Fix realtime entropy patching (#26)
Some checks are pending
Lint with Black / lint (push) Waiting to run
Lint with isort / lint (push) Waiting to run

* allow loading of the entropy model directly

* remove unused argument

* remove spammy warning

* allow patch_batch_size to be adjusted in the forward() method

* revert to original patcher style, fix warning

* allow grads when calculating entropies

* fix grad flow

* return preds from calculate_entropies()

* remove legacy arg

* fix an error with monotonicity and small sequence lengths

* ensure patcher is serializable

* revert patcher to original

* remove unused import
This commit is contained in:
Ink 2025-01-21 18:34:23 -06:00 committed by GitHub
parent 6ffeb66b53
commit 392117bff2
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 26 additions and 12 deletions

View file

@ -2,6 +2,7 @@
import math
import time
from collections import defaultdict
from contextlib import nullcontext
from enum import Enum
import torch
@ -58,7 +59,11 @@ def entropy(scores):
def calculate_entropies(
tokens: torch.tensor, entropy_model, patching_batch_size, device: str | None = None
tokens: torch.tensor,
entropy_model,
patching_batch_size,
device: str | None = None,
enable_grad: bool = False,
):
"""
tokens: 2D tensor of shape [batch_size, seq_len]
@ -67,8 +72,12 @@ def calculate_entropies(
Splits the tokens into chunks of size max_length and calculates entropies for each chunk.
Entropy model can be executed on cpu or gpu, specify either 'cuda' or 'cpu' in the device argument.
"""
with torch.no_grad():
grad_context = nullcontext() if enable_grad else torch.no_grad()
with grad_context:
entropies = []
preds = []
max_length = getattr(entropy_model, "max_length", 8192)
batch_numel = max_length * patching_batch_size
splits = torch.split(tokens.flatten(), batch_numel)
@ -86,12 +95,15 @@ def calculate_entropies(
pred = pred.reshape(-1, pred.shape[-1])[
: split.numel() - pad_size, :
] # [batch_size * seq_len, vocab]
preds.append(pred)
pred_entropies = entropy(pred)
entropies.append(pred_entropies)
concat_entropies = torch.cat(entropies, dim=0)
concat_entropies = concat_entropies.reshape(tokens.shape)
return concat_entropies
concat_preds = torch.cat(preds, dim=0)
concat_preds = concat_preds.reshape(tokens.shape[0], tokens.shape[1], -1)
return concat_entropies, concat_preds
def patch_start_mask_from_entropy_with_monotonicity(entropies, t):
@ -101,6 +113,10 @@ def patch_start_mask_from_entropy_with_monotonicity(entropies, t):
returns [bs, seq_len] mask where True indicates the start of a patch
"""
bs, seq_len = entropies.shape
if seq_len == 0:
return entropies > t
mask = torch.zeros_like(entropies, dtype=torch.bool)
mask[:, 0] = True
@ -123,6 +139,10 @@ def patch_start_mask_global_and_monotonicity(entropies, t, t_add=0):
returns [bs, seq_len] mask where True indicates the start of a patch
"""
bs, seq_len = entropies.shape
if seq_len == 0:
return entropies > t
mask = torch.zeros_like(entropies, dtype=torch.bool)
mask[:, 0] = True
@ -521,12 +541,12 @@ class Patcher:
if self.log_time:
s = time.time()
if entropies is not None:
scores = torch.tensor(entropies, dtype=torch.float32)
scores = entropies.to(dtype=torch.float32)
elif preds is not None:
scores = entropy(preds)
else:
start_entropies = time.time()
scores = calculate_entropies(
scores, _ = calculate_entropies(
tokens,
self.entropy_model,
self.patching_batch_size,

View file

@ -199,9 +199,6 @@ class LocalModelBase(nn.Module):
class LocalEncoder(LocalModelBase):
def __init__(self, args: LocalModelArgs):
super().__init__(args)
self.output_proj = (
args.patching_mode in ["entropy", "probmax"]
) and args.entropy_model_checkpoint_dir is None
self.apply_transformer = args.use_local_encoder_transformer
self.downsampling_by_pooling = args.downsampling_by_pooling

View file

@ -162,9 +162,6 @@ def create_causal_mask(
return "causal"
if BLT_SUPPRESS_ATTN_ERROR == 1:
logging.warning(
"SDPA attention being used, which doesn't have specialized attention implementations for block_causal and local_block_causal attention. Allowing model to run since BLT_SUPPRESS_ATTN_ERROR=1"
)
return "causal"
else:
raise ValueError(

View file

@ -117,7 +117,7 @@ def main(
text = get_text(doc)
tokens = torch.tensor(tokenizer.encode(text))
patch_start = time.time()
scores = calculate_entropies(
scores, _ = calculate_entropies(
tokens,
entropy_model,
patching_batch_size,