Public/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-09-19 09:15:04 +02:00
Code Issues Packages Projects Releases Wiki Activity

FACP (Fewer Acronym Classes Please)

Browse Source
This commit is contained in:
Sigbjørn Skjæret
2026-08-27 12:15:18 +02:00
parent 69a1269f70
commit e5227ab28e
2 changed files with 10 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files
gguf-py/gguf/constants.py
+1 -1
View File
@@ -228,7 +228,7 @@ class Keys:
# absent means the mix projection is full rank (DeepSeek-V4 behaviour)
LOW_RANK = "{arch}.hyper_connection.low_rank"
class PLE:
class PerLayerEmbedding:
LAYERS = "{arch}.ple.layers"
NGRAM_SIZE = "{arch}.ple.ngram_size"
HEADS_PER_NGRAM = "{arch}.ple.heads_per_ngram"
gguf-py/gguf/gguf_writer.py
+9 -9
View File
@@ -1033,35 +1033,35 @@ class GGUFWriter:
self.add_uint32(Keys.HyperConnection.LOW_RANK.format(arch=self.arch), value)
def add_ple_layers(self, values: Sequence[int]) -> None:
self.add_array(Keys.PLE.LAYERS.format(arch=self.arch), values)
self.add_array(Keys.PerLayerEmbedding.LAYERS.format(arch=self.arch), values)
def add_ple_ngram_size(self, value: int) -> None:
self.add_uint32(Keys.PLE.NGRAM_SIZE.format(arch=self.arch), value)
self.add_uint32(Keys.PerLayerEmbedding.NGRAM_SIZE.format(arch=self.arch), value)
def add_ple_heads_per_ngram(self, value: int) -> None:
self.add_uint32(Keys.PLE.HEADS_PER_NGRAM.format(arch=self.arch), value)
self.add_uint32(Keys.PerLayerEmbedding.HEADS_PER_NGRAM.format(arch=self.arch), value)
def add_ple_conv_kernel(self, value: int) -> None:
self.add_uint32(Keys.PLE.CONV_KERNEL.format(arch=self.arch), value)
self.add_uint32(Keys.PerLayerEmbedding.CONV_KERNEL.format(arch=self.arch), value)
# multipliers reach ~2.4e13; default INT32 inference would truncate them
def _add_u64_array(self, key: str, values: Sequence[int]) -> None:
self.add_key_value(key, list(values), GGUFValueType.ARRAY, GGUFValueType.UINT64)
def add_ple_layer_multipliers(self, values: Sequence[int]) -> None:
self._add_u64_array(Keys.PLE.LAYER_MULTIPLIERS.format(arch=self.arch), values)
self._add_u64_array(Keys.PerLayerEmbedding.LAYER_MULTIPLIERS.format(arch=self.arch), values)
def add_ple_head_offsets(self, values: Sequence[int]) -> None:
self._add_u64_array(Keys.PLE.HEAD_OFFSETS.format(arch=self.arch), values)
self._add_u64_array(Keys.PerLayerEmbedding.HEAD_OFFSETS.format(arch=self.arch), values)
def add_ple_head_vocab_sizes(self, values: Sequence[int]) -> None:
self._add_u64_array(Keys.PLE.HEAD_VOCAB_SIZES.format(arch=self.arch), values)
self._add_u64_array(Keys.PerLayerEmbedding.HEAD_VOCAB_SIZES.format(arch=self.arch), values)
def add_ple_eos_token_id(self, value: int) -> None:
self.add_uint32(Keys.PLE.EOS_TOKEN_ID.format(arch=self.arch), value)
self.add_uint32(Keys.PerLayerEmbedding.EOS_TOKEN_ID.format(arch=self.arch), value)
def add_ple_image_token_id(self, value: int) -> None:
self.add_uint32(Keys.PLE.IMAGE_TOKEN_ID.format(arch=self.arch), value)
self.add_uint32(Keys.PerLayerEmbedding.IMAGE_TOKEN_ID.format(arch=self.arch), value)
def add_attention_scale(self, value: float) -> None:
self.add_float32(Keys.Attention.SCALE.format(arch=self.arch), value)