Encoder B loads from snapshot directories, as the reference vector did
Measured on Yggdrasil: loading by repo id leaves the model's _name_or_path equal to the Llama repo id, so prepare_for_tokenization wraps the text in the chat header and the vector matches assembly A (cosine 1.0, 0.909 to the reference B). Loading from the snapshot paths reproduces the reference. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
parent
6535cda025
commit
ffa2c7bf58
3 changed files with 31 additions and 12 deletions
|
|
@ -15,7 +15,7 @@ import msgpack
|
|||
import torch
|
||||
|
||||
sys.path.insert(0, str(Path(__file__).resolve().parent))
|
||||
from text_encoding import ASSEMBLY, EMBEDDING_DIM, LLAMA_BASE, apply_thread_cap # noqa: E402
|
||||
from text_encoding import ASSEMBLY, EMBEDDING_DIM, LLAMA_BASE, apply_thread_cap, snapshot_path # noqa: E402
|
||||
|
||||
from peft import PeftModel # noqa: E402
|
||||
from transformers import AutoConfig, AutoTokenizer # noqa: E402
|
||||
|
|
@ -24,20 +24,23 @@ from kimodo.model.llm2vec.llm2vec import LLM2Vec # noqa: E402
|
|||
from kimodo.model.load_model import TEXT_ENCODER_PRESETS # noqa: E402
|
||||
|
||||
|
||||
def load_encoder(mntp_id, supervised_id, dtype):
|
||||
def load_encoder(hf_home, mntp_id, supervised_id, dtype):
|
||||
"""Assembly B: the Llama base with the MNTP adapter merged in, then the supervised adapter on top.
|
||||
|
||||
Kimodo's own LLM2VecEncoder loads the MNTP repo as the base and never applies the MNTP weights under the
|
||||
pinned transformers (assembly A); this assembly is ours (motion-pipeline:ruling:encoder-assembly-b). The
|
||||
identity carries the Kimodo commit and this file's sha256, so a changed loader misses the cache.
|
||||
Kimodo's own LLM2VecEncoder loads the MNTP repo as the base (assembly A); this assembly is ours
|
||||
(motion-pipeline:ruling:encoder-assembly-b). Everything loads from the local snapshot directories, as the
|
||||
research's reference vector did: the model's _name_or_path is then a path, not the repo id, so
|
||||
LLM2Vec.prepare_for_tokenization does not wrap the text in Llama's chat header, which is what separates B's
|
||||
vector from A's. The identity carries the Kimodo commit and this file's sha256, so a changed loader misses.
|
||||
"""
|
||||
tokenizer = AutoTokenizer.from_pretrained(LLAMA_BASE)
|
||||
llama, mntp, supervised = (snapshot_path(hf_home, repo) for repo in (LLAMA_BASE, mntp_id, supervised_id))
|
||||
tokenizer = AutoTokenizer.from_pretrained(llama)
|
||||
tokenizer.pad_token = tokenizer.eos_token
|
||||
tokenizer.padding_side = "left"
|
||||
model_class = LLM2Vec._get_model_class(AutoConfig.from_pretrained(LLAMA_BASE).__class__.__name__, enable_bidirectional=True)
|
||||
model = model_class.from_pretrained(LLAMA_BASE, torch_dtype=dtype)
|
||||
model = PeftModel.from_pretrained(model, mntp_id).merge_and_unload()
|
||||
model = PeftModel.from_pretrained(model, supervised_id)
|
||||
model_class = LLM2Vec._get_model_class(AutoConfig.from_pretrained(llama).__class__.__name__, enable_bidirectional=True)
|
||||
model = model_class.from_pretrained(llama, torch_dtype=dtype)
|
||||
model = PeftModel.from_pretrained(model, mntp).merge_and_unload()
|
||||
model = PeftModel.from_pretrained(model, supervised)
|
||||
return LLM2Vec(model=model, tokenizer=tokenizer)
|
||||
|
||||
|
||||
|
|
@ -48,7 +51,7 @@ def main() -> int:
|
|||
preset = TEXT_ENCODER_PRESETS["llm2vec"]["kwargs"]
|
||||
|
||||
started = time.perf_counter()
|
||||
model = load_encoder(preset["base_model_name_or_path"], preset["peft_model_name_or_path"], getattr(torch, preset["dtype"]))
|
||||
model = load_encoder(os.environ["HF_HOME"], preset["base_model_name_or_path"], preset["peft_model_name_or_path"], getattr(torch, preset["dtype"]))
|
||||
model.eval()
|
||||
for p in model.parameters():
|
||||
p.requires_grad = False
|
||||
|
|
|
|||
Loading…
Reference in a new issue