diff options
| -rw-r--r-- | TRANSFORMER_CROSSOVER.md | 16 | ||||
| -rw-r--r-- | experiments/analyze_transformer_crossover_t2.py | 323 | ||||
| -rwxr-xr-x | experiments/finalize_accept.sh | 4 | ||||
| -rw-r--r-- | experiments/transformer_crossover_native.py | 1 | ||||
| -rw-r--r-- | experiments/transformer_crossover_t2.py | 392 | ||||
| -rw-r--r-- | experiments/transformer_crossover_t2_smoke.py | 65 |
6 files changed, 801 insertions, 0 deletions
diff --git a/TRANSFORMER_CROSSOVER.md b/TRANSFORMER_CROSSOVER.md index 8be1e72..019c0e6 100644 --- a/TRANSFORMER_CROSSOVER.md +++ b/TRANSFORMER_CROSSOVER.md @@ -181,6 +181,22 @@ T2 is a single-seed scaling panel, not the final uncertainty estimate. An untouched multi-seed T3 registry must contain every one of the same 27 cells; winner-only or SDIL-only replication cannot support the scaling claim. +The executable T2 contract is +`experiments/transformer_crossover_t2.py`. It refuses an incomplete T1 +selector, requires the selector to be committed, binds its hash, selected +rates, clean source, environment freeze, and GPU5/7 allowlist into a new launch +lock, and enumerates all 27 method--depth cells before sharding. Expensive EP, +Dual Propagation, and Forward--Forward cells are interleaved across shards. +`experiments/analyze_transformer_crossover_t2.py` requires all manifests, +retains timeout and nonfinite outcomes, rejects data/source/architecture/test +drift, and checks token presentations, candidate evaluations, relaxation +passes, local VJPs, logical queries, and enumerated MACs. The immutable +registry can be checked before T1 finishes with: + +```bash +python experiments/transformer_crossover_t2_smoke.py +``` + ## Metrics and cost accounting The primary task metric is full-validation next-token NLL; perplexity and diff --git a/experiments/analyze_transformer_crossover_t2.py b/experiments/analyze_transformer_crossover_t2.py new file mode 100644 index 0000000..5eeb226 --- /dev/null +++ b/experiments/analyze_transformer_crossover_t2.py @@ -0,0 +1,323 @@ +#!/usr/bin/env python3 +"""Audit the complete failure-retaining 27-cell Transformer T2 panel.""" +import argparse +import hashlib +import json +import math +import os +import sys + + +ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__))) +sys.path.insert(0, ROOT) +from experiments.transformer_crossover_t2 import ( + DEFAULT_SELECTOR, + DEPTHS, + METHODS, + registry_sha256, + selector_report, + t2_jobs, +) + + +TRAIN_HASH = ( + "6ec305602a99ac2802745a134e1f5e33e2231b4855525b00b9aebb730ac2626f" +) +VALIDATION_HASH = ( + "d37d30cc0c8327c270d493299c3dca54135f6d5f1c9ef60cda78076e311204b1" +) +EXPECTED_TRAIN_TOKENS = 5000 * 32 * 64 +EXPECTED_VALIDATION_TOKENS = 1742 * 64 +PARAMETERS = {4: 813_568, 8: 1_602_048, 12: 2_390_528} + + +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 read_json(path): + with open(path, encoding="utf-8") as handle: + return json.load(handle) + + +def expected_work(record, job): + work = record["work"] + method = job["method"] + depth = job["depth"] + train = EXPECTED_TRAIN_TOKENS + validation = EXPECTED_VALIDATION_TOKENS + presentations = 2 * train if method in ("pepita", "ff") else train + candidate = 65 * validation if method == "ff" else 0 + relaxation = 0 + local_vjp = 0 + if method == "dualprop": + relaxation = 16 * train + local_vjp = train * (depth + 1) * 16 + elif method == "ep": + relaxation = 24 * train + 20 * validation + local_vjp = train * (depth + 1) * 24 + assert work["ordinary_training_tokens"] == train + assert work["ordinary_validation_tokens"] == validation + assert work["training_token_presentations"] == presentations + assert work["candidate_token_presentations"] == candidate + assert work["relaxation_token_passes"] == relaxation + assert work["local_vjp_token_evaluations"] == local_vjp + assert work["logical_task_loss_queries"] == 