File size: 5,250 Bytes
307b208 | 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 | import asyncio
import json
import shutil
from pathlib import Path
import numpy as np
import torch
from PIL import Image
from transformers import AutoModel, AutoProcessor
from coreai_embeddinggemma2 import EmbeddingGemma2CoreAI
def cases():
wave = (0.2 * np.sin(2 * np.pi * 440 * np.arange(16000) / 16000)).astype(np.float32)
image = Image.new("RGB", (320, 192), (20, 80, 210))
gradient = Image.fromarray(
np.tile(np.arange(256, dtype=np.uint8)[None, :, None], (128, 1, 3))
)
frames = np.stack(
[
np.asarray(image.resize((128, 96))),
np.asarray(Image.new("RGB", (128, 96), (210, 50, 20))),
]
)
return [
(
"query",
{"text": "What causes the northern lights?", "prompt_name": "SearchQuery"},
),
(
"document",
{
"text": "The northern lights are caused by charged particles from the sun.",
"prompt_name": "Document",
},
),
("unprefixed", {"text": "An apple a day keeps the doctor away."}),
(
"code",
{"text": "def add(a, b): return a + b", "prompt_name": "CodeRetrieval"},
),
("spanish", {"text": "La biblioteca está abierta hasta las ocho."}),
("japanese", {"text": "桜の花が春に咲きます。"}),
("short", {"text": "cat"}),
("long", {"text": "The oceans contain diverse ecosystems. " * 90}),
("near-capacity", {"text": "The oceans contain diverse ecosystems. " * 160}),
("image", {"images": image}),
("gradient", {"images": gradient}),
("image-text", {"text": "<|image|> A blue rectangle.", "images": image}),
("two-images", {"images": [image, gradient]}),
("audio", {"audio": wave}),
("audio-text", {"text": "<|audio|> A musical tone.", "audio": wave}),
("audio-short", {"audio": wave[:3200]}),
("audio-long", {"audio": np.tile(wave, 5)}),
("audio-limit", {"audio": np.pad(np.tile(wave, 11), (0, 3200))}),
("video", {"videos": frames}),
("video-text", {"text": "<|video|> Colors changing.", "videos": frames}),
(
"mixed",
{
"text": "<|image|> Blue. <|audio|> Tone. <|video|> Colors.",
"images": image,
"audio": wave,
"videos": frames,
},
),
]
async def main():
torch.set_num_threads(4)
root = Path("bundles/embeddinggemma-2-CoreAI")
source = Path("upstream/embeddinggemma-2")
for f in source.iterdir():
if f.is_file() and f.suffix in [".json", ".jinja", ".model"]:
shutil.copy2(f, root / f.name)
m = AutoModel.from_pretrained(
source,
dtype=torch.float32,
attn_implementation={
"language_model": "sdpa",
"vision_tower": "sdpa",
"audio_tower": "eager",
},
).eval()
p = AutoProcessor.from_pretrained(source)
print("Loading Core AI", flush=True)
runtime = await EmbeddingGemma2CoreAI.load(root)
print("Loaded Core AI", flush=True)
records = []
for name, args in cases():
data = dict(args)
prompt = data.pop("prompt_name", None)
if prompt:
data["text"] = runtime.prompts[prompt] + data["text"]
if "audio" in data:
data["sampling_rate"] = 16000
if "videos" in data:
data["videos_kwargs"] = {"do_sample_frames": False}
e = p(**data, return_tensors="pt")
with torch.no_grad():
h = m(**e).last_hidden_state
mask = e["attention_mask"][:, :, None].float()
reference = torch.nn.functional.normalize(
(h * mask).sum(1) / mask.sum(1), dim=-1
)[0].numpy()
actual = await runtime.encode_processed({k: v.numpy() for k, v in e.items()})
record = {
"name": name,
"tokens": e["input_ids"].shape[1],
"max_abs_error": float(np.max(np.abs(actual - reference))),
"cosine": float(actual @ reference),
"finite": bool(np.isfinite(actual).all()),
"unit_norm": float(np.linalg.norm(actual)),
}
print(record, flush=True)
records.append(record)
assert (
record["finite"]
and record["cosine"] > 0.99999
and record["max_abs_error"] < 0.00005
), record
for dim in [128, 256, 512]:
actual = await runtime.encode("The sky is blue.", truncate_dim=dim)
assert actual.shape == (dim,) and abs(np.linalg.norm(actual) - 1) < 1e-5
try:
await runtime.encode("word " * 3000)
except ValueError:
pass
else:
raise AssertionError("Over-capacity input accepted")
(root / "validation.json").write_text(
json.dumps(
{
"thresholds": {"min_cosine": 0.99999, "max_abs_error": 0.00005},
"cases": records,
"truncation_dims": [128, 256, 512],
"oversized_input_rejected": True,
},
indent=2,
)
+ "\n"
)
if __name__ == "__main__":
asyncio.run(main())
|