Merge branch 'master' into dr-support-pip-cm

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -8,13 +8,15 @@ body:
         Before submitting a **Bug Report**, please ensure the following:
 
         - **1:** You are running the latest version of ComfyUI.
-        - **2:** You have looked at the existing bug reports and made sure this isn't already reported.
+        - **2:** You have your ComfyUI logs and relevant workflow on hand and will post them in this bug report.
         - **3:** You confirmed that the bug is not caused by a custom node. You can disable all custom nodes by passing
-        `--disable-all-custom-nodes` command line argument.
+        `--disable-all-custom-nodes` command line argument. If you have custom node try updating them to the latest version.
         - **4:** This is an actual bug in ComfyUI, not just a support question. A bug is when you can specify exact
         steps to replicate what went wrong and others will be able to repeat your steps and see the same issue happen.
 
-        If unsure, ask on the [ComfyUI Matrix Space](https://app.element.io/#/room/%23comfyui_space%3Amatrix.org) or the [Comfy Org Discord](https://discord.gg/comfyorg) first.
+        ## Very Important
+
+        Please make sure that you post ALL your ComfyUI logs in the bug report. A bug report without logs will likely be ignored.
   - type: checkboxes
     id: custom-nodes-test
     attributes:

README.md

@@ -112,10 +112,11 @@ Workflow examples can be found on the [Examples page](https://comfyanonymous.git
 
 ## Release Process
 
-ComfyUI follows a weekly release cycle targeting Friday but this regularly changes because of model releases or large changes to the codebase. There are three interconnected repositories:
+ComfyUI follows a weekly release cycle targeting Monday but this regularly changes because of model releases or large changes to the codebase. There are three interconnected repositories:
 
 1. **[ComfyUI Core](https://github.com/comfyanonymous/ComfyUI)**
-   - Releases a new stable version (e.g., v0.7.0)
+   - Releases a new stable version (e.g., v0.7.0) roughly every week.
+   - Commits outside of the stable release tags may be very unstable and break many custom nodes.
    - Serves as the foundation for the desktop release
 
 2. **[ComfyUI Desktop](https://github.com/Comfy-Org/desktop)**

comfy/cli_args.py

@@ -151,10 +151,11 @@ class PerformanceFeature(enum.Enum):
     Fp8MatrixMultiplication = "fp8_matrix_mult"
     CublasOps = "cublas_ops"
     AutoTune = "autotune"
-    PinnedMem = "pinned_memory"
 
 parser.add_argument("--fast", nargs="*", type=PerformanceFeature, help="Enable some untested and potentially quality deteriorating optimizations. This is used to test new features so using it might crash your comfyui. --fast with no arguments enables everything. You can pass a list specific optimizations if you only want to enable specific ones. Current valid optimizations: {}".format(" ".join(map(lambda c: c.value, PerformanceFeature))))
 
+parser.add_argument("--disable-pinned-memory", action="store_true", help="Disable pinned memory use.")
+
 parser.add_argument("--mmap-torch-files", action="store_true", help="Use mmap when loading ckpt/pt files.")
 parser.add_argument("--disable-mmap", action="store_true", help="Don't use mmap when loading safetensors.")
 

comfy/ldm/flux/model.py

@@ -210,7 +210,7 @@ class Flux(nn.Module):
         img = self.final_layer(img, vec)  # (N, T, patch_size ** 2 * out_channels)
         return img
 
-    def process_img(self, x, index=0, h_offset=0, w_offset=0):
+    def process_img(self, x, index=0, h_offset=0, w_offset=0, transformer_options={}):
         bs, c, h, w = x.shape
         patch_size = self.patch_size
         x = comfy.ldm.common_dit.pad_to_patch_size(x, (patch_size, patch_size))
@@ -222,10 +222,22 @@ class Flux(nn.Module):
         h_offset = ((h_offset + (patch_size // 2)) // patch_size)
         w_offset = ((w_offset + (patch_size // 2)) // patch_size)
 
