vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

fix(convert_hf_to_gguf/gguf-py): _multiplier -> _scale

Browse source 
The transformers names with _multiplier will now be converted to the _scale
equivalent during conversion.

Branch: GraniteLM

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
This commit is contained in:
Gabe Goodhart 2024-09-16 08:56:56 -06:00
parent 0bdf04e7b5
commit 65c5bb91ab
3 changed files with 20 additions and 18 deletions
Download patch file Download diff file Expand all files Collapse all files

20
convert_hf_to_gguf.py
View file

@ -4090,20 +4090,22 @@ class GraniteModel(LlamaModel):
- No head_dim support - No head_dim support
- New multiplier params: - New multiplier params:
- attention_multiplier - attention_scale
- embedding_multiplier - embedding_scale
- residual_multiplier - residual_scale
- logits_scaling - logits_scaling
""" """
if head_dim := self.hparams.pop("head_dim", None): if head_dim := self.hparams.pop("head_dim", None):
logger.warning("Ignoring head_dim (%s) from config for Granite", head_dim) logger.warning("Ignoring head_dim (%s) from config for Granite", head_dim)
super().set_gguf_parameters() super().set_gguf_parameters()
if attention_multiplier := self.hparams.get("attention_multiplier"): # NOTE: Convert _multiplier params to _scale params for naming
self.gguf_writer.add_attention_multiplier(attention_multiplier) # consistency
if embedding_multiplier := self.hparams.get("embedding_multiplier"): if attention_scale := self.hparams.get("attention_multiplier"):
self.gguf_writer.add_embedding_multiplier(embedding_multiplier) self.gguf_writer.add_attention_scale(attention_scale)
if residual_multiplier := self.hparams.get("residual_multiplier"): if embedding_scale := self.hparams.get("embedding_multiplier"):
self.gguf_writer.add_residual_multiplier(residual_multiplier) self.gguf_writer.add_embedding_scale(embedding_scale)
if residual_scale := self.hparams.get("residual_multiplier"):
self.gguf_writer.add_residual_scale(residual_scale)
if logits_scaling := self.hparams.get("logits_scaling"): if logits_scaling := self.hparams.get("logits_scaling"):
self.gguf_writer.add_logit_scale(logits_scaling) self.gguf_writer.add_logit_scale(logits_scaling)

6
gguf-py/gguf/constants.py
View file

@ -97,8 +97,8 @@ class Keys:
RESCALE_EVERY_N_LAYERS = "{arch}.rescale_every_n_layers" RESCALE_EVERY_N_LAYERS = "{arch}.rescale_every_n_layers"
TIME_MIX_EXTRA_DIM = "{arch}.time_mix_extra_dim" TIME_MIX_EXTRA_DIM = "{arch}.time_mix_extra_dim"
TIME_DECAY_EXTRA_DIM = "{arch}.time_decay_extra_dim" TIME_DECAY_EXTRA_DIM = "{arch}.time_decay_extra_dim"
RESIDUAL_MULTIPLIER = "{arch}.residual_multiplier" RESIDUAL_SCALE = "{arch}.residual_scale"
EMBEDDING_MULTIPLIER = "{arch}.embedding_multiplier" EMBEDDING_SCALE = "{arch}.embedding_scale"
class Attention: class Attention:
HEAD_COUNT = "{arch}.attention.head_count" HEAD_COUNT = "{arch}.attention.head_count"
@ -114,7 +114,7 @@ class Keys:
KV_LORA_RANK = "{arch}.attention.kv_lora_rank" KV_LORA_RANK = "{arch}.attention.kv_lora_rank"
REL_BUCKETS_COUNT = "{arch}.attention.relative_buckets_count" REL_BUCKETS_COUNT = "{arch}.attention.relative_buckets_count"
SLIDING_WINDOW = "{arch}.attention.sliding_window" SLIDING_WINDOW = "{arch}.attention.sliding_window"
MULTIPLIER = "{arch}.attention.multiplier" SCALE = "{arch}.attention.scale"
class Rope: class Rope:
DIMENSION_COUNT = "{arch}.rope.dimension_count" DIMENSION_COUNT = "{arch}.rope.dimension_count"

12
gguf-py/gguf/gguf_writer.py
View file

@ -679,11 +679,11 @@ class GGUFWriter:
def add_time_decay_extra_dim(self, dim: int) -> None: def add_time_decay_extra_dim(self, dim: int) -> None:
self.add_uint32(Keys.LLM.TIME_DECAY_EXTRA_DIM.format(arch=self.arch), dim) self.add_uint32(Keys.LLM.TIME_DECAY_EXTRA_DIM.format(arch=self.arch), dim)
def add_residual_multiplier(self, value: float) -> None: def add_residual_scale(self, value: float) -> None:
self.add_float32(Keys.LLM.RESIDUAL_MULTIPLIER.format(arch=self.arch), value) self.add_float32(Keys.LLM.RESIDUAL_SCALE.format(arch=self.arch), value)
def add_embedding_multiplier(self, value: float) -> None: def add_embedding_scale(self, value: float) -> None:
self.add_float32(Keys.LLM.EMBEDDING_MULTIPLIER.format(arch=self.arch), value) self.add_float32(Keys.LLM.EMBEDDING_SCALE.format(arch=self.arch), value)
def add_wkv_head_size(self, size: int) -> None: def add_wkv_head_size(self, size: int) -> None:
self.add_uint32(Keys.WKV.HEAD_SIZE.format(arch=self.arch), size) self.add_uint32(Keys.WKV.HEAD_SIZE.format(arch=self.arch), size)
@ -709,8 +709,8 @@ class GGUFWriter:
def add_sliding_window(self, value: int) -> None: def add_sliding_window(self, value: int) -> None:
self.add_uint32(Keys.Attention.SLIDING_WINDOW.format(arch=self.arch), value) self.add_uint32(Keys.Attention.SLIDING_WINDOW.format(arch=self.arch), value)
def add_attention_multiplier(self, value: float) -> None: def add_attention_scale(self, value: float) -> None:
self.add_float32(Keys.Attention.MULTIPLIER.format(arch=self.arch), value) self.add_float32(Keys.Attention.SCALE.format(arch=self.arch), value)
def add_pooling_type(self, value: PoolingType) -> None: def add_pooling_type(self, value: PoolingType) -> None:
self.add_uint32(Keys.LLM.POOLING_TYPE.format(arch=self.arch), value.value) self.add_uint32(Keys.LLM.POOLING_TYPE.format(arch=self.arch), value.value)