Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-20 01:31:13 +02:00
Code Issues Packages Projects Releases Wiki Activity

update xet and patch qk_norm defaults

Browse Source 
Signed-off-by: Vladimir Mandic <mandic00@live.com>
This commit is contained in:
Vladimir Mandic
2026-05-16 08:27:11 +02:00
parent 84cbc37659
commit 4db62cc771
9 changed files with 41 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files
CHANGELOG.md
+3
View File
@@ -9,6 +9,8 @@
- **Captioning** new feature: analyze existing images for prompt adherence
*tip*: image analysis requires larger VLM model to produce quality output
new api endpoint: `/sdapi/v1/analyze`
- **HF download** use `XET` by default
see *settings -> huggingface -> download method* for options
- **AI**
- Cognitive analysis and improvements to *all* AI prompts
- Automated fixes using `/check-` skills
@@ -18,6 +20,7 @@
- `gradio` initial hijack
- `SmolVLM` captioning
- `gradio` temp files guard against large image
- `diffusers` patch custom pipelines for `qk_norm`
## Update for 2026-05-13
modules/models_hf.py
+10 -13
View File
@@ -1,7 +1,6 @@
import os
import time
import gradio as gr
from installer import install
from modules.logger import log
from modules.shared import opts
@@ -19,22 +18,20 @@ def hf_init():
os.environ.setdefault('HF_ENABLE_PARALLEL_LOADING', 'true' if opts.sd_parallel_load else 'false')
os.environ.setdefault('HF_HUB_CACHE', opts.hfcache_dir)
os.environ.setdefault('HF_XET_CACHE', opts.xetcache_dir)
if opts.hf_transfer_mode == 'requests':
if opts.hf_transfer_mode == 'HTTP':
os.environ.setdefault('HF_XET_HIGH_PERFORMANCE', 'false')
os.environ.setdefault('HF_HUB_ENABLE_HF_TRANSFER', 'false')
os.environ.setdefault('HF_HUB_DISABLE_XET', 'true')
elif opts.hf_transfer_mode == 'rust':
install('hf_transfer')
elif opts.hf_transfer_mode == 'XET':
os.environ.setdefault('HF_XET_HIGH_PERFORMANCE', 'false')
os.environ.setdefault('HF_HUB_ENABLE_HF_TRANSFER', 'true')
os.environ.setdefault('HF_HUB_DISABLE_XET', 'true')
elif opts.hf_transfer_mode == 'xet':
install('hf_xet')
import huggingface_hub
huggingface_hub.utils._runtime.is_xet_available = lambda: True # pylint: disable=protected-access
os.environ.setdefault('HF_XET_HIGH_PERFORMANCE', 'true')
os.environ.setdefault('HF_HUB_ENABLE_HF_TRANSFER', 'true')
os.environ.setdefault('HF_HUB_DISABLE_XET', 'false')
elif opts.hf_transfer_mode == 'XET HighPerformance':
os.environ.setdefault('HF_XET_HIGH_PERFORMANCE', 'true')
os.environ.setdefault('HF_HUB_DISABLE_XET', 'false')
os.environ.setdefault('HF_XET_RECONSTRUCT_WRITE_SEQUENTIALLY', 'false')
elif opts.hf_transfer_mode == 'XET Sequential':
os.environ.setdefault('HF_XET_HIGH_PERFORMANCE', 'false')
os.environ.setdefault('HF_HUB_DISABLE_XET', 'false')
os.environ.setdefault('HF_XET_RECONSTRUCT_WRITE_SEQUENTIALLY', 'true')
obfuscated_token = None
if len(opts.huggingface_token) > 0 and opts.huggingface_token.startswith('hf_'):
modules/sd_hijack_hfhub.py
+4 -4
View File
@@ -9,14 +9,14 @@ orig_xet_get = None
def http_get_hijack(*args, **kwargs):
from modules.shared import state
from modules.shared import state, opts
if len(args) > 0 and isinstance(args[0], str) and args[0].endswith(".json"):
return orig_http_get(*args, **kwargs)
jobid = state.begin('Download')
fn = kwargs.get("displayed_filename", None)