-        img_ids = torch.zeros((h_len, w_len, 3), device=x.device, dtype=x.dtype)
+        steps_h = h_len
+        steps_w = w_len
+
+        rope_options = transformer_options.get("rope_options", None)
+        if rope_options is not None:
+            h_len = (h_len - 1.0) * rope_options.get("scale_y", 1.0) + 1.0
+            w_len = (w_len - 1.0) * rope_options.get("scale_x", 1.0) + 1.0
+
+            index += rope_options.get("shift_t", 0.0)
+            h_offset += rope_options.get("shift_y", 0.0)
+            w_offset += rope_options.get("shift_x", 0.0)
+
+        img_ids = torch.zeros((steps_h, steps_w, 3), device=x.device, dtype=x.dtype)
         img_ids[:, :, 0] = img_ids[:, :, 1] + index
-        img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1)
+        img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=steps_h, device=x.device, dtype=x.dtype).unsqueeze(1)
-        img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0)
+        img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=steps_w, device=x.device, dtype=x.dtype).unsqueeze(0)
         return img, repeat(img_ids, "h w c -> b (h w) c", b=bs)
 
     def forward(self, x, timestep, context, y=None, guidance=None, ref_latents=None, control=None, transformer_options={}, **kwargs):
@@ -241,7 +253,7 @@ class Flux(nn.Module):
 
         h_len = ((h_orig + (patch_size // 2)) // patch_size)
         w_len = ((w_orig + (patch_size // 2)) // patch_size)
-        img, img_ids = self.process_img(x)
+        img, img_ids = self.process_img(x, transformer_options=transformer_options)
         img_tokens = img.shape[1]
         if ref_latents is not None:
             h = 0

comfy/ldm/qwen_image/controlnet.py

@@ -44,7 +44,7 @@ class QwenImageControlNetModel(QwenImageTransformer2DModel):
         txt_start = round(max(((x.shape[-1] + (self.patch_size // 2)) // self.patch_size) // 2, ((x.shape[-2] + (self.patch_size // 2)) // self.patch_size) // 2))
         txt_ids = torch.arange(txt_start, txt_start + context.shape[1], device=x.device).reshape(1, -1, 1).repeat(x.shape[0], 1, 3)
         ids = torch.cat((txt_ids, img_ids), dim=1)
-        image_rotary_emb = self.pe_embedder(ids).squeeze(1).unsqueeze(2).to(x.dtype)
+        image_rotary_emb = self.pe_embedder(ids).to(x.dtype).contiguous()
         del ids, txt_ids, img_ids
 
         hidden_states = self.img_in(hidden_states) + self.controlnet_x_embedder(hint)

comfy/model_management.py

@@ -503,7 +503,11 @@ class LoadedModel:
         use_more_vram = lowvram_model_memory
         if use_more_vram == 0:
             use_more_vram = 1e32
-        self.model_use_more_vram(use_more_vram, force_patch_weights=force_patch_weights)
+        if use_more_vram > 0:
+            self.model_use_more_vram(use_more_vram, force_patch_weights=force_patch_weights)
+        else:
+            self.model.partially_unload(self.model.offload_device, -use_more_vram, force_patch_weights=force_patch_weights)
+
         real_model = self.model.model
 
         if is_intel_xpu() and not args.disable_ipex_optimize and 'ipex' in globals() and real_model is not None:
@@ -689,7 +693,10 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
             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(0.1, lowvram_model_memory - loaded_memory)
+            lowvram_model_memory = lowvram_model_memory - loaded_memory
+
+            if lowvram_model_memory == 0:
+                lowvram_model_memory = 0.1
 
         if vram_set_state == VRAMState.NO_VRAM:
             lowvram_model_memory = 0.1
@@ -1085,23 +1092,28 @@ def cast_to_device(tensor, device, dtype, copy=False):
 
