnect to prevent triggering during loading
with warnings.catch_warnings():
warnings.filterwarnings('ignore')
self.metadata_table.itemChanged.disconnect(self.on_metadata_changed)
if self.on_metadata_changed_is_connected:
with warnings.catch_warnings():
warnings.filterwarnings('ignore')
self.metadata_table.itemChanged.disconnect(self.on_metadata_changed)
self.on_metadata_changed_is_connected = False
for i, (key, field) in enumerate(self.reader.fields.items()):
self.metadata_table.insertRow(i)
@@ -1021,6 +1024,7 @@ class GGUFEditorWindow(QMainWindow):
# Reconnect after loading
self.metadata_table.itemChanged.connect(self.on_metadata_changed)
self.on_metadata_changed_is_connected = True
def extract_array_values(self, field: ReaderField) -> list:
"""Extract all values from an array field."""
@@ -1517,19 +1521,21 @@ class GGUFEditorWindow(QMainWindow):
continue
# Apply changes if any
sub_type = None
if field.name in self.metadata_changes:
value_type, value = self.metadata_changes[field.name]
if value_type == GGUFValueType.ARRAY:
# Handle array values
element_type, array_values = value
writer.add_array(field.name, array_values)
else:
writer.add_key_value(field.name, value, value_type)
sub_type, value = value
else:
# Copy original value
value = field.contents()
if value is not None and field.types:
writer.add_key_value(field.name, value, field.types[0])
value_type = field.types[0]
if value_type == GGUFValueType.ARRAY:
sub_type = field.types[-1]
if value is not None:
writer.add_key_value(field.name, value, value_type, sub_type=sub_type)
# Add new metadata
for key, (value_type, value) in self.metadata_changes.items():
@@ -1537,7 +1543,12 @@ class GGUFEditorWindow(QMainWindow):
if self.reader.get_field(key) is not None:
continue
writer.add_key_value(key, value, value_type)
sub_type = None
if value_type == GGUFValueType.ARRAY:
# Handle array values
sub_type, value = value
writer.add_key_value(key, value, value_type, sub_type=sub_type)
# Add tensors (including data)
for tensor in self.reader.tensors:

7
gguf-py/gguf/scripts/gguf_new_metadata.py
View File

@@ -24,6 +24,7 @@ class MetadataDetails(NamedTuple):
type: gguf.GGUFValueType
value: Any
description: str = ''
sub_type: gguf.GGUFValueType | None = None
def get_field_data(reader: gguf.GGUFReader, key: str) -> Any:
@@ -57,7 +58,9 @@ def copy_with_new_metadata(reader: gguf.GGUFReader, writer: gguf.GGUFWriter, new
logger.debug(f'Removing {field.name}')
continue
old_val = MetadataDetails(field.types[0], field.contents())
val_type = field.types[0]
sub_type = field.types[-1] if val_type == gguf.GGUFValueType.ARRAY else None
old_val = MetadataDetails(val_type, field.contents(), sub_type=sub_type)
val = new_metadata.get(field.name, old_val)
if field.name in new_metadata:
@@ -67,7 +70,7 @@ def copy_with_new_metadata(reader: gguf.GGUFReader, writer: gguf.GGUFWriter, new
logger.debug(f'Copying {field.name}')
