perf: add BFD bin-packing and custom attention mask to IFD batch scoring
This commit is contained in:
parent
4d3c9341c1
commit
c7158418dd
|
|
@ -18,29 +18,108 @@ from astrai.model import AutoModel
|
|||
from astrai.tokenize import AutoTokenizer
|
||||
|
||||
|
||||
def _score(context_ids, resp_ids, model, device):
|
||||
"""Core IFD computation: context → L_cond, response alone → L_uncond."""
|
||||
if not resp_ids:
|
||||
return None
|
||||
full_ids = context_ids + resp_ids
|
||||
inp_full = torch.tensor([full_ids], device=device, dtype=torch.long)
|
||||
inp_resp = torch.tensor([resp_ids], device=device, dtype=torch.long)
|
||||
logits_full = model(inp_full)["logits"][0]
|
||||
logits_resp = model(inp_resp)["logits"][0]
|
||||
ctx_len = len(context_ids)
|
||||
resp_logits = logits_full[ctx_len - 1 : -1]
|
||||
resp_targets = torch.tensor(resp_ids, device=device, dtype=torch.long)
|
||||
L_cond = F.cross_entropy(resp_logits, resp_targets, reduction="mean").item()
|
||||
unp_logits = logits_resp[:-1]
|
||||
unp_targets = torch.tensor(resp_ids[1:], device=device, dtype=torch.long)
|
||||
L_uncond = F.cross_entropy(unp_logits, unp_targets, reduction="mean").item()
|
||||
ifd = L_cond / L_uncond if L_uncond > 0 else None
|
||||
return {
|
||||
"L_cond": round(L_cond, 6),
|
||||
"L_uncond": round(L_uncond, 6),
|
||||
"ifd": round(ifd, 6) if ifd is not None else None,
|
||||
"resp_len": len(resp_ids),
|
||||
}
|
||||
def _pack_bins(pairs, max_len):
|
||||
"""BFD bin packing: pack (c+r) into bins of max total length."""
|
||||
indexed = sorted(enumerate(pairs), key=lambda x: -(len(x[1][0]) + len(x[1][1])))
|
||||
bins = [] # each bin: list of (orig_idx, ctx_ids, resp_ids)
|
||||
lengths = []
|
||||
for orig_idx, (c, r) in indexed:
|
||||
size = len(c) + len(r)
|
||||
best_bin = -1
|
||||
for bi, rem in enumerate(lengths):
|
||||
if rem >= size:
|
||||
if best_bin < 0 or rem < lengths[best_bin]:
|
||||
best_bin = bi
|
||||
if best_bin >= 0:
|
||||
bins[best_bin].append((orig_idx, c, r))
|
||||
lengths[best_bin] -= size
|
||||
else:
|
||||
bins.append([(orig_idx, c, r)])
|
||||
lengths.append(max_len - size)
|
||||
return bins
|
||||
|
||||
|
||||
@torch.inference_mode()
|
||||
def _score_batch(pairs, model, device, max_len=2048):
|
||||
"""BFD-packed IFD: pack items into bins, one forward pass per bin."""
|
||||
if not pairs:
|
||||
return []
|
||||
bins = _pack_bins(pairs, max_len)
|
||||
|
||||
result = [None] * len(pairs)
|
||||
|
||||
for bin_items in bins:
|
||||
seq_ids = []
|
||||
global_pos = [] # doc-reset position IDs for RoPE
|
||||
doc_ids = [] # document index for attention mask
|
||||
doc_offsets = []
|
||||
|
||||
for di, (orig_idx, c, r) in enumerate(bin_items):
|
||||
ctx_len = len(c)
|
||||
start = len(seq_ids)
|
||||
item_len = len(c) + len(r)
|
||||
seq_ids.extend(c)
|
||||
seq_ids.extend(r)
|
||||
end = len(seq_ids)
|
||||
global_pos.extend(range(item_len))
|
||||
doc_ids.extend([di] * item_len)
|
||||
doc_offsets.append((start, end, orig_idx, ctx_len))
|
||||
|
||||
full_ids = torch.tensor([seq_ids], device=device, dtype=torch.long)
|
||||
pos_ids = torch.tensor([global_pos], device=device, dtype=torch.long)
|
||||
T = len(seq_ids)
|
||||
causal = torch.tril(torch.ones(T, T, dtype=torch.bool, device=device))
|
||||
doc_t = torch.tensor([doc_ids], device=device)
|
||||
doc_mask = doc_t.unsqueeze(-1) == doc_t.unsqueeze(-2)
|
||||
attn_mask = (causal & doc_mask[0]).unsqueeze(0).unsqueeze(0)
|
||||
logits_full = model(full_ids, position_ids=pos_ids, input_mask=attn_mask)[
|
||||
"logits"
|
||||
][0]
|
||||
|
||||
for start, end, orig_idx, ctx_len in doc_offsets:
|
||||
rl = end - start - ctx_len
|
||||
if rl < 2:
|
||||
continue
|
||||
resp_start = start + ctx_len - 1
|
||||
resp_logits = logits_full[resp_start : end - 1]
|
||||
resp_targets = torch.tensor(
|
||||
seq_ids[start + ctx_len : end], device=device, dtype=torch.long
|
||||
)
|
||||
L_cond = F.cross_entropy(resp_logits, resp_targets, reduction="mean").item()
|
||||
result[orig_idx] = (L_cond, rl)
|
||||
|
||||
# unconditional pass: batch all responses separately (sorted by length)
|
||||
resp_seqs = [
|
||||
(i, result[i][1], pairs[i][1])
|
||||
for i in range(len(pairs))
|
||||
if result[i] is not None
|
||||
]
|
||||
if resp_seqs:
|
||||
resp_seqs.sort(key=lambda x: -x[1])
|
||||
r_batch = torch.zeros(
|
||||
len(resp_seqs),
|
||||
max(len(r) for _, _, r in resp_seqs),
|
||||
dtype=torch.long,
|
||||
device=device,
|
||||
)
|
||||
for ri, (_, rl, r_ids) in enumerate(resp_seqs):
|
||||
r_batch[ri, :rl] = torch.tensor(r_ids, dtype=torch.long)
|
||||
logits_resp = model(r_batch)["logits"]
|
||||
|
||||
for ri, (orig_idx, rl, _) in enumerate(resp_seqs):
|
||||
L_cond = result[orig_idx][0]
|
||||
unp_logits = logits_resp[ri, : rl - 1]
|
||||
unp_targets = r_batch[ri, 1:rl]
|
||||
L_uncond = F.cross_entropy(unp_logits, unp_targets, reduction="mean").item()
|
||||
ifd = L_cond / L_uncond if L_uncond > 0 else None
|
||||
result[orig_idx] = {
|
||||
"L_cond": round(L_cond, 6),
|
||||
"L_uncond": round(L_uncond, 6),
|
||||
"ifd": round(ifd, 6) if ifd is not None else None,
|
||||
"resp_len": rl,
|
||||
}
|
||||
|
||||
return result
|
||||
|
||||
|
||||
def _trim(context_ids, resp_ids, max_len):
|
||||
|
|
@ -63,7 +142,7 @@ def score_plain(model, tokenizer, instruction, response, device, max_len=2048):
|
|||
ctx_ids, resp_ids = _trim(ctx_ids, resp_ids, max_len)
|
||||
if not ctx_ids or not resp_ids:
|
||||
return {"L_cond": None, "L_uncond": None, "ifd": None, "error": "empty"}
|
||||
return _score(ctx_ids, resp_ids, model, device)
|
||||
return _score_batch([(ctx_ids, resp_ids)], model, device, max_len)[0]
|
||||
|
||||
|
||||
def score_messages(model, tokenizer, messages, device, max_len=2048):
|
||||
|
|
@ -77,21 +156,21 @@ def score_messages(model, tokenizer, messages, device, max_len=2048):
|
|||
resp_ids = tokenizer.encode(msg["content"], add_special_tokens=False)
|
||||
ctx_ids, resp_ids = _trim(ctx_ids, resp_ids, max_len)
|
||||
if ctx_ids and resp_ids:
|
||||
turns.append(_score(ctx_ids, resp_ids, model, device))
|
||||
turns.append((ctx_ids, resp_ids))
|
||||
if not turns:
|
||||
return None
|
||||
valid = [t for t in turns if t is not None and t["ifd"] is not None]
|
||||
raw_scores = _score_batch(turns, model, device, max_len)
|
||||
valid = [s for s in raw_scores if s is not None and s["ifd"] is not None]
|
||||
if not valid:
|
||||
return {"ifd": None, "ifd_turns": turns}
|
||||
avg = sum(t["ifd"] for t in valid) / len(valid)
|
||||
return {"ifd": None, "ifd_turns": raw_scores}
|
||||
avg = sum(s["ifd"] for s in valid) / len(valid)
|
||||
return {
|
||||
"ifd": avg,
|
||||
"ifd_detail": valid[0] if len(valid) == 1 else None,
|
||||
"ifd_turns": turns,
|
||||
"ifd_turns": raw_scores,
|
||||
}
|
||||
|
||||
|
||||
@torch.inference_mode()
|
||||
def process_file(
|
||||
param_path,
|
||||
input_file,
|
||||
|
|
@ -100,9 +179,12 @@ def process_file(
|
|||
resp_key,
|
||||
max_len=2048,
|
||||
data_format="plain",
|
||||
batch_size=1,
|
||||
device=None,
|
||||
):
|
||||
device = "cuda" if torch.cuda.is_available() else "cpu"
|
||||
dtype = torch.bfloat16 if device == "cuda" else torch.float32
|
||||
if device is None:
|
||||
device = "cuda" if torch.cuda.is_available() else "cpu"
|
||||
dtype = torch.bfloat16 if "cuda" in device else torch.float32
|
||||
|
||||
model = AutoModel.from_pretrained(param_path)
|
||||
tokenizer = AutoTokenizer.from_pretrained(param_path)
|
||||
|
|
@ -114,23 +196,38 @@ def process_file(
|
|||
|
||||
results = []
|
||||
all_ifds = []
|
||||
buffer = []
|
||||
|
||||
for item in tqdm.tqdm(data, desc="Computing IFD", unit="sample"):
|
||||
if data_format == "messages":
|
||||
scores = score_messages(
|
||||
model, tokenizer, item.get("messages", []), device, max_len
|
||||
)
|
||||
if scores is None:
|
||||
turns = []
|
||||
for i, msg in enumerate(item.get("messages", [])):
|
||||
if msg.get("role") != "assistant":
|
||||
continue
|
||||
ctx_text = "\n\n".join(m["content"] for m in item["messages"][:i])
|
||||
ctx_ids = tokenizer.encode(ctx_text)
|
||||
resp_ids = tokenizer.encode(msg["content"], add_special_tokens=False)
|
||||
ctx_ids, resp_ids = _trim(ctx_ids, resp_ids, max_len)
|
||||
if ctx_ids and resp_ids:
|
||||
turns.append((ctx_ids, resp_ids))
|
||||
if not turns:
|
||||
results.append({**item, "ifd": None, "ifd_turns": []})
|
||||
else:
|
||||
all_ifds.append(scores["ifd"])
|
||||
results.append({**item, **scores})
|
||||
continue
|
||||
buffer.append((item, turns, "messages"))
|
||||
else:
|
||||
scores = score_plain(
|
||||
model, tokenizer, item[instr_key], item[resp_key], device, max_len
|
||||
)
|
||||
all_ifds.append(scores["ifd"])
|
||||
results.append({**item, "ifd": scores["ifd"], "ifd_detail": scores})
|
||||
ctx_ids = tokenizer.encode(item[instr_key], add_special_tokens=False)
|
||||
resp_ids = tokenizer.encode(item[resp_key], add_special_tokens=False)
|
||||
ctx_ids, resp_ids = _trim(ctx_ids, resp_ids, max_len)
|
||||
if not ctx_ids or not resp_ids:
|
||||
results.append({**item, "ifd": None, "ifd_detail": {"error": "empty"}})
|
||||
continue
|
||||
buffer.append((item, [(ctx_ids, resp_ids)], "plain"))
|
||||
|
||||
if len(buffer) >= batch_size:
|
||||
_flush_buffer(buffer, results, all_ifds, model, device, max_len)
|
||||
|
||||
if buffer:
|
||||
_flush_buffer(buffer, results, all_ifds, model, device, max_len)
|
||||
|
||||
with open(output_file, "w", encoding="utf-8") as f:
|
||||
for item in results:
|
||||
|
|
@ -151,6 +248,41 @@ def process_file(
|
|||
print(f"Results saved to {output_file}")
|
||||
|
||||
|
||||
def _flush_buffer(buffer, results, all_ifds, model, device, max_len=2048):
|
||||
all_pairs = []
|
||||
indices = []
|
||||
for item, turns, fmt in buffer:
|
||||
start = len(all_pairs)
|
||||
all_pairs.extend(turns)
|
||||
indices.append((item, turns, fmt, start, len(all_pairs)))
|
||||
|
||||
raw = _score_batch(all_pairs, model, device, max_len)
|
||||
|
||||
for item, turns, fmt, start, end in indices:
|
||||
turn_scores = raw[start:end]
|
||||
if fmt == "messages":
|
||||
valid = [s for s in turn_scores if s is not None and s["ifd"] is not None]
|
||||
if not valid:
|
||||
results.append({**item, "ifd": None, "ifd_turns": turn_scores})
|
||||
else:
|
||||
avg = sum(s["ifd"] for s in valid) / len(valid)
|
||||
all_ifds.append(avg)
|
||||
results.append(
|
||||
{
|
||||
**item,
|
||||
"ifd": avg,
|
||||
"ifd_detail": valid[0] if len(valid) == 1 else None,
|
||||
"ifd_turns": turn_scores,
|
||||
}
|
||||
)
|
||||
else:
|
||||
score = turn_scores[0]
|
||||
all_ifds.append(score["ifd"])
|
||||
results.append({**item, "ifd": score["ifd"], "ifd_detail": score})
|
||||
|
||||
buffer.clear()
|
||||
|
||||
|
||||
def main():
|
||||
parser = argparse.ArgumentParser(
|
||||
description="Compute IFD scores for instruction-response data"
|
||||
|
|
@ -164,7 +296,7 @@ def main():
|
|||
type=str,
|
||||
default="plain",
|
||||
choices=["plain", "messages"],
|
||||
help="Input format: 'plain' for instr_key+resp_key, 'messages' for messages array",
|
||||
help="Input format",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--instr_key", type=str, default="instruction", help="Key for instruction field"
|
||||
|
|
@ -172,6 +304,10 @@ def main():
|
|||
parser.add_argument(
|
||||
"--resp_key", type=str, default="response", help="Key for response field"
|
||||
)
|
||||
parser.add_argument(
|
||||
"--batch_size", type=int, default=8, help="Batch size for model forward passes"
|
||||
)
|
||||
parser.add_argument("--device", type=str, default=None, help="Device (e.g. cuda:0)")
|
||||
args = parser.parse_args()
|
||||
|
||||
process_file(
|
||||
|
|
@ -182,6 +318,8 @@ def main():
|
|||
args.resp_key,
|
||||
args.max_len,
|
||||
data_format=args.format,
|
||||
batch_size=args.batch_size,
|
||||
device=args.device,
|
||||
)
|
||||
|
||||
|
||||
|
|
|
|||
Loading…
Reference in New Issue