Add free-GPU structural smoke runner
Browse files
training/run_fable_router_structural_smoke.py
ADDED
|
@@ -0,0 +1,479 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python
|
| 2 |
+
# /// script
|
| 3 |
+
# requires-python = ">=3.10"
|
| 4 |
+
# dependencies = [
|
| 5 |
+
# "torch>=2.4",
|
| 6 |
+
# "transformers>=4.56",
|
| 7 |
+
# "accelerate>=1.2",
|
| 8 |
+
# "safetensors>=0.5",
|
| 9 |
+
# "huggingface_hub>=0.34",
|
| 10 |
+
# "trackio>=0.3",
|
| 11 |
+
# ]
|
| 12 |
+
# ///
|
| 13 |
+
"""Run one provenance-bound, non-routing Fable donor-bank structural smoke.
|
| 14 |
+
|
| 15 |
+
This is deliberately not a quality training run. It validates all 2,880
|
| 16 |
+
bank tensors but attaches only the final host layer, keeping a real gradient
|
| 17 |
+
and counterfactual test feasible on a free 16 GiB GPU.
|
| 18 |
+
"""
|
| 19 |
+
from __future__ import annotations
|
| 20 |
+
|
| 21 |
+
import argparse
|
| 22 |
+
import json
|
| 23 |
+
import os
|
| 24 |
+
import platform
|
| 25 |
+
import shutil
|
| 26 |
+
import sys
|
| 27 |
+
import time
|
| 28 |
+
import traceback
|
| 29 |
+
from datetime import datetime, timezone
|
| 30 |
+
from pathlib import Path
|
| 31 |
+
from typing import Any
|
| 32 |
+
|
| 33 |
+
import torch
|
| 34 |
+
import torch.nn.functional as F
|
| 35 |
+
from huggingface_hub import HfApi, hf_hub_download
|
| 36 |
+
from safetensors.torch import load_file, save_file
|
| 37 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 38 |
+
|
| 39 |
+
HERE = Path(__file__).resolve().parent
|
| 40 |
+
if str(HERE) not in sys.path:
|
| 41 |
+
sys.path.insert(0, str(HERE))
|
| 42 |
+
|
| 43 |
+
from fable_router_common import ( # noqa: E402
|
| 44 |
+
iter_jsonl,
|
| 45 |
+
read_json,
|
| 46 |
+
selected_expert_ids,
|
| 47 |
+
sha256,
|
| 48 |
+
validate_bank_header,
|
| 49 |
+
validate_curriculum_row,
|
| 50 |
+
verify_file,
|
| 51 |
+
)
|
| 52 |
+
from fable_router_hybrid import ( # noqa: E402
|
| 53 |
+
FrozenExpertRouterBlock,
|
| 54 |
+
assert_trainable_isolation,
|
| 55 |
+
attach_router_block,
|
| 56 |
+
benefit_targets,
|
| 57 |
+
benefit_weighted_router_loss,
|
| 58 |
+
freeze_except_routers,
|
| 59 |
+
load_router_state_dict,
|
| 60 |
+
router_state_dict,
|
| 61 |
+
)
|
| 62 |
+
|
| 63 |
+
|
| 64 |
+
def now() -> str:
|
| 65 |
+
return datetime.now(timezone.utc).isoformat()
|
| 66 |
+
|
| 67 |
+
|
| 68 |
+
def download(repo: str, revision: str, filename: str, repo_type: str, token: str | None) -> Path:
|
| 69 |
+
return Path(
|
| 70 |
+
hf_hub_download(
|
| 71 |
+
repo_id=repo,
|
| 72 |
+
revision=revision,
|
| 73 |
+
filename=filename,
|
| 74 |
+
repo_type=repo_type,
|
| 75 |
+
token=token,
|
| 76 |
+
)
|
| 77 |
+
)
|
| 78 |
+
|
| 79 |
+
|
| 80 |
+
def render_and_tokenize(tokenizer: Any, messages: list[dict[str, Any]]) -> list[int]:
