Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-26 22:30:32 +01:00
Code Issues Packages Projects Releases Wiki Activity

Refine Model Hyperparameters and Params Class

Browse Source 
- Updated type annotations to use `Optional` for clarity.
- Improved method names and attribute consistency.
- Removed unnecessary variables for better code readability.

Additional Notes:

- Highlighted the use of `Optional` for clearer intent.
- Ensured backward and forward compatibility.
This commit is contained in:
teleprint-me 2024-01-07 19:25:07 -05:00
parent acf8f4b20f
commit 15e18973da
No known key found for this signature in database
GPG Key ID: B0D11345E65C4D48
1 changed files with 99 additions and 77 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

84
convert.py
View File

@ -189,37 +189,57 @@ class Params:
n_ff: int n_ff: int
n_head: int n_head: int
n_head_kv: int n_head_kv: int
n_experts: int | None = None f_norm_eps: Optional[float] = None
n_experts_used: int | None = None n_experts: Optional[int] = None
f_norm_eps: float | None = None n_experts_used: Optional[int] = None
rope_scaling_type: gguf.RopeScalingType | None = None rope_scaling_type: Optional[gguf.RopeScalingType] = None
f_rope_freq_base: float | None = None f_rope_freq_base: Optional[float] = None
f_rope_scale: float | None = None f_rope_scale: Optional[float] = None
n_orig_ctx: int | None = None n_orig_ctx: Optional[int] = None
rope_finetuned: bool | None = None rope_finetuned: Optional[bool] = None
ftype: GGMLFileType | None = None ftype: Optional[GGMLFileType] = None
# path to the directory containing the model files # path to the directory containing the model files
path_model: Path | None = None path_model: Optional[Path] = None
@staticmethod @staticmethod
def guessed(model: LazyModel) -> Params: def guessed(model: LazyModel) -> "Params":
# try transformer naming first # try transformer naming first
n_vocab, n_embd = model["model.embed_tokens.weight"].shape if "model.embed_tokens.weight" in model else model["tok_embeddings.weight"].shape n_vocab, n_embd = (
model["model.embed_tokens.weight"].shape
if "model.embed_tokens.weight" in model
else model["tok_embeddings.weight"].shape
)
# try transformer naming first # try transformer naming first
if "model.layers.0.self_attn.q_proj.weight" in model: if "model.layers.0.self_attn.q_proj.weight" in model:
n_layer = next(i for i in itertools.count() if f"model.layers.{i}.self_attn.q_proj.weight" not in model) n_layer = next(
elif "model.layers.0.self_attn.W_pack.weight" in model: # next: try baichuan naming i
n_layer = next(i for i in itertools.count() if f"model.layers.{i}.self_attn.W_pack.weight" not in model) for i in itertools.count()
if f"model.layers.{i}.self_attn.q_proj.weight" not in model
)
elif (
"model.layers.0.self_attn.W_pack.weight" in model
): # next: try baichuan naming
n_layer = next(
i
for i in itertools.count()
if f"model.layers.{i}.self_attn.W_pack.weight" not in model
)
else: else:
n_layer = next(i for i in itertools.count() if f"layers.{i}.attention.wq.weight" not in model) n_layer = next(
i
for i in itertools.count()
if f"layers.{i}.attention.wq.weight" not in model
)
if n_layer < 1: if n_layer < 1:
raise Exception("failed to guess 'n_layer'. This model is unknown or unsupported.\n" raise Exception(
"Suggestion: provide 'config.json' of the model in the same directory containing model files.") "failed to guess 'n_layer'. This model is unknown or unsupported.\n"
"Suggestion: provide 'config.json' of the model in the same directory containing model files."
)
n_head = n_embd // 128 # guessed n_head = n_embd // 128 # guessed
n_mult = 256 # guessed n_mult = 256 # guessed
@ -240,7 +260,7 @@ class Params:
) )
@staticmethod @staticmethod
def loadHFTransformerJson(model: LazyModel, config_path: Path) -> Params: def load_transformers_config(model: LazyModel, config_path: Path) -> "Params":