size = kwargs.get("expected_size", None)
if fn and not fn.endswith(".json") and size is not None and size > 10240:
log.debug(f'Download: type=http fn="{fn}" size={size}')
log.debug(f'Download: type=http mode="{opts.hf_transfer_mode}" fn="{fn}" size={size}')
debug(f'Download start: type=http args={args} kwargs={kwargs}')
t0 = time.time()
res = orig_http_get(*args, **kwargs)
@@ -27,14 +27,14 @@ def http_get_hijack(*args, **kwargs):
def xet_get_hijack(*args, **kwargs):
from modules.shared import state
from modules.shared import state, opts
if len(args) > 0 and isinstance(args[0], str) and args[0].endswith(".json"):
return orig_xet_get(*args, **kwargs)
jobid = state.begin('Download')
fn = kwargs.get("displayed_filename", None)
size = kwargs.get("expected_size", None)
if fn and not fn.endswith(".json"):
log.debug(f'Download: type=xet fn="{fn}" size={size}')
log.debug(f'Download: type=xet mode="{opts.hf_transfer_mode}" fn="{fn}" size={size}')
debug(f'Download start: type=xet args={args} kwargs={kwargs}')
res = orig_xet_get(*args, **kwargs)
debug(f'Download end: type=xet res={res}')
modules/ui_definitions.py
+1 -1
View File
@@ -606,7 +606,7 @@ def create_settings(cmd_opts):
"huggingface_sep": OptionInfo("<h2>Huggingface</h2>", "", gr.HTML),
"diffuser_cache_config": OptionInfo(True, "Use cached model config when available"),
"huggingface_token": OptionInfo('', 'HuggingFace token', gr.Textbox, {"lines": 2}, secret=True, env_var='HF_TOKEN'),
"hf_transfer_mode": OptionInfo("rust", "HuggingFace download method", gr.Radio, {"choices": ['requests', 'rust', 'xet']}),
"hf_transfer_mode": OptionInfo("XET", "HuggingFace download method", gr.Dropdown, {"choices": ['HTTP', 'XET', 'XET HighPerformance', 'XET Sequential']}),
"huggingface_mirror": OptionInfo('', 'HuggingFace mirror', gr.Textbox),
"offline_mode": OptionInfo(False, 'Force offline mode', gr.Checkbox),
package.json
+1 -1
View File
@@ -18,7 +18,7 @@
"venv": ". venv/bin/activate",
"start": ". venv/bin/activate; python launch.py --debug",
"localize": "node cli/localize.js",
"packages": ". venv/bin/activate && pip install --upgrade accelerate huggingface_hub safetensors tokenizers peft pytorch_lightning pylint ruff",
"packages": ". venv/bin/activate && pip install --upgrade accelerate huggingface_hub hf_xet safetensors tokenizers peft pytorch_lightning pylint ruff",
"format": ". venv/bin/activate && pre-commit run --all-files",
"format-win": "venv\\scripts\\activate && pre-commit run --all-files",
"eslint": "eslint javascript/",
pipelines/bria/transformer_block.py
+3 -1
View File
@@ -106,9 +106,11 @@ class FluxSingleTransformerBlock(nn.Module):
@maybe_allow_in_graph
class FluxTransformerBlock(nn.Module):
def __init__(
self, dim: int, num_attention_heads: int, attention_head_dim: int, qk_norm: str = "rms_norm", eps: float = 1e-6
self, dim: int, num_attention_heads: int, attention_head_dim: int, qk_norm: str | bool | None = "rms_norm", eps: float = 1e-6
):
super().__init__()
if isinstance(qk_norm, bool):
qk_norm = "rms_norm" if qk_norm else None
self.norm1 = AdaLayerNormZero(dim)
self.norm1_context = AdaLayerNormZero(dim)
pipelines/meissonic/transformer.py
+3 -1
View File
@@ -440,8 +440,10 @@ class TransformerBlock(nn.Module):
processing of `context` conditions.
