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

llama : add CodeShell support (#5016)

Browse source 
* llama: add codeshell support

* llama.cpp: fix codeshell with NeoX rope

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
chiranko 2024-01-19 17:07:27 +08:00 • committed by GitHub
parent 993fba8180
commit 2b3b999cac
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 268 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

67
convert-hf-to-gguf.py
View file

@ -197,6 +197,8 @@ class Model:
return Phi2Model
if model_architecture == "PlamoForCausalLM":
return PlamoModel
if model_architecture == "CodeShellForCausalLM":
return CodeShellModel
return Model
def _is_model_safetensors(self) -> bool:
@ -242,6 +244,8 @@ class Model:
return gguf.MODEL_ARCH.PHI2
if arch == "PlamoForCausalLM":
return gguf.MODEL_ARCH.PLAMO
if arch == "CodeShellForCausalLM":
return gguf.MODEL_ARCH.CODESHELL
raise NotImplementedError(f'Architecture "{arch}" not supported!')
@ -1175,6 +1179,69 @@ class PlamoModel(Model):
self.gguf_writer.add_tensor(new_name, data)
class CodeShellModel(Model):
def set_gguf_parameters(self):
block_count = self.hparams["n_layer"]
self.gguf_writer.add_name("CodeShell")
self.gguf_writer.add_context_length(self.hparams["n_positions"])
self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
self.gguf_writer.add_feed_forward_length(4 * self.hparams["n_embd"])
self.gguf_writer.add_block_count(block_count)
self.gguf_writer.add_head_count(self.hparams["n_head"])
self.gguf_writer.add_head_count_kv(self.hparams["num_query_groups"])
self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
self.gguf_writer.add_file_type(self.ftype)
self.gguf_writer.add_rope_freq_base(10000.0)
self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
self.gguf_writer.add_rope_scaling_factor(1.0)
def write_tensors(self):
block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
tensors = dict(self.get_tensors())
has_lm_head = "lm_head.weight" in tensors.keys() or "output.weight" in tensors.keys()
for name, data_torch in tensors.items():
# we don't need these
if name.endswith((".attn.rotary_emb.inv_freq")):
continue
old_dtype = data_torch.dtype
# convert any unsupported data types to float32
if data_torch.dtype not in (torch.float16, torch.float32):
data_torch = data_torch.to(torch.float32)
data = data_torch.squeeze().numpy()
# map tensor names
new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
if new_name is None:
print(f"Can not map tensor {name!r}")
sys.exit()
n_dims = len(data.shape)
data_dtype = data.dtype
# if f32 desired, convert any float16 to float32
if self.ftype == 0 and data_dtype == np.float16:
data = data.astype(np.float32)
# TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
data = data.astype(np.float32)
# if f16 desired, convert any float32 2-dim weight tensors to float16
if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data)
if not has_lm_head and name == "transformer.wte.weight":
self.gguf_writer.add_tensor("output.weight", data)
print(name, f"=> output.weight, shape = {data.shape}, {old_dtype} --> {data.dtype}")
###### CONVERSION LOGIC ######