fable-mini / configuration_fable.py
dev-sandakelum's picture
Upload folder using huggingface_hub
feb6b86 verified
Raw History Blame Contribute Delete
2.91 kB
"""FableConfig — Hugging Face configuration for the `fable` miniature.
Mirrors `fable_lm.model.ModelConfig` (vocab_size=2048, context_len=512, n_layers=4,
d_model=256, n_heads=4, d_ff=768) using standard HF configuration names so the
model is loadable with `AutoConfig` / `trust_remote_code`.
This file is intentionally self-contained (no `fable_lm` import): it is copied
verbatim into the Hub repo, where only the files next to `config.json` exist.
"""
from __future__ import annotations
from transformers import PretrainedConfig
class FableConfig(PretrainedConfig):
model_type = "fable"
def __init__(
self,
vocab_size: int = 2048,
max_position_embeddings: int = 512,
num_hidden_layers: int = 4,
hidden_size: int = 256,
num_attention_heads: int = 4,
intermediate_size: int = 768,
dropout: float = 0.0,
pad_token_id: int = 256, # <|pad|>
bos_token_id: int = 258, # <|doc|> doubles as BOS (document separator)
eos_token_id: int = 257, # <|eos|>
**kwargs,
):
# architecture hyperparameters (names follow HF conventions)
self.vocab_size = vocab_size
self.max_position_embeddings = max_position_embeddings
self.num_hidden_layers = num_hidden_layers
self.hidden_size = hidden_size
self.num_attention_heads = num_attention_heads
self.intermediate_size = intermediate_size
self.dropout = dropout
# token embeddings are tied into the LM head (FableLM: head.weight = tok_emb.weight)
self.tie_word_embeddings = True
super().__init__(
pad_token_id=pad_token_id,
bos_token_id=bos_token_id,
eos_token_id=eos_token_id,
**kwargs,
)
# ---- read-only aliases for the original repo-style names ----------------
# Properties (not attributes) so they never appear in to_dict()/config.json;
# serialization always uses the canonical HF names above.
@property
def context_len(self) -> int:
return self.max_position_embeddings
@property
def n_layers(self) -> int:
return self.num_hidden_layers
@property
def d_model(self) -> int:
return self.hidden_size
@property
def n_heads(self) -> int:
return self.num_attention_heads
@property
def d_ff(self) -> int:
return self.intermediate_size
try:
# Local (non-remote-code) use: `import hf_fable` makes AutoConfig aware of
# "fable" without trust_remote_code. exist_ok: the module may also be
# imported under a second name (hf_fable.configuration_fable vs a plain
# configuration_fable copied next to config.json).
from transformers import AutoConfig
AutoConfig.register("fable", FableConfig, exist_ok=True)
except Exception: # pragma: no cover - registration must never break import
pass