Merge 2d3a8d523d into dd86b15521

This is log clutter and concerning to users. Its a false alarm.

comfy/model_management.py

@@ -1202,27 +1202,36 @@ def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False, str
         assert r is None
         assert stream is None
 
-        r = torch.empty_like(weight, dtype=weight._model_dtype, device=device)
+        cast_geometry = comfy.memory_management.tensors_to_geometries([ weight ])
+
+        if dtype is None:
+            dtype = weight._model_dtype
+
+        r = torch.empty_like(weight, dtype=dtype, device=device)
 
         signature = comfy_aimdo.model_vbar.vbar_fault(weight._v)
         if signature is not None:
             raw_tensor = comfy_aimdo.torch.aimdo_to_tensor(weight._v, device)
-            v_tensor = comfy.memory_management.interpret_gathered_like([r], raw_tensor)[0]
-
-        if comfy_aimdo.model_vbar.vbar_signature_compare(signature, weight._v_signature):
+            v_tensor = comfy.memory_management.interpret_gathered_like(cast_geometry, raw_tensor)[0]
+            if not comfy_aimdo.model_vbar.vbar_signature_compare(signature, weight._v_signature):
+                weight._v_signature = signature
+                #Send it over
+                v_tensor.copy_(weight, non_blocking=non_blocking)
             #always take a deep copy even if _v is good, as we have no reasonable point to unpin
             #a non comfy weight
             r.copy_(v_tensor)
             comfy_aimdo.model_vbar.vbar_unpin(weight._v)
             return r
 
+        if weight.dtype != r.dtype and weight.dtype != weight._model_dtype:
+            #Offloaded casting could skip this, however it would make the quantizations
+            #inconsistent between loaded and offloaded weights. So force the double casting
+            #that would happen in regular flow to make offload deterministic.
+            cast_buffer = torch.empty_like(weight, dtype=weight._model_dtype, device=device)
+            cast_buffer.copy_(weight, non_blocking=non_blocking)
+            weight = cast_buffer
         r.copy_(weight, non_blocking=non_blocking)
 
-        if signature is not None:
-            weight._v_signature = signature
-            v_tensor.copy_(r)
-            comfy_aimdo.model_vbar.vbar_unpin(weight._v)
-
         return r
 
     if device is None or weight.device == device:

comfy/model_patcher.py

@@ -161,6 +161,11 @@ def get_key_weight(model, key):
 
     return weight, set_func, convert_func
 
+def key_param_name_to_key(key, param):
+    if len(key) == 0:
+        return param
+    return "{}.{}".format(key, param)
+
 class AutoPatcherEjector:
     def __init__(self, model: 'ModelPatcher', skip_and_inject_on_exit_only=False):
         self.model = model
@@ -795,7 +800,7 @@ class ModelPatcher:
                         continue
 
                 for param in params:
-                    key = "{}.{}".format(n, param)
+                    key = key_param_name_to_key(n, param)
                     self.unpin_weight(key)
                     self.patch_weight_to_device(key, device_to=device_to)
                 if comfy.model_management.is_device_cuda(device_to):
@@ -811,7 +816,7 @@ class ModelPatcher:
                 n = x[1]
                 params = x[3]
                 for param in params:
-                    self.pin_weight_to_device("{}.{}".format(n, param))
+                    self.pin_weight_to_device(key_param_name_to_key(n, param))
 
             usable_stat = "{:.2f} MB usable,".format(lowvram_model_memory / (1024 * 1024)) if lowvram_model_memory < 1e32 else ""
             if lowvram_counter > 0:
@@ -917,7 +922,7 @@ class ModelPatcher:
                 if hasattr(m, "comfy_patched_weights") and m.comfy_patched_weights == True:
                     move_weight = True
                     for param in params:
-                        key = "{}.{}".format(n, param)
+                        key = key_param_name_to_key(n, param)
                         bk = self.backup.get(key, None)
                         if bk is not None:
                             if not lowvram_possible:
@@ -968,7 +973,7 @@ class ModelPatcher:
                         logging.debug("freed {}".format(n))
 
                         for param in params:
-                            self.pin_weight_to_device("{}.{}".format(n, param))
+                            self.pin_weight_to_device(key_param_name_to_key(n, param))
 
 
             self.model.model_lowvram = True
@@ -1501,7 +1506,7 @@ class ModelPatcherDynamic(ModelPatcher):
 
                 def setup_param(self, m, n, param_key):
                     nonlocal num_patches
-                    key = "{}.{}".format(n, param_key)
+                    key = key_param_name_to_key(n, param_key)
 
                     weight_function = []
 
@@ -1540,7 +1545,7 @@ class ModelPatcherDynamic(ModelPatcher):
 
                 else:
                     for param in params:
-                        key = "{}.{}".format(n, param)
+                        key = key_param_name_to_key(n, param)
                         weight, _, _ = get_key_weight(self.model, key)
                         weight.seed_key = key
                         set_dirty(weight, dirty)

comfy/text_encoders/anima.py

@@ -8,7 +8,7 @@ import torch
 class Qwen3Tokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer")
-        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=1024, embedding_key='qwen3_06b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, pad_token=151643, tokenizer_data=tokenizer_data)
+        super().__init__(tokenizer_path, pad_with_end=False, embedding_directory=embedding_directory, embedding_size=1024, embedding_key='qwen3_06b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, pad_token=151643, tokenizer_data=tokenizer_data)
 
 class T5XXLTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):

