diff options
Diffstat (limited to 'gguf-py')
| -rw-r--r-- | gguf-py/gguf/constants.py | 26 | ||||
| -rw-r--r-- | gguf-py/gguf/tensor_mapping.py | 3 | ||||
| -rw-r--r-- | gguf-py/gguf/vocab.py | 33 |
3 files changed, 58 insertions, 4 deletions
diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index d6feb9ea..e2f4eb1a 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -181,6 +181,7 @@ class GGUFType: class MODEL_ARCH(IntEnum): LLAMA = auto() + DECI = auto() FALCON = auto() BAICHUAN = auto() GROK = auto() @@ -316,6 +317,7 @@ class MODEL_TENSOR(IntEnum): MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.LLAMA: "llama", + MODEL_ARCH.DECI: "deci", MODEL_ARCH.FALCON: "falcon", MODEL_ARCH.BAICHUAN: "baichuan", MODEL_ARCH.GROK: "grok", @@ -470,6 +472,26 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_DOWN_EXP, MODEL_TENSOR.FFN_UP_EXP, ], + MODEL_ARCH.DECI: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_ROT_EMBD, + MODEL_TENSOR.FFN_GATE_INP, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.FFN_GATE_EXP, + MODEL_TENSOR.FFN_DOWN_EXP, + MODEL_TENSOR.FFN_UP_EXP, + ], MODEL_ARCH.GROK: [ MODEL_TENSOR.TOKEN_EMBD, MODEL_TENSOR.OUTPUT_NORM, @@ -1149,6 +1171,10 @@ MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.ROPE_FREQS, MODEL_TENSOR.ATTN_ROT_EMBD, ], + MODEL_ARCH.DECI: [ + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_ROT_EMBD, + ], MODEL_ARCH.BAICHUAN: [ MODEL_TENSOR.ROPE_FREQS, MODEL_TENSOR.ATTN_ROT_EMBD, diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py index 1dea6a82..3ff70cd7 100644 --- a/gguf-py/gguf/tensor_mapping.py +++ b/gguf-py/gguf/tensor_mapping.py @@ -176,7 +176,8 @@ class TensorNameMap: "transformer.blocks.{bid}.attn.out_proj", # mpt "transformer.h.{bid}.self_attention.dense", # falcon "h.{bid}.self_attention.dense", # bloom - "model.layers.{bid}.self_attn.o_proj", # llama-hf + "model.layers.{bid}.self_attn.o_proj", # llama-hf nemotron olmoe olmo2 + "model.layers.{bid}.self_attn.linear_attn", # deci "layers.{bid}.attention.wo", # llama-pth "encoder.layer.{bid}.attention.output.dense", # bert "transformer.h.{bid}.attn.out_proj", # gpt-j diff --git a/gguf-py/gguf/vocab.py b/gguf-py/gguf/vocab.py index dc574991..cca09798 100644 --- a/gguf-py/gguf/vocab.py +++ b/gguf-py/gguf/vocab.py @@ -122,8 +122,30 @@ class SpecialVocab: tokenizer = json.load(f) if self.load_merges: merges = tokenizer.get('model', {}).get('merges') - if isinstance(merges, list) and merges and isinstance(merges[0], str): - self.merges = merges + if isinstance(merges, list) and merges: + if isinstance(merges[0], str): + self.merges = merges + elif isinstance(merges[0], list) and len(merges[0]) == 2 and isinstance(merges[0][0], str): + # New format since transformers 4.45 to support spaces in merges + # ref: https://github.com/ggml-org/llama.cpp/issues/9692 + # TODO: internally store as the new format instead of converting to old + if any(' ' in s for pair in merges for s in pair): + logger.warning(f'Spaces in merges detected, encoding as {chr(ord(" ") + 256)!r}') + self.merges = [ + ' '.join( + [ + # ensure the spaces are properly encoded + ''.join( + chr(ord(c) + 256) if c == ' ' else c + for c in part + ) + for part in pair + ] + ) + for pair in merges + ] + else: + raise ValueError("Unknown tokenizer merges format") added_tokens = tokenizer.get('added_tokens', {}) else: added_tokens = {} @@ -132,7 +154,12 @@ class SpecialVocab: return True with open(tokenizer_config_file, encoding = 'utf-8') as f: tokenizer_config = json.load(f) - chat_template = tokenizer_config.get('chat_template') + chat_template_alt = None + chat_template_file = path / 'chat_template.json' + if chat_template_file.is_file(): + with open(chat_template_file, encoding = 'utf-8') as f: + chat_template_alt = json.load(f).get('chat_template') + chat_template = tokenizer_config.get('chat_template', chat_template_alt) if chat_template is None or isinstance(chat_template, (str, list)): self.chat_template = chat_template else: |