|
| 81 |
+
# Transformers 5 can return a BatchEncoding from apply_chat_template with
|
| 82 |
+
# tokenize=True, whose len() is the number of fields rather than tokens.
|
| 83 |
+
rendered = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=False)
|
| 84 |
+
return list(tokenizer(rendered, add_special_tokens=False)["input_ids"])
|
| 85 |
+
|
| 86 |
+
|
| 87 |
+
def select_complete_rows(path: Path, tokenizer: Any, maximum_tokens: int) -> list[dict[str, Any]]:
|
| 88 |
+
wanted = ("host_preservation", "verified_expert")
|
| 89 |
+
best: dict[str, tuple[int, dict[str, Any], list[int]]] = {}
|
| 90 |
+
for row in iter_jsonl(path):
|
| 91 |
+
validate_curriculum_row(row, "train")
|
| 92 |
+
lane = str(row["lane"])
|
| 93 |
+
if lane not in wanted:
|
| 94 |
+
continue
|
| 95 |
+
ids = render_and_tokenize(tokenizer, row["messages"])
|
| 96 |
+
if len(ids) > maximum_tokens:
|
| 97 |
+
continue
|
| 98 |
+
previous = best.get(lane)
|
| 99 |
+
if previous is None or len(ids) < previous[0]:
|
| 100 |
+
best[lane] = (len(ids), row, ids)
|
| 101 |
+
missing = sorted(set(wanted) - set(best))
|
| 102 |
+
if missing:
|
| 103 |
+
raise RuntimeError(f"no complete <= {maximum_tokens}-token rows for lanes {missing}")
|
| 104 |
+
return [
|
| 105 |
+
{"id": best[lane][1]["id"], "lane": lane, "inputIds": best[lane][2]}
|
| 106 |
+
for lane in wanted
|
| 107 |
+
]
|
| 108 |
+
|
| 109 |
+
|
| 110 |
+
def batch_rows(rows: list[dict[str, Any]], pad_token_id: int, device: torch.device) -> dict[str, torch.Tensor]:
|
| 111 |
+
maximum = max(len(row["inputIds"]) for row in rows)
|
| 112 |
+
input_ids, attention_mask, labels = [], [], []
|
| 113 |
+
for row in rows:
|
| 114 |
+
ids = list(row["inputIds"])
|
| 115 |
+
padding = maximum - len(ids)
|
| 116 |
+
input_ids.append(ids + [pad_token_id] * padding)
|
| 117 |
+
attention_mask.append([1] * len(ids) + [0] * padding)
|
| 118 |
+
labels.append(ids + [-100] * padding)
|
| 119 |
+
return {
|
| 120 |
+
"input_ids": torch.tensor(input_ids, dtype=torch.long, device=device),
|
| 121 |
+
"attention_mask": torch.tensor(attention_mask, dtype=torch.long, device=device),
|
| 122 |
+
"labels": torch.tensor(labels, dtype=torch.long, device=device),
|
| 123 |
+
}
|
| 124 |
+
|
| 125 |
+
|
| 126 |
+
def token_nll(logits: torch.Tensor, labels: torch.Tensor) -> torch.Tensor:
|
| 127 |
+
shifted_logits = logits[:, :-1, :].float()
|
| 128 |
+
shifted_labels = labels[:, 1:]
|
| 129 |
+
losses = F.cross_entropy(
|
| 130 |
+
shifted_logits.reshape(-1, shifted_logits.shape[-1]),
|
| 131 |
+
shifted_labels.reshape(-1),
|
| 132 |
+
reduction="none",
|
| 133 |
+
ignore_index=-100,
|
| 134 |
+
).reshape(shifted_labels.shape)
|
| 135 |
+
return losses
|
| 136 |
+
|
| 137 |
+
|
| 138 |
+
def gpu_facts(minimum_vram_gib: float) -> dict[str, Any]:
|
| 139 |
+
if not torch.cuda.is_available():
|
| 140 |
+
raise RuntimeError("a CUDA GPU is required for the structural smoke")
|
| 141 |
+
properties = torch.cuda.get_device_properties(0)
|
| 142 |
+
total_gib = properties.total_memory / 2**30
|
| 143 |
+
if total_gib < minimum_vram_gib:
|
| 144 |
+
raise RuntimeError(f"GPU has {total_gib:.2f} GiB, requires at least {minimum_vram_gib:.2f} GiB")
|
| 145 |
+
return {
|
| 146 |
+
"name": properties.name,
|
| 147 |
+
"totalVramGiB": total_gib,
|
| 148 |
+
"capability": list(torch.cuda.get_device_capability(0)),
|
| 149 |
+
"bf16Supported": bool(torch.cuda.is_bf16_supported()),
|
| 150 |
+
}
|
| 151 |
+
|
| 152 |
+
|
| 153 |
+
def upload_evidence(output: Path, repo_id: str, token: str) -> dict[str, Any]:
|
| 154 |
+
api = HfApi(token=token)
|
| 155 |
+
api.create_repo(repo_id, repo_type="dataset", private=True, exist_ok=True)
|
| 156 |
+
info = api.repo_info(repo_id, repo_type="dataset")
|
| 157 |
+
if not bool(info.private):
|
| 158 |
+
raise RuntimeError(f"refusing to upload smoke evidence to public repo {repo_id}")
|
| 159 |
+
prefix = output.name
|
| 160 |
+
for item in output.iterdir():
|
| 161 |
+
if item.is_file():
|
| 162 |
+
api.upload_file(
|
| 163 |
+
path_or_fileobj=str(item),
|
| 164 |
+
path_in_repo=f"runs/{prefix}/{item.name}",
|
| 165 |
+
repo_id=repo_id,
|
| 166 |
+
repo_type="dataset",
|
| 167 |
+
commit_message=f"Persist non-routing structural smoke {prefix}",
|
| 168 |
+
)
|
| 169 |
+
return {"repo": repo_id, "private": True, "path": f"runs/{prefix}"}
|
| 170 |
+
|
| 171 |
+
|
| 172 |
+
def run(args: argparse.Namespace, result: dict[str, Any]) -> None:
|
| 173 |
+
config_path = args.config.resolve()
|
| 174 |
+
config = read_json(config_path)
|
| 175 |
+
if config.get("trainingAuthorized") is not False or config.get("nonRouting") is not True:
|
| 176 |
+
raise RuntimeError("cloud smoke config is not explicitly non-routing/training-disabled")
|
| 177 |
+
smoke = config["smoke"]
|
| 178 |
+
bank_config = config["banks"]["artifacts"].get(args.bank)
|
| 179 |
+
if not bank_config:
|
| 180 |
+
raise RuntimeError(f"unknown bank {args.bank}")
|
| 181 |
+
token = os.environ.get("HF_TOKEN")
|
| 182 |
+
if not token:
|
| 183 |
+
raise RuntimeError("HF_TOKEN is required to read private curriculum and persist evidence")
|
| 184 |
+
|
| 185 |
+
result["config"] = {"path": str(config_path), "sha256": sha256(config_path)}
|
| 186 |
+
result["bank"] = args.bank
|
| 187 |
+
result["platform"] = args.platform
|
| 188 |
+
result["gpu"] = gpu_facts(14.0)
|
| 189 |
+
free_disk_gib = shutil.disk_usage(args.work_dir).free / 2**30
|
| 190 |
+
result["freeDiskGiBBeforeDownloads"] = free_disk_gib
|
| 191 |
+
if free_disk_gib < float(smoke["minimumFreeDiskGiBBeforeBankDownload"]):
|
| 192 |
+
raise RuntimeError(
|
| 193 |
+
f"only {free_disk_gib:.2f} GiB free; requires "
|
| 194 |
+
f"{smoke['minimumFreeDiskGiBBeforeBankDownload']} GiB before downloads"
|
| 195 |
+
)
|
| 196 |
+
|
| 197 |
+
bank_repo = config["banks"]
|
| 198 |
+
curriculum = config["curriculum"]
|
| 199 |
+
bank_manifest = download(
|
