File size: 5,799 Bytes
0810902 | 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 | #!/usr/bin/env python3
"""Diagnose degenerate generation: is it the checkpoint or the placement?
The composed checkpoint emits a constant token under `device_map="auto"` with
CPU offload. This script isolates the cause by checking, for one short prompt:
* whether hidden states or logits contain NaN/Inf
* where in the layer stack the first non-finite value appears
* whether the result changes without CPU offload (4-bit, fully resident)
Run it against the composed checkpoint and against the language-only source
checkpoint to tell a composition bug apart from an environment bug.
"""
from __future__ import annotations
import argparse
import json
from pathlib import Path
from typing import Any
import torch
def load(model_path: str, mode: str) -> tuple[Any, Any]:
from transformers import AutoModelForMultimodalLM, AutoTokenizer
kwargs: dict[str, Any] = {"dtype": torch.bfloat16}
if mode == "offload":
kwargs["device_map"] = "auto"
elif mode == "gpu4bit":
from transformers import BitsAndBytesConfig
kwargs["quantization_config"] = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_quant_type="nf4",
bnb_4bit_compute_dtype=torch.bfloat16,
bnb_4bit_use_double_quant=True,
)
kwargs["device_map"] = {"": 0}
elif mode == "cpu":
kwargs["device_map"] = {"": "cpu"}
kwargs["dtype"] = torch.float32
else:
raise ValueError(mode)
model = AutoModelForMultimodalLM.from_pretrained(model_path, **kwargs)
model.eval()
tokenizer = AutoTokenizer.from_pretrained(model_path)
return model, tokenizer
def encode(tokenizer: Any, prompt: str, device: Any) -> torch.Tensor:
"""Return a plain input_ids tensor across transformers 4.x / 5.x behaviour."""
encoded = tokenizer.apply_chat_template(
[{"role": "user", "content": prompt}],
add_generation_prompt=True,
return_tensors="pt",
)
if not isinstance(encoded, torch.Tensor):
encoded = encoded["input_ids"]
return encoded.to(device)
def probe(model: Any, tokenizer: Any, prompt: str) -> dict[str, Any]:
ids = encode(tokenizer, prompt, model.device)
with torch.inference_mode():
out = model(input_ids=ids, output_hidden_states=True, use_cache=False)
logits = out.logits[0, -1].float()
hidden = out.hidden_states
first_bad = None
layer_stats = []
for index, state in enumerate(hidden):
tensor = state.float()
finite = bool(torch.isfinite(tensor).all())
layer_stats.append(
{
"layer": index,
"finite": finite,
"absmax": float(tensor.abs().max()) if finite else None,
"std": float(tensor.std()) if finite else None,
}
)
if not finite and first_bad is None:
first_bad = index
top = torch.topk(logits, 5)
return {
"prompt": prompt,
"prompt_tokens": int(ids.shape[1]),
"logits_finite": bool(torch.isfinite(logits).all()),
"logits_absmax": float(logits.abs().max()) if bool(torch.isfinite(logits).all()) else None,
"logits_nan_count": int(torch.isnan(logits).sum()),
"logits_inf_count": int(torch.isinf(logits).sum()),
"first_nonfinite_hidden_layer": first_bad,
"hidden_layer_count": len(hidden),
"top5_token_ids": [int(i) for i in top.indices],
"top5_tokens": [tokenizer.decode([int(i)]) for i in top.indices],
"top5_logits": [float(v) for v in top.values],
"layer_stats": layer_stats,
}
def short_generate(model: Any, tokenizer: Any, prompt: str, n: int = 24) -> str:
ids = encode(tokenizer, prompt, model.device)
with torch.inference_mode():
out = model.generate(ids, max_new_tokens=n, do_sample=False)
return tokenizer.decode(out[0][ids.shape[1] :], skip_special_tokens=True).strip()
def main() -> None:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--model", required=True)
parser.add_argument("--mode", default="gpu4bit", choices=["offload", "gpu4bit", "cpu"])
parser.add_argument("--label", default=None)
parser.add_argument("--prompt", default="What is the capital of France?")
parser.add_argument("--output", type=Path, default=Path("reports/generation_diagnosis.json"))
args = parser.parse_args()
print(f"Loading {args.model} [{args.mode}]", flush=True)
model, tokenizer = load(args.model, args.mode)
record = {
"model": args.model,
"label": args.label or Path(args.model).name,
"mode": args.mode,
"device_map": str(getattr(model, "hf_device_map", None)),
"probe": probe(model, tokenizer, args.prompt),
"generation": short_generate(model, tokenizer, args.prompt),
}
print(json.dumps({k: v for k, v in record.items() if k != "probe"}, indent=2))
p = record["probe"]
print(
f"logits finite={p['logits_finite']} nan={p['logits_nan_count']} "
f"inf={p['logits_inf_count']} first_bad_hidden={p['first_nonfinite_hidden_layer']}"
)
pairs = list(zip(p["top5_tokens"], [round(v, 2) for v in p["top5_logits"]], strict=True))
print(f"top5: {pairs}")
args.output.parent.mkdir(parents=True, exist_ok=True)
existing = []
if args.output.is_file():
existing = json.loads(args.output.read_text(encoding="utf-8"))
existing.append(record)
args.output.write_text(json.dumps(existing, indent=2) + "\n", encoding="utf-8")
print(f"\nAppended to {args.output}")
if __name__ == "__main__":
main()
|