File size: 2,127 Bytes
12496fc | 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 | """CPU dynamic INT8 experiment on the reference model only."""
from pathlib import Path
import copy
import json
import sys
import time
import warnings
sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
import numpy as np
import torch
from safetensors.torch import load_file
from nexora.model import NexoraLM, ModelConfig
from nexora.training import batch
from nexora.evaluation import percentiles
def main():
torch.set_num_threads(4)
cfg = ModelConfig(**json.loads(Path("artifacts/tiny/config.json").read_text()))
model = NexoraLM(cfg).eval()
model.load_state_dict(load_file("artifacts/tiny/model.safetensors"))
with warnings.catch_warnings():
warnings.simplefilter("ignore", DeprecationWarning)
quant = torch.ao.quantization.quantize_dynamic(copy.deepcopy(model), {torch.nn.Linear}, dtype=torch.qint8)
ids = torch.from_numpy(np.load("artifacts/data/validation.npy").astype(np.int64))
x, y = batch(ids, 4, 128, torch.Generator().manual_seed(9), "cpu")
results = {}
logits = {}
for name, m in [("fp32", model), ("dynamic_int8_linear_only", quant)]:
times = []
with torch.no_grad():
m(x, y)
for _ in range(20):
start = time.perf_counter()
output, loss = m(x, y)
times.append(time.perf_counter()-start)
logits[name] = output
results[name] = {"validation_loss": loss.item(), "batch_latency_seconds": percentiles(times), "samples": len(times)}
results["max_logit_difference"] = (logits["fp32"]-logits["dynamic_int8_linear_only"]).abs().max().item()
results["argmax_agreement"] = (logits["fp32"].argmax(-1) == logits["dynamic_int8_linear_only"].argmax(-1)).float().mean().item()
results["limitations"] = "Tiny model, one validation batch, CPU linear-layer INT8 only. Not INT8 embeddings, BF16/FP8/INT4 deployment, coding/reasoning/context/tool quality validation. Do not infer production speedup."
Path("reports/quantization.json").write_text(json.dumps(results, indent=2))
print(json.dumps(results))
if __name__ == "__main__":
main()
|