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

llama : add support for Chameleon (#8543)

Browse source 
* convert chameleon hf to gguf

* add chameleon tokenizer tests

* fix lint

* implement chameleon graph

* add swin norm param

* return qk norm weights and biases to original format

* implement swin norm

* suppress image token output

* rem tabs

* add comment to conversion

* fix ci

* check for k norm separately

* adapt to new lora implementation

* fix layer input for swin norm

* move swin_norm in gguf writer

* add comment regarding special token regex in chameleon pre-tokenizer

* Update src/llama.cpp

Co-authored-by: compilade <git@compilade.net>

* fix punctuation regex in chameleon pre-tokenizer (@compilade)

Co-authored-by: compilade <git@compilade.net>

* fix lint

* trigger ci

---------

Co-authored-by: compilade <git@compilade.net>
This commit is contained in:
nopperl 2024-09-28 12:08:43 +00:00 • committed by GitHub
parent 43bcdd9703
commit 9a913110cf
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
10 changed files with 505 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

44
convert_hf_to_gguf.py
View file

@ -640,6 +640,9 @@ class Model:
if chkhsh == "fcace8b9cac38ce847670c970cd5892031a753a1ef381abd1d9af00f713da085":
# ref: https://huggingface.co/microsoft/phi-2
res = "phi-2"
if chkhsh == "60824e3c0d9401f89943cbb2fff727f0e2d4c545ba4df2d6e4f09a6db0f5b450":
# ref: https://huggingface.co/facebook/chameleon-7b
res = "chameleon"
if res is None:
logger.warning("\n")
@ -4138,6 +4141,47 @@ class GraniteMoeModel(GraniteModel):
return super().modify_tensors(data_torch, name, bid)
@Model.register("ChameleonForCausalLM")
class ChameleonModel(Model):
model_arch = gguf.MODEL_ARCH.CHAMELEON
def set_gguf_parameters(self):
super().set_gguf_parameters()
self.gguf_writer.add_swin_norm(self.hparams.get("swin_norm", False))
def set_vocab(self):
self._set_vocab_gpt2()
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
# ignore image tokenizer for now
# TODO: remove this once image support is implemented for Chameleon
if name.startswith("model.vqmodel"):
return []
n_head = self.hparams["num_attention_heads"]
n_kv_head = self.hparams.get("num_key_value_heads")
hidden_dim = self.hparams.get("hidden_size")
if name.endswith(("q_proj.weight", "q_proj.bias")):
data_torch = LlamaModel.permute(data_torch, n_head, n_head)
if name.endswith(("k_proj.weight", "k_proj.bias")):
data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head)
if name.endswith(("q_norm.weight", "q_norm.bias")):
data_torch = ChameleonModel._reverse_hf_permute(data_torch, n_head, hidden_dim)
if name.endswith(("k_norm.weight", "k_norm.bias")):
data_torch = ChameleonModel._reverse_hf_permute(data_torch, n_kv_head, hidden_dim)
return [(self.map_tensor_name(name), data_torch)]
# see: https://github.com/huggingface/transformers/blob/72fb02c47dbbe1999ae105319f24631cad6e2e00/src/transformers/models/chameleon/convert_chameleon_weights_to_hf.py#L176-L203
@staticmethod
def _reverse_hf_permute(data_torch, n_heads, hidden_dim):
head_dim = hidden_dim // n_heads
data_torch = data_torch[0].view(2, head_dim // 2).t().reshape(1, -1)
data_torch = data_torch.repeat_interleave(n_heads, 0)
return data_torch
###### CONVERSION LOGIC ######