0 + assert work["completed_optimizer_steps"] == 5000 + assert work["forward_parameter_count"] == PARAMETERS[depth] + assert work["full_forward_token_passes"] > 0 + assert work["enumerated_full_forward_macs"] > 0 + if method in ("dfa", "pepita", "ff", "dualprop", "ep"): + assert ( + work["local_block_token_evaluations"] > 0 + or method == "ff" + ) + return work + + +def audit_completed(job, manifest, source): + assert manifest["output_exists"] is True + assert manifest["output_sha256"] == sha256(job["output"]) + record = read_json(job["output"]) + provenance = record["provenance"] + assert provenance["git_commit"] == source["git_commit"] + assert provenance["git_tracked_dirty"] is False + args = record["args"] + expected_args = { + "method": job["method"], + "depth": job["depth"], + "width": 128, + "heads": 4, + "mlp_ratio": 4, + "context_length": 64, + "batch_size": 32, + "eval_batch_size": 32, + "train_steps": 5000, + "lr": job["rate"], + "schedule": "cosine", + "min_lr": job["rate"] * 0.1, + "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, + "max_val_batches": 0, + } + for key, expected in expected_args.items(): + assert args[key] == expected, ( + f"{job['experiment_name']}: argument drift in {key}" + ) + assert record["protocol_family"] == ( + "transformer_local_learning_crossover" + ) + dataset = record["dataset"] + assert dataset["train_sha256"] == TRAIN_HASH + assert dataset["validation_sha256"] == VALIDATION_HASH + architecture = record["architecture"] + assert architecture["depth"] == job["depth"] + assert architecture["width"] == 128 + assert architecture["heads"] == 4 + assert architecture["context_length"] == 64 + assert architecture["forward_parameter_count"] == PARAMETERS[job["depth"]] + evaluation = record["evaluation_protocol"] + assert evaluation["split"] == "validation" + assert evaluation["test_evaluations"] == 0 + assert evaluation["test_used_for_selection"] is False + assert len(record["validation"]) == 1 + final = record["final"] + assert final["tokens"] == EXPECTED_VALIDATION_TOKENS + assert final["step"] <= 5000 + first_nonfinite = record["first_nonfinite_step"] + nll = float(final["nll"]) + perplexity = float(final["perplexity"]) + accuracy = float(final["accuracy"]) + is_finite = ( + first_nonfinite is None + and final["step"] == 5000 + and math.isfinite(nll) + and math.isfinite(perplexity) + and math.isfinite(accuracy) + ) + work = None + if is_finite: + work = expected_work(record, job) + else: + assert record["work"]["ordinary_training_tokens"] <= ( + EXPECTED_TRAIN_TOKENS + ) + assert record["work"]["logical_task_loss_queries"] == 0 + hardware = record["hardware"] + assert hardware["cuda_visible_devices"] in ("5", "7") + assert hardware["device_name"] == "NVIDIA GeForce GTX 1080" + assert hardware["peak_memory_allocated_bytes"] is not None + assert hardware["peak_memory_reserved_bytes"] is not None + return { + "finite": is_finite, + "first_nonfinite_step": first_nonfinite, + "completed_optimizer_steps": + int(record["work"]["completed_optimizer_steps"]), + "final_validation_nll": nll if math.isfinite(nll) else None, + "final_validation_perplexity": + perplexity if math.isfinite(perplexity) else None, + "final_validation_accuracy": + accuracy if math.isfinite(accuracy) else None, + "forward_parameter_count": + int(architecture["forward_parameter_count"]), + "feedback_parameter_count": + int(architecture["feedback_parameter_count"]), + "peak_memory_allocated_bytes": + int(hardware["peak_memory_allocated_bytes"]), + "total_wall_seconds": float(record["total_wall_seconds"]), + "work": work if work is not None else record["work"], + } + + +def audit_job(job, source, selector): + manifest_path = job["output"] + ".manifest.json" + if not os.path.isfile(manifest_path): + raise AssertionError(f"missing T2 manifest: {job['experiment_name']}") + manifest = read_json(manifest_path) + for key in ( + "stage", + "method", + "architecture", + "depth", + "rate", + "experiment_name", + "output", + "timeout_seconds", + "command", + ): + assert manifest[key] == job[key], ( + f"{job['experiment_name']}: manifest drift in {key}" + ) + assert manifest["source"] == source + assert manifest["selector"] == selector + hardware = manifest["hardware_lock"] + assert hardware["physical_gpu_index"] in (5, 7) + assert hardware["physical_gpu_uuid"] + common = { + "cell_id": f"transformer{job['depth']}::{job['method']}", + "method": job["method"], + "depth": job["depth"], + "rate": job["rate"], + "status": manifest["status"], + "manifest": os.path.relpath(manifest_path, ROOT), + "driver_wall_seconds": float(manifest["driver_wall_seconds"]), + "physical_gpu_index": hardware["physical_gpu_index"], + "physical_gpu_uuid": hardware["physical_gpu_uuid"], + "output_sha256": manifest["output_sha256"], + } + if manifest["status"] == "completed": + return { + **common, + **audit_completed(job, manifest, source), + } + assert manifest["status"] in { + "timeout", "nonzero_exit", "missing_output" + } + if manifest["output_exists"]: + assert manifest["output_sha256"] == sha256(job["output"]) + else: + assert manifest["output_sha256"] is None + return { + **common, + "finite": False, + "first_nonfinite_step": None, + "completed_optimizer_steps": None, + "final_validation_nll": None, + "final_validation_perplexity": None, + "final_validation_accuracy": None, + "forward_parameter_count": None, + "feedback_parameter_count": None, + "peak_memory_allocated_bytes": None, + "total_wall_seconds": None, + "work": None, + } + + +def main(): + parser = argparse.ArgumentParser() + parser.add_argument("--selector", default=DEFAULT_SELECTOR) + parser.add_argument( + "--out", default="results/transformer_crossover/t2_audit.json" + ) + args = parser.parse_args() + selector = selector_report(args.selector) + jobs = t2_jobs(selector["selected_rates"]) + launch_path = os.path.join( + ROOT, "results", "transformer_crossover", "t2_launch.json" + ) + assert os.path.isfile(launch_path), "missing Transformer T2 launch lock" + launch = read_json(launch_path) + assert launch["stage"] == "t2" + assert launch["selector"] == selector + assert launch["registry_sha256"] == registry_sha256(jobs) + assert launch["num_jobs"] == 27 + assert launch["allowed_physical_gpus"] == [5, 7] + source = launch["source"] + records = [audit_job(job, source, selector) for job in jobs] + assert len(records) == 27 + assert len({record["cell_id"] for record in records}) == 27 + assert { + (record["method"], record["depth"]) for record in records + } == {(method, depth) for method in METHODS for depth in DEPTHS} + failures = [ + record["cell_id"] for record in records if not record["finite"] + ] + report = { + "audit_status": "passed", + "stage": "transformer_crossover_t2", + "complete_grid": True, + "failure_retaining": True, + "num_expected_cells": 27, + "num_audited_cells": len(records), + "num_finite_cells": len(records) - len(failures), + "failed_or_incomplete_cells": failures, + "test_policy": "none", + "source": source, + "selector": selector, + "launch_lock": { + "path": os.path.relpath(launch_path, ROOT), + "sha256": sha256(launch_path), + "registry_sha256": launch["registry_sha256"], + }, + "records": records, + } + os.makedirs(os.path.dirname(os.path.abspath(args.out)), exist_ok=True) + with open(args.out, "w", encoding="utf-8") as handle: + json.dump(report, handle, indent=2, sort_keys=True) + handle.write("\n") + print( + json.dumps( + { + "audit_status": report["audit_status"], + "num_audited_cells": len(records), + "num_finite_cells": report["num_finite_cells"], + "failed_or_incomplete_cells": failures, + }, + indent=2, + sort_keys=True, + ) + ) + + +if __name__ == "__main__": + main() diff --git a/experiments/finalize_accept.sh b/experiments/finalize_accept.sh index 9a511b9..e445caa 100755 --- a/experiments/finalize_accept.sh +++ b/experiments/finalize_accept.sh @@ -25,6 +25,8 @@ experiments/finalize_claims.sh experiments/oral_a_dynamic_scaling_smoke.py /home/yurenh2/miniconda3/envs/ep_pascal/bin/python3 \ experiments/resnet_crossover_r2_smoke.py +/home/yurenh2/miniconda3/envs/ep_pascal/bin/python3 \ + experiments/transformer_crossover_t2_smoke.py /home/yurenh2/miniconda3/envs/ep_pascal/bin/python3 -m py_compile \ experiments/bci_td_run.py experiments/analyze_bci_td_development.py \ experiments/bci_td_confirmation.py \ @@ -42,6 +44,8 @@ experiments/finalize_claims.sh experiments/analyze_oral_a_dynamic_scaling_v2.py \ experiments/resnet_crossover_r2.py \ experiments/analyze_resnet_crossover_r2.py \ + experiments/transformer_crossover_t2.py \ + experiments/analyze_transformer_crossover_t2.py \ experiments/plot_resnet_confirmation.py \ experiments/plot_oral_a_scaling.py \ experiments/plot_bci_v2_confirmation.py \ diff --git a/experiments/transformer_crossover_native.py b/experiments/transformer_crossover_native.py index 7ccb8fb..ca47751 100644 --- a/experiments/transformer_crossover_native.py +++ b/experiments/transformer_crossover_native.py @@ -365,6 +365,7 @@ def work_report( "candidate_token_presentations": sum( row["candidate_token_presentations"] for row in validation_evaluations), + "logical_task_loss_queries": 0, "relaxation_token_passes": relaxation, "local_vjp_token_evaluations": local_vjps, "full_forward_token_passes": full_forward_token_passes, diff --git a/experiments/transformer_crossover_t2.py b/experiments/transformer_crossover_t2.py new file mode 100644 index 0000000..17f61a2 --- /dev/null +++ b/experiments/transformer_crossover_t2.py @@ -0,0 +1,392 @@ +#!/usr/bin/env python3 +"""Immutable driver for the complete 27-cell Transformer T2 crossover.""" +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") +DEFAULT_SELECTOR = os.path.join(RESULT_ROOT, "p1_selector.json") +METHODS = ( + "bp", "fa", "dfa", "pepita", "ff", "ep", "dualprop", + "clean_kp", "sdil", +) +DEPTHS = (4, 8, 12) +ORDER = ( + ("ep", 12), + ("dualprop", 12), + ("ff", 12), + ("ep", 8), + ("dualprop", 8), + ("ff", 8), + ("ep", 4), + ("dualprop", 4), + ("ff", 4), + ("pepita", 12), + ("sdil", 12), + ("clean_kp", 12), + ("fa", 12), + ("dfa", 12), + ("bp", 12), + ("pepita", 8), + ("sdil", 8), + ("clean_kp", 8), + ("fa", 8), + ("dfa", 8), + ("bp", 8), + ("pepita", 4), + ("sdil", 4), + ("clean_kp", 4), + ("fa", 4), + ("dfa", 4), + ("bp", 4), +) + + +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 read_json(path): + with open(path, encoding="utf-8") as handle: + return json.load(handle) + + +def git_output(*args): + return subprocess.run( + ["git", *args], + cwd=ROOT, + check=True, + capture_output=True, + text=True, + ).stdout.strip() + + +def rate_tag(rate): + return f"{rate:g}".replace(".", "p") + + +def selector_report(path): + selector = read_json(path) + if not ( + selector.get("gate") == "pass" + and selector.get("stage") == "transformer_crossover_p1" + and selector.get("num_expected_records") == 27 + and selector.get("num_audited_records") == 27 + and selector.get("missing_experiments") == [] + and selector.get("test_policy") == "none" + and set(selector.get("selected", {})) == set(METHODS) + ): + raise RuntimeError( + "Transformer T2 requires the complete passed T1 selector" + ) + rates = {} + for method in METHODS: + rate = float(selector["selected"][method]["rate"]) + if not rate > 0: + raise RuntimeError(f"invalid selected Transformer rate: {method}") + rates[method] = rate + return { + "path": os.path.relpath(os.path.abspath(path), ROOT), + "sha256": sha256(path), + "p1_source": selector["source"], + "selected_rates": rates, + } + + +def source_report(selector_path): + if git_output("status", "--porcelain", "--untracked-files=no"): + raise RuntimeError("Transformer T2 requires clean tracked source") + paths = [ + PROTOCOL, + os.path.abspath(__file__), + os.path.join( + ROOT, "experiments", "analyze_transformer_crossover_t2.