if val.value is not None:
writer.add_key_value(field.name, val.value, val.type)
writer.add_key_value(field.name, val.value, val.type, sub_type=sub_type if val.sub_type is None else val.sub_type)
if gguf.Keys.Tokenizer.CHAT_TEMPLATE in new_metadata:
logger.debug('Adding chat template(s)')

203
gguf-py/gguf/tensor_mapping.py
View File

@@ -68,7 +68,7 @@ class TensorNameMap:
"output_layer", # chatglm
"head", # rwkv
"head.out", # wavtokenizer
"language_model.lm_head", # llama4
"lm_head", # llama4
),
# Output norm
@@ -91,7 +91,7 @@ class TensorNameMap:
"rwkv.ln_out", # rwkv6
"model.ln_out", # rwkv7
"backbone.final_layer_norm", # wavtokenizer
"language_model.model.norm", # llama4
"model.norm", # llama4
),
# Rope frequencies
@@ -133,7 +133,7 @@ class TensorNameMap:
"transformer.layers.{bid}.attn_norm", # openelm
"rwkv.blocks.{bid}.ln1", # rwkv6
"model.layers.{bid}.ln1", # rwkv7
"language_model.model.layers.{bid}.input_layernorm", # llama4
"model.layers.{bid}.input_layernorm", # llama4
),
# Attention norm 2
@@ -157,6 +157,7 @@ class TensorNameMap:
"h.{bid}.attn.c_attn", # gpt2
"transformer.h.{bid}.mixer.Wqkv", # phi2
"encoder.layers.{bid}.attn.Wqkv", # nomic-bert
"encoder.layers.{bid}.mixer.Wqkv", # jina
"model.layers.{bid}.self_attn.qkv_proj", # phi3
"encoder.layers.{bid}.self_attention.query_key_value", # chatglm
"transformer.layers.{bid}.attn.qkv_proj", # openelm
@@ -168,12 +169,13 @@ class TensorNameMap:
"model.layers.{bid}.self_attn.q_proj_no_perm", # llama-custom
"layers.{bid}.attention.wq", # llama-pth
"encoder.layer.{bid}.attention.self.query", # bert
"transformer.layer.{bid}.attention.q_lin", # distillbert
"transformer.h.{bid}.attn.q_proj", # gpt-j
"model.layers.layers.{bid}.self_attn.q_proj", # plamo
"model.layers.{bid}.attention.wq", # internlm2
"transformer.decoder_layer.{bid}.multi_head_attention.query",# Grok
"transformer.h.{bid}.attn.attention.q_proj", # exaone
"language_model.model.layers.{bid}.self_attn.q_proj", # llama4
"model.layers.{bid}.self_attn.q_proj", # llama4
),
# Attention key
@@ -182,13 +184,14 @@ class TensorNameMap:
"model.layers.{bid}.self_attn.k_proj_no_perm", # llama-custom
"layers.{bid}.attention.wk", # llama-pth
"encoder.layer.{bid}.attention.self.key", # bert
"transformer.layer.{bid}.attention.k_lin", # distillbert
"transformer.h.{bid}.attn.k_proj", # gpt-j
"transformer.h.{bid}.attn.k", # refact
"model.layers.layers.{bid}.self_attn.k_proj", # plamo
"model.layers.{bid}.attention.wk", # internlm2
"transformer.decoder_layer.{bid}.multi_head_attention.key",# Grok
"transformer.h.{bid}.attn.attention.k_proj", # exaone
"language_model.model.layers.{bid}.self_attn.k_proj", # llama4
"model.layers.{bid}.self_attn.k_proj", # llama4
),
# Attention value
@@ -196,13 +199,14 @@ class TensorNameMap:
"model.layers.{bid}.self_attn.v_proj", # llama-hf nemotron olmoe olmo2 phimoe
"layers.{bid}.attention.wv", # llama-pth
