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

llama: Support MiniCPM-1B (with & w/o longrope) (#10559)

Browse source
This commit is contained in:
JFLFY2255 2024-12-04 17:42:50 +08:00 • committed by GitHub
parent 2759916d86
commit 8d0cfd554a
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 61 additions and 183 deletions
Download patch file Download diff file Expand all files Collapse all files

57
convert_hf_to_gguf.py
View file

@ -1831,29 +1831,40 @@ class MiniCPMModel(Model):
model_arch = gguf.MODEL_ARCH.MINICPM
def set_gguf_parameters(self):
block_count = self.hparams["num_hidden_layers"]
self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"])
self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
self.gguf_writer.add_block_count(block_count)
self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
self.gguf_writer.add_rope_dimension_count(self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
self.gguf_writer.add_head_count(self.hparams["num_attention_heads"])
self.gguf_writer.add_head_count_kv(self.hparams["num_key_value_heads"])
self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
self.gguf_writer.add_file_type(self.ftype)
super().set_gguf_parameters()
embedding_scale = float(self.hparams["scale_emb"])
self.gguf_writer.add_embedding_scale(embedding_scale)
logger.info(f"gguf: (minicpm) embedding_scale = {embedding_scale}")
residual_scale = self.hparams["scale_depth"] / self.hparams["num_hidden_layers"] ** 0.5
self.gguf_writer.add_residual_scale(residual_scale)
logger.info(f"gguf: (minicpm) residual_scale = {residual_scale}")
logit_scale = self.hparams["hidden_size"] / self.hparams["dim_model_base"]
self.gguf_writer.add_logit_scale(logit_scale)
logger.info(f"gguf: (minicpm) logit_scale = {logit_scale}")
if self.hparams.get("rope_scaling") is not None:
if self.hparams["rope_scaling"].get("type") == "longrope":
self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LONGROPE)
logger.info(f"gguf: (minicpm) rope_scaling_type = {gguf.RopeScalingType.LONGROPE}")
def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
rope_dims = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
rope_scaling = self.find_hparam(['rope_scaling'], True)
if rope_scaling is not None:
long_factors = rope_scaling.get('long_factor', None)
short_factors = rope_scaling.get('short_factor', None)
if long_factors is None or short_factors is None:
raise KeyError('Missing the required key rope_scaling.long_factor or rope_scaling_short_factor')
if len(long_factors) != len(short_factors) or len(long_factors) != rope_dims / 2:
raise ValueError(f'The length of rope long and short factors must be {rope_dims / 2}')
yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FACTORS_LONG), torch.tensor(long_factors, dtype=torch.float32))
yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT), torch.tensor(short_factors, dtype=torch.float32))
def set_vocab(self):
self._set_vocab_llama_hf()
def _reverse_hf_permute(self, weights: Tensor, n_head: int, n_kv_head: int | None = None) -> Tensor:
if n_kv_head is not None and n_head != n_kv_head:
n_head //= n_kv_head
return (
weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
.swapaxes(1, 2)
.reshape(weights.shape)
)
self._set_vocab_sentencepiece()
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
del bid # unused
@ -1863,9 +1874,9 @@ class MiniCPMModel(Model):
# HF models permute some of the tensors, so we need to undo that
if name.endswith(("q_proj.weight")):
data_torch = self._reverse_hf_permute(data_torch, n_head, n_head)
data_torch = LlamaModel.permute(data_torch, n_head, n_head)
if name.endswith(("k_proj.weight")):
data_torch = self._reverse_hf_permute(data_torch, n_head, n_kv_head)
data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head)
return [(self.map_tensor_name(name), data_torch)]