File size: 6,212 Bytes
1d60d59 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 | #!/usr/bin/env python3
"""Run one source-free STRATA Native LM v1 authoritative read."""
from __future__ import annotations
import argparse
import json
import os
from pathlib import Path
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from strata.data.native_lm_integration import NativeLMExample, address_codes
from strata.eval.native_lm_frame_separated_copy import frame_separated_generate
from strata.memory_model.codec import FrozenMemoryCodec
from strata.modeling.exact_payload_realizer import PayloadAuthority
from strata.modeling.native_lm_integration import (
QualifiedP0M2Reader,
StrataMemoryConditionedLM,
)
from strata.modeling.structural_copy import StructuralCopyActionHead
from strata.training.native_lm_integration import compact_state_table
def load_model(root: Path, base_model: str, device: torch.device):
config = json.loads((root / "configs/strata_native_lm_system_v1.json").read_text())
m1_config = json.loads(
(root / "configs/strata_native_lm_integration_m1.json").read_text()
)
model_config = m1_config["model"]
tokenizer = AutoTokenizer.from_pretrained(base_model, local_files_only=True)
if tokenizer.pad_token_id is None:
tokenizer.pad_token = tokenizer.eos_token
backbone = AutoModelForCausalLM.from_pretrained(
base_model,
local_files_only=True,
torch_dtype=torch.bfloat16,
attn_implementation=model_config["attention_implementation"],
).to(device)
backbone.config.use_cache = False
codec = FrozenMemoryCodec(
checkpoint_path=root / "checkpoints/P0_M2_CHECKPOINT_FINAL.pt",
config_path=root / "configs/strata_native_lm_p0_m2_v1.json",
device="cpu",
)
model = StrataMemoryConditionedLM(
backbone,
qualified_reader=QualifiedP0M2Reader(codec.model),
layer_indices=model_config["memory_port_layers"],
compact_width=int(model_config["compact_width"]),
address_width=int(model_config["address_width"]),
payload_width=int(m1_config["substrate"]["payload_width"]),
memory_width=int(model_config["memory_width"]),
memory_tokens=int(model_config["memory_tokens"]),
attention_width=int(model_config["attention_width"]),
heads=int(model_config["attention_heads"]),
adapter_rank=int(model_config["adapter_rank"]),
payload_classes=int(model_config["payload_classes"]),
auxiliary_payload_loss_weight=float(
model_config["auxiliary_payload_loss_weight"]
),
).to(device)
checkpoint = torch.load(
root / "checkpoints/MEMORY_PATH_FINAL.pt",
map_location="cpu",
weights_only=False,
)
model.load_trainable_state_dict(checkpoint["state"])
model.eval()
for parameter in model.parameters():
parameter.requires_grad_(False)
head_checkpoint = torch.load(
root / "checkpoints/ACTION_HEAD_FINAL.pt",
map_location="cpu",
weights_only=False,
)
head = StructuralCopyActionHead(int(head_checkpoint["hidden_size"])).to(device)
head.load_state_dict(head_checkpoint["state"], strict=True)
head.eval()
for parameter in head.parameters():
parameter.requires_grad_(False)
return config, tokenizer, model, head, codec
def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument(
"--base-model",
default=os.environ.get("STRATA_BASE_MODEL"),
help="Local Qwen3-4B-Instruct-2507 snapshot",
)
parser.add_argument("--event", required=True)
parser.add_argument("--predicate", required=True)
parser.add_argument("--role", required=True)
parser.add_argument("--payload-handle", type=int, required=True)
parser.add_argument("--payload", required=True)
parser.add_argument("--query", required=True)
parser.add_argument("--event-version", type=int, default=1)
parser.add_argument("--device", default="cuda:0")
args = parser.parse_args()
if not args.base_model:
parser.error("--base-model or STRATA_BASE_MODEL is required")
if not 1 <= args.payload_handle <= 255:
parser.error("--payload-handle must be in [1,255]")
root = Path(__file__).resolve().parent
device = torch.device(args.device)
config, tokenizer, model, head, codec = load_model(root, args.base_model, device)
row = NativeLMExample(
example_id="release-request",
split="release",
schema=args.event.split(":", 1)[0],
field=args.role,
event=args.event,
predicate=args.predicate,
role=args.role,
value_type="authoritative",
payload_handle=args.payload_handle,
value=args.payload,
address_codes=address_codes(args.event, args.predicate, args.role),
query=args.query,
full_history_query=args.query,
answer=f"The {args.role.replace('_', ' ')} is {args.payload}.",
operation="point",
age_windows=0,
)
authority = PayloadAuthority.issue(
event=args.event,
predicate=args.predicate,
role=args.role,
handle=args.payload_handle,
payload=args.payload,
version=args.event_version,
)
frame = config["frame"]
outputs, timing = frame_separated_generate(
model,
head,
tokenizer,
[row],
compact_state_table(codec),
[[authority]],
[0],
batch_size=1,
max_actions=int(config["evaluation"]["max_actions"]),
frame_handle=int(frame["canonical_frame_handle"]),
frame_surrogate=str(frame["canonical_frame_surrogate"]),
terminator=str(frame["structural_terminator"]),
current_versions=[args.event_version],
)
result = outputs[0]
print(
json.dumps(
{
"answer": result.text,
"frame": result.frame,
"status": result.status,
"payload_handle": result.controller_handle,
"receipt": authority.receipt,
"timing": timing,
},
ensure_ascii=False,
sort_keys=True,
)
)
if __name__ == "__main__":
main()
|