"encoder.layer.{bid}.attention.self.value", # bert
"transformer.layer.{bid}.attention.v_lin", # distillbert
"transformer.h.{bid}.attn.v_proj", # gpt-j
"transformer.h.{bid}.attn.v", # refact
"model.layers.layers.{bid}.self_attn.v_proj", # plamo
"model.layers.{bid}.attention.wv", # internlm2
"transformer.decoder_layer.{bid}.multi_head_attention.value",# Grok
"transformer.h.{bid}.attn.attention.v_proj", # exaone
"language_model.model.layers.{bid}.self_attn.v_proj", # llama4
"model.layers.{bid}.self_attn.v_proj", # llama4
),
# Attention output
@@ -216,6 +220,7 @@ class TensorNameMap:
"model.layers.{bid}.self_attn.linear_attn", # deci
"layers.{bid}.attention.wo", # llama-pth
"encoder.layer.{bid}.attention.output.dense", # bert
"transformer.layer.{bid}.attention.out_lin", # distillbert
"transformer.h.{bid}.attn.out_proj", # gpt-j
"language_model.encoder.layers.{bid}.self_attention.dense", # persimmon
"model.layers.{bid}.self_attn.dense", # persimmon
@@ -224,17 +229,19 @@ class TensorNameMap:
"model.layers.layers.{bid}.self_attn.o_proj", # plamo
"model.layers.{bid}.attention.wo", # internlm2
"encoder.layers.{bid}.attn.out_proj", # nomic-bert
"encoder.layers.{bid}.mixer.out_proj", # jina
"transformer.decoder_layer.{bid}.multi_head_attention.linear", # Grok
"transformer.blocks.{bid}.norm_attn_norm.attn.out_proj", # dbrx
"encoder.layers.{bid}.self_attention.dense", # chatglm
"transformer.layers.{bid}.attn.out_proj", # openelm
"transformer.h.{bid}.attn.attention.out_proj", # exaone
"language_model.model.layers.{bid}.self_attn.o_proj", # llama4
"model.layers.{bid}.self_attn.o_proj", # llama4
),
# Attention output norm
MODEL_TENSOR.ATTN_OUT_NORM: (
"encoder.layer.{bid}.attention.output.LayerNorm", # bert
"transformer.layer.{bid}.sa_layer_norm", # distillbert
"encoder.layers.{bid}.norm1", # nomic-bert
"transformer.decoder_layer.{bid}.rms_norm_1", # Grok
"transformer.blocks.{bid}.norm_attn_norm.norm_2", # dbrx
@@ -268,7 +275,7 @@ class TensorNameMap:
"transformer.decoder_layer.{bid}.rms_norm_2", # Grok
"encoder.layers.{bid}.post_attention_layernorm", # chatglm
"transformer.layers.{bid}.ffn_norm", # openelm
"language_model.model.layers.{bid}.post_attention_layernorm", # llama4
"model.layers.{bid}.post_attention_layernorm", # llama4
),
# Post feed-forward norm
@@ -289,7 +296,7 @@ class TensorNameMap:
"transformer.decoder_layer.{bid}.router", # Grok
"transformer.blocks.{bid}.ffn.router.layer", # dbrx
"model.layers.{bid}.block_sparse_moe.router.layer", # granitemoe
"language_model.model.layers.{bid}.feed_forward.router", # llama4
"model.layers.{bid}.feed_forward.router", # llama4
"encoder.layers.{bid}.mlp.router.layer", # nomic-bert-moe
),
@@ -311,6 +318,7 @@ class TensorNameMap:
"model.layers.{bid}.mlp.up_proj", # llama-hf refact nemotron olmo2
"layers.{bid}.feed_forward.w3", # llama-pth
"encoder.layer.{bid}.intermediate.dense", # bert
"transformer.layer.{bid}.ffn.lin1", # distillbert
