File size: 18,585 Bytes
87a0bf5 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 | #!/usr/bin/env python3
from __future__ import annotations
import argparse
import shutil
from pathlib import Path
PROJECT_ROOT = Path(__file__).resolve().parents[3]
SOURCE_REPO = PROJECT_ROOT / "external" / "hsd"
DEFAULT_WORKSPACE = PROJECT_ROOT / "experiments" / "hsd_inference" / "workspace"
def copy_workspace(src: Path, dst: Path) -> None:
dst.parent.mkdir(parents=True, exist_ok=True)
shutil.copytree(
src,
dst,
dirs_exist_ok=True,
ignore=shutil.ignore_patterns(".git", ".idea", "__pycache__", "*.pyc"),
)
def replace_once_or_skip(text: str, old: str, new: str, path: Path) -> str:
if new in text:
return text
if old in text:
return text.replace(old, new, 1)
raise RuntimeError(f"Expected text not found while patching {path}: {old}")
def patch_eval_script(path: Path) -> None:
text = path.read_text()
replacements = {
"os.environ['HF_HOME'] = '/home/c02yuzh/CISPA-projects/rectified_softmax_ml-2023/hf_home'": (
"from pathlib import Path\n"
"PROMPT_DIR = Path(__file__).resolve().parent / 'lib_prompt'\n"
'os.environ.setdefault("HF_HOME", os.environ.get("SHARED_CACHE_ROOT", "/project/mpese/tigersec/runw/workdirs/speculative-decoding-robustness/cache") + "/huggingface")'
),
"parser.add_argument('--target-model', default='Qwen/Qwen2.5-72B-Instruct-GPTQ-Int8', help='must be complex or original')": (
"parser.add_argument('--target-model', default='Qwen/Qwen2.5-72B-Instruct-GPTQ-Int8', help='must be complex or original')\n"
"parser.add_argument('--draft-model', default='Qwen/Qwen2.5-0.5B-Instruct-GPTQ-Int8')\n"
"parser.add_argument('--num-samples', type=int, default=None)\n"
"parser.add_argument('--prompt-suffix', dest='prompt_suffix', default='')\n"
"parser.add_argument('--prompt-tag', default='')\n"
"parser.add_argument('--seed', type=int, default=2027)\n"
"parser.add_argument('--max-new-tokens', type=int, default=512)\n"
"parser.add_argument('--min-new-tokens', type=int, default=0)\n"
"parser.add_argument('--sample-offset', type=int, default=0)"
),
"args = parser.parse_args()": (
"args = parser.parse_args()\n"
"random.seed(args.seed)\n"
"np.random.seed(args.seed)\n"
"torch.manual_seed(args.seed)\n"
"if torch.cuda.is_available():\n"
" torch.cuda.manual_seed_all(args.seed)"
),
'model1_name = "Qwen/Qwen2.5-0.5B-Instruct-GPTQ-Int8"': "model1_name = args.draft_model",
'model_size = model2_name.split("/")[1].split("-")[1]': (
'model_ref = os.path.basename(model2_name.rstrip("/"))\n'
'size_match = re.search(r"(\\d+(?:\\.\\d+)?B)", model_ref)\n'
'if size_match is None:\n'
' raise ValueError(f"Could not infer model size from target model path: {model2_name}")\n'
'model_size = size_match.group(1)'
),
"num_samples = len(gsm8k_test['question'])//5": (
"num_samples = args.num_samples if args.num_samples is not None else len(gsm8k_test['question'])//5\n"
"if args.sample_offset < 0 or args.sample_offset + num_samples > len(gsm8k_test['question']):\n"
" raise ValueError('sample-offset/num-samples exceed the GSM8K split')"
),
"for q, a in tqdm(zip(gsm8k_test['question'][:num_samples], gsm8k_test['answer'][:num_samples]),\n"
" total=num_samples):": (
"sample_end = args.sample_offset + num_samples\n"
" for q, a in tqdm(zip(gsm8k_test['question'][args.sample_offset:sample_end], "
"gsm8k_test['answer'][args.sample_offset:sample_end]),\n"
" total=num_samples):"
),
