File size: 8,001 Bytes
7845694 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 | """Compare shared typed readout with compact ordered-array generation."""
from __future__ import annotations
import argparse
import hashlib
import json
import statistics
import time
from pathlib import Path
from semif_phase1.core import load_causal_model
from semif_phase1.shared import score_shared
class TimelineStreamer:
def __init__(self, tokenizer, started: float):
self.tokenizer = tokenizer
self.started = started
self.initial = True
self.events = []
def put(self, value):
if self.initial: # generate() first sends the input IDs
self.initial = False
return
elapsed = time.perf_counter() - self.started
for token_id in value.detach().cpu().reshape(-1).tolist():
self.events.append(
{
"seconds": elapsed,
"token_id": token_id,
"text": self.tokenizer.decode([token_id], skip_special_tokens=False),
}
)
def end(self):
pass
def compact_messages(state: str, rows: list[dict]) -> list[dict]:
request = {
"state": state,
"criteria_in_output_order": [row["question"] for row in rows],
"allowed_values": ["yes", "no"],
}
return [
{
"role": "system",
"content": (
"Apply each criterion independently to the state. Return exactly one JSON array "
"containing one lowercase yes or no string per criterion, in the supplied order. "
"Emit no keys, confidence values, markdown, or explanation."
),
},
{"role": "user", "content": json.dumps(request, ensure_ascii=False)},
]
def run_generation(model, tokenizer, state: str, rows: list[dict], max_new_tokens: int) -> dict:
import torch
started = time.perf_counter()
prompt = tokenizer.apply_chat_template(
compact_messages(state, rows),
tokenize=False,
add_generation_prompt=True,
enable_thinking=False,
)
inputs = tokenizer(prompt, return_tensors="pt").to(next(model.parameters()).device)
input_tokens = int(inputs["input_ids"].shape[-1])
streamer = TimelineStreamer(tokenizer, started)
with torch.inference_mode():
output = model.generate(
**inputs,
do_sample=False,
max_new_tokens=max_new_tokens,
streamer=streamer,
use_cache=True,
)
if next(model.parameters()).device.type == "cuda":
torch.cuda.synchronize()
total = time.perf_counter() - started
generated = output[0, input_tokens:].detach().cpu().tolist()
text = tokenizer.decode(generated, skip_special_tokens=True)
parsed = None
try:
candidate = json.loads(text)
if (
isinstance(candidate, list)
and len(candidate) == len(rows)
and all(choice in {"yes", "no"} for choice in candidate)
):
parsed = candidate
except json.JSONDecodeError:
pass
return {
"prompt_sha256": hashlib.sha256(prompt.encode()).hexdigest(),
"input_tokens": input_tokens,
"output_tokens": len(generated),
"time_to_first_token_seconds": streamer.events[0]["seconds"] if streamer.events else None,
"total_seconds": total,
"output_text": text,
"valid_complete_array": parsed is not None,
"choices": parsed,
"timeline": streamer.events,
}
def main() -> None:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--model", required=True)
parser.add_argument("--revision", required=True)
parser.add_argument("--input", type=Path, required=True)
parser.add_argument("--output", type=Path, required=True)
parser.add_argument("--repeats", type=int, default=3)
parser.add_argument("--max-new-tokens", type=int, default=128)
args = parser.parse_args()
if args.output.exists() or args.repeats < 1 or args.max_new_tokens < 1:
parser.error("Output must be new and numeric limits must be positive")
rows = [json.loads(line) for line in args.input.read_text().splitlines() if line.strip()][:21]
if len(rows) != 21 or len({row["state"] for row in rows}) != 1:
parser.error("Input must begin with one complete 21-question shared-state group")
model, tokenizer, metadata = load_causal_model(args.model, args.revision)
import torch
# Warm both paths; warmup is excluded from every reported duration.
score_shared(model, tokenizer, rows, metadata)
warmup = run_generation(model, tokenizer, rows[0]["state"], rows[:1], 16)
if not warmup["timeline"]:
raise RuntimeError("Generation warmup emitted no token events")
direct_runs = []
direct_outputs = None
for _ in range(args.repeats):
torch.cuda.reset_peak_memory_stats()
direct_outputs, timing = score_shared(model, tokenizer, rows, metadata)
direct_runs.append({**timing, "peak_cuda_bytes": torch.cuda.max_memory_allocated()})
generation_runs = []
for _ in range(args.repeats):
torch.cuda.reset_peak_memory_stats()
run = run_generation(model, tokenizer, rows[0]["state"], rows, args.max_new_tokens)
run["peak_cuda_bytes"] = torch.cuda.max_memory_allocated()
generation_runs.append(run)
direct_choices = [
row["option_ids"][max(range(len(row["probabilities"])), key=row["probabilities"].__getitem__)]
for row in direct_outputs
]
complete = [run for run in generation_runs if run["choices"] is not None]
agreement = None
if complete:
agreement = sum(a == b for a, b in zip(direct_choices, complete[0]["choices"])) / len(rows)
report = {
"version": "decision-vs-compact-generation-v2",
"model": metadata,
"hardware": torch.cuda.get_device_name(0),
"input": {
"path": str(args.input),
"sha256": hashlib.sha256(args.input.read_bytes()).hexdigest(),
"states": 1,
"criteria": 21,
},
"scope": (
"Same frozen model, exact state, criteria, hardware and process. Direct path returns "
"two-option distributions; generation emits only an ordered yes/no JSON array."
),
"direct_parallel": {
"runs": direct_runs,
"median_total_seconds": statistics.median(run["total_seconds"] for run in direct_runs),
"outputs": direct_outputs,
},
"compact_generation": {
"prompt_messages": compact_messages(rows[0]["state"], rows),
"runs": generation_runs,
"median_total_seconds": statistics.median(run["total_seconds"] for run in generation_runs),
"median_output_tokens": statistics.median(run["output_tokens"] for run in generation_runs),
"all_runs_valid_complete_arrays": all(run["valid_complete_array"] for run in generation_runs),
"agreement_with_direct_argmax_first_run": agreement,
},
}
report["median_wall_ratio"] = (
report["compact_generation"]["median_total_seconds"]
/ report["direct_parallel"]["median_total_seconds"]
)
args.output.parent.mkdir(parents=True, exist_ok=True)
args.output.write_text(json.dumps(report, indent=2, ensure_ascii=False, allow_nan=False) + "\n")
print(
json.dumps(
{
"direct_median_seconds": report["direct_parallel"]["median_total_seconds"],
"compact_generation_median_seconds": report["compact_generation"]["median_total_seconds"],
"ratio": report["median_wall_ratio"],
"median_output_tokens": report["compact_generation"]["median_output_tokens"],
"valid": report["compact_generation"]["all_runs_valid_complete_arrays"],
"agreement": agreement,
}
)
)
if __name__ == "__main__":
main()
|