"transformer.h.{bid}.mlp.fc_in", # gpt-j
"transformer.h.{bid}.mlp.linear_3", # refact
"language_model.encoder.layers.{bid}.mlp.dense_h_to_4h", # persimmon
@@ -325,11 +333,13 @@ class TensorNameMap:
"encoder.layers.{bid}.mlp.fc11", # nomic-bert
"encoder.layers.{bid}.mlp.fc1", # nomic-bert-moe
"model.layers.{bid}.mlp.c_fc", # starcoder2
"encoder.layer.{bid}.mlp.gated_layers_v", # jina-bert-v2
"encoder.layer.{bid}.mlp.gated_layers_v", # jina-bert-v2 (split up/gate, no longer used)
"encoder.layer.{bid}.mlp.gated_layers", # jina-bert-v2 (GEGLU)
"encoder.layer.{bid}.mlp.up_gated_layer", # jina-v2-code (GEGLU)
"model.layers.{bid}.residual_mlp.w3", # arctic
"encoder.layers.{bid}.mlp.dense_h_to_4h", # chatglm
"transformer.h.{bid}.mlp.c_fc_1", # exaone
"language_model.model.layers.{bid}.feed_forward.up_proj", # llama4
"model.layers.{bid}.feed_forward.up_proj", # llama4
),
MODEL_TENSOR.FFN_UP_EXP: (
@@ -338,14 +348,14 @@ class TensorNameMap:
"transformer.blocks.{bid}.ffn.experts.mlp.v1", # dbrx
"model.layers.{bid}.mlp.experts.up_proj", # qwen2moe olmoe (merged)
"model.layers.{bid}.block_sparse_moe.experts.w3", # phimoe (merged)
"language_model.model.layers.{bid}.feed_forward.experts.up_proj", # llama4
"model.layers.{bid}.feed_forward.experts.up_proj", # llama4
"encoder.layers.{bid}.mlp.experts.mlp.w1", # nomic-bert-moe
),
MODEL_TENSOR.FFN_UP_SHEXP: (
"model.layers.{bid}.mlp.shared_expert.up_proj", # qwen2moe
"model.layers.{bid}.mlp.shared_experts.up_proj", # deepseek deepseek2
"language_model.model.layers.{bid}.feed_forward.shared_expert.up_proj", # llama4
"model.layers.{bid}.mlp.shared_expert.up_proj", # qwen2moe
"model.layers.{bid}.mlp.shared_experts.up_proj", # deepseek deepseek2
"model.layers.{bid}.feed_forward.shared_expert.up_proj", # llama4
),
# AWQ-activation gate
@@ -362,26 +372,26 @@ class TensorNameMap:
"model.layers.layers.{bid}.mlp.gate_proj", # plamo
"model.layers.{bid}.feed_forward.w1", # internlm2
"encoder.layers.{bid}.mlp.fc12", # nomic-bert
"encoder.layer.{bid}.mlp.gated_layers_w", # jina-bert-v2
"encoder.layer.{bid}.mlp.gated_layers_w", # jina-bert-v2 (split up/gate, no longer used)
"transformer.h.{bid}.mlp.linear_1", # refact
"model.layers.{bid}.residual_mlp.w1", # arctic
"transformer.h.{bid}.mlp.c_fc_0", # exaone
"language_model.model.layers.{bid}.feed_forward.gate_proj", # llama4
"model.layers.{bid}.feed_forward.gate_proj", # llama4
),
MODEL_TENSOR.FFN_GATE_EXP: (
"layers.{bid}.feed_forward.experts.w1", # mixtral (merged)
"transformer.decoder_layer.{bid}.moe.linear", # Grok (merged)
"transformer.blocks.{bid}.ffn.experts.mlp.w1", # dbrx
"model.layers.{bid}.mlp.experts.gate_proj", # qwen2moe olmoe (merged)
"model.layers.{bid}.block_sparse_moe.experts.w1", # phimoe (merged)
"language_model.model.layers.{bid}.feed_forward.experts.gate_proj", # llama4
"layers.{bid}.feed_forward.experts.w1", # mixtral (merged)