 PINNED_MEMORY = {}
 TOTAL_PINNED_MEMORY = 0
-if PerformanceFeature.PinnedMem in args.fast:
+MAX_PINNED_MEMORY = -1
-    if WINDOWS:
+if not args.disable_pinned_memory:
-        MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.45  # Windows limit is apparently 50%
+    if is_nvidia() or is_amd():
-    else:
+        if WINDOWS:
-        MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.95
+            MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.45  # Windows limit is apparently 50%
-else:
+        else:
-    MAX_PINNED_MEMORY = -1
+            MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.95
+        logging.info("Enabled pinned memory {}".format(MAX_PINNED_MEMORY // (1024 * 1024)))
+
 
 def pin_memory(tensor):
     global TOTAL_PINNED_MEMORY
     if MAX_PINNED_MEMORY <= 0:
         return False
 
-    if not is_nvidia():
+    if not is_device_cpu(tensor.device):
         return False
 
-    if not is_device_cpu(tensor.device):
+    if tensor.is_pinned():
+        #NOTE: Cuda does detect when a tensor is already pinned and would
+        #error below, but there are proven cases where this also queues an error
+        #on the GPU async. So dont trust the CUDA API and guard here
         return False
 
     size = tensor.numel() * tensor.element_size()
@@ -1121,13 +1133,21 @@ def unpin_memory(tensor):
     if MAX_PINNED_MEMORY <= 0:
         return False
 
-    if not is_nvidia():
-        return False
-
     if not is_device_cpu(tensor.device):
         return False
 
     ptr = tensor.data_ptr()
+    size = tensor.numel() * tensor.element_size()
+
+    size_stored = PINNED_MEMORY.get(ptr, None)
+    if size_stored is None:
+        logging.warning("Tried to unpin tensor not pinned by ComfyUI")
+        return False
+
+    if size != size_stored:
+        logging.warning("Size of pinned tensor changed")
+        return False
+
     if torch.cuda.cudart().cudaHostUnregister(ptr) == 0:
         TOTAL_PINNED_MEMORY -= PINNED_MEMORY.pop(ptr)
         if len(PINNED_MEMORY) == 0:

comfy/model_patcher.py

@@ -843,7 +843,7 @@ class ModelPatcher:
 
         self.object_patches_backup.clear()
 
-    def partially_unload(self, device_to, memory_to_free=0):
+    def partially_unload(self, device_to, memory_to_free=0, force_patch_weights=False):
         with self.use_ejected():
             hooks_unpatched = False
             memory_freed = 0
@@ -887,13 +887,19 @@ class ModelPatcher:
                         module_mem += move_weight_functions(m, device_to)
                         if lowvram_possible:
                             if weight_key in self.patches:
-                                _, set_func, convert_func = get_key_weight(self.model, weight_key)
+                                if force_patch_weights:
-                                m.weight_function.append(LowVramPatch(weight_key, self.patches, convert_func, set_func))
+                                    self.patch_weight_to_device(weight_key)
-                                patch_counter += 1
+                                else:
+                                    _, set_func, convert_func = get_key_weight(self.model, weight_key)
+                                    m.weight_function.append(LowVramPatch(weight_key, self.patches, convert_func, set_func))
+                                    patch_counter += 1
                             if bias_key in self.patches:
-                                _, set_func, convert_func = get_key_weight(self.model, bias_key)
+                                if force_patch_weights:
-                                m.bias_function.append(LowVramPatch(bias_key, self.patches, convert_func, set_func))
+                                    self.patch_weight_to_device(bias_key)
-                                patch_counter += 1
+                                else:
+                                    _, set_func, convert_func = get_key_weight(self.model, bias_key)
+                                    m.bias_function.append(LowVramPatch(bias_key, self.patches, convert_func, set_func))
+                                    patch_counter += 1
                             cast_weight = True
 
                         if cast_weight:
@@ -909,6 +915,7 @@ class ModelPatcher:
             self.model.model_lowvram = True
             self.model.lowvram_patch_counter += patch_counter
             self.model.model_loaded_weight_memory -= memory_freed
+            logging.info("loaded partially: {:.2f} MB loaded, lowvram patches: {}".format(self.model.model_loaded_weight_memory / (1024 * 1024), self.model.lowvram_patch_counter))
             return memory_freed
 
     def partially_load(self, device_to, extra_memory=0, force_patch_weights=False):