root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-11-11 13:30:35 +00:00
Code Issues Actions 11 Packages Projects Releases Wiki Activity

convert : adapt MiniCPM3 to separate rope_freqs insertion

Browse Source 
MiniCPM3's tokenizer is treated as a SentencePiece tokenizer to avoid
having to run its custom Python code which mixes tokenization
in the same file as tool calls.

gguf-py : add long and short RoPE factors to tensor mappings

Empty, but the key names are used to populate the mappings.
This commit is contained in:
Francis Couture-Harpin 2024-09-16 12:05:29 -04:00
parent ed0f2c4ab1
commit e83d2707d3
3 changed files with 17 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
convert_hf_to_gguf.py
View File

@ -1862,8 +1862,6 @@ class MiniCPM3Model(Model):
def set_gguf_parameters(self):
hparams = self.hparams
rope_dims = hparams["qk_rope_head_dim"]
self.gguf_writer.add_file_type(self.ftype)
self.gguf_writer.add_context_length(hparams["max_position_embeddings"])
self.gguf_writer.add_embedding_length(hparams["hidden_size"])
@ -1879,9 +1877,10 @@ class MiniCPM3Model(Model):
self.gguf_writer.add_key_length(hparams["qk_nope_head_dim"] + hparams["qk_rope_head_dim"])
self.gguf_writer.add_rope_dimension_count(hparams["qk_rope_head_dim"])
def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
rope_scaling = self.find_hparam(['rope_scaling'], True)
if rope_scaling is None:
return
if rope_scaling is not None:
rope_dims = self.hparams["qk_rope_head_dim"]
long_factors = rope_scaling.get('long_factor', None)
short_factors = rope_scaling.get('short_factor', None)
@ -1892,11 +1891,11 @@ class MiniCPM3Model(Model):
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}')
self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_LONG] + ".weight", np.array(long_factors, dtype=np.float32))
self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT] + ".weight", np.array(short_factors, dtype=np.float32))
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()
self._set_vocab_sentencepiece()
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:

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

@ -877,6 +877,8 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ROPE_FACTORS_LONG,
MODEL_TENSOR.ROPE_FACTORS_SHORT,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_Q_A,
MODEL_TENSOR.ATTN_Q_B,

3
gguf-py/gguf/tensor_mapping.py
View File

@ -87,6 +87,9 @@ class TensorNameMap:
"rope.freqs", # llama-pth
"rotary_pos_emb.inv_freq", # chatglm
),
MODEL_TENSOR.ROPE_FACTORS_LONG: (),
MODEL_TENSOR.ROPE_FACTORS_SHORT: (),
}
block_mappings_cfg: dict[MODEL_TENSOR, tuple[str, ...]] = {