"transformer.decoder_layer.{bid}.moe.linear", # Grok (merged)
"transformer.blocks.{bid}.ffn.experts.mlp.w1", # dbrx
"model.layers.{bid}.mlp.experts.gate_proj", # qwen2moe olmoe (merged)
"model.layers.{bid}.block_sparse_moe.experts.w1", # phimoe (merged)
"model.layers.{bid}.feed_forward.experts.gate_proj", # llama4
),
MODEL_TENSOR.FFN_GATE_SHEXP: (
"model.layers.{bid}.mlp.shared_expert.gate_proj", # qwen2moe
"model.layers.{bid}.mlp.shared_experts.gate_proj", # deepseek deepseek2
"language_model.model.layers.{bid}.feed_forward.shared_expert.gate_proj", # llama4
"model.layers.{bid}.mlp.shared_expert.gate_proj", # qwen2moe
"model.layers.{bid}.mlp.shared_experts.gate_proj", # deepseek deepseek2
"model.layers.{bid}.feed_forward.shared_expert.gate_proj", # llama4
),
# Feed-forward down
@@ -394,6 +404,7 @@ class TensorNameMap:
"model.layers.{bid}.mlp.down_proj", # llama-hf nemotron olmo2
"layers.{bid}.feed_forward.w2", # llama-pth
"encoder.layer.{bid}.output.dense", # bert
"transformer.layer.{bid}.ffn.lin2", # distillbert
"transformer.h.{bid}.mlp.fc_out", # gpt-j
"language_model.encoder.layers.{bid}.mlp.dense_4h_to_h", # persimmon
"model.layers.{bid}.mlp.dense_4h_to_h", # persimmon
@@ -410,7 +421,7 @@ class TensorNameMap:
"encoder.layer.{bid}.mlp.down_layer", # jina-bert-v2
"encoder.layers.{bid}.mlp.dense_4h_to_h", # chatglm
"model.layers.h.{bid}.mlp.c_proj", # exaone
"language_model.model.layers.{bid}.feed_forward.down_proj", # llama4
"model.layers.{bid}.feed_forward.down_proj", # llama4
),
MODEL_TENSOR.FFN_DOWN_EXP: (
@@ -420,14 +431,15 @@ class TensorNameMap:
"model.layers.{bid}.mlp.experts.down_proj", # qwen2moe olmoe (merged)
"model.layers.{bid}.block_sparse_moe.output_linear", # granitemoe
"model.layers.{bid}.block_sparse_moe.experts.w2", # phimoe (merged)
"language_model.model.layers.{bid}.feed_forward.experts.down_proj", # llama4
"model.layers.{bid}.feed_forward.experts.down_proj", # llama4
"encoder.layers.{bid}.mlp.experts.mlp.w2", # nomic-bert-moe
),
MODEL_TENSOR.FFN_DOWN_SHEXP: (
"model.layers.{bid}.mlp.shared_expert.down_proj", # qwen2moe
"model.layers.{bid}.mlp.shared_experts.down_proj", # deepseek deepseek2
"language_model.model.layers.{bid}.feed_forward.shared_expert.down_proj", # llama4
"model.layers.{bid}.mlp.shared_expert.down_proj", # qwen2moe
"model.layers.{bid}.mlp.shared_experts.down_proj", # deepseek deepseek2
"model.layers.{bid}.feed_forward.shared_expert.down_proj", # llama4
"model.layers.{bid}.shared_mlp.output_linear", # granitemoe
),
MODEL_TENSOR.ATTN_Q_NORM: (
@@ -454,6 +466,7 @@ class TensorNameMap:
MODEL_TENSOR.LAYER_OUT_NORM: (
"encoder.layer.{bid}.output.LayerNorm", # bert
"transformer.layer.{bid}.output_layer_norm", # distillbert
"encoder.layers.{bid}.norm2", # nomic-bert
"transformer.decoder_layer.{bid}.rms_norm_3", # Grok
"encoder.layer.{bid}.mlp.layernorm", # jina-bert-v2
@@ -828,6 +841,7 @@ class TensorNameMap:
MODEL_TENSOR.CLS: (
