Merge 6f4d889053 into 451af70154

Adds a new command-line argument `--total-ram` to limit the amount of
system RAM that ComfyUI considers available, allowing users to simulate
lower memory environments. This enables more predictable behavior when
testing or running on systems with limited resources.

Rationale:

I run Comfy inside a Docker container. Using `mem_limit` doesn't hide
total system RAM from Comfy, so OOM can occur easily. Cache pressure
limits cause frequent out-of-memory errors. Adding this flag allows
precise control over visible memory.

Signed-off-by: blob42 <contact@blob42.xyz>

comfy/cli_args.py

@@ -90,6 +90,7 @@ parser.add_argument("--directml", type=int, nargs="?", metavar="DIRECTML_DEVICE"
 parser.add_argument("--oneapi-device-selector", type=str, default=None, metavar="SELECTOR_STRING", help="Sets the oneAPI device(s) this instance will use.")
 parser.add_argument("--disable-ipex-optimize", action="store_true", help="Disables ipex.optimize default when loading models with Intel's Extension for Pytorch.")
 parser.add_argument("--supports-fp8-compute", action="store_true", help="ComfyUI will act like if the device supports fp8 compute.")
+parser.add_argument("--total-ram", type=float, default=0, help="Maximum system RAM visible to comfy in GB (default 0: all)")
 
 class LatentPreviewMethod(enum.Enum):
     NoPreviews = "none"

comfy/model_management.py

@@ -192,8 +192,12 @@ def get_total_memory(dev=None, torch_total_too=False):
     if dev is None:
         dev = get_torch_device()
 
-    if hasattr(dev, 'type') and (dev.type == 'cpu' or dev.type == 'mps'):
-        mem_total = psutil.virtual_memory().total
+    if hasattr(dev, "type") and (dev.type == "cpu" or dev.type == "mps"):
+        mem_total = 0
+        if args.total_ram != 0:
+            mem_total = args.total_ram * 1024 * 1024
+        else:
+            mem_total = psutil.virtual_memory().total
         mem_total_torch = mem_total
     else:
         if directml_enabled:
@@ -236,8 +240,15 @@ def mac_version():
         return None
 
 total_vram = get_total_memory(get_torch_device()) / (1024 * 1024)
-total_ram = psutil.virtual_memory().total / (1024 * 1024)
-logging.info("Total VRAM {:0.0f} MB, total RAM {:0.0f} MB".format(total_vram, total_ram))
+
+total_ram = 0
+if args.total_ram != 0:
+    total_ram = args.total_ram * (1024)  # arg in GB
+else:
+    total_ram = psutil.virtual_memory().total / (1024 * 1024)
+logging.info(
+    "Total VRAM {:0.0f} MB, total RAM {:0.0f} MB".format(total_vram, total_ram)
+)
 
 try:
     logging.info("pytorch version: {}".format(torch_version))

comfy/text_encoders/llama.py

@@ -77,6 +77,28 @@ class Qwen25_3BConfig:
     rope_scale = None
     final_norm: bool = True
 
+@dataclass
+class Qwen3_06BConfig:
+    vocab_size: int = 151936
+    hidden_size: int = 1024
+    intermediate_size: int = 3072
+    num_hidden_layers: int = 28
+    num_attention_heads: int = 16
+    num_key_value_heads: int = 8
+    max_position_embeddings: int = 32768
+    rms_norm_eps: float = 1e-6
+    rope_theta: float = 1000000.0
+    transformer_type: str = "llama"
+    head_dim = 128
+    rms_norm_add = False
+    mlp_activation = "silu"
+    qkv_bias = False
+    rope_dims = None
+    q_norm = "gemma3"
+    k_norm = "gemma3"
+    rope_scale = None
+    final_norm: bool = True
+
 @dataclass
 class Qwen3_4BConfig:
     vocab_size: int = 151936
@@ -641,6 +663,15 @@ class Qwen25_3B(BaseLlama, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
+class Qwen3_06B(BaseLlama, torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        config = Qwen3_06BConfig(**config_dict)
+        self.num_layers = config.num_hidden_layers
+
+        self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
+        self.dtype = dtype
+
 class Qwen3_4B(BaseLlama, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()

comfy/utils.py

@@ -611,6 +611,14 @@ def flux_to_diffusers(mmdit_config, output_prefix=""):
                         "ff_context.net.0.proj.bias": "txt_mlp.0.bias",
                         "ff_context.net.2.weight": "txt_mlp.2.weight",
                         "ff_context.net.2.bias": "txt_mlp.2.bias",
+                        "ff.linear_in.weight": "img_mlp.0.weight",  # LyCoris LoKr
+                        "ff.linear_in.bias": "img_mlp.0.bias",
+                        "ff.linear_out.weight": "img_mlp.2.weight",
+                        "ff.linear_out.bias": "img_mlp.2.bias",
+                        "ff_context.linear_in.weight": "txt_mlp.0.weight",
+                        "ff_context.linear_in.bias": "txt_mlp.0.bias",
+                        "ff_context.linear_out.weight": "txt_mlp.2.weight",
+                        "ff_context.linear_out.bias": "txt_mlp.2.bias",
                         "attn.norm_q.weight": "img_attn.norm.query_norm.scale",
                         "attn.norm_k.weight": "img_attn.norm.key_norm.scale",
                         "attn.norm_added_q.weight": "txt_attn.norm.query_norm.scale",

comfy_api_nodes/nodes_vidu.py

@@ -703,7 +703,7 @@ class Vidu2ReferenceVideoNode(IO.ComfyNode):
                     "subjects",
                     template=IO.Autogrow.TemplateNames(
                         IO.Image.Input("reference_images"),
-                        names=["subject1", "subject2", "subject3"],
+                        names=["subject1", "subject2", "subject3", "subject4", "subject5", "subject6", "subject7"],
                         min=1,
                     ),
                     tooltip="For each subject, provide up to 3 reference images (7 images total across all subjects). "
@@ -738,7 +738,7 @@ class Vidu2ReferenceVideoNode(IO.ComfyNode):
                     control_after_generate=True,
                 ),
                 IO.Combo.Input("aspect_ratio", options=["16:9", "9:16", "4:3", "3:4", "1:1"]),
-                IO.Combo.Input("resolution", options=["720p"]),
+                IO.Combo.Input("resolution", options=["720p", "1080p"]),
                 IO.Combo.Input(
                     "movement_amplitude",
                     options=["auto", "small", "medium", "large"],