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"), + os.path.abspath(selector_path), + ] + for path in paths: + relative = os.path.relpath(path, ROOT) + tracked = subprocess.run( + ["git", "ls-files", "--error-unmatch", relative], + cwd=ROOT, + capture_output=True, + ).returncode == 0 + if not tracked: + raise RuntimeError(f"untracked Transformer T2 source: {relative}") + 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 t2_command(method, depth, rate): + name = f"transformer-t2-{method}-d{depth}-lr{rate_tag(rate)}" + output = os.path.join(RESULT_ROOT, "t2", name + ".json") + command = [ + sys.executable, + "experiments/transformer_crossover_native.py", + "--method", method, + "--out", output, + "--device", "cuda", + "--depth", str(depth), + "--width", "128", + "--heads", "4", + "--mlp_ratio", "4", + "--context_length", "64", + "--batch_size", "32", + "--eval_batch_size", "32", + "--train_steps", "5000", + "--lr", str(rate), + "--schedule", "cosine", + "--min_lr", str(rate * 0.1), + "--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": "t2", + "method": method, + "architecture": f"transformer{depth}", + "depth": depth, + "rate": rate, + "experiment_name": name, + "output": output, + "timeout_seconds": 48 * 60 * 60, + "command": command, + } + + +def t2_jobs(selected_rates): + if set(selected_rates) != set(METHODS): + raise ValueError("T2 requires one selected rate for every method") + jobs = [ + t2_command(method, depth, float(selected_rates[method])) + for method, depth in ORDER + ] + identifiers = [job["experiment_name"] for job in jobs] + cells = {(job["method"], job["depth"]) for job in jobs} + if ( + len(jobs) != 27 + or len(set(identifiers)) != 27 + or cells != {(method, depth) for method in METHODS for depth in DEPTHS} + ): + raise AssertionError( + "Transformer T2 registry must contain 27 unique cells" + ) + return jobs + + +def registry_sha256(jobs): + encoded = json.dumps( + jobs, sort_keys=True, separators=(",", ":") + ).encode("utf-8") + return hashlib.sha256(encoded).hexdigest() + + +def ensure_launch(source, selector, jobs): + path = os.path.join(RESULT_ROOT, "t2_launch.json") + expected = { + "stage": "t2", + "source": source, + "selector": selector, + "registry_sha256": registry_sha256(jobs), + "num_jobs": len(jobs), + "allowed_physical_gpus": [5, 7], + } + if os.path.isfile(path): + if read_json(path) != expected: + raise RuntimeError("Transformer T2 launch lock drift") + return path + os.makedirs(os.path.dirname(path), exist_ok=True) + with open(path, "w", encoding="utf-8") as handle: + json.dump(expected, handle, indent=2, sort_keys=True) + handle.write("\n") + return path + + +def assert_source_unchanged(source): + if git_output("status", "--porcelain", "--untracked-files=no"): + raise RuntimeError("tracked source changed after Transformer T2 launch") + if git_output("rev-parse", "HEAD") != source["git_commit"]: + raise RuntimeError("source commit changed after Transformer T2 launch") + for relative, expected_hash in source["tracked_files"].items(): + if sha256(os.path.join(ROOT, relative)) != expected_hash: + raise RuntimeError(f"Transformer T2 source drift: {relative}") + freeze = subprocess.run( + [sys.executable, "-m", "pip", "freeze"], + check=True, + capture_output=True, + text=True, + ).stdout + if text_sha256(freeze) != source["environment_freeze_sha256"]: + raise RuntimeError("Transformer T2 environment drift") + + +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("Transformer T2 requires physical GPU 