"classifier", # jina
"classifier.dense", # roberta
"pre_classifier", # distillbert
),
MODEL_TENSOR.CLS_OUT: (
@@ -900,6 +914,7 @@ class TensorNameMap:
MODEL_TENSOR.V_MMPROJ: (
"multi_modal_projector.linear_{bid}",
"visual.merger.mlp.{bid}", # qwen2vl
),
MODEL_TENSOR.V_MMPROJ_FC: (
@@ -908,6 +923,8 @@ class TensorNameMap:
MODEL_TENSOR.V_MMPROJ_MLP: (
"model.mm_projector.mlp.mlp.{bid}",
"vision_model.vision_adapter.mlp.fc{bid}", # llama 4
"mlp1.{bid}", # InternVL
),
MODEL_TENSOR.V_MMPROJ_PEG: (
@@ -916,6 +933,7 @@ class TensorNameMap:
MODEL_TENSOR.V_ENC_EMBD_CLS: (
"vision_tower.vision_model.embeddings.class_embedding",
"vision_model.class_embedding", # llama 4
),
MODEL_TENSOR.V_ENC_EMBD_PATCH: (
@@ -923,82 +941,126 @@ class TensorNameMap:
"vpm.embeddings.patch_embedding",
"model.vision_model.embeddings.patch_embedding", # SmolVLM
"vision_tower.patch_conv", # pixtral
"vision_model.patch_embedding.linear", # llama 4
"visual.patch_embed.proj", # qwen2vl
),
MODEL_TENSOR.V_ENC_EMBD_POS: (
"vision_tower.vision_model.embeddings.position_embedding",
"vpm.embeddings.position_embedding",
"model.vision_model.embeddings.position_embedding", # SmolVLM
"vision_model.positional_embedding_vlm", # llama 4
),
MODEL_TENSOR.V_ENC_ATTN_Q: (
"vision_tower.vision_model.encoder.layers.{bid}.self_attn.q_proj",
"vpm.encoder.layers.{bid}.self_attn.q_proj",
"model.vision_model.encoder.layers.{bid}.self_attn.q_proj", # SmolVLM
"vision_model.model.layers.{bid}.self_attn.q_proj", # llama4
"vision_tower.transformer.layers.{bid}.attention.q_proj", # pixtral
"visual.blocks.{bid}.attn.q", # qwen2vl, generated
),
MODEL_TENSOR.V_ENC_ATTN_Q_NORM: (
"vision_tower.vision_model.encoder.layers.{bid}.attn.q_norm", # InternVL
),
MODEL_TENSOR.V_ENC_ATTN_K: (
"vision_tower.vision_model.encoder.layers.{bid}.self_attn.k_proj",
"vpm.encoder.layers.{bid}.self_attn.k_proj",
"model.vision_model.encoder.layers.{bid}.self_attn.k_proj", # SmolVLM
"vision_model.model.layers.{bid}.self_attn.k_proj", # llama4
"vision_tower.transformer.layers.{bid}.attention.k_proj", # pixtral
"visual.blocks.{bid}.attn.k", # qwen2vl, generated
),
MODEL_TENSOR.V_ENC_ATTN_K_NORM: (
"vision_tower.vision_model.encoder.layers.{bid}.attn.k_norm", # InternVL
),
MODEL_TENSOR.V_ENC_ATTN_V: (
"vision_tower.vision_model.encoder.layers.{bid}.self_attn.v_proj",
"vpm.encoder.layers.{bid}.self_attn.v_proj",
"model.vision_model.encoder.layers.{bid}.self_attn.v_proj", # SmolVLM
"vision_model.model.layers.{bid}.self_attn.v_proj", # llama4
"vision_tower.transformer.layers.{bid}.attention.v_proj", # pixtral
"visual.blocks.{bid}.attn.v", # qwen2vl, generated
),
MODEL_TENSOR.V_ENC_INPUT_NORM: (
"vision_tower.vision_model.encoder.layers.{bid}.layer_norm1",
"vision_tower.vision_model.encoder.layers.{bid}.norm1", # InternVL
"vpm.encoder.layers.{bid}.layer_norm1",
"model.vision_model.encoder.layers.{bid}.layer_norm1", # SmolVLM
