Merge 019eaab4c9 into fcd9a236b0

This logic was checking comfy_cast_weights, and going straight to
to the forward_comfy_cast_weights implementation without
attempting to downscale input to fp8 in the event comfy_cast_weights
is set.

The main reason comfy_cast_weights would be set would be for async
offload, which is not a good reason to nix FP8MM.

So instead, and together the underlying exclusions for FP8MM which
are:

* having a weight_function (usually LowVramPatch)
* force_cast_weights (compute dtype override)
* the weight is not Quantized
* the input is already quantized
* the model or layer has MM explictily disabled.

If you get past all of those exclusions, quantize the input tensor.
Then hand the new input, quantized or not off to
forward_comfy_cast_weights to handle it. If the weight is offloaded
but input is quantized you will get an offloaded MM8.

.github/workflows/stable-release.yml

@@ -117,7 +117,7 @@ jobs:
           ./python.exe get-pip.py
           ./python.exe -s -m pip install ../${{ inputs.cache_tag }}_python_deps/*
 
-          grep comfyui ../ComfyUI/requirements.txt > ./requirements_comfyui.txt
+          grep comfy ../ComfyUI/requirements.txt > ./requirements_comfyui.txt
           ./python.exe -s -m pip install -r requirements_comfyui.txt
           rm requirements_comfyui.txt
 

app/user_manager.py

@@ -377,8 +377,22 @@ class UserManager():
             try:
                 body = await request.read()
 
-                with open(path, "wb") as f:
-                    f.write(body)
+                # Pretty print JSON files for better source control
+                if path.lower().endswith('.json'):
+                    try:
+                        # Parse JSON and re-serialize with indentation
+                        json_data = json.loads(body.decode('utf-8'))
+                        formatted_json = json.dumps(json_data, indent=2)
+                        with open(path, "w", encoding='utf-8') as f:
+                            f.write(formatted_json)
+                    except (json.JSONDecodeError, UnicodeDecodeError):
+                        # If JSON parsing fails, save as-is
+                        with open(path, "wb") as f:
+                            f.write(body)
+                else:
+                    # Non-JSON files are saved as-is
+                    with open(path, "wb") as f:
+                        f.write(body)
             except OSError as e:
                 logging.warning(f"Error saving file '{path}': {e}")
                 return web.Response(

comfy/model_patcher.py

@@ -718,6 +718,7 @@ class ModelPatcher:
                             continue
 
                 cast_weight = self.force_cast_weights
+                m.comfy_force_cast_weights = self.force_cast_weights
                 if lowvram_weight:
                     if hasattr(m, "comfy_cast_weights"):
                         m.weight_function = []
@@ -790,11 +791,12 @@ class ModelPatcher:
                 for param in params:
                     self.pin_weight_to_device("{}.{}".format(n, param))
 
+            usable_stat = "{:.2f} MB usable,".format(lowvram_model_memory / (1024 * 1024)) if lowvram_model_memory < 1e32 else ""
             if lowvram_counter > 0:
-                logging.info("loaded partially; {:.2f} MB usable, {:.2f} MB loaded, {:.2f} MB offloaded, {:.2f} MB buffer reserved, lowvram patches: {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), lowvram_mem_counter / (1024 * 1024), offload_buffer / (1024 * 1024), patch_counter))
+                logging.info("loaded partially; {} {:.2f} MB loaded, {:.2f} MB offloaded, {:.2f} MB buffer reserved, lowvram patches: {}".format(usable_stat, mem_counter / (1024 * 1024), lowvram_mem_counter / (1024 * 1024), offload_buffer / (1024 * 1024), patch_counter))
                 self.model.model_lowvram = True
             else:
-                logging.info("loaded completely; {:.2f} MB usable, {:.2f} MB loaded, full load: {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), full_load))
+                logging.info("loaded completely; {} {:.2f} MB loaded, full load: {}".format(usable_stat, mem_counter / (1024 * 1024), full_load))
                 self.model.model_lowvram = False
                 if full_load:
                     self.model.to(device_to)

comfy/ops.py

@@ -654,29 +654,29 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
                 run_every_op()
 
                 input_shape = input.shape
-                tensor_3d = input.ndim == 3
-
-                if self._full_precision_mm or self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0:
-                    return self.forward_comfy_cast_weights(input, *args, **kwargs)
+                reshaped_3d = False
 
                 if (getattr(self, 'layout_type', None) is not None and
-                    not isinstance(input, QuantizedTensor)):
+                    not isinstance(input, QuantizedTensor) and not self._full_precision_mm and
+                    not getattr(self, 'comfy_force_cast_weights', False) and
+                    len(self.weight_function) == 0 and len(self.bias_function) == 0):
 
