File size: 9,733 Bytes
506a341 49ad2ef 506a341 49ad2ef 506a341 49ad2ef 506a341 49ad2ef 506a341 49ad2ef 506a341 49ad2ef 506a341 49ad2ef 506a341 49ad2ef 506a341 49ad2ef 506a341 49ad2ef 506a341 49ad2ef | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 | """Compare an exported end-to-end ONNX bundle with its PyTorch checkpoint."""
from __future__ import annotations
import argparse
import json
from pathlib import Path
import numpy as np
import onnxruntime as ort
import torch
from transformers import AutoModel, AutoTokenizer
from vons.models import ModelConfig, build_model, cosine_beta_schedule
def _model_config_from_checkpoint(checkpoint: dict[str, object]) -> ModelConfig:
config = checkpoint["config"]
if not isinstance(config, dict):
raise TypeError("checkpoint config must be a mapping")
backend = str(checkpoint["backend"])
conditioning = str(config.get("diffusion_conditioning", "pooled"))
token_conditioned = backend == "diffusion" and conditioning == "token_cross_attention"
return ModelConfig(
hidden_size=int(checkpoint["hidden_size"]),
latent_size=int(config.get("latent_size", 64 if token_conditioned else 32)),
diffusion_steps=int(config.get("diffusion_steps", 64 if token_conditioned else 32)),
inference_steps=int(config.get("inference_steps", 8 if token_conditioned else 4)),
abstain_threshold=float(config.get("abstain_threshold", 0.55)),
diffusion_conditioning=conditioning,
attention_heads=int(config.get("attention_heads", 4)),
)
def verify(checkpoint_path: Path, bundle_path: Path, *, max_length: int = 32) -> dict[str, object]:
checkpoint = torch.load(checkpoint_path, map_location="cpu")
if checkpoint["backend"] not in {"direct", "diffusion"}:
raise ValueError("unsupported checkpoint backend")
encoder_config = checkpoint["config"]["encoder"]
tokenizer = AutoTokenizer.from_pretrained(encoder_config["name"], revision=encoder_config.get("revision"))
encoder = AutoModel.from_pretrained(encoder_config["name"], revision=encoder_config.get("revision")).eval()
backend = checkpoint["backend"]
conditioning = str(checkpoint["config"].get("diffusion_conditioning", "pooled"))
head = build_model(
backend,
option_count=checkpoint["option_count"],
config=_model_config_from_checkpoint(checkpoint),
)
head.load_state_dict(checkpoint["state_dict"])
head.eval()
texts = [
"The user asked for weather in Seoul.\nQuestion: Which action?\nCandidate: call_weather",
"The user asked for weather in Seoul.\nQuestion: Which action?\nCandidate: clarify",
]
token_length = max_length if backend == "direct" else int(encoder_config.get("max_tokens", 512))
encoded = tokenizer(texts, padding="max_length", truncation=True, max_length=token_length, return_tensors="pt")
encoded.setdefault("token_type_ids", torch.zeros_like(encoded["input_ids"]))
options = (
checkpoint["option_count"]
if backend == "diffusion" and conditioning == "pooled"
else len(texts)
)
if options > len(texts):
for name, tensor in encoded.items():
encoded[name] = torch.nn.functional.pad(tensor, (0, 0, 0, options - len(texts)))
batched = {name: tensor.unsqueeze(0) for name, tensor in encoded.items()}
option_mask = torch.zeros((1, options), dtype=torch.bool)
option_mask[:, : len(texts)] = True
with torch.no_grad():
if backend == "direct":
candidate_embeddings = encoder(
input_ids=batched["input_ids"].reshape(options, token_length),
attention_mask=batched["attention_mask"].reshape(options, token_length),
token_type_ids=batched["token_type_ids"].reshape(options, token_length),
).last_hidden_state[:, 0, :].reshape(1, options, -1)
pooled = (candidate_embeddings * option_mask.unsqueeze(-1)).sum(dim=1) / option_mask.sum(
dim=1, keepdim=True
).clamp_min(1)
torch_output = head(candidate_embeddings, pooled, option_mask)
elif conditioning == "token_cross_attention":
sequence_hidden = encoder(
input_ids=batched["input_ids"].reshape(options, token_length),
attention_mask=batched["attention_mask"].reshape(options, token_length),
token_type_ids=batched["token_type_ids"].reshape(options, token_length),
).last_hidden_state.reshape(1, options, token_length, -1)
attention_mask = batched["attention_mask"]
diffusion_steps = int(checkpoint["config"].get("diffusion_steps", 64))
inference_steps = int(checkpoint["config"].get("inference_steps", 8))
