Merge branch 'master' into update-frontend-1.32.10

Author: christian-byrne

.ci/windows_amd_base_files/README_VERY_IMPORTANT.txt

@@ -1,5 +1,5 @@
-As of the time of writing this you need this preview driver for best results:
-https://www.amd.com/en/resources/support-articles/release-notes/RN-AMDGPU-WINDOWS-PYTORCH-PREVIEW.html
+As of the time of writing this you need this driver for best results:
+https://www.amd.com/en/resources/support-articles/release-notes/RN-AMDGPU-WINDOWS-PYTORCH-7-1-1.html
 
 HOW TO RUN:
 
@@ -25,3 +25,4 @@ In the ComfyUI directory you will find a file: extra_model_paths.yaml.example
 Rename this file to: extra_model_paths.yaml and edit it with your favorite text editor.
 
 
+

.github/workflows/release-stable-all.yml

@@ -65,11 +65,11 @@ jobs:
       contents: "write"
       packages: "write"
       pull-requests: "read"
-    name: "Release AMD ROCm 6.4.4"
+    name: "Release AMD ROCm 7.1.1"
     uses: ./.github/workflows/stable-release.yml
     with:
       git_tag: ${{ inputs.git_tag }}
-      cache_tag: "rocm644"
+      cache_tag: "rocm711"
       python_minor: "12"
       python_patch: "10"
       rel_name: "amd"

app/user_manager.py

@@ -59,22 +59,26 @@ def get_request_user_id(self, request):
         user = "default"
         if args.multi_user and "comfy-user" in request.headers:
             user = request.headers["comfy-user"]
+            # Block System Users (use same error message to prevent probing)
+            if user.startswith(folder_paths.SYSTEM_USER_PREFIX):
+                raise KeyError("Unknown user: " + user)
 
         if user not in self.users:
             raise KeyError("Unknown user: " + user)
 
         return user
 
     def get_request_user_filepath(self, request, file, type="userdata", create_dir=True):
-        user_directory = folder_paths.get_user_directory()
-
         if type == "userdata":
-            root_dir = user_directory
+            root_dir = folder_paths.get_user_directory()
         else:
             raise KeyError("Unknown filepath type:" + type)
 
         user = self.get_request_user_id(request)
-        path = user_root = os.path.abspath(os.path.join(root_dir, user))
+        user_root = folder_paths.get_public_user_directory(user)
+        if user_root is None:
+            return None
+        path = user_root
 
         # prevent leaving /{type}
         if os.path.commonpath((root_dir, user_root)) != root_dir:
@@ -101,7 +105,11 @@ def add_user(self, name):
         name = name.strip()
         if not name:
             raise ValueError("username not provided")
+        if name.startswith(folder_paths.SYSTEM_USER_PREFIX):
+            raise ValueError("System User prefix not allowed")
         user_id = re.sub("[^a-zA-Z0-9-_]+", '-', name)
+        if user_id.startswith(folder_paths.SYSTEM_USER_PREFIX):
+            raise ValueError("System User prefix not allowed")
         user_id = user_id + "_" + str(uuid.uuid4())
 
         self.users[user_id] = name
@@ -132,7 +140,10 @@ async def post_users(request):
             if username in self.users.values():
                 return web.json_response({"error": "Duplicate username."}, status=400)
 
-            user_id = self.add_user(username)
+            try:
+                user_id = self.add_user(username)
+            except ValueError as e:
+                return web.json_response({"error": str(e)}, status=400)
             return web.json_response(user_id)
 
         @routes.get("/userdata")
@@ -424,7 +435,7 @@ async def move_userdata(request):
                 return source
 
             dest = get_user_data_path(request, check_exists=False, param="dest")
-            if not isinstance(source, str):
+            if not isinstance(dest, str):
                 return dest
 
             overwrite = request.query.get("overwrite", 'true') != "false"

comfy/cli_args.py

@@ -131,7 +131,8 @@ class LatentPreviewMethod(enum.Enum):
 
 parser.add_argument("--reserve-vram", type=float, default=None, help="Set the amount of vram in GB you want to reserve for use by your OS/other software. By default some amount is reserved depending on your OS.")
 
-parser.add_argument("--async-offload", action="store_true", help="Use async weight offloading.")
+parser.add_argument("--async-offload", nargs='?', const=2, type=int, default=None, metavar="NUM_STREAMS", help="Use async weight offloading. An optional argument controls the amount of offload streams. Default is 2. Enabled by default on Nvidia.")
+parser.add_argument("--disable-async-offload", action="store_true", help="Disable async weight offloading.")
 
 parser.add_argument("--force-non-blocking", action="store_true", help="Force ComfyUI to use non-blocking operations for all applicable tensors. This may improve performance on some non-Nvidia systems but can cause issues with some workflows.")
 