                     # Reshape 3D tensors to 2D for quantization (needed for NVFP4 and others)
-                    if tensor_3d:
-                        input = input.reshape(-1, input_shape[2])
+                    input_reshaped = input.reshape(-1, input_shape[2]) if input.ndim == 3 else input
 
-                    if input.ndim != 2:
-                        # Fall back to comfy_cast_weights for non-2D tensors
-                        return self.forward_comfy_cast_weights(input.reshape(input_shape), *args, **kwargs)
+                    # Fall back to non-quantized for non-2D tensors
+                    if input_reshaped.ndim == 2:
+                        reshaped_3d = input.ndim == 3
+                        # dtype is now implicit in the layout class
+                        scale = getattr(self, 'input_scale', None)
+                        if scale is not None:
+                            scale = comfy.model_management.cast_to_device(scale, input.device, None)
+                        input = QuantizedTensor.from_float(input_reshaped, self.layout_type, scale=scale)
 
-                    # dtype is now implicit in the layout class
-                    input = QuantizedTensor.from_float(input, self.layout_type, scale=getattr(self, 'input_scale', None))
-
-                output = self._forward(input, self.weight, self.bias)
+                output = self.forward_comfy_cast_weights(input)
 
                 # Reshape output back to 3D if input was 3D
-                if tensor_3d:
+                if reshaped_3d:
                     output = output.reshape((input_shape[0], input_shape[1], self.weight.shape[0]))
 
                 return output

comfy/quant_ops.py

@@ -19,6 +19,7 @@ try:
         cuda_version = tuple(map(int, str(torch.version.cuda).split('.')))
         if cuda_version < (13,):
             ck.registry.disable("cuda")
+            logging.warning("WARNING: You need pytorch with cu130 or higher to use optimized CUDA operations.")
 
     ck.registry.disable("triton")
     for k, v in ck.list_backends().items():

comfy/sd.py

@@ -218,7 +218,7 @@ class CLIP:
             if unprojected:
                 self.cond_stage_model.set_clip_options({"projected_pooled": False})
 
-            self.load_model()
+            self.load_model(tokens)
             self.cond_stage_model.set_clip_options({"execution_device": self.patcher.load_device})
             all_hooks.reset()
             self.patcher.patch_hooks(None)
@@ -266,7 +266,7 @@ class CLIP:
         if return_pooled == "unprojected":
             self.cond_stage_model.set_clip_options({"projected_pooled": False})
 
-        self.load_model()
+        self.load_model(tokens)
         self.cond_stage_model.set_clip_options({"execution_device": self.patcher.load_device})
         o = self.cond_stage_model.encode_token_weights(tokens)
         cond, pooled = o[:2]
@@ -299,8 +299,11 @@ class CLIP:
             sd_clip[k] = sd_tokenizer[k]
         return sd_clip
 
-    def load_model(self):
-        model_management.load_model_gpu(self.patcher)
+    def load_model(self, tokens={}):
+        memory_used = 0
+        if hasattr(self.cond_stage_model, "memory_estimation_function"):
+            memory_used = self.cond_stage_model.memory_estimation_function(tokens, device=self.patcher.load_device)
+        model_management.load_models_gpu([self.patcher], memory_required=memory_used)
         return self.patcher
 
     def get_key_patches(self):

comfy/supported_models.py

@@ -845,7 +845,7 @@ class LTXAV(LTXV):
 
     def __init__(self, unet_config):
         super().__init__(unet_config)
-        self.memory_usage_factor = 0.055  # TODO
+        self.memory_usage_factor = 0.061  # TODO
 
     def get_model(self, state_dict, prefix="", device=None):
         out = model_base.LTXAV(self, device=device)

comfy/text_encoders/lt.py

@@ -36,10 +36,10 @@ class LTXAVGemmaTokenizer(sd1_clip.SD1Tokenizer):
 
 class Gemma3_12BModel(sd1_clip.SDClipModel):
     def __init__(self, device="cpu", layer="all", layer_idx=None, dtype=None, attention_mask=True, model_options={}):
-        llama_scaled_fp8 = model_options.get("gemma_scaled_fp8", None)
-        if llama_scaled_fp8 is not None:
+        llama_quantization_metadata = model_options.get("llama_quantization_metadata", None)
+        if llama_quantization_metadata is not None:
             model_options = model_options.copy()
-            model_options["scaled_fp8"] = llama_scaled_fp8
+            model_options["quantization_metadata"] = llama_quantization_metadata
 
         super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"start": 2, "pad": 0}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Gemma3_12B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
 
