| """ |
| scripts/model_info.py |
| |
| Step 13 requirement: report the parameter count via an explicit command, |
| never just claim a number. Run with a config path OR a Models.create_model() |
| key (o-mini, 250m, 500m, gpt2-small, gpt2-plus, deep64, mla, ...) to see the |
| real breakdown for that architecture. |
| |
| Usage: |
| PYTHONPATH=. python3 scripts/model_info.py configs/model/brain_0_2_30m.yaml |
| PYTHONPATH=. python3 scripts/model_info.py deep64 |
| """ |
|
|
| import sys |
| from pathlib import Path |
|
|
| from brain.core.config import BrainConfig |
| from brain.model.brain_model import BrainForCausalLM |
|
|
|
|
| def main() -> None: |
| arg = sys.argv[1] if len(sys.argv) > 1 else "configs/model/brain_0_2_30m.yaml" |
| if Path(arg).exists(): |
| config = BrainConfig.from_yaml(arg) |
| model = BrainForCausalLM(config) |
| label = arg |
| else: |
| from Models.Foundation.model_factory import create_model |
| model = create_model(arg) |
| config = model.config |
| label = f"create_model({arg!r})" |
| counts = model.count_parameters() |
|
|
| print(f"Brain config: {config.name} ({label})") |
| print("-" * 50) |
| for k, v in config.to_dict().items(): |
| print(f" {k}: {v}") |
| print("-" * 50) |
| print("Parameter breakdown:") |
| for k, v in counts.items(): |
| pct = f" ({100 * v / counts['total']:.1f}%)" if k != "total" and k != "trainable" else "" |
| print(f" {k:14s}: {v:>12,}{pct}") |
| print("-" * 50) |
| print(f"Total: {counts['total']:,} ({counts['total']/1e6:.2f}M)") |
|
|
| if model.rope.effective_theta != config.rope_theta: |
| print("-" * 50) |
| print( |
| f"NOTE: RoPE scaling ({config.rope_scaling_type} x{config.rope_scaling_factor}) " |
| f"is applied unconditionally from construction, not only once generation exceeds " |
| f"max_position_embeddings -- the theta actually running is " |
| f"{model.rope.effective_theta:,.0f}, not the configured {config.rope_theta:,.0f}." |
| ) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
