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

Merge pull request #4917 from QualiaRain/fix/model-vae-fixes

Browse Source 
Fix wrong-variable checks in TE/quant loading and remote-VAE response handling
This commit is contained in:
Vladimir Mandic
2026-06-12 19:28:33 +02:00
committed by GitHub
parent dfc1fc70ab 023c829179
commit 8f4ea879b9
6 changed files with 42 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/lora/lora_load.py
+3 -1
View File
@@ -98,9 +98,11 @@ def load_safetensors(name, network_on_disk: network.NetworkOnDisk) -> network.Ne
continue
if parts[0] in ["clip_l","clip_g","t5","unet","transformer"]:
network_part = []
while parts[-1] in ["alpha","weight","lora_up","lora_down"]:
while parts and parts[-1] in ["alpha","weight","lora_up","lora_down"]:
network_part.insert(0,parts[-1])
parts = parts[0:-1]
if not parts:
continue
network_part = ".".join(network_part)
key_network_without_network_parts = "_".join(parts)
if key_network_without_network_parts.startswith("unet") or key_network_without_network_parts.startswith("transformer"):
modules/memstats.py
+5 -1
View File
@@ -54,7 +54,11 @@ def ram_stats():
res = process.memory_info()
if 'total' not in ram:
process = psutil.Process(os.getpid())
ram_total = 100 * res.rss / process.memory_percent()
mem_percent = process.memory_percent()
if mem_percent > 0:
ram_total = 100 * res.rss / mem_percent
else:
ram_total = res.rss
ram_total = min(ram_total, get_docker_limit(), get_runpod_limit())
ram['total'] = gb(ram_total)
ram['rss'] = gb(res.rss)
modules/merging/modules_sdxl.py
+1 -1
View File
@@ -251,7 +251,7 @@ def get_metadata():
"modelspec.license": recipe.license,
"modelspec.usage_hint": recipe.hint,
"modelspec.prediction_type": recipe.prediction,
"modelspec.dtype": str(recipe.dtype).split('.')[1],
"modelspec.dtype": str(recipe.dtype).split('.')[-1] if '.' in str(recipe.dtype) else str(recipe.dtype),
"modelspec.hash_sha256": "",
"modelspec.thumbnail": get_thumbnail(),
"recipe": json.dumps({
modules/model_quant.py
+14 -2
View File
@@ -445,7 +445,13 @@ def sdnq_quantize_model(model, op=None, sd_model=None, do_gc: bool = True, weigh
if quant_last_model_name is not None:
if "." in quant_last_model_name:
last_model_names = quant_last_model_name.split(".")
getattr(getattr(sd_model, last_model_names[0]), last_model_names[1]).to(quant_last_model_device)
if len(last_model_names) >= 2:
try:
parent = getattr(sd_model, last_model_names[0], None)
if parent is not None:
getattr(parent, last_model_names[1]).to(quant_last_model_device)
except (AttributeError, TypeError):
log.warning(f'Quantization: failed to access {quant_last_model_name}')
else:
getattr(sd_model, quant_last_model_name).to(quant_last_model_device)
if do_gc:
@@ -479,7 +485,13 @@ def sdnq_quantize_weights(sd_model):
if quant_last_model_name is not None:
if "." in quant_last_model_name:
last_model_names = quant_last_model_name.split(".")
getattr(getattr(sd_model, last_model_names[0]), last_model_names[1]).to(quant_last_model_device)
if len(last_model_names) >= 2:
try:
parent = getattr(sd_model, last_model_names[0], None)
if parent is not None:
getattr(parent, last_model_names[1]).to(quant_last_model_device)
except (AttributeError, TypeError):
log.warning(f'Quantization: failed to access {quant_last_model_name}')
else:
getattr(sd_model, quant_last_model_name).to(quant_last_model_device)
devices.torch_gc(force=True, reason='sdnq')
modules/model_te.py
+1 -1
View File
@@ -77,7 +77,7 @@ def load_t5(name=None, cache_dir=None):
elif '/' in name:
log.debug(f'Load model: type=T5 repo={name}')
quant_config = model_quant.create_config(module='TE')
if quantization_config is not None:
if quant_config is not None:
t5 = transformers.T5EncoderModel.from_pretrained(name, cache_dir=cache_dir, torch_dtype=devices.dtype, **quant_config)
else:
modules/vae/sd_vae_remote.py
+18 -6
View File
@@ -64,9 +64,9 @@ def remote_decode(latents: torch.Tensor, width: int = 0, height: int = 0, model_
t0 = time.time()
modelloader.hf_login()
latent_copy = latents.detach().clone().to(device=devices.cpu, dtype=devices.dtype)
latent_copy = latents.unsqueeze(0) if len(latents.shape) == 3 else latents
latent_copy = latent_copy.unsqueeze(0) if len(latent_copy.shape) == 3 else latent_copy
if model_type == 'hunyuanvideo':
latent_copy = latent_copy.unsqueeze(0) if len(latents.shape) == 4 else latents
latent_copy = latent_copy.unsqueeze(0) if len(latent_copy.shape) == 4 else latent_copy
for i in range(latent_copy.shape[0]):
params = {}
@@ -112,13 +112,25 @@ def remote_decode(latents: torch.Tensor, width: int = 0, height: int = 0, model_
timeout=300,
)
if not response.ok:
log.error(f'Decode: type="remote" model={model_type} code={response.status_code} shape={latent.shape} url="{url}" args={params} headers={response.headers} response={response.json()}')
try:
resp_json = response.json()
except Exception:
resp_json = response.text
log.error(f'Decode: type="remote" model={model_type} code={response.status_code} shape={latent.shape} url="{url}" args={params} headers={response.headers} response={resp_json}')
else:
content += len(response.content)
if shared.opts.remote_vae_type == 'raw' or 'video' in model_type:
shape = json.loads(response.headers["shape"])
dtype = response.headers["dtype"]
tensor = torch.frombuffer(bytearray(response.content), dtype=dtypes[dtype]).reshape(shape)
try:
shape = json.loads(response.headers.get("shape", "[]"))
dtype = response.headers.get("dtype", "float32")
if dtype in dtypes:
tensor = torch.frombuffer(bytearray(response.content), dtype=dtypes[dtype]).reshape(shape)
else:
log.warning(f'Decode: unknown dtype {dtype}')
continue
except (json.JSONDecodeError, KeyError, ValueError) as e:
log.warning(f'Decode: shape/dtype parsing error {e}')
continue
tensors.append(tensor)
elif shared.opts.remote_vae_type == 'jpg' or shared.opts.remote_vae_type == 'png':
image = Image.open(io.BytesIO(response.content)).convert("RGB")