diff options
Diffstat (limited to 'experiments/transformer_crossover_grid.py')
| -rw-r--r-- | experiments/transformer_crossover_grid.py | 258 |
1 files changed, 258 insertions, 0 deletions
diff --git a/experiments/transformer_crossover_grid.py b/experiments/transformer_crossover_grid.py new file mode 100644 index 0000000..0fa8ba8 --- /dev/null +++ b/experiments/transformer_crossover_grid.py @@ -0,0 +1,258 @@ +#!/usr/bin/env python3 +"""Immutable Transformer selector registry and failure-continuing driver.""" +import argparse +import hashlib +import json +import os +import subprocess +import sys +import time + + +ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__))) +PROTOCOL = os.path.join(ROOT, "TRANSFORMER_CROSSOVER.md") +RESULT_ROOT = os.path.join(ROOT, "results", "transformer_crossover") +METHODS = ( + "bp", "fa", "dfa", "pepita", "ff", "ep", "dualprop", + "clean_kp", "sdil", +) +RATE_GRIDS = { + "bp": (0.0003, 0.001, 0.003), + "fa": (0.0003, 0.001, 0.003), + "dfa": (0.0003, 0.001, 0.003), + "pepita": (0.0001, 0.0003, 0.001), + "ff": (0.0001, 0.0003, 0.001), + "ep": (0.0001, 0.0003, 0.001), + "dualprop": (0.0001, 0.0003, 0.001), + "clean_kp": (0.0003, 0.001, 0.003), + "sdil": (0.0003, 0.001, 0.003), +} + + +def sha256(path): + digest = hashlib.sha256() + with open(path, "rb") as handle: + for chunk in iter(lambda: handle.read(1024 * 1024), b""): + digest.update(chunk) + return digest.hexdigest() + + +def text_sha256(text): + return hashlib.sha256(text.encode("utf-8")).hexdigest() + + +def git_output(*args): + return subprocess.run( + ["git", *args], cwd=ROOT, check=True, capture_output=True, + text=True).stdout.strip() + + +def source_report(): + dirty = git_output( + "status", "--porcelain", "--untracked-files=no") + if dirty: + raise RuntimeError( + "Transformer crossover requires clean tracked source") + paths = [ + PROTOCOL, + os.path.abspath(__file__), + os.path.join( + ROOT, "experiments", "analyze_transformer_crossover_p1.py"), + os.path.join( + ROOT, "experiments", "transformer_crossover_native.py"), + os.path.join( + ROOT, "experiments", "transformer_feedback_smoke.py"), + os.path.join(ROOT, "sdil", "transformer.py"), + ] + for path in paths: + tracked = subprocess.run( + ["git", "ls-files", "--error-unmatch", + os.path.relpath(path, ROOT)], + cwd=ROOT, capture_output=True).returncode == 0 + if not tracked: + raise RuntimeError(f"untracked crossover source: {path}") + freeze = subprocess.run( + [sys.executable, "-m", "pip", "freeze"], + check=True, capture_output=True, text=True).stdout + return { + "git_commit": git_output("rev-parse", "HEAD"), + "protocol_path": PROTOCOL, + "protocol_sha256": sha256(PROTOCOL), + "tracked_files": { + os.path.relpath(path, ROOT): sha256(path) for path in paths}, + "python": sys.executable, + "environment_freeze_sha256": text_sha256(freeze), + "environment_freeze": freeze.splitlines(), + } + + +def rate_tag(rate): + return f"{rate:g}".replace(".", "p") + + +def p1_command(method, rate): + name = f"transformer-p1-{method}-d4-lr{rate_tag(rate)}" + output = os.path.join(RESULT_ROOT, "p1", name + ".json") + command = [ + sys.executable, + "experiments/transformer_crossover_native.py", + "--method", method, + "--out", output, + "--device", "cuda", + "--depth", "4", + "--width", "128", + "--heads", "4", + "--mlp_ratio", "4", + "--context_length", "64", + "--batch_size", "32", + "--eval_batch_size", "32", + "--train_steps", "1000", + "--lr", str(rate), + "--schedule", "constant", + "--min_lr", "0.0001", + "--warmup_steps", "100", + "--weight_decay", "0.1", + "--run_seed", "0", + "--model_seed", "2027", + "--loader_seed", "0", + "--negative_seed", "5001", + "--ep_sign_seed", "5002", + "--traffic_ratio", "4.0", + "--ff_threshold", "2.0", + "--ff_score_from_layer", "1", + "--ep_beta", "0.5", + "--ep_dt", "0.5", + "--ep_free_steps", "20", + "--ep_nudge_steps", "4", + "--dp_alpha", "0.0", + "--dp_beta", "0.1", + "--dp_inference_passes", "16", + "--eval_every", "0", + "--log_every", "50", + "--max_val_batches", "0", + ] + return { + "stage": "p1", + "method": method, + "architecture": "transformer4", + "rate": rate, + "experiment_name": name, + "output": output, + "timeout_seconds": 48 * 60 * 60, + "command": command, + } + + +def p1_jobs(): + # Interleave the expensive methods so two shards do not inherit all state + # jobs from contiguous method groups. + order = ( + "ep", "dualprop", "ff", "pepita", "dfa", "fa", + "clean_kp", "sdil", "bp", + ) + jobs = [] + for grid_index in range(3): + for method in order: + jobs.append(p1_command( + method, RATE_GRIDS[method][grid_index])) + return jobs + + +def physical_gpu_report(dry_run=False): + visible = os.environ.get("CUDA_VISIBLE_DEVICES") + if dry_run: + return { + "cuda_visible_devices": visible, + "physical_gpu_index": None, + "physical_gpu_uuid": None, + } + if visible not in {"5", "7"}: + raise RuntimeError( + "formal Transformer jobs require CUDA_VISIBLE_DEVICES=5 or 7") + query = subprocess.run([ + "nvidia-smi", + "--query-gpu=index,uuid,name", + "--format=csv,noheader,nounits", + ], check=True, capture_output=True, text=True).stdout.splitlines() + rows = {} + for line in query: + index, uuid, name = [value.strip() for value in line.split(",", 2)] + rows[index] = {"uuid": uuid, "name": name} + if visible not in rows: + raise RuntimeError(f"physical GPU {visible} not found") + return { + "cuda_visible_devices": visible, + "physical_gpu_index": int(visible), + "physical_gpu_uuid": rows[visible]["uuid"], + "physical_gpu_name": rows[visible]["name"], + } + + +def run_job(job, source, gpu, dry_run): + manifest_path = job["output"] + ".manifest.json" + if os.path.exists(manifest_path): + print(f"preserving {job['experiment_name']}", flush=True) + return + if os.path.exists(job["output"]): + raise RuntimeError( + f"orphaned output requires audit: {job['output']}") + print("RUN", " ".join(job["command"]), flush=True) + if dry_run: + return + os.makedirs(os.path.dirname(job["output"]), exist_ok=True) + started = time.time() + try: + result = subprocess.run( + job["command"], cwd=ROOT, timeout=job["timeout_seconds"]) + return_code = result.returncode + status = "completed" if return_code == 0 else "nonzero_exit" + except subprocess.TimeoutExpired: + return_code = None + status = "timeout" + output_exists = os.path.isfile(job["output"]) + if status == "completed" and not output_exists: + status = "missing_output" + manifest = { + **job, + "source": source, + "hardware_lock": gpu, + "status": status, + "return_code": return_code, + "output_exists": output_exists, + "output_sha256": sha256(job["output"]) if output_exists else None, + "driver_wall_seconds": time.time() - started, + "completed_unix_time": time.time(), + } + with open(manifest_path, "w", encoding="utf-8") as handle: + json.dump(manifest, handle, indent=2, sort_keys=True) + handle.write("\n") + print( + f"DONE status={status} {job['experiment_name']}", flush=True) + + +def main(): + parser = argparse.ArgumentParser() + parser.add_argument("--stage", choices=("p1",), default="p1") + parser.add_argument("--shard-index", type=int, default=0) + parser.add_argument("--num-shards", type=int, default=1) + parser.add_argument("--method", choices=METHODS) + parser.add_argument("--dry-run", action="store_true") + args = parser.parse_args() + if not 0 <= args.shard_index < args.num_shards: + raise ValueError("invalid shard") + source = source_report() + gpu = physical_gpu_report(args.dry_run) + jobs = p1_jobs() + if args.method: + jobs = [job for job in jobs if job["method"] == args.method] + jobs = [ + job for index, job in enumerate(jobs) + if index % args.num_shards == args.shard_index] + if not jobs: + raise ValueError("no jobs selected") + for job in jobs: + run_job(job, source, gpu, args.dry_run) + + +if __name__ == "__main__": + main() |