masahiroid's picture
Publish validated original-checkpoint conversion with provenance and usage
9a0cc30 verified
Raw History Blame Contribute Delete
3.7 kB
import argparse,json,shutil,traceback
from pathlib import Path
import numpy as np
import torch
from common import *
def metadata(name,fmt,model):
p=ROOT/'publish'/f'{name}-{fmt}'; p.mkdir(parents=True,exist_ok=True)
model.config.save_pretrained(p)
if name!='chronos-bolt-small': AutoTokenizer.from_pretrained(ROOT/'sources'/name,use_fast=name!='mdeberta-v3-base').save_pretrained(p)
return p
def build(name,fmt):
print('START',name,fmt,flush=True)
model=load(name); wrapper=Wrapper(model,name).eval()
p=metadata(name,fmt,model); tests=cases(name); refs=[reference(name,c,wrapper) for c in tests]
if fmt in ('mlx','gguf'):
import mlx.core as mx
from runtime_mlx import Model
weights={k:v.detach().float().numpy() for k,v in model.state_dict().items()}
if fmt=='mlx': mx.save_safetensors(str(p/'model.safetensors'),{k:mx.array(v) for k,v in weights.items()})
else:
from gguf import GGUFWriter
writer=GGUFWriter(str(p/'model-f32.gguf'),'hf-native-mlx-v1')
writer.add_name(name); writer.add_string('runtime','runtime_mlx.py (custom; not llama.cpp)')
for k,v in weights.items(): writer.add_tensor(k,v)
writer.write_header_to_file(); writer.write_kv_data_to_file(); writer.write_tensors_to_file(); writer.close()
shutil.copy(ROOT/'runtime_mlx.py',p/'runtime_mlx.py')
converted=Model(p,gguf=fmt=='gguf')
outs=[]
for c in tests:
inp={k:v for k,v in c.items() if k in ('context','input_ids','attention_mask')}
y=converted(**inp); mx.eval(y); outs.append(np.array(y))
elif fmt=='coreml':
import coremltools as ct
if name=='mdeberta-v3-base':
import math
import transformers.models.deberta_v2.modeling_deberta_v2 as d
# Fixed head dimension: fold the scripted integer sqrt into an FP32 constant.
d.scaled_size_sqrt=lambda q,f: torch.tensor(math.sqrt(64*f),dtype=q.dtype,device=q.device)
# Self-attention has equal query/key lengths; fold the scripted branch.
d.build_rpos=lambda q,k,r,*unused:r
args=args_for(name,tests[0]); traced=torch.jit.trace(ExportWrapper(model,name).eval(),args,strict=False,check_trace=False)
names=['context'] if name=='chronos-bolt-small' else ['input_ids','attention_mask']
inputs=[ct.TensorType(name=k,shape=tuple(v.shape),dtype=np.float32 if k=='context' else np.int32) for k,v in zip(names,args)]
converted=ct.convert(traced,inputs=inputs,outputs=[ct.TensorType(name='output')],convert_to='mlprogram',
compute_precision=ct.precision.FLOAT32,minimum_deployment_target=ct.target.macOS14)
converted.save(str(p/'model.mlpackage'))
outs=[converted.predict({k:c[k].astype(np.float32 if k=='context' else np.int32) for k in names})['output'] for c in tests]
else: raise ValueError(fmt)
result=[metrics(r,y) for r,y in zip(refs,outs)]
print('METRICS',result,flush=True)
assert all(m['max_abs']<0.02 and m['cosine']>0.999 for m in result), result
(p/'validation.json').write_text(json.dumps(dict(status='passed',cases=result,input_shape=list(next(iter(tests[0].values())).shape),precision='float32'),indent=2))
np.savez(p/'validation_vectors.npz',**{f'reference_{i}':v for i,v in enumerate(refs)},**{f'input_{i}_{k}':v for i,c in enumerate(tests) for k,v in c.items() if k in ('context','input_ids','attention_mask')})
print('PASS',name,fmt,flush=True)
if __name__=='__main__':
a=argparse.ArgumentParser(); a.add_argument('name',choices=NAMES); a.add_argument('format'); x=a.parse_args(); build(x.name,x.format)