Sparse-AST-BWM / latest /model.py
Chaman1234's picture
Add verified step-93213 SafeTensors release and update README
d3c3fac verified
Raw History Blame Contribute Delete
2.65 kB
"""SafeTensors-only loader for the step-93213 Sparse-AST old/new checkpoint."""
from pathlib import Path
import json
import torch
from safetensors.torch import load_file
from test_models import SparseAST
from family_oldnew_architecture import SparseASTFamilyOldNew
def checkpoint_directory(path_or_repo, revision=None, training=False):
p=Path(path_or_repo)
if p.is_dir():return p
from huggingface_hub import snapshot_download
patterns=['training_state.json','training_state.safetensors'] if training else ['config.json','model.safetensors']
return Path(snapshot_download(repo_id=str(path_or_repo),revision=revision,allow_patterns=['latest/'+n for n in patterns]))/'latest'
def load_model(path_or_repo,device='cpu',revision=None):
root=checkpoint_directory(path_or_repo,revision);cfg=json.loads((root/'config.json').read_text())
with torch.device('meta'):
base=SparseAST(d=cfg['d_model'],h=cfg['d_hidden'],layers=cfg['num_layers'],al_hidden=cfg['al_hidden'],seq_len=cfg['position_embedding_capacity'],vocab=cfg['vocab_size'])
model=SparseASTFamilyOldNew(base)
state=load_file(str(root/'model.safetensors'),device='cpu')
for alias,canonical in cfg['tensor_aliases'].items():state[alias]=state[canonical]
model.load_state_dict(state,strict=True,assign=True)
model.to(device);model.eval();model.training_context_bytes=cfg['training_context_bytes']
return model
def decode_state(obj,tensors):
if isinstance(obj,dict) and obj.get('__kind__')=='tensor':return tensors[obj['key']]
if isinstance(obj,dict) and obj.get('__kind__')=='dict':return {decode_state(k,tensors):decode_state(v,tensors) for k,v in obj['items']}
if isinstance(obj,dict) and obj.get('__kind__')=='tuple':return tuple(decode_state(v,tensors) for v in obj['items'])
if isinstance(obj,list):return [decode_state(v,tensors) for v in obj]
return obj
def load_training_state(path_or_repo,revision=None):
root=checkpoint_directory(path_or_repo,revision,training=True)
return decode_state(json.loads((root/'training_state.json').read_text()),load_file(str(root/'training_state.safetensors'),device='cpu'))
@torch.inference_mode()
def generate(model,prompt,max_new_bytes=128,context_bytes=128):
if not prompt:raise ValueError('Provide a nonempty UTF-8 prompt')
if not 1<=context_bytes<=model.seq_len:raise ValueError('Context exceeds position embedding capacity')
encoded=list(prompt.encode('utf-8'));device=next(model.parameters()).device
for _ in range(max_new_bytes):
logits=model(torch.tensor([encoded[-context_bytes:]],dtype=torch.long,device=device))[0,-1,:256]
encoded.append(int(logits.argmax()))
return bytes(encoded).decode('utf-8',errors='replace')