comfy/latent_formats.py

@@ -431,6 +431,7 @@ class HunyuanVideo(LatentFormat):
     ]
 
     latent_rgb_factors_bias = [ 0.0259, -0.0192, -0.0761]
+    taesd_decoder_name = "taehv"
 
 class Cosmos1CV8x8x8(LatentFormat):
     latent_channels = 16
@@ -494,7 +495,7 @@ def __init__(self):
         ]).view(1, self.latent_channels, 1, 1, 1)
 
 
-        self.taesd_decoder_name = None #TODO
+        self.taesd_decoder_name = "lighttaew2_1"
 
     def process_in(self, latent):
         latents_mean = self.latents_mean.to(latent.device, latent.dtype)
@@ -565,6 +566,7 @@ class Wan22(Wan21):
 
     def __init__(self):
         self.scale_factor = 1.0
+        self.taesd_decoder_name = "lighttaew2_2"
         self.latents_mean = torch.tensor([
                 -0.2289, -0.0052, -0.1323, -0.2339, -0.2799, 0.0174, 0.1838, 0.1557,
                 -0.1382, 0.0542, 0.2813, 0.0891, 0.1570, -0.0098, 0.0375, -0.1825,
@@ -719,6 +721,7 @@ class HunyuanVideo15(LatentFormat):
     latent_channels = 32
     latent_dimensions = 3
     scale_factor = 1.03682
+    taesd_decoder_name = "lighttaehy1_5"
 
 class Hunyuan3Dv2(LatentFormat):
     latent_channels = 64

comfy/ldm/flux/model.py

@@ -171,7 +171,10 @@ def forward_orig(
             pe = None
 
         blocks_replace = patches_replace.get("dit", {})
+        transformer_options["total_blocks"] = len(self.double_blocks)
+        transformer_options["block_type"] = "double"
         for i, block in enumerate(self.double_blocks):
+            transformer_options["block_index"] = i
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
@@ -215,7 +218,10 @@ def block_wrap(args):
         if self.params.global_modulation:
             vec, _ = self.single_stream_modulation(vec_orig)
 
+        transformer_options["total_blocks"] = len(self.single_blocks)
+        transformer_options["block_type"] = "single"
         for i, block in enumerate(self.single_blocks):
+            transformer_options["block_index"] = i
             if ("single_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}

comfy/ldm/lumina/model.py

@@ -509,23 +509,35 @@ def patchify_and_embed(
 
         if self.pad_tokens_multiple is not None:
             pad_extra = (-cap_feats.shape[1]) % self.pad_tokens_multiple
-            cap_feats = torch.cat((cap_feats, self.cap_pad_token.to(device=cap_feats.device, dtype=cap_feats.dtype).unsqueeze(0).repeat(cap_feats.shape[0], pad_extra, 1)), dim=1)
+            cap_feats = torch.cat((cap_feats, self.cap_pad_token.to(device=cap_feats.device, dtype=cap_feats.dtype, copy=True).unsqueeze(0).repeat(cap_feats.shape[0], pad_extra, 1)), dim=1)
 
         cap_pos_ids = torch.zeros(bsz, cap_feats.shape[1], 3, dtype=torch.float32, device=device)
         cap_pos_ids[:, :, 0] = torch.arange(cap_feats.shape[1], dtype=torch.float32, device=device) + 1.0
 
         B, C, H, W = x.shape
         x = self.x_embedder(x.view(B, C, H // pH, pH, W // pW, pW).permute(0, 2, 4, 3, 5, 1).flatten(3).flatten(1, 2))
 
+        rope_options = transformer_options.get("rope_options", None)
+        h_scale = 1.0
+        w_scale = 1.0
+        h_start = 0
+        w_start = 0
+        if rope_options is not None:
+            h_scale = rope_options.get("scale_y", 1.0)
+            w_scale = rope_options.get("scale_x", 1.0)
+
+            h_start = rope_options.get("shift_y", 0.0)
+            w_start = rope_options.get("shift_x", 0.0)
+
         H_tokens, W_tokens = H // pH, W // pW
         x_pos_ids = torch.zeros((bsz, x.shape[1], 3), dtype=torch.float32, device=device)
         x_pos_ids[:, :, 0] = cap_feats.shape[1] + 1
-        x_pos_ids[:, :, 1] = torch.arange(H_tokens, dtype=torch.float32, device=device).view(-1, 1).repeat(1, W_tokens).flatten()
-        x_pos_ids[:, :, 2] = torch.arange(W_tokens, dtype=torch.float32, device=device).view(1, -1).repeat(H_tokens, 1).flatten()
+        x_pos_ids[:, :, 1] = (torch.arange(H_tokens, dtype=torch.float32, device=device) * h_scale + h_start).view(-1, 1).repeat(1, W_tokens).flatten()
+        x_pos_ids[:, :, 2] = (torch.arange(W_tokens, dtype=torch.float32, device=device) * w_scale + w_start).view(1, -1).repeat(H_tokens, 1).flatten()
 