comfy/text_encoders/flux.py

@@ -118,7 +118,7 @@ class MistralTokenizerClass:
 class Mistral3Tokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         self.tekken_data = tokenizer_data.get("tekken_model", None)
-        super().__init__("", pad_with_end=False, embedding_size=5120, embedding_key='mistral3_24b', tokenizer_class=MistralTokenizerClass, has_end_token=False, pad_to_max_length=False, pad_token=11, start_token=1, max_length=99999999, min_length=1, pad_left=True, tokenizer_args=load_mistral_tokenizer(self.tekken_data), tokenizer_data=tokenizer_data)
+        super().__init__("", pad_with_end=False, embedding_directory=embedding_directory, embedding_size=5120, embedding_key='mistral3_24b', tokenizer_class=MistralTokenizerClass, has_end_token=False, pad_to_max_length=False, pad_token=11, start_token=1, max_length=99999999, min_length=1, pad_left=True, tokenizer_args=load_mistral_tokenizer(self.tekken_data), tokenizer_data=tokenizer_data)
 
     def state_dict(self):
         return {"tekken_model": self.tekken_data}
@@ -176,12 +176,12 @@ def flux2_te(dtype_llama=None, llama_quantization_metadata=None, pruned=False):
 class Qwen3Tokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer")
-        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=2560, embedding_key='qwen3_4b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=512, pad_token=151643, tokenizer_data=tokenizer_data)
+        super().__init__(tokenizer_path, pad_with_end=False, embedding_directory=embedding_directory, embedding_size=2560, embedding_key='qwen3_4b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=512, pad_token=151643, tokenizer_data=tokenizer_data)
 
 class Qwen3Tokenizer8B(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer")
-        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=4096, embedding_key='qwen3_8b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=512, pad_token=151643, tokenizer_data=tokenizer_data)
+        super().__init__(tokenizer_path, pad_with_end=False, embedding_directory=embedding_directory, embedding_size=4096, embedding_key='qwen3_8b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=512, pad_token=151643, tokenizer_data=tokenizer_data)
 
 class KleinTokenizer(sd1_clip.SD1Tokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}, name="qwen3_4b"):

comfy/text_encoders/z_image.py

@@ -6,7 +6,7 @@ import os
 class Qwen3Tokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer")
-        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=2560, embedding_key='qwen3_4b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, pad_token=151643, tokenizer_data=tokenizer_data)
+        super().__init__(tokenizer_path, pad_with_end=False, embedding_directory=embedding_directory, embedding_size=2560, embedding_key='qwen3_4b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, pad_token=151643, tokenizer_data=tokenizer_data)
 
 
 class ZImageTokenizer(sd1_clip.SD1Tokenizer):

comfy/utils.py

@@ -32,6 +32,7 @@ from comfy.cli_args import args, enables_dynamic_vram
 import json
 import time
 import mmap
+import warnings
 
 MMAP_TORCH_FILES = args.mmap_torch_files
 DISABLE_MMAP = args.disable_mmap
@@ -85,7 +86,10 @@ def load_safetensors(ckpt):
     header_size = struct.unpack("<Q", mapping[:8])[0]
     header = json.loads(mapping[8:8+header_size].decode("utf-8"))
 
-    data_area = torch.frombuffer(mapping, dtype=torch.uint8)[8 + header_size:]
+    with warnings.catch_warnings():
+        #We are working with read-only RAM by design
+        warnings.filterwarnings("ignore", message="The given buffer is not writable")
+        data_area = torch.frombuffer(mapping, dtype=torch.uint8)[8 + header_size:]
 
     sd = {}
     for name, info in header.items():

requirements.txt

@@ -22,7 +22,7 @@ alembic
 SQLAlchemy
 av>=14.2.0
 comfy-kitchen>=0.2.7
-comfy-aimdo>=0.1.6
+comfy-aimdo>=0.1.7
 requests
 
 #non essential dependencies:

server.py

@@ -521,7 +521,7 @@ class PromptServer():
                             buffer.seek(0)
 
                             return web.Response(body=buffer.read(), content_type=f'image/{image_format}',
-                                                headers={"Content-Disposition": f"filename=\"{filename}\""})
+                                                headers={"Content-Disposition": f"inline; filename=\"{filename}\""})
 
                     if 'channel' not in request.rel_url.query:
                         channel = 'rgba'
@@ -541,7 +541,7 @@ class PromptServer():
                             buffer.seek(0)
 
                             return web.Response(body=buffer.read(), content_type='image/png',
-                                                headers={"Content-Disposition": f"filename=\"{filename}\""})
+                                                headers={"Content-Disposition": f"inline; filename=\"{filename}\""})
 
                     elif channel == 'a':
                         with Image.open(file) as img:
@@ -558,7 +558,7 @@ class PromptServer():
                             alpha_buffer.seek(0)
 
                             return web.Response(body=alpha_buffer.read(), content_type='image/png',
-                                                headers={"Content-Disposition": f"filename=\"{filename}\""})
+                                                headers={"Content-Disposition": f"inline; filename=\"{filename}\""})
                     else:
                         # Get content type from mimetype, defaulting to 'application/octet-stream'
                         content_type = mimetypes.guess_type(filename)[0] or 'application/octet-stream'
@@ -570,7 +570,7 @@ class PromptServer():
                         return web.FileResponse(
                             file,
                             headers={
-                                "Content-Disposition": f"filename=\"{filename}\"",
+                                "Content-Disposition": f"inline; filename=\"{filename}\"",
                                 "Content-Type": content_type
                             }
                         )