alpha_bars = torch.cumprod(1.0 - cosine_beta_schedule(diffusion_steps), dim=0)
timesteps = torch.linspace(diffusion_steps - 1, 0, inference_steps).round().long()
sample = torch.zeros((1, options))
torch_output = {}
for index, timestep in enumerate(timesteps):
timestep_batch = torch.full((1,), int(timestep), dtype=torch.long)
result = head(sample, sequence_hidden, attention_mask, timestep_batch, option_mask)
predicted_x0 = torch.softmax(result["logits"], dim=-1)
alpha = alpha_bars[timestep]
predicted_noise = (sample - torch.sqrt(alpha) * predicted_x0) / torch.sqrt(
1 - alpha
).clamp_min(1e-6)
previous_alpha = (
alpha_bars[timesteps[index + 1]]
if index + 1 < len(timesteps)
else sample.new_tensor(1.0)
)
sample = torch.sqrt(previous_alpha) * predicted_x0 + torch.sqrt(
1 - previous_alpha
) * predicted_noise
torch_output = result
torch_output = {
"scores": torch_output["logits"],
"answerability": torch_output["answerability"],
}
else:
candidate_embeddings = encoder(
input_ids=batched["input_ids"].reshape(options, token_length),
attention_mask=batched["attention_mask"].reshape(options, token_length),
token_type_ids=batched["token_type_ids"].reshape(options, token_length),
).last_hidden_state[:, 0, :].reshape(1, options, -1)
weights = option_mask.to(candidate_embeddings.dtype).unsqueeze(-1)
pooled = (candidate_embeddings * weights).sum(dim=1) / weights.sum(dim=1).clamp_min(1)
diffusion_steps = int(checkpoint["config"].get("diffusion_steps", 32))
inference_steps = int(checkpoint["config"].get("inference_steps", 4))
betas = torch.linspace(1e-4, 0.02, diffusion_steps)
alpha_bars = torch.cumprod(1.0 - betas, dim=0)
timesteps = torch.linspace(len(alpha_bars) - 1, 0, inference_steps).round().long()
sample = torch.zeros((1, checkpoint["option_count"]))
for index, timestep in enumerate(timesteps):
predicted_noise = head(sample, pooled, torch.full((1,), int(timestep), dtype=torch.long))["noise"]
alpha = alpha_bars[timestep]
previous_alpha = alpha_bars[timesteps[index + 1]] if index + 1 < len(timesteps) else sample.new_tensor(1.0)
predicted_x0 = (sample - torch.sqrt(1 - alpha) * predicted_noise) / torch.sqrt(alpha)
sample = torch.sqrt(previous_alpha) * predicted_x0 + torch.sqrt(1 - previous_alpha) * predicted_noise
torch_output = {
"scores": sample.masked_fill(~option_mask, float("-inf")),
"answerability": head(
torch.zeros_like(sample), pooled, torch.zeros((1,), dtype=torch.long)
)["answerability"],
}
session = ort.InferenceSession(str(bundle_path), providers=["CPUExecutionProvider"])
ort_feed = {
"input_ids": batched["input_ids"].numpy(),
"attention_mask": batched["attention_mask"].numpy(),
"token_type_ids": batched["token_type_ids"].numpy(),
"option_mask": option_mask.numpy(),
}
if backend == "diffusion":
ort_feed["initial_noise"] = np.zeros((1, options), dtype=np.float32)
ort_output = session.run(None, ort_feed)
score_name = "logits" if backend == "direct" else "scores"
score_delta = torch_output[score_name].numpy() - ort_output[0]
finite_scores = np.isfinite(torch_output[score_name].numpy()) & np.isfinite(ort_output[0])
max_score_delta = float(np.max(np.abs(score_delta[finite_scores]))) if finite_scores.any() else 0.0
result = {
"checkpoint": str(checkpoint_path),
"bundle": str(bundle_path),
"provider": "CPUExecutionProvider",
"backend": backend,
"max_abs_logits": max_score_delta,
"max_abs_answerability": float(np.max(np.abs(torch_output["answerability"].numpy() - ort_output[1]))),
"pass": bool(
np.allclose(torch_output[score_name].numpy(), ort_output[0], atol=1e-5, rtol=1e-5, equal_nan=True)
and np.allclose(torch_output["answerability"].numpy(), ort_output[1], atol=1e-5, rtol=1e-5)
),
}
return result
def main() -> int:
parser = argparse.ArgumentParser()
parser.add_argument("--checkpoint", required=True, type=Path)
parser.add_argument("--bundle", required=True, type=Path)
parser.add_argument("--output", required=True, type=Path)
args = parser.parse_args()
result = verify(args.checkpoint, args.bundle)
args.output.parent.mkdir(parents=True, exist_ok=True)
args.output.write_text(json.dumps(result, indent=2, sort_keys=True) + "\n", encoding="utf-8")
print(json.dumps(result, indent=2, sort_keys=True))
return 0 if result["pass"] else 1
if __name__ == "__main__":
raise SystemExit(main())
|