| 200 |
+
bank_repo["repo"], bank_repo["revision"], bank_repo["manifestPath"], "model", token
|
| 201 |
+
)
|
| 202 |
+
if sha256(bank_manifest) != bank_repo["manifestSha256"]:
|
| 203 |
+
raise RuntimeError("frozen selected-bank manifest hash mismatch")
|
| 204 |
+
manifest = read_json(bank_manifest)
|
| 205 |
+
banks = {row["id"]: row for row in manifest["banks"]}
|
| 206 |
+
bank_definition = banks[args.bank]
|
| 207 |
+
|
| 208 |
+
warm_config = bank_repo["routerWarmstart"]
|
| 209 |
+
warmstart = download(bank_repo["repo"], bank_repo["revision"], warm_config["path"], "model", token)
|
| 210 |
+
verify_file(warmstart, int(warm_config["bytes"]), warm_config["sha256"])
|
| 211 |
+
bank_path = download(bank_repo["repo"], bank_repo["revision"], bank_config["path"], "model", token)
|
| 212 |
+
bank_artifact = verify_file(bank_path, int(bank_config["bytes"]), bank_config["sha256"])
|
| 213 |
+
bank_validation = validate_bank_header(bank_path, bank_definition)
|
| 214 |
+
result["bankArtifact"] = bank_artifact
|
| 215 |
+
result["bankValidation"] = bank_validation.as_dict()
|
| 216 |
+
|
| 217 |
+
train_config = curriculum["sftTrain"]
|
| 218 |
+
train_path = download(
|
| 219 |
+
curriculum["repo"], curriculum["revision"], train_config["path"], "dataset", token
|
| 220 |
+
)
|
| 221 |
+
result["curriculumArtifact"] = verify_file(
|
| 222 |
+
train_path, int(train_config["bytes"]), train_config["sha256"]
|
| 223 |
+
)
|
| 224 |
+
|
| 225 |
+
host = config["host"]
|
| 226 |
+
tokenizer = AutoTokenizer.from_pretrained(
|
| 227 |
+
host["repo"], revision=host["revision"], token=token, trust_remote_code=True,
|
| 228 |
+
fix_mistral_regex=True,
|
| 229 |
+
)
|
| 230 |
+
rows = select_complete_rows(train_path, tokenizer, int(smoke["maximumTokens"]))
|
| 231 |
+
result["retokenization"] = {
|
| 232 |
+
"maximumTokens": int(smoke["maximumTokens"]),
|
| 233 |
+
"truncated": False,
|
| 234 |
+
"rows": [{"id": row["id"], "lane": row["lane"], "tokens": len(row["inputIds"])} for row in rows],
|
| 235 |
+
}
|
| 236 |
+
|
| 237 |
+
device = torch.device("cuda:0")
|
| 238 |
+
dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
|
| 239 |
+
torch.cuda.reset_peak_memory_stats(device)
|
| 240 |
+
load_started = time.perf_counter()
|
| 241 |
+
model = AutoModelForCausalLM.from_pretrained(
|
| 242 |
+
host["repo"], revision=host["revision"], token=token, trust_remote_code=True,
|
| 243 |
+
torch_dtype=dtype, low_cpu_mem_usage=True, attn_implementation="sdpa",
|
| 244 |
+
).to(device)
|
| 245 |
+
model.config.use_cache = False
|
| 246 |
+
model.eval()
|
| 247 |
+
batch = batch_rows(rows, int(tokenizer.pad_token_id), device)
|
| 248 |
+
|
| 249 |
+
with torch.inference_mode():
|
| 250 |
+
baseline_logits = model(
|
| 251 |
+
input_ids=batch["input_ids"], attention_mask=batch["attention_mask"], use_cache=False
|
| 252 |
+
).logits.detach().cpu()
|
| 253 |
+
|
| 254 |
+
layer = int(smoke["attachedLayer"])
|
| 255 |
+
