MCPcopy Create free account
hub / github.com/DeusData/codebase-memory-mcp / main

Function main

scripts/extract_nomic_vectors.py:405–573  ·  view source on GitHub ↗
()

Source from the content-addressed store, hash-verified

403# ── Main ─────────────────────────────────────────────────────────────
404
405def main():
406 parser = argparse.ArgumentParser(description="Extract nomic-embed-code token embeddings")
407 parser.add_argument("--output-dir", default="vendored/nomic",
408 help="Output directory (default: vendored/nomic)")
409 parser.add_argument("--device", default=None,
410 help="Device: cuda, mps, cpu (auto-detected)")
411 parser.add_argument("--skip-attention", action="store_true",
412 help="Skip simulated attention (faster, lower quality)")
413 parser.add_argument("--batch-size", type=int, default=BATCH_SIZE,
414 help=f"Batch size (default: {BATCH_SIZE})")
415 parser.add_argument("--checkpoint", default=None,
416 help="Checkpoint file path (auto: <output-dir>/checkpoint.npz)")
417 args = parser.parse_args()
418
419 batch_size = args.batch_size
420
421 # Auto-detect device
422 # Prefer CPU for 7B models on Apple Silicon — MPS shares unified memory
423 # with the system and can cause OOM/crashes. CPU keeps allocation predictable.
424 # Use --device mps to override if you have enough headroom (32GB+).
425 if args.device:
426 device = args.device
427 elif torch.cuda.is_available():
428 device = "cuda"
429 else:
430 device = "cpu"
431
432 # Force line-buffered stdout so tee/log sees output immediately
433 sys.stdout.reconfigure(line_buffering=True)
434
435 print(f"device={device}")
436 print(f"threads={torch.get_num_threads()}")
437 print(f"model={MODEL_NAME}")
438 print(f"output_dim={OUTPUT_DIM}")
439 print()
440
441 # Create output dir
442 out_dir = Path(args.output_dir)
443 out_dir.mkdir(parents=True, exist_ok=True)
444
445 checkpoint_path = args.checkpoint or str(out_dir / "checkpoint.npz")
446
447 # ── Step 1: Load model + tokenizer ──
448 print("step 1: loading model + tokenizer...")
449 t0 = time.time()
450 tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, trust_remote_code=True)
451 model = AutoModel.from_pretrained(
452 MODEL_NAME,
453 trust_remote_code=True,
454 dtype=torch.float16, # 7B×2B = ~14GB (vs 28GB float32)
455 low_cpu_mem_usage=True, # Stream weights, no 2x peak during load
456 )
457 model = model.to(device)
458 print(f" loaded in {time.time() - t0:.1f}s")
459 print(f" hidden_size={model.config.hidden_size}")
460 print(f" vocab_size={tokenizer.vocab_size}")
461 print()
462

Callers 1

Calls 12

strFunction · 0.85
setFunction · 0.85
is_code_relevantFunction · 0.85
clean_tokenFunction · 0.85
extract_embeddingsFunction · 0.85
simulated_attentionFunction · 0.85
write_binFunction · 0.85
write_tokens_txtFunction · 0.85
write_tokens_hFunction · 0.85
write_vectors_hFunction · 0.85
write_blob_sFunction · 0.85
removeMethod · 0.80

Tested by

no test coverage detected