@@ -98,10 +98,13 @@ class LTXAVTEModel(torch.nn.Module):
 
         out, pooled, extra = self.gemma3_12b.encode_token_weights(token_weight_pairs)
         out_device = out.device
+        if comfy.model_management.should_use_bf16(self.execution_device):
+            out = out.to(device=self.execution_device, dtype=torch.bfloat16)
         out = out.movedim(1, -1).to(self.execution_device)
         out = 8.0 * (out - out.mean(dim=(1, 2), keepdim=True)) / (out.amax(dim=(1, 2), keepdim=True) - out.amin(dim=(1, 2), keepdim=True) + 1e-6)
         out = out.reshape((out.shape[0], out.shape[1], -1))
         out = self.text_embedding_projection(out)
+        out = out.float()
         out_vid = self.video_embeddings_connector(out)[0]
         out_audio = self.audio_embeddings_connector(out)[0]
         out = torch.concat((out_vid, out_audio), dim=-1)
@@ -118,13 +121,21 @@ class LTXAVTEModel(torch.nn.Module):
 
             return self.load_state_dict(sdo, strict=False)
 
+    def memory_estimation_function(self, token_weight_pairs, device=None):
+        constant = 6.0
+        if comfy.model_management.should_use_bf16(device):
+            constant /= 2.0
 
-def ltxav_te(dtype_llama=None, llama_scaled_fp8=None):
+        token_weight_pairs = token_weight_pairs.get("gemma3_12b", [])
+        num_tokens = sum(map(lambda a: len(a), token_weight_pairs))
+        return num_tokens * constant * 1024 * 1024
+
+def ltxav_te(dtype_llama=None, llama_quantization_metadata=None):
     class LTXAVTEModel_(LTXAVTEModel):
         def __init__(self, device="cpu", dtype=None, model_options={}):
-            if llama_scaled_fp8 is not None and "llama_scaled_fp8" not in model_options:
+            if llama_quantization_metadata is not None:
                 model_options = model_options.copy()
-                model_options["llama_scaled_fp8"] = llama_scaled_fp8
+                model_options["llama_quantization_metadata"] = llama_quantization_metadata
             if dtype_llama is not None:
                 dtype = dtype_llama
             super().__init__(dtype_llama=dtype_llama, device=device, dtype=dtype, model_options=model_options)

comfy_extras/nodes_lt_audio.py

@@ -185,6 +185,10 @@ class LTXAVTextEncoderLoader(io.ComfyNode):
                 io.Combo.Input(
                     "ckpt_name",
                     options=folder_paths.get_filename_list("checkpoints"),
+                ),
+                io.Combo.Input(
+                    "device",
+                    options=["default", "cpu"],
                 )
             ],
             outputs=[io.Clip.Output()],
@@ -197,7 +201,11 @@ class LTXAVTextEncoderLoader(io.ComfyNode):
         clip_path1 = folder_paths.get_full_path_or_raise("text_encoders", text_encoder)
         clip_path2 = folder_paths.get_full_path_or_raise("checkpoints", ckpt_name)
 
-        clip = comfy.sd.load_clip(ckpt_paths=[clip_path1, clip_path2], embedding_directory=folder_paths.get_folder_paths("embeddings"), clip_type=clip_type)
+        model_options = {}
+        if device == "cpu":
+            model_options["load_device"] = model_options["offload_device"] = torch.device("cpu")
+
+        clip = comfy.sd.load_clip(ckpt_paths=[clip_path1, clip_path2], embedding_directory=folder_paths.get_folder_paths("embeddings"), clip_type=clip_type, model_options=model_options)
         return io.NodeOutput(clip)
 
