clairic / clairic_runtime /readout.py
areneau's picture
Update Clairic Transformers runtime
524fd9d verified
Raw History Blame Contribute Delete
3.31 kB
"""Read typed fields from backbone hidden states."""
from __future__ import annotations
import inspect
import json
from pathlib import Path
from typing import Any
import torch
from safetensors.torch import load_file
from . import model as cm
def load_head(path: Path, *, match_bfloat16: bool = False) -> cm.ClairicHead:
configuration = json.loads((path / "clair_head_config.json").read_text())
accepted = inspect.signature(cm.ClairicHead).parameters
head = cm.ClairicHead(
**{name: value for name, value in configuration.items() if name in accepted}
)
state = load_file(path / "clair_head.safetensors")
if match_bfloat16:
# Serving constructs the head in BF16 before widening it to FP32.
# The length bias has an explicit FP32 initialization.
state = {
name: value.bfloat16().float()
if value.is_floating_point() and name != "length_bias"
else value
for name, value in state.items()
}
head.load_state_dict(state, strict=True)
return head.float().eval()
def read_slots(
states: torch.Tensor, output_weight: torch.Tensor, chunk: int = 256
) -> tuple[torch.Tensor, torch.Tensor]:
ids, probabilities = [], []
for rows in states.split(chunk):
logits = (rows @ output_weight.t()).float()
best = logits.max(dim=-1)
ids.append(best.indices)
probabilities.append(best.values - torch.logsumexp(logits, dim=-1))
return torch.cat(ids), torch.cat(probabilities)
@torch.inference_mode()
def read_record(
head: cm.ClairicHead,
states: torch.Tensor,
item: cm.EncodedRecord,
output_weight: torch.Tensor,
) -> dict[str, Any]:
start, length, kinds, queries, items, slot_fields, slots = (
cm.parse_prediction_layout(
cm.prediction_layout(item),
len(item.input_ids),
with_slots=True,
)
)
views = cm.prediction_views(kinds, queries, items, states.device)
output = head.forward_states(states, start, start + length, views)
if slots:
width = max(map(len, slots))
positions = torch.tensor(
[row + [row[-1]] * (width - len(row)) for row in slots],
dtype=torch.long,
device=states.device,
)
if head.answer_slot_routing:
field_queries = views["query_positions"][slot_fields]
slot_kinds = torch.tensor(
[
cm.ROUTE_KINDS.index(
"tags" if kinds[row] == "text_list" else kinds[row]
)
for row in slot_fields
],
dtype=torch.long,
device=states.device,
)
lengths = torch.tensor(
[len(row) for row in slots], dtype=torch.long, device=states.device
)
slot_states = head.answer_slot_states(
states, field_queries, slot_kinds, positions, lengths, start + length
).flatten(0, 1)
else:
slot_states = states[positions.flatten()]
ids, logprobs = read_slots(slot_states, output_weight)
output["slot_ids"] = ids.view(positions.shape)
output["slot_log_probs"] = logprobs.view(positions.shape)
return output