'if args.prompt == "original":\n prompt = open(\'lib_prompt/prompt_original.txt\').read()\nelse:\n prompt = open(\'lib_prompt/prompt_hardest.txt\').read()': (
'if args.prompt == "original":\n prompt = (PROMPT_DIR / \'prompt_original.txt\').read_text()\nelse:\n prompt = (PROMPT_DIR / \'prompt_hardest.txt\').read_text()\n\n'
"if args.prompt_suffix:\n"
" print('prompt_suffix:', args.prompt_suffix.encode('unicode_escape').decode('ascii'))"
),
' if args.parallel:\n sd += "_parallel"\n': (
' if args.parallel:\n sd += "_parallel"\n\n'
' if args.prompt_tag:\n sd += f"_{args.prompt_tag}"\n'
),
" input_text = tokenizer2.apply_chat_template(\n"
" messages,\n"
" tokenize=False,\n"
" add_generation_prompt=True\n"
" )\n": (
" input_text = tokenizer2.apply_chat_template(\n"
" messages,\n"
" tokenize=False,\n"
" add_generation_prompt=True\n"
" )\n"
"\n"
" if args.prompt_suffix:\n"
" input_text += args.prompt_suffix\n"
),
'total_counts["time"].append(start-end)': 'total_counts["time"].append(end-start)',
# Flush after every sample so NFS sees partial results and they survive SIGTERM/preemption
" fd.write('Q: %s\\nA_model:\\n%s\\nA:\\n%s\\n\\n' % (q, ans_, a))\n": (
" fd.write('Q: %s\\nA_model:\\n%s\\nA:\\n%s\\n\\n' % (q, ans_, a))\n"
" fd.flush() # flush after every sample so NFS sees progress and partial results survive SIGTERM\n"
),
' "step_back_probs":[], "p_i":[], "q_i":[], "hist_lengths": [], "time":[], "ids":[]}': (
' "step_back_probs":[], "p_i":[], "q_i":[], "hist_lengths": [], "time":[], "ids":[],\n'
' "aggregate_acceptance_per_sample": [], "block_efficiency_per_sample": [],\n'
' "accepted_draft_tokens_per_sample": [], "drafted_tokens_per_sample": [],\n'
' "target_steps_per_sample": [], "first_token_acceptance_per_sample": [],\n'
' "acceptance_trace_per_sample": [], "fallback_triggered_per_sample": [],\n'
' "fallback_trigger_block_per_sample": [], "fallback_generated_tokens_per_sample": [],\n'
' "generated_tokens_per_sample": []}'
),
' return_probs=args.backward or args.blockwise,': ' return_probs=True,',
' total_counts["sample_length"].append(len(outputs[0] - len(input_ids[0])))': (
' generated_length = int(outputs.shape[-1] - input_ids.shape[-1])\n'
' total_counts["sample_length"].append(generated_length)\n'
' total_counts["generated_tokens_per_sample"].append(generated_length)\n'
' total_counts["target_steps_per_sample"].append(float(generated_length))'
),
' total_counts["ids"].append(counts["ids"])\n': (
' total_counts["ids"].append(counts["ids"])\n'
'\n'
' drafted_tokens = float(sum(counts["draft_eval"]))\n'
' accepted_draft_tokens = float(sum(counts.get("accepted_draft_tokens", [\n'
' min(int(d), max(int(s) - int(t), 0))\n'
' for d, s, t in zip(counts["draft_eval"], counts["sample_length"], counts["target_eval"])\n'
' ])))\n'
' generated_tokens = float(sum(counts["sample_length"]))\n'
' target_steps = float(sum(counts["total_step"]))\n'
' aggregate_acceptance = accepted_draft_tokens / drafted_tokens if drafted_tokens > 0 else None\n'
' block_efficiency = generated_tokens / target_steps if target_steps > 0 else None\n'
'\n'
' first_token_acceptance = None\n'
' if counts["p_i"] and counts["q_i"] and len(counts["p_i"]) > 0 and len(counts["q_i"]) > 0:\n'
' valid_prob_pairs = []\n'
' for p_val, q_val in zip(counts["p_i"], counts["q_i"]):\n'
' if p_val is None or q_val is None:\n'
' continue\n'
' try:\n'
' p_float = float(p_val)\n'
' q_float = float(q_val)\n'
' except (TypeError, ValueError):\n'
' continue\n'
' valid_prob_pairs.append((p_float, q_float))\n'