"""
def __init__(self, dim, num_attention_heads, attention_head_dim, qk_norm="rms_norm", eps=1e-6):
def __init__(self, dim, num_attention_heads, attention_head_dim, qk_norm: str | bool | None = "rms_norm", eps=1e-6):
super().__init__()
if isinstance(qk_norm, bool):
qk_norm = "rms_norm" if qk_norm else None
self.norm1 = AdaLayerNormZero(dim)
pipelines/vibe/vibe_sana_editing.py
+14 -4
View File
@@ -23,6 +23,14 @@ from .edit_head import MetaConnector
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
def _normalize_qk_norm(qk_norm: str | bool | None) -> str | None:
"""Normalize legacy boolean qk_norm values to Diffusers-compatible strings."""
if isinstance(qk_norm, bool):
# Legacy checkpoints can store qk_norm as bool while current diffusers expects a string.
return "rms_norm" if qk_norm else None
return qk_norm
class SanaLinearAttnProcessor2_0:
r"""Processor for implementing scaled dot-product linear attention."""
@@ -112,7 +120,7 @@ class SanaTransformerBlock(nn.Module):
cross_attention_dim: int = 2240,
norm_eps: float = 1e-6,
mlp_ratio: float = 2.5,
qk_norm: str | None = None,
qk_norm: str | bool | None = None,
*,
attention_out_bias: bool = True,
attention_bias: bool = True,
@@ -130,12 +138,13 @@ class SanaTransformerBlock(nn.Module):
cross_attention_dim (int): The dimension of the cross-attention context.
norm_eps (float): Epsilon value for layer normalization stability.
mlp_ratio (float): Expansion ratio for the feed-forward network hidden dimension.
qk_norm (str | None): Normalization method for Query/Key vectors.
qk_norm (str | bool | None): Normalization method for Query/Key vectors.
attention_out_bias (bool): Whether to include bias in the attention output projection.
attention_bias (bool): Whether to include bias in the attention Q/K/V projections.
norm_elementwise_affine (bool): Whether to learn affine parameters for normalization.
"""
super().__init__()
qk_norm = _normalize_qk_norm(qk_norm)
# 1. Self Attention
self.norm1 = nn.LayerNorm(dim, elementwise_affine=False, eps=norm_eps)
@@ -269,7 +278,7 @@ class VIBESanaEditingModel(SanaTransformer2DModel):
patch_size: int = 1,
norm_eps: float = 1e-6,
interpolation_scale: int | None = None,
qk_norm: str | None = None,
qk_norm: str | bool | None = None,
timestep_scale: float = 1.0,
input_condition_type: str = "channel_cat",
edit_head_input_dim: int = 2048,
@@ -299,7 +308,7 @@ class VIBESanaEditingModel(SanaTransformer2DModel):
patch_size (int): Size of the patches extracted from the input latent.
norm_eps (float): Epsilon for layer normalization.
interpolation_scale (int | None): Scale factor for positional embedding interpolation.
qk_norm (str | None): Normalization type for Query/Key (e.g., 'rms_norm').
qk_norm (str | bool | None): Normalization type for Query/Key (e.g., 'rms_norm').
timestep_scale (float): Scale factor for the timestep.
input_condition_type (str): Method for conditioning on the input image. Options: 'channel_cat', 'seq_cat'.
edit_head_input_dim (int): Input dimension for the `MetaConnector` edit head.
@@ -311,6 +320,7 @@ class VIBESanaEditingModel(SanaTransformer2DModel):
guidance_embeds (bool): Whether to use additional guidance embeddings.
"""
super(SanaTransformer2DModel, self).__init__()
qk_norm = _normalize_qk_norm(qk_norm)
out_channels = out_channels or in_channels
inner_dim = num_attention_heads * attention_head_dim
requirements.txt
+2 -1
View File
@@ -30,7 +30,8 @@ requests==2.32.3
tqdm==4.67.3
accelerate==1.13.0
einops==0.8.1
huggingface_hub==1.14.0
huggingface_hub==1.15.0
hf_xet==1.5.0
numpy==2.1.2
pandas==2.3.1
protobuf==6.33.5