         if self.pad_tokens_multiple is not None:
             pad_extra = (-x.shape[1]) % self.pad_tokens_multiple
-            x = torch.cat((x, self.x_pad_token.to(device=x.device, dtype=x.dtype).unsqueeze(0).repeat(x.shape[0], pad_extra, 1)), dim=1)
+            x = torch.cat((x, self.x_pad_token.to(device=x.device, dtype=x.dtype, copy=True).unsqueeze(0).repeat(x.shape[0], pad_extra, 1)), dim=1)
             x_pos_ids = torch.nn.functional.pad(x_pos_ids, (0, 0, 0, pad_extra))
 
         freqs_cis = self.rope_embedder(torch.cat((cap_pos_ids, x_pos_ids), dim=1)).movedim(1, 2)

comfy/lora.py

@@ -313,6 +313,15 @@ def model_lora_keys_unet(model, key_map={}):
                 key_map["transformer.{}".format(key_lora)] = k
                 key_map["lycoris_{}".format(key_lora.replace(".", "_"))] = k #SimpleTuner lycoris format
 
+    if isinstance(model, comfy.model_base.Lumina2):
+        diffusers_keys = comfy.utils.z_image_to_diffusers(model.model_config.unet_config, output_prefix="diffusion_model.")
+        for k in diffusers_keys:
+            if k.endswith(".weight"):
+                to = diffusers_keys[k]
+                key_lora = k[:-len(".weight")]
+                key_map["diffusion_model.{}".format(key_lora)] = to
+                key_map["lycoris_{}".format(key_lora.replace(".", "_"))] = to
+
     return key_map
 
 

comfy/model_management.py

@@ -689,7 +689,7 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
             loaded_memory = loaded_model.model_loaded_memory()
             current_free_mem = get_free_memory(torch_dev) + loaded_memory
 
-            lowvram_model_memory = max(128 * 1024 * 1024, (current_free_mem - minimum_memory_required), min(current_free_mem * MIN_WEIGHT_MEMORY_RATIO, current_free_mem - minimum_inference_memory()))
+            lowvram_model_memory = max(0, (current_free_mem - minimum_memory_required), min(current_free_mem * MIN_WEIGHT_MEMORY_RATIO, current_free_mem - minimum_inference_memory()))
             lowvram_model_memory = lowvram_model_memory - loaded_memory
 
             if lowvram_model_memory == 0:
@@ -1012,9 +1012,18 @@ def force_channels_last():
 
 
 STREAMS = {}
-NUM_STREAMS = 1
-if args.async_offload:
-    NUM_STREAMS = 2
+NUM_STREAMS = 0
+if args.async_offload is not None:
+    NUM_STREAMS = args.async_offload
+else:
+    #  Enable by default on Nvidia
+    if is_nvidia():
+        NUM_STREAMS = 2
+
+if args.disable_async_offload:
+    NUM_STREAMS = 0
+
+if NUM_STREAMS > 0:
     logging.info("Using async weight offloading with {} streams".format(NUM_STREAMS))
 
 def current_stream(device):
@@ -1030,7 +1039,10 @@ def current_stream(device):
 stream_counters = {}
 def get_offload_stream(device):
     stream_counter = stream_counters.get(device, 0)
-    if NUM_STREAMS <= 1:
+    if NUM_STREAMS == 0:
+        return None
+
+    if torch.compiler.is_compiling():
         return None
 
     if device in STREAMS:
@@ -1043,15 +1055,19 @@ def get_offload_stream(device):
     elif is_device_cuda(device):
         ss = []
         for k in range(NUM_STREAMS):
-            ss.append(torch.cuda.Stream(device=device, priority=0))
+            s1 = torch.cuda.Stream(device=device, priority=0)
+            s1.as_context = torch.cuda.stream
+            ss.append(s1)
         STREAMS[device] = ss
         s = ss[stream_counter]
         stream_counters[device] = stream_counter
         return s
     elif is_device_xpu(device):
         ss = []
         for k in range(NUM_STREAMS):
-            ss.append(torch.xpu.Stream(device=device, priority=0))
+            s1 = torch.xpu.Stream(device=device, priority=0)
+            s1.as_context = torch.xpu.stream
+            ss.append(s1)
         STREAMS[device] = ss
         s = ss[stream_counter]
         stream_counters[device] = stream_counter
@@ -1069,12 +1085,19 @@ def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False, str
             if dtype is None or weight.dtype == dtype:
                 return weight
         if stream is not None:
-            with stream:
+            wf_context = stream
+            if hasattr(wf_context, "as_context"):
+                wf_context = wf_context.as_context(stream)
+            with wf_context:
                 return weight.to(dtype=dtype, copy=copy)
         return weight.to(dtype=dtype, copy=copy)
 
+
     if stream is not None:
-        with stream:
+        wf_context = stream
+        if hasattr(wf_context, "as_context"):
+            wf_context = wf_context.as_context(stream)
+        with wf_context:
             r = torch.empty_like(weight, dtype=dtype, device=device)
             r.copy_(weight, non_blocking=non_blocking)
     else: