mirror of
https://github.com/hanzoai/insights.git
synced 2026-10-30 16:15:31 +00:00
zeekay
10e6fec4c0
The rule table mapped HOG_ -> FN_ ahead of any POSTHOG rule, so POSTHOG_FOO split into POST + HOG_ and came out POSTFN_FOO. Nothing sets that. The live pods set INSIGHTS_SKIP_MIGRATION_CHECKS, which bin/migrate-check reads on the container entrypoint, so the misspelling does not degrade a feature -- it runs migrate --check against a database whose history this tree does not share and the container never starts. Both replacements are measured against the shipped tree rather than chosen: every HOG_ name with a counterpart there is spelled INSIGHTS_ (HOG_QL -> INSIGHTS_QL, HOG_FUNCTION_TEMPLATES -> INSIGHTS_FUNCTION_TEMPLATES, 12 of 12 determinate cases), and FN_ appears in it nowhere. Upstream carries 95 HOG_ tokens and zero FN_ ones, so repairing FN_ back out is exactly the inverse of the bad rule and cannot collide with an upstream name. The rule is fixed for the next convergence, and the tree it already produced is repaired in the same commit -- the codemod is idempotent against its own output, so it cannot undo this itself.
60 lines
2.1 KiB
Python
Executable File
60 lines
2.1 KiB
Python
Executable File
#!/usr/bin/env python3
|
|
"""Download and export the sentiment ONNX model for LLM Analytics."""
|
|
|
|
import os
|
|
import sys
|
|
from pathlib import Path
|
|
|
|
CACHE_DIR = Path(os.environ.get("INSIGHTS_SENTIMENT_MODEL_CACHE", "/tmp/insights-sentiment-onnx-cache"))
|
|
REQUIRED_FILES = ("model.onnx", "tokenizer.json", "config.json")
|
|
|
|
MODEL_NAME = "cardiffnlp/twitter-roberta-base-sentiment-latest"
|
|
MODEL_REVISION = "3216a57f2a0d9c45a2e6c20157c20c49fb4bf9c7"
|
|
|
|
|
|
def main() -> None:
|
|
if all((CACHE_DIR / f).exists() for f in REQUIRED_FILES):
|
|
print(f"Sentiment ONNX model already exists at {CACHE_DIR}")
|
|
sys.exit(0)
|
|
|
|
print(f"Downloading and exporting sentiment ONNX model to {CACHE_DIR}...")
|
|
|
|
import torch
|
|
from optimum.onnxruntime import ORTModelForSequenceClassification
|
|
from transformers import AutoTokenizer, pipeline
|
|
|
|
CACHE_DIR.mkdir(parents=True, exist_ok=True)
|
|
|
|
print(f"Downloading {MODEL_NAME} (revision {MODEL_REVISION[:12]})")
|
|
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, revision=MODEL_REVISION)
|
|
|
|
# PyTorch 2.9+ defaults torch.onnx.export to dynamo=True which breaks optimum
|
|
original_export = torch.onnx.export
|
|
|
|
def export_with_dynamo_disabled(*args, **kwargs):
|
|
kwargs.setdefault("dynamo", False)
|
|
return original_export(*args, **kwargs)
|
|
|
|
torch.onnx.export = export_with_dynamo_disabled
|
|
|
|
print("Exporting to ONNX...")
|
|
model = ORTModelForSequenceClassification.from_pretrained(MODEL_NAME, revision=MODEL_REVISION, export=True)
|
|
torch.onnx.export = original_export
|
|
|
|
model.save_pretrained(CACHE_DIR)
|
|
tokenizer.save_pretrained(CACHE_DIR)
|
|
|
|
print("Running smoke test...")
|
|
pipe = pipeline(
|
|
"sentiment-analysis", model=model, tokenizer=tokenizer, top_k=None, truncation=True, max_length=512
|
|
)
|
|
result = pipe("I love this product")[0]
|
|
labels = {r["label"]: r["score"] for r in result}
|
|
assert labels.get("positive", 0) > 0.5, f"Smoke test failed: expected positive > 0.5, got {labels}"
|
|
print(f"Smoke test passed: positive={labels['positive']:.4f}")
|
|
print(f"Model saved to {CACHE_DIR}")
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|