block = FrozenExpertRouterBlock(
|
| 256 |
+
selected_expert_ids(bank_definition, layer),
|
| 257 |
+
top_k=int(smoke["topK"]),
|
| 258 |
+
initial_scale=float(smoke["initialExpertScale"]),
|
| 259 |
+
maximum_scale=float(smoke["maximumExpertScale"]),
|
| 260 |
+
)
|
| 261 |
+
wrapper = attach_router_block(model, layer, block)
|
| 262 |
+
block.router.to(device=device, dtype=dtype)
|
| 263 |
+
block.expert_scale.data = block.expert_scale.data.to(device=device)
|
| 264 |
+
block.materialize_experts(bank_path, layer, device=device, dtype=dtype)
|
| 265 |
+
warm_tensors = load_file(str(warmstart), device="cpu")
|
| 266 |
+
block.load_router_warmstart(warm_tensors[f"layers.{layer}.router.gate.weight"])
|
| 267 |
+
block.enabled = False
|
| 268 |
+
with torch.inference_mode():
|
| 269 |
+
wrapped_disabled = model(
|
| 270 |
+
input_ids=batch["input_ids"], attention_mask=batch["attention_mask"], use_cache=False
|
| 271 |
+
).logits.detach().cpu()
|
| 272 |
+
identity_equal = torch.equal(baseline_logits, wrapped_disabled)
|
| 273 |
+
identity_max_difference = (baseline_logits.float() - wrapped_disabled.float()).abs().max().item()
|
| 274 |
+
result["hostDisabledIdentity"] = {
|
| 275 |
+
"exact": identity_equal,
|
| 276 |
+
"maximumAbsoluteDifference": identity_max_difference,
|
| 277 |
+
}
|
| 278 |
+
if not identity_equal:
|
| 279 |
+
raise RuntimeError(f"host-disabled identity failed: max difference {identity_max_difference}")
|
| 280 |
+
del baseline_logits, wrapped_disabled
|
| 281 |
+
|
| 282 |
+
counts = freeze_except_routers(model)
|
| 283 |
+
trainable_names = assert_trainable_isolation(model)
|
| 284 |
+
result["trainableIsolation"] = {"counts": counts, "names": trainable_names}
|
| 285 |
+
block.enabled = True
|
| 286 |
+
model.zero_grad(set_to_none=True)
|
| 287 |
+
gradient_output = model(**batch, use_cache=False)
|
| 288 |
+
gradient_output.loss.backward()
|
| 289 |
+
router_gradient = block.router.gate.weight.grad
|
| 290 |
+
scale_gradient = block.expert_scale.grad
|
| 291 |
+
gradient_gate = {
|
| 292 |
+
"routerFinite": bool(router_gradient is not None and torch.isfinite(router_gradient).all()),
|
| 293 |
+
"routerL1": float(router_gradient.float().abs().sum()) if router_gradient is not None else 0.0,
|
| 294 |
+
"scaleFinite": bool(scale_gradient is not None and torch.isfinite(scale_gradient).all()),
|
| 295 |
+
"scaleAbsolute": float(scale_gradient.float().abs()) if scale_gradient is not None else 0.0,
|
| 296 |
+
}
|
| 297 |
+
result["gradientGate"] = gradient_gate
|
| 298 |
+
if not all((gradient_gate["routerFinite"], gradient_gate["scaleFinite"])) or min(
|
| 299 |
+
gradient_gate["routerL1"], gradient_gate["scaleAbsolute"]
|
| 300 |
+
) <= 0:
|
| 301 |
+
raise RuntimeError(f"router gradient gate failed: {gradient_gate}")
|
| 302 |
+
|
| 303 |
+
expert_row = next(row for row in rows if row["lane"] == "verified_expert")
|
| 304 |
+
oracle_batch = batch_rows([expert_row], int(tokenizer.pad_token_id), device)
