Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-19 09:14:35 +02:00
Code Issues Packages Projects Releases Wiki Activity

add fix-te

Browse Source 
Signed-off-by: Vladimir Mandic <mandic00@live.com>
This commit is contained in:
Vladimir Mandic
2024-10-01 12:12:10 -04:00
parent 684263c122
commit 046d588de2
4 changed files with 28 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/prompt_parser_diffusers.py
+23 -3
View File
@@ -5,7 +5,7 @@ import typing
import torch
from compel.embeddings_provider import BaseTextualInversionManager, EmbeddingsProvider
from transformers import PreTrainedTokenizer
from modules import shared, prompt_parser, devices, sd_models
from modules import shared, prompt_parser, devices, sd_models, errors
from modules.prompt_parser_xhinker import get_weighted_text_embeddings_sd15, get_weighted_text_embeddings_sdxl_2p, get_weighted_text_embeddings_sd3, get_weighted_text_embeddings_flux1
debug_enabled = os.environ.get('SD_PROMPT_DEBUG', None)
@@ -17,10 +17,28 @@ token_type = None # used by helper get_tokens
cache = {}
def fix_position_ids(pipe):
# position_ids are created on te creation and are simple index cache
# but somehow can be corrupt in CLIPTextEmbeddings forward call
# see transformers/models/clip/modeling_clip.py:CLIPTextEmbeddings
# reproduction: load sdxl model -> generate -> generate -> load sdxl model -> generate -> generate -> load sdxl model -> generate -> generate
if hasattr(pipe, 'text_encoder') and pipe.text_encoder.text_model.embeddings.position_ids[0][0] > 0:
shared.log.warning(f'TE1 fix: ids={pipe.text_encoder.text_model.embeddings.position_ids}')
pipe.text_encoder.text_model.embeddings.position_ids = torch.arange(pipe.text_encoder.config.max_position_embeddings).expand((1, -1)).to(pipe.text_encoder.device)
if hasattr(pipe, 'text_encoder_2') and pipe.text_encoder_2.text_model.embeddings.position_ids[0][0] > 0:
shared.log.warning(f'TE2 fix: ids={pipe.text_encoder_2.text_model.embeddings.position_ids}')
pipe.text_encoder_2.text_model.embeddings.position_ids = torch.arange(pipe.text_encoder_2.config.max_position_embeddings).expand((1, -1)).to(pipe.text_encoder_2.device)
def compel_hijack(self, token_ids: torch.Tensor,
attention_mask: typing.Optional[torch.Tensor] = None) -> torch.Tensor:
needs_hidden_states = self.returned_embeddings_type != 1
text_encoder_output = self.text_encoder(token_ids, attention_mask, output_hidden_states=needs_hidden_states, return_dict=True)
try: # can crash in ATen/native/cuda/Indexing since position_ids are corrupt so index lookup fails, but its not compel specific, happens with fixed attention as well
text_encoder_output = self.text_encoder(token_ids, attention_mask, output_hidden_states=needs_hidden_states, return_dict=True)
except Exception as e: # its a non-recoverable error as cuda state is corrupt
shared.log.error(f'TE: class={self.text_encoder.__class__} device={self.text_encoder.device} dtype={self.text_encoder.dtype} {e}')
errors.display(e, 'TE:')
if not needs_hidden_states:
return text_encoder_output.last_hidden_state
try:
@@ -147,6 +165,7 @@ def get_tokens(msg, prompt):
def encode_prompts(pipe, p, prompts: list, negative_prompts: list, steps: int, clip_skip: typing.Optional[int] = None):
params_match = prompts == cache.get('prompts', None) and negative_prompts == cache.get('negative_prompts', None) and clip_skip == cache.get('clip_skip', None) and steps == cache.get('steps', None)
if (
'StableDiffusion' not in pipe.__class__.__name__ and
'DemoFusion' not in pipe.__class__.__name__ and
@@ -155,7 +174,7 @@ def encode_prompts(pipe, p, prompts: list, negative_prompts: list, steps: int, c
):
shared.log.warning(f"Prompt parser not supported: {pipe.__class__.__name__}")
return
elif shared.opts.sd_textencoder_cache and prompts == cache.get('prompts', None) and negative_prompts == cache.get('negative_prompts', None) and clip_skip == cache.get('clip_skip', None) and cache.get('model_type', None) == shared.sd_model_type and steps == cache.get('steps', None):
elif shared.opts.sd_textencoder_cache and cache.get('model_type', None) == shared.sd_model_type and params_match and False:
p.prompt_embeds = cache.get('prompt_embeds', None)
p.positive_pooleds = cache.get('positive_pooleds', None)
p.negative_embeds = cache.get('negative_embeds', None)
@@ -171,6 +190,7 @@ def encode_prompts(pipe, p, prompts: list, negative_prompts: list, steps: int, c
pipe.maybe_free_model_hooks()
devices.torch_gc()
fix_position_ids(pipe)
prompt_embeds, positive_pooleds, negative_embeds, negative_pooleds = [], [], [], []
last_prompt, last_negative = None, None
for prompt, negative in zip(prompts, negative_prompts):