config = json.load(open(config_path)) config = json.load(open(config_path))
rope_scaling_type = f_rope_scale = n_orig_ctx = rope_finetuned = None rope_scaling_type = f_rope_scale = n_orig_ctx = rope_finetuned = None
@ -253,18 +273,20 @@ class Params:
rope_scaling_type = gguf.RopeScalingType.LINEAR rope_scaling_type = gguf.RopeScalingType.LINEAR
elif typ == "yarn": elif typ == "yarn":
rope_scaling_type = gguf.RopeScalingType.YARN rope_scaling_type = gguf.RopeScalingType.YARN
n_orig_ctx = rope_scaling['original_max_position_embeddings'] n_orig_ctx = rope_scaling["original_max_position_embeddings"]
rope_finetuned = rope_scaling['finetuned'] rope_finetuned = rope_scaling["finetuned"]
else: else:
raise NotImplementedError(f'Unknown rope scaling type: {typ}') raise NotImplementedError(f"Unknown rope scaling type: {typ}")
if "max_sequence_length" in config: if "max_sequence_length" in config:
n_ctx = config["max_sequence_length"] n_ctx = config["max_sequence_length"]
elif "max_position_embeddings" in config: elif "max_position_embeddings" in config:
n_ctx = config["max_position_embeddings"] n_ctx = config["max_position_embeddings"]
else: else:
raise Exception("failed to guess 'n_ctx'. This model is unknown or unsupported.\n" raise Exception(
"Suggestion: provide 'config.json' of the model in the same directory containing model files.") "failed to guess 'n_ctx'. This model is unknown or unsupported.\n"
"Suggestion: provide 'config.json' of the model in the same directory containing model files."
)
n_experts = None n_experts = None
n_experts_used = None n_experts_used = None
@ -294,7 +316,7 @@ class Params:
# LLaMA v2 70B params.json # LLaMA v2 70B params.json
# {"dim": 8192, "multiple_of": 4096, "ffn_dim_multiplier": 1.3, "n_heads": 64, "n_kv_heads": 8, "n_layers": 80, "norm_eps": 1e-05, "vocab_size": -1} # {"dim": 8192, "multiple_of": 4096, "ffn_dim_multiplier": 1.3, "n_heads": 64, "n_kv_heads": 8, "n_layers": 80, "norm_eps": 1e-05, "vocab_size": -1}
@staticmethod @staticmethod
def loadOriginalParamsJson(model: LazyModel, config_path: Path) -> Params: def load_torch_params(model: LazyModel, config_path: Path) -> "Params":
config = json.load(open(config_path)) config = json.load(open(config_path))
n_experts = None n_experts = None
@ -325,7 +347,7 @@ class Params:
f_rope_freq_base = 1e6 f_rope_freq_base = 1e6
return Params( return Params(
n_vocab = model["tok_embeddings.weight"].shape[0], n_vocab=config.get("vocab_size", model["tok_embeddings.weight"].shape[0]),
n_embd=config["dim"], n_embd=config["dim"],
n_layer=config["n_layers"], n_layer=config["n_layers"],
n_ctx=n_ctx, n_ctx=n_ctx,
@ -339,18 +361,18 @@ class Params:
) )
@staticmethod @staticmethod
def load(model_plus: ModelPlus) -> Params: def load(model_plus: ModelPlus) -> "Params":
hf_config_path = model_plus.paths[0].parent / "config.json" hf_config_path = model_plus.paths[0].parent / "config.json"
orig_config_path = model_plus.paths[0].parent / "params.json" orig_config_path = model_plus.paths[0].parent / "params.json"
if hf_config_path.exists(): if hf_config_path.exists():
params = Params.loadHFTransformerJson(model_plus.model, hf_config_path) params = Params.load_transformers_config(model_plus.model, hf_config_path)
elif orig_config_path.exists(): elif orig_config_path.exists():
params = Params.loadOriginalParamsJson(model_plus.model, orig_config_path) params = Params.load_torch_params(model_plus.model, orig_config_path)
elif model_plus.format != 'none': elif model_plus.format != "none":
params = Params.guessed(model_plus.model) params = Params.guessed(model_plus.model)
else: else:
raise ValueError('Cannot guess params when model format is none') raise ValueError("Cannot guess params when model format is none")
params.path_model = model_plus.paths[0].parent params.path_model = model_plus.paths[0].parent