' if valid_prob_pairs:\n'
' p0, q0 = valid_prob_pairs[0]\n'
' first_token_acceptance = min(1.0, p0 / max(q0, 1e-12))\n'
'\n'
' total_counts["accepted_draft_tokens_per_sample"].append(accepted_draft_tokens)\n'
' total_counts["drafted_tokens_per_sample"].append(drafted_tokens)\n'
' total_counts["target_steps_per_sample"].append(target_steps)\n'
' total_counts["aggregate_acceptance_per_sample"].append(aggregate_acceptance)\n'
' total_counts["block_efficiency_per_sample"].append(block_efficiency)\n'
' total_counts["first_token_acceptance_per_sample"].append(first_token_acceptance)\n'
' total_counts["acceptance_trace_per_sample"].append(counts.get("acceptance_trace", []))\n'
' total_counts["fallback_triggered_per_sample"].append(bool(counts.get("fallback_triggered", False)))\n'
' total_counts["fallback_trigger_block_per_sample"].append(counts.get("fallback_trigger_block"))\n'
' total_counts["fallback_generated_tokens_per_sample"].append(int(counts.get("fallback_generated_tokens", 0)))\n'
),
' with open(f"{sd}_total_counts.json", "w") as f:\n json.dump(total_counts, f)\n': (
' with open(f"{sd}_total_counts.json", "w") as f:\n'
' valid_acceptance = [x for x in total_counts["aggregate_acceptance_per_sample"] if x is not None]\n'
' valid_block_eff = [x for x in total_counts["block_efficiency_per_sample"] if x is not None]\n'
' valid_first_alpha = [x for x in total_counts["first_token_acceptance_per_sample"] if x is not None]\n'
'\n'
' total_counts["summary_metrics"] = {\n'
' "aggregate_acceptance_mean": float(np.mean(valid_acceptance)) if valid_acceptance else None,\n'
' "block_efficiency_mean": float(np.mean(valid_block_eff)) if valid_block_eff else None,\n'
' "first_token_acceptance_mean": float(np.mean(valid_first_alpha)) if valid_first_alpha else None,\n'
' }\n'
' json.dump(total_counts, f)\n'
),
}
for old, new in replacements.items():
text = replace_once_or_skip(text, old, new, path)
text = text.replace("max_new_tokens=512", "max_new_tokens=args.max_new_tokens")
text = text.replace(
"max_new_tokens=args.max_new_tokens,",
"max_new_tokens=args.max_new_tokens,\n min_new_tokens=args.min_new_tokens,",
)
path.write_text(text)
def patch_generation_utils(path: Path) -> None:
text = path.read_text()
fallback_init = (
' fallback_threshold_raw = os.environ.get("ADSD_ACCEPTANCE_FALLBACK_THRESHOLD", "").strip()\n'
' fallback_threshold = float(fallback_threshold_raw) if fallback_threshold_raw else None\n'
' fallback_window = int(os.environ.get("ADSD_ACCEPTANCE_FALLBACK_WINDOW", "8"))\n'
' if fallback_threshold is not None and not 0.0 <= fallback_threshold <= 1.0:\n'
' raise ValueError("ADSD_ACCEPTANCE_FALLBACK_THRESHOLD must be in [0, 1]")\n'
' if fallback_window <= 0:\n'
' raise ValueError("ADSD_ACCEPTANCE_FALLBACK_WINDOW must be positive")\n'
' fallback_active = False\n'
' fallback_trigger_block = None\n'
' fallback_generated_tokens = 0\n'
' acceptance_trace = []\n'
' acceptance_window = []\n'
'\n'
' counts = {"draft_eval": [], "target_eval": [], "total_step": [], "sample_length": [],\n'
' "step_back_probs": [], "p_i": [], "q_i": [], "hist_lengths": [], "ids": [],\n'
' "accepted_draft_tokens": [], "acceptance_trace": acceptance_trace,\n'
' "fallback_threshold": fallback_threshold, "fallback_window": fallback_window,\n'
' "fallback_triggered": False, "fallback_trigger_block": None,\n'
' "fallback_generated_tokens": 0}'
)
replacements = {
" if backward:\n": (