|
| 305 |
+
block.enabled = False
|
| 306 |
+
with torch.inference_mode():
|
| 307 |
+
host_logits = model(
|
| 308 |
+
input_ids=oracle_batch["input_ids"], attention_mask=oracle_batch["attention_mask"], use_cache=False
|
| 309 |
+
).logits
|
| 310 |
+
host_nll = token_nll(host_logits, oracle_batch["labels"])[0]
|
| 311 |
+
valid_tokens = oracle_batch["labels"][0, 1:] != -100
|
| 312 |
+
host_nll = host_nll[valid_tokens].detach().cpu()
|
| 313 |
+
|
| 314 |
+
candidate_indices = list(range(0, 32, max(1, 32 // int(smoke["oracleCandidateExperts"]))))[
|
| 315 |
+
: int(smoke["oracleCandidateExperts"])
|
| 316 |
+
]
|
| 317 |
+
best_candidate_nll = []
|
| 318 |
+
oracle_rows = []
|
| 319 |
+
block.enabled = True
|
| 320 |
+
for candidate in candidate_indices:
|
| 321 |
+
scale_losses = []
|
| 322 |
+
for scale in smoke["oracleProbeScales"]:
|
| 323 |
+
with block.forced_route(candidate, float(scale)), torch.inference_mode():
|
| 324 |
+
logits = model(
|
| 325 |
+
input_ids=oracle_batch["input_ids"],
|
| 326 |
+
attention_mask=oracle_batch["attention_mask"],
|
| 327 |
+
use_cache=False,
|
| 328 |
+
).logits
|
| 329 |
+
losses = token_nll(logits, oracle_batch["labels"])[0][valid_tokens].detach().cpu()
|
| 330 |
+
scale_losses.append(losses)
|
| 331 |
+
oracle_rows.append(
|
| 332 |
+
{
|
| 333 |
+
"localExpert": candidate,
|
| 334 |
+
"globalExpert": block.expert_ids[candidate],
|
| 335 |
+
"scale": float(scale),
|
| 336 |
+
"meanNll": float(losses.mean()),
|
| 337 |
+
"improvedTokenCount": int((losses < host_nll).sum()),
|
| 338 |
+
}
|
| 339 |
+
)
|
| 340 |
+
best_candidate_nll.append(torch.stack(scale_losses).min(dim=0).values)
|
| 341 |
+
candidate_matrix = torch.stack(best_candidate_nll, dim=-1)
|
| 342 |
+
targets = benefit_targets(host_nll, candidate_matrix, float(smoke["benefitMarginNats"]))
|
| 343 |
+
positive = targets != candidate_matrix.shape[-1]
|
| 344 |
+
|
| 345 |
+
block.enabled = True
|
| 346 |
+
with torch.no_grad():
|
| 347 |
+
_ = model(
|
| 348 |
+
input_ids=oracle_batch["input_ids"], attention_mask=oracle_batch["attention_mask"], use_cache=False
|
| 349 |
+
)
|
| 350 |
+
if block.last_trace is None:
|
| 351 |
+
raise RuntimeError("router did not retain a main-path trace")
|
| 352 |
+
router_logits = block.last_trace.logits[: targets.numel()].float()
|
| 353 |
+
ranking_loss = benefit_weighted_router_loss(router_logits, targets.to(device), host_nll.to(device), candidate_matrix.to(device))
|
| 354 |
+
result["counterfactualDiscovery"] = {
|
| 355 |
+
"hostOnlyClass": block.off_class_index,
|
| 356 |
+
"candidateLocalExperts": candidate_indices,
|
| 357 |
+
"probes": oracle_rows,
|
| 358 |
+
"eligibleTokens": int(targets.numel()),
|
| 359 |
+
"positiveBenefitTokens": int(positive.sum()),
|
| 360 |
+
"hostOnlyTargets": int((~positive).sum()),
|
| 361 |
+
"rankingLossFinite": bool(torch.isfinite(ranking_loss)),
|
| 362 |
+
}
|
| 363 |
+