 

requirements.txt

@@ -1,5 +1,5 @@
 comfyui-frontend-package==1.35.9
-comfyui-workflow-templates==0.7.67
+comfyui-workflow-templates==0.7.69
 comfyui-embedded-docs==0.3.1
 torch
 torchsde
@@ -21,7 +21,7 @@ psutil
 alembic
 SQLAlchemy
 av>=14.2.0
-comfy-kitchen>=0.2.3
+comfy-kitchen>=0.2.5
 
 #non essential dependencies:
 kornia>=0.7.1

server.py

@@ -518,7 +518,7 @@ 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"attachment; filename=\"{filename}\""})
 
                     if 'channel' not in request.rel_url.query:
                         channel = 'rgba'
@@ -538,7 +538,7 @@ 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"attachment; filename=\"{filename}\""})
 
                     elif channel == 'a':
                         with Image.open(file) as img:
@@ -555,7 +555,7 @@ 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"attachment; filename=\"{filename}\""})
                     else:
                         # Get content type from mimetype, defaulting to 'application/octet-stream'
                         content_type = mimetypes.guess_type(filename)[0] or 'application/octet-stream'
@@ -567,7 +567,7 @@ class PromptServer():
                         return web.FileResponse(
                             file,
                             headers={
-                                "Content-Disposition": f"filename=\"{filename}\"",
+                                "Content-Disposition": f"attachment; filename=\"{filename}\"",
                                 "Content-Type": content_type
                             }
                         )

tests-unit/prompt_server_test/user_manager_test.py

@@ -287,3 +287,72 @@ async def test_listuserdata_v2_url_encoded_path(aiohttp_client, app, tmp_path):
     assert entry["name"] == "file.txt"
     # Ensure the path is correctly decoded and uses forward slash
     assert entry["path"] == "my dir/file.txt"
+
+
+async def test_post_userdata_json_pretty_print(aiohttp_client, app, tmp_path):
+    """Test that JSON files are saved with pretty printing (indentation)"""
+    import json
+
+    client = await aiohttp_client(app)
+
+    # Create a compact JSON workflow
+    workflow_data = {
+        "nodes": [
+            {"id": "1", "type": "LoadImage", "inputs": {"image": "test.png"}},
+            {"id": "2", "type": "SaveImage", "inputs": {"images": ["1", 0]}}
+        ],
+        "metadata": {"version": "1.0", "author": "test"}
+    }
+    compact_json = json.dumps(workflow_data).encode('utf-8')
+
+    # Save as JSON file
+    resp = await client.post("/userdata/workflow.json", data=compact_json)
+    assert resp.status == 200
+
+    # Read the saved file and verify it's pretty-printed
+    with open(tmp_path / "workflow.json", "r", encoding='utf-8') as f:
+        saved_content = f.read()
+
+    # Verify the file contains indentation (pretty-printed)
+    assert "  " in saved_content  # Should have 2-space indentation
+    assert "\n" in saved_content  # Should have newlines
+
+    # Verify the content is still valid JSON and matches original data
+    saved_data = json.loads(saved_content)
+    assert saved_data == workflow_data
+
+    # Verify it's actually formatted (not compact)
+    # Compact JSON would be much shorter
+    assert len(saved_content) > len(compact_json)
+
+
+async def test_post_userdata_json_invalid_fallback(aiohttp_client, app, tmp_path):
+    """Test that invalid JSON is saved as-is without error"""
+    client = await aiohttp_client(app)
+
+    # Create invalid JSON content
+    invalid_json = b'{"invalid": json content}'
+
+    # Save as JSON file - should not fail
+    resp = await client.post("/userdata/invalid.json", data=invalid_json)
+    assert resp.status == 200
+
+    # Verify file was saved as-is
+    with open(tmp_path / "invalid.json", "rb") as f:
+        assert f.read() == invalid_json
+
+
+async def test_post_userdata_non_json_unchanged(aiohttp_client, app, tmp_path):
+    """Test that non-JSON files are saved unchanged"""
+    client = await aiohttp_client(app)
+
+    # Create binary content
+    binary_content = b'\x00\x01\x02\x03\x04\x05'
+
+    # Save as non-JSON file
+    resp = await client.post("/userdata/test.bin", data=binary_content)
+    assert resp.status == 200
+
+    # Verify file was saved exactly as-is
+    with open(tmp_path / "test.bin", "rb") as f:
+        assert f.read() == binary_content