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, selector, gpu, dry_run): + if not dry_run: + assert_source_unchanged(source) + if sha256(os.path.join(ROOT, selector["path"])) != selector["sha256"]: + raise RuntimeError("Transformer T2 selector changed after launch") + 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 T2 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, + "selector": selector, + "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("--selector", default=DEFAULT_SELECTOR) + 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("--depth", type=int, choices=DEPTHS) + parser.add_argument("--dry-run", action="store_true") + args = parser.parse_args() + if not 0 <= args.shard_index < args.num_shards: + raise ValueError("invalid Transformer T2 shard") + selector = selector_report(args.selector) + source = source_report(args.selector) + gpu = physical_gpu_report(args.dry_run) + complete_jobs = t2_jobs(selector["selected_rates"]) + if not args.dry_run: + launch = ensure_launch(source, selector, complete_jobs) + print(f"T2 launch lock: {launch}", flush=True) + jobs = complete_jobs + if args.method: + jobs = [job for job in jobs if job["method"] == args.method] + if args.depth is not None: + jobs = [job for job in jobs if job["depth"] == args.depth] + jobs = [ + job + for index, job in enumerate(jobs) + if index % args.num_shards == args.shard_index + ] + if not jobs: + raise ValueError("no Transformer T2 jobs selected") + for job in jobs: + run_job(job, source, selector, gpu, args.dry_run) + + +if __name__ == "__main__": + main() diff --git a/experiments/transformer_crossover_t2_smoke.py b/experiments/transformer_crossover_t2_smoke.py new file mode 100644 index 0000000..febca40 --- /dev/null +++ b/experiments/transformer_crossover_t2_smoke.py @@ -0,0 +1,65 @@ +#!/usr/bin/env python3 +"""Deterministic audit of the complete Transformer T2 registry.""" +import os +import sys + +sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__)))) +from experiments.transformer_crossover_t2 import DEPTHS, METHODS, t2_jobs + + +RATES = { + "bp": 0.001, + "fa": 0.001, + "dfa": 0.001, + "pepita": 0.0003, + "ff": 0.0003, + "ep": 0.0003, + "dualprop": 0.0003, + "clean_kp": 0.001, + "sdil": 0.001, +} + + +def value(command, flag): + index = command.index(flag) + return command[index + 1] + + +def main(): + jobs = t2_jobs(RATES) + assert len(jobs) == 27 + assert {(job["method"], job["depth"]) for job in jobs} == { + (method, depth) for method in METHODS for depth in DEPTHS + } + assert len({job["experiment_name"] for job in jobs}) == 27 + assert len({job["output"] for job in jobs}) == 27 + for job in jobs: + command = job["command"] + rate = RATES[job["method"]] + assert value(command, "--depth") == str(job["depth"]) + assert value(command, "--train_steps") == "5000" + assert value(command, "--batch_size") == "32" + assert value(command, "--context_length") == "64" + assert value(command, "--lr") == str(rate) + assert value(command, "--schedule") == "cosine" + assert value(command, "--min_lr") == str(rate * 0.1) + assert value(command, "--warmup_steps") == "100" + assert value(command, "--eval_every") == "0" + assert value(command, "--max_val_batches") == "0" + assert job["timeout_seconds"] == 48 * 60 * 60 + assert "--test" not in command + assert value(command, "--traffic_ratio") == "4.0" + assert value(command, "--ep_free_steps") == "20" + assert value(command, "--ep_nudge_steps") == "4" + assert value(command, "--dp_inference_passes") == "16" + shard0 = jobs[0::2] + shard1 = jobs[1::2] + assert len(shard0) == 14 and len(shard1) == 13 + for method in ("ep", "dualprop", "ff"): + assert any(job["method"] == method for job in shard0) + assert any(job["method"] == method for job in shard1) + print("Transformer T2 registry: 27/27 cells; schedules and shards exact") + + +if __name__ == "__main__": + main() |