if not torch.isfinite(ranking_loss) or int(positive.sum()) == 0 or int((~positive).sum()) == 0:
|
| 364 |
+
raise RuntimeError("counterfactual smoke did not exercise both positive expert and host-only targets")
|
| 365 |
+
|
| 366 |
+
output = Path(result["output"])
|
| 367 |
+
checkpoint = output / "router-checkpoint.safetensors"
|
| 368 |
+
saved_state = router_state_dict(model)
|
| 369 |
+
save_file(saved_state, str(checkpoint), metadata={"autonoma": "non-routing-structural-smoke"})
|
| 370 |
+
with torch.inference_mode():
|
| 371 |
+
before_reload = model(
|
| 372 |
+
input_ids=oracle_batch["input_ids"], attention_mask=oracle_batch["attention_mask"], use_cache=False
|
| 373 |
+
).logits.detach().cpu()
|
| 374 |
+
with torch.no_grad():
|
| 375 |
+
block.router.gate.weight.zero_()
|
| 376 |
+
load_router_state_dict(model, load_file(str(checkpoint), device="cpu"))
|
| 377 |
+
with torch.inference_mode():
|
| 378 |
+
after_reload = model(
|
| 379 |
+
input_ids=oracle_batch["input_ids"], attention_mask=oracle_batch["attention_mask"], use_cache=False
|
| 380 |
+
).logits.detach().cpu()
|
| 381 |
+
reload_equal = torch.equal(before_reload, after_reload)
|
| 382 |
+
result["checkpointParity"] = {
|
| 383 |
+
"exact": reload_equal,
|
| 384 |
+
"path": str(checkpoint),
|
| 385 |
+
"bytes": checkpoint.stat().st_size,
|
| 386 |
+
"sha256": sha256(checkpoint),
|
| 387 |
+
}
|
| 388 |
+
if not reload_equal:
|
| 389 |
+
raise RuntimeError("router checkpoint save/reload parity failed")
|
| 390 |
+
|
| 391 |
+
result["runtime"] = {
|
| 392 |
+
"loadAndSmokeSeconds": time.perf_counter() - load_started,
|
| 393 |
+
"peakAllocatedVramMiB": torch.cuda.max_memory_allocated(device) / 2**20,
|
| 394 |
+
"peakReservedVramMiB": torch.cuda.max_memory_reserved(device) / 2**20,
|
| 395 |
+
"computeDtype": str(dtype),
|
| 396 |
+
"torch": torch.__version__,
|
| 397 |
+
}
|
| 398 |
+
result["gates"] = {gate: True for gate in smoke["requiredGates"]}
|
| 399 |
+
|
| 400 |
+
|
| 401 |
+
def main() -> int:
|
| 402 |
+
parser = argparse.ArgumentParser(description=__doc__)
|
| 403 |
+
parser.add_argument("--owner-execute", action="store_true")
|
| 404 |
+
parser.add_argument("--config", type=Path, required=True)
|
| 405 |
+
parser.add_argument("--bank", required=True)
|
| 406 |
+
parser.add_argument("--platform", choices=("colab", "kaggle", "other"), default="other")
|
| 407 |
+
parser.add_argument("--work-dir", type=Path, default=Path("/content/autonoma-fable-smoke"))
|
| 408 |
+
args = parser.parse_args()
|
| 409 |
+
if not args.owner_execute:
|
| 410 |
+
raise SystemExit("refusing GPU/model execution without --owner-execute")
|
| 411 |
+
args.work_dir.mkdir(parents=True, exist_ok=True)
|
| 412 |
+
stamp = datetime.now(timezone.utc).strftime("%Y%m%dT%H%M%SZ")
|
| 413 |
+
output = args.work_dir / f"fable-router-structural-smoke-{args.bank}-{stamp}"
|
| 414 |
+
output.mkdir(parents=True, exist_ok=False)
|
| 415 |
+
result: dict[str, Any] = {
|
| 416 |
+
"schema": "AutonomaFableRouterStructuralSmoke.v1",
|
| 417 |
+
"status": "running_nonrouting",
|
| 418 |
+
"nonRouting": True,
|
| 419 |
+
"trainingAuthorized": False,
|
| 420 |
+
"startedAt": now(),
|
| 421 |
+
"output": str(output),
|
| 422 |
+
"system": {"python": sys.version, "platform": platform.platform()},
|
| 423 |
+
}
|
| 424 |
+
tracking: dict[str, Any] = {"status": "not_started"}
|
| 425 |
+
try:
|
| 426 |
+
import trackio
|
| 427 |
+
|
| 428 |
+
trackio.init(
|
| 429 |
+
project="autonoma-fable-router-smoke",
|
| 430 |
+
name=f"{args.bank}-{stamp}",
|
| 431 |
+
space_id="Swordnael/trackio",
|
| 432 |
+
config={"bank": args.bank, "platform": args.platform, "mode": "structural-smoke"},
|
| 433 |
+
)
|
| 434 |
+
tracking = {"status": "started", "space": "Swordnael/trackio"}
|
| 435 |
+
run(args, result)
|
| 436 |
+
result["status"] = "structural_smoke_passed_nonrouting"
|
| 437 |
+
result["passed"] = True
|
| 438 |
+
trackio.log(
|
| 439 |
+
{
|
| 440 |
+
"peak_vram_mib": result["runtime"]["peakAllocatedVramMiB"],
|
| 441 |
+
"positive_oracle_tokens": result["counterfactualDiscovery"]["positiveBenefitTokens"],
|
| 442 |
+
"router_gradient_l1": result["gradientGate"]["routerL1"],
|
| 443 |
+
}
|
| 444 |
+
)
|
| 445 |
+
except BaseException as exc:
|
| 446 |
+
result["status"] = "structural_smoke_failed_nonrouting"
|
| 447 |
+
result["passed"] = False
|
| 448 |
+
result["error"] = {"type": type(exc).__name__, "message": str(exc), "traceback": traceback.format_exc()}
|
| 449 |
+
finally:
|
| 450 |
+
result["finishedAt"] = now()
|
| 451 |
+
result["tracking"] = tracking
|
| 452 |
+
result_path = output / "result.json"
|
| 453 |
+
result_path.write_text(json.dumps(result, indent=2) + "\n", encoding="utf-8")
|
| 454 |
+
try:
|
| 455 |
+
if tracking["status"] == "started":
|
| 456 |
+
import trackio
|
| 457 |
+
|
| 458 |
+
trackio.finish()
|
| 459 |
+
tracking["status"] = "finished"
|
| 460 |
+
result_path.write_text(json.dumps(result, indent=2) + "\n", encoding="utf-8")
|
| 461 |
+
except BaseException as exc:
|
| 462 |
+
tracking["finishError"] = str(exc)
|
| 463 |
+
result_path.write_text(json.dumps(result, indent=2) + "\n", encoding="utf-8")
|
| 464 |
+
token = os.environ.get("HF_TOKEN")
|
| 465 |
+
try:
|
| 466 |
+
config = read_json(args.config.resolve())
|
| 467 |
+
if token:
|
| 468 |
+
result["evidenceUpload"] = upload_evidence(output, config["smoke"]["resultRepo"], token)
|
| 469 |
+
result_path.write_text(json.dumps(result, indent=2) + "\n", encoding="utf-8")
|
| 470 |
+
except BaseException as exc:
|
| 471 |
+
result["evidenceUpload"] = {"status": "failed", "error": str(exc)}
|
| 472 |
+
result_path.write_text(json.dumps(result, indent=2) + "\n", encoding="utf-8")
|
| 473 |
+
print(result_path.resolve())
|
| 474 |
+
return 0 if result.get("passed") else 1
|
| 475 |
+
|
| 476 |
+
|
| 477 |
+
if __name__ == "__main__":
|
| 478 |
+
raise SystemExit(main())
|
| 479 |
+
|