" step_back_probs, p_next, q_next, ids = None, None, None, None\n"
" if backward:\n"
),
" if not return_probs:\n return valid_tokens, n_matches, ind\n else:\n return valid_tokens, n_matches, None, None, None, None, ind\n": (
" if not return_probs:\n"
" return valid_tokens, n_matches, ind\n"
" else:\n"
" return valid_tokens, n_matches, None, p_i.cpu().numpy().tolist(), q_i.cpu().numpy().tolist(), "
"new_candidate_input_ids.cpu().numpy().tolist(), ind\n"
),
' counts = {"draft_eval": [], "target_eval": [], "total_step": [], "sample_length": [],\n'
' "step_back_probs": [], "p_i": [], "q_i": [], "hist_lengths": [], "ids": []}': (
fallback_init
),
" num_residual_tokens = num_assistant_tokens\n": (
" if fallback_active:\n"
" candidate_generator.num_assistant_tokens = 0\n"
" num_assistant_tokens = 0\n"
"\n"
" num_residual_tokens = num_assistant_tokens\n"
),
" candidate_generator.update_candidate_strategy(input_ids, new_logits, n_matches)\n"
" # update the num_assistant_tokens if needed\n"
" num_assistant_tokens = candidate_generator.num_assistant_tokens\n": (
" candidate_generator.update_candidate_strategy(input_ids, new_logits, n_matches)\n"
" # update the num_assistant_tokens if needed\n"
" num_assistant_tokens = candidate_generator.num_assistant_tokens\n"
" if fallback_active:\n"
" candidate_generator.num_assistant_tokens = 0\n"
" num_assistant_tokens = 0\n"
),
' counts["hist_lengths"].append(hist_lengths)\n': (
' counts["hist_lengths"].append(hist_lengths)\n'
' accepted_draft_tokens = min(int(draft_eval), max(int(sample_length) - int(target_eval), 0))\n'
' counts["accepted_draft_tokens"].append(accepted_draft_tokens)\n'
' if fallback_active:\n'
' fallback_generated_tokens += int(sample_length)\n'
' elif draft_eval > 0:\n'
' block_acceptance = float(accepted_draft_tokens) / float(draft_eval)\n'
' acceptance_trace.append(block_acceptance)\n'
' acceptance_window.append(block_acceptance)\n'
' if len(acceptance_window) > fallback_window:\n'
' acceptance_window.pop(0)\n'
' if (\n'
' fallback_threshold is not None\n'
' and len(acceptance_window) == fallback_window\n'
' and sum(acceptance_window) / len(acceptance_window) < fallback_threshold\n'
' ):\n'
' fallback_active = True\n'
' fallback_trigger_block = len(counts["total_step"])\n'
' counts["fallback_triggered"] = True\n'
' counts["fallback_trigger_block"] = fallback_trigger_block\n'
' candidate_generator.num_assistant_tokens = 0\n'
' num_assistant_tokens = 0\n'
' counts["fallback_generated_tokens"] = fallback_generated_tokens\n'
),
}
for old, new in replacements.items():
text = replace_once_or_skip(text, old, new, path)
path.write_text(text)
def ensure_outputs_dir(workspace: Path) -> None:
outputs_dir = workspace / "chain-of-thought-hub" / "gsm8k" / "outputs"
outputs_dir.mkdir(parents=True, exist_ok=True)
def main() -> int:
parser = argparse.ArgumentParser(description="Prepare a patched local HSD workspace")
parser.add_argument("--workspace", type=Path, default=DEFAULT_WORKSPACE)
args = parser.parse_args()
if not SOURCE_REPO.exists():
raise SystemExit(f"Missing source repo: {SOURCE_REPO}")
copy_workspace(SOURCE_REPO, args.workspace)
patch_eval_script(args.workspace / "chain-of-thought-hub" / "gsm8k" / "eval_speculative_decoding_llm.py")
patch_generation_utils(args.workspace / "transformers" / "generation" / "utils.py")
ensure_outputs_dir(args.workspace)
print(args.workspace)
return 0
if __name__ == "__main__":
raise SystemExit(main())
|