"vision_tower.transformer.layers.{bid}.attention_norm", # pixtral
"vision_model.model.layers.{bid}.input_layernorm", # llama4
"visual.blocks.{bid}.norm1", # qwen2vl
),
MODEL_TENSOR.V_ENC_OUTPUT: (
MODEL_TENSOR.V_ENC_ATTN_O: (
"vision_tower.vision_model.encoder.layers.{bid}.self_attn.out_proj",
"vision_tower.vision_model.encoder.layers.{bid}.attn.proj", # InternVL
"vpm.encoder.layers.{bid}.self_attn.out_proj",
"model.vision_model.encoder.layers.{bid}.self_attn.out_proj", # SmolVLM
"vision_model.model.layers.{bid}.self_attn.o_proj", # llama4
"vision_tower.transformer.layers.{bid}.attention.o_proj", # pixtral
"visual.blocks.{bid}.attn.proj", # qwen2vl
),
MODEL_TENSOR.V_ENC_OUTPUT_NORM: (
MODEL_TENSOR.V_ENC_POST_ATTN_NORM: (
"vision_tower.vision_model.encoder.layers.{bid}.layer_norm2",
"vision_tower.vision_model.encoder.layers.{bid}.norm2", # InternVL
"vpm.encoder.layers.{bid}.layer_norm2",
"model.vision_model.encoder.layers.{bid}.layer_norm2", # SmolVLM
"vision_model.model.layers.{bid}.post_attention_layernorm", # llama4
"vision_tower.transformer.layers.{bid}.ffn_norm", # pixtral
"visual.blocks.{bid}.norm2", # qwen2vl
),
MODEL_TENSOR.V_ENC_FFN_UP: (
"vision_tower.vision_model.encoder.layers.{bid}.mlp.fc1",
"vpm.encoder.layers.{bid}.mlp.fc1",
"model.vision_model.encoder.layers.{bid}.mlp.fc2", # SmolVLM, gemma3 (note: name is swapped)
"model.vision_model.encoder.layers.{bid}.mlp.fc1", # SmolVLM, gemma3
"vision_tower.transformer.layers.{bid}.feed_forward.up_proj", # pixtral
"vision_model.model.layers.{bid}.mlp.fc1", # llama4
"visual.blocks.{bid}.mlp.fc1", # qwen2vl
"visual.blocks.{bid}.mlp.up_proj", # qwen2.5vl
),
MODEL_TENSOR.V_ENC_FFN_GATE: (
"vision_tower.transformer.layers.{bid}.feed_forward.gate_proj", # pixtral
"visual.blocks.{bid}.mlp.gate_proj", # qwen2.5vl
),
MODEL_TENSOR.V_ENC_FFN_DOWN: (
"vision_tower.vision_model.encoder.layers.{bid}.mlp.fc2",
"vpm.encoder.layers.{bid}.mlp.fc2",
"model.vision_model.encoder.layers.{bid}.mlp.fc1", # SmolVLM, gemma3 (note: name is swapped)
"model.vision_model.encoder.layers.{bid}.mlp.fc2", # SmolVLM, gemma3
"vision_tower.transformer.layers.{bid}.feed_forward.down_proj", # pixtral
"vision_model.model.layers.{bid}.mlp.fc2", # llama4
"visual.blocks.{bid}.mlp.fc2", # qwen2vl
"visual.blocks.{bid}.mlp.down_proj", # qwen2.5vl
),
MODEL_TENSOR.V_LAYER_SCALE_1: (
"vision_tower.vision_model.encoder.layers.{bid}.ls1", # InternVL
),
MODEL_TENSOR.V_LAYER_SCALE_2: (
"vision_tower.vision_model.encoder.layers.{bid}.ls2", # InternVL
),
MODEL_TENSOR.V_PRE_NORM: (
"vision_tower.vision_model.pre_layrnorm",
"vision_tower.ln_pre", # pixtral
"vision_model.layernorm_pre", # llama4
),
MODEL_TENSOR.V_POST_NORM: (
"vision_tower.vision_model.post_layernorm",
"model.vision_model.post_layernorm", # SmolVLM
"vision_model.layernorm_post", # llama4
"visual.merger.ln_q", # qwen2vl
),
MODEL_TENSOR.V_MM_INP_PROJ: (
@@ -1064,6 +1126,77 @@ class TensorNameMap:
MODEL_TENSOR.V_MM_PATCH_MERGER: (
"multi_modal_projector.patch_merger.merging_layer", # mistral small 3.1
),
# audio (mtmd)
MODEL_TENSOR.A_ENC_EMBD_POS: (
"audio_tower.embed_positions", # ultravox
),
MODEL_TENSOR.A_ENC_CONV1D: (
"audio_tower.conv{bid}", # ultravox
),
MODEL_TENSOR.A_PRE_NORM: (),
MODEL_TENSOR.A_POST_NORM: (
"audio_tower.layer_norm", # ultravox
"audio_tower.ln_post", # qwen2omni
),
MODEL_TENSOR.A_ENC_ATTN_Q: (
"audio_tower.layers.{bid}.self_attn.q_proj", # ultravox
),
MODEL_TENSOR.A_ENC_ATTN_K: (
"audio_tower.layers.{bid}.self_attn.k_proj", # ultravox
),
MODEL_TENSOR.A_ENC_ATTN_V: (
"audio_tower.layers.{bid}.self_attn.v_proj", # ultravox
),
MODEL_TENSOR.A_ENC_INPUT_NORM: (
"audio_tower.layers.{bid}.self_attn_layer_norm", # ultravox
),
MODEL_TENSOR.A_ENC_OUTPUT: (
"audio_tower.layers.{bid}.self_attn.out_proj", # ultravox
),
MODEL_TENSOR.A_ENC_OUTPUT_NORM: (
"audio_tower.layers.{bid}.final_layer_norm", # ultravox
),
MODEL_TENSOR.A_ENC_FFN_UP: (
"audio_tower.layers.{bid}.fc1", # ultravox
),
MODEL_TENSOR.A_ENC_FFN_GATE: (),
MODEL_TENSOR.A_ENC_FFN_DOWN: (
"audio_tower.layers.{bid}.fc2", # ultravox
),
# note: some tensors below has "audio." pseudo-prefix, to prevent conflicts with vision tensors
# this prefix is added in the conversion code in modify_tensors()
MODEL_TENSOR.A_MMPROJ: (
"audio.multi_modal_projector.linear_{bid}", # ultravox
),
MODEL_TENSOR.A_MMPROJ_FC: (
"audio.multi_modal_projector.linear", # qwen2audio
"audio_tower.proj", # qwen2omni
),
MODEL_TENSOR.A_MM_NORM_PRE: (
"audio.multi_modal_projector.ln_pre", # ultravox
),
MODEL_TENSOR.A_MM_NORM_MID: (
"audio.multi_modal_projector.ln_mid", # ultravox
),
}
# architecture-specific block mappings

2
gguf-py/gguf/utility.py
View File

@@ -231,7 +231,7 @@ class SafetensorRemote:
response.raise_for_status()
# Get raw byte data
return response.content[:size]
return response.content[slice(size if size > -1 else None)]
@classmethod
def check_file_exist(cls, url: str) -> bool:

12
gguf-py/pyproject.toml
View File

@@ -1,6 +1,6 @@
[tool.poetry]
name = "gguf"
version = "0.16.2"
version = "0.17.0"
description = "Read and write ML models in GGUF for GGML"
authors = ["GGML <ggml@ggml.ai>"]
packages = [
@@ -36,8 +36,8 @@ requires = ["poetry-core>=1.0.0"]
build-backend = "poetry.core.masonry.api"
[tool.poetry.scripts]
gguf-convert-endian = "gguf.scripts:gguf_convert_endian_entrypoint"
gguf-dump = "gguf.scripts:gguf_dump_entrypoint"
gguf-set-metadata = "gguf.scripts:gguf_set_metadata_entrypoint"
gguf-new-metadata = "gguf.scripts:gguf_new_metadata_entrypoint"
gguf-editor-gui = "gguf.scripts:gguf_editor_gui_entrypoint"
gguf-convert-endian = "gguf.scripts.gguf_convert_endian:main"
gguf-dump = "gguf.scripts.gguf_dump:main"
gguf-set-metadata = "gguf.scripts.gguf_set_metadata:main"
gguf-new-metadata = "gguf.scripts.gguf_new_metadata:main"
gguf-editor-gui = "gguf.scripts.gguf_editor_gui:main"