From 97f198e4215680a83749ba95849f3cdcfa7aa64a Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 5 Nov 2025 15:07:35 -0800 Subject: [PATCH 1/3] Fix qwen controlnet regression. (#10657) --- comfy/ldm/qwen_image/controlnet.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/ldm/qwen_image/controlnet.py b/comfy/ldm/qwen_image/controlnet.py index 92ac3cf0a..a6d408104 100644 --- a/comfy/ldm/qwen_image/controlnet.py +++ b/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) From 1d69245981f9fb3861018613246042296d887dd3 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 5 Nov 2025 15:08:13 -0800 Subject: [PATCH 2/3] Enable pinned memory by default on Nvidia. (#10656) Removed the --fast pinned_memory flag. You can use --disable-pinned-memory to disable it. Please report if it causes any issues. --- comfy/cli_args.py | 3 ++- comfy/model_management.py | 22 +++++++++------------- 2 files changed, 11 insertions(+), 14 deletions(-) diff --git a/comfy/cli_args.py b/comfy/cli_args.py index 3947e62a8..2f30b72d2 100644 --- a/comfy/cli_args.py +++ b/comfy/cli_args.py @@ -145,10 +145,11 @@ 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.") diff --git a/comfy/model_management.py b/comfy/model_management.py index 0d040e55e..4d13c52c1 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -1085,22 +1085,21 @@ def cast_to_device(tensor, device, dtype, copy=False): PINNED_MEMORY = {} TOTAL_PINNED_MEMORY = 0 -if PerformanceFeature.PinnedMem in args.fast: - if WINDOWS: - MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.45 # Windows limit is apparently 50% - else: - MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.95 -else: - MAX_PINNED_MEMORY = -1 +MAX_PINNED_MEMORY = -1 +if not args.disable_pinned_memory: + if is_nvidia(): + if WINDOWS: + MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.45 # Windows limit is apparently 50% + else: + 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(): - return False - if not is_device_cpu(tensor.device): return False @@ -1121,9 +1120,6 @@ 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 From 09dc24c8a982776abd5cb2f71e3d041139e1d5b2 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 5 Nov 2025 16:11:15 -0800 Subject: [PATCH 3/3] Pinned mem also seems to work on AMD. (#10658) --- comfy/model_management.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index 4d13c52c1..7a30c4bec 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -1087,7 +1087,7 @@ PINNED_MEMORY = {} TOTAL_PINNED_MEMORY = 0 MAX_PINNED_MEMORY = -1 if not args.disable_pinned_memory: - if is_nvidia(): + if is_nvidia() or is_amd(): if WINDOWS: MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.45 # Windows limit is apparently 50% else: