From ea6cdd2631fbca6ed81b95796150c32c9a029f0d Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 5 Sep 2025 22:05:05 -0700 Subject: [PATCH 1/7] Print all fast options in --help (#9737) --- comfy/cli_args.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/cli_args.py b/comfy/cli_args.py index 72eeaea9a..cc1f12482 100644 --- a/comfy/cli_args.py +++ b/comfy/cli_args.py @@ -145,7 +145,7 @@ class PerformanceFeature(enum.Enum): CublasOps = "cublas_ops" AutoTune = "autotune" -parser.add_argument("--fast", nargs="*", type=PerformanceFeature, help="Enable some untested and potentially quality deteriorating optimizations. --fast with no arguments enables everything. You can pass a list specific optimizations if you only want to enable specific ones. Current valid optimizations: fp16_accumulation fp8_matrix_mult cublas_ops") +parser.add_argument("--fast", nargs="*", type=PerformanceFeature, help="Enable some untested and potentially quality deteriorating optimizations. --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("--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.") From 27a0fcccc376fef6f035ed97664db8aa7e2e6117 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sat, 6 Sep 2025 20:25:22 -0700 Subject: [PATCH 2/7] Enable bf16 VAE on RDNA4. (#9746) --- comfy/model_management.py | 19 ++++++++++++++++++- 1 file changed, 18 insertions(+), 1 deletion(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index d08aee1fe..17516b6ed 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -289,6 +289,21 @@ def is_amd(): return True return False +def amd_min_version(device=None, min_rdna_version=0): + if not is_amd(): + return False + + arch = torch.cuda.get_device_properties(device).gcnArchName + if arch.startswith('gfx') and len(arch) == 7: + try: + cmp_rdna_version = int(arch[4]) + 2 + except: + cmp_rdna_version = 0 + if cmp_rdna_version >= min_rdna_version: + return True + + return False + MIN_WEIGHT_MEMORY_RATIO = 0.4 if is_nvidia(): MIN_WEIGHT_MEMORY_RATIO = 0.0 @@ -905,7 +920,9 @@ def vae_dtype(device=None, allowed_dtypes=[]): # NOTE: bfloat16 seems to work on AMD for the VAE but is extremely slow in some cases compared to fp32 # slowness still a problem on pytorch nightly 2.9.0.dev20250720+rocm6.4 tested on RDNA3 - if d == torch.bfloat16 and (not is_amd()) and should_use_bf16(device): + # also a problem on RDNA4 except fp32 is also slow there. + # This is due to large bf16 convolutions being extremely slow. + if d == torch.bfloat16 and ((not is_amd()) or amd_min_version(device, min_rdna_version=4)) and should_use_bf16(device): return d return torch.float32 From bcbd7884e3af5ee8b6ab848da2a3123f247d6114 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sat, 6 Sep 2025 21:29:38 -0700 Subject: [PATCH 3/7] Don't enable pytorch attention on AMD if triton isn't available. (#9747) --- comfy/model_management.py | 14 ++++++++------ 1 file changed, 8 insertions(+), 6 deletions(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index 17516b6ed..cb6580f73 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -22,6 +22,7 @@ from enum import Enum from comfy.cli_args import args, PerformanceFeature import torch import sys +import importlib import platform import weakref import gc @@ -336,12 +337,13 @@ try: logging.info("AMD arch: {}".format(arch)) logging.info("ROCm version: {}".format(rocm_version)) if args.use_split_cross_attention == False and args.use_quad_cross_attention == False: - if torch_version_numeric >= (2, 7): # works on 2.6 but doesn't actually seem to improve much - if any((a in arch) for a in ["gfx90a", "gfx942", "gfx1100", "gfx1101", "gfx1151"]): # TODO: more arches, TODO: gfx950 - ENABLE_PYTORCH_ATTENTION = True -# if torch_version_numeric >= (2, 8): -# if any((a in arch) for a in ["gfx1201"]): -# ENABLE_PYTORCH_ATTENTION = True + if importlib.util.find_spec('triton') is not None: # AMD efficient attention implementation depends on triton. TODO: better way of detecting if it's compiled in or not. + if torch_version_numeric >= (2, 7): # works on 2.6 but doesn't actually seem to improve much + if any((a in arch) for a in ["gfx90a", "gfx942", "gfx1100", "gfx1101", "gfx1151"]): # TODO: more arches, TODO: gfx950 + ENABLE_PYTORCH_ATTENTION = True +# if torch_version_numeric >= (2, 8): +# if any((a in arch) for a in ["gfx1201"]): +# ENABLE_PYTORCH_ATTENTION = True if torch_version_numeric >= (2, 7) and rocm_version >= (6, 4): if any((a in arch) for a in ["gfx1201", "gfx942", "gfx950"]): # TODO: more arches SUPPORT_FP8_OPS = True From fb763d43332aaf15e96350cf1c25e2a1927423f1 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sun, 7 Sep 2025 18:16:29 -0700 Subject: [PATCH 4/7] Fix amd_min_version crash when cpu device. (#9754) --- comfy/model_management.py | 3 +++ 1 file changed, 3 insertions(+) diff --git a/comfy/model_management.py b/comfy/model_management.py index cb6580f73..bbfc3c7a1 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -294,6 +294,9 @@ def amd_min_version(device=None, min_rdna_version=0): if not is_amd(): return False + if is_device_cpu(device): + return False + arch = torch.cuda.get_device_properties(device).gcnArchName if arch.startswith('gfx') and len(arch) == 7: try: From bd1d9bcd5fcdb8379ce5a8020cb2b8f42de1b7c7 Mon Sep 17 00:00:00 2001 From: Jedrzej Kosinski Date: Mon, 8 Sep 2025 12:07:04 -0700 Subject: [PATCH 5/7] Add ZeroDivisionError catch for EasyCache logging statement (#9768) --- comfy_extras/nodes_easycache.py | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-) diff --git a/comfy_extras/nodes_easycache.py b/comfy_extras/nodes_easycache.py index 9d2988f5f..c170e9fd9 100644 --- a/comfy_extras/nodes_easycache.py +++ b/comfy_extras/nodes_easycache.py @@ -162,7 +162,12 @@ def easycache_sample_wrapper(executor, *args, **kwargs): logging.info(f"{easycache.name} [verbose] - output_change_rates {len(output_change_rates)}: {output_change_rates}") logging.info(f"{easycache.name} [verbose] - approx_output_change_rates {len(approx_output_change_rates)}: {approx_output_change_rates}") total_steps = len(args[3])-1 - logging.info(f"{easycache.name} - skipped {easycache.total_steps_skipped}/{total_steps} steps ({total_steps/(total_steps-easycache.total_steps_skipped):.2f}x speedup).") + # catch division by zero for log statement; sucks to crash after all sampling is done + try: + speedup = total_steps/(total_steps-easycache.total_steps_skipped) + except ZeroDivisionError: + speedup = 1.0 + logging.info(f"{easycache.name} - skipped {easycache.total_steps_skipped}/{total_steps} steps ({speedup:.2f}x speedup).") easycache.reset() guider.model_options = orig_model_options From 97652d26b81f83fc9a3675be55ede7762fafb7bd Mon Sep 17 00:00:00 2001 From: contentis Date: Mon, 8 Sep 2025 21:08:18 +0200 Subject: [PATCH 6/7] Add explicit casting in apply_rope for Qwen VL (#9759) --- comfy/text_encoders/llama.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/comfy/text_encoders/llama.py b/comfy/text_encoders/llama.py index 4c976058f..5e11956b5 100644 --- a/comfy/text_encoders/llama.py +++ b/comfy/text_encoders/llama.py @@ -128,11 +128,12 @@ def precompute_freqs_cis(head_dim, position_ids, theta, rope_dims=None, device=N def apply_rope(xq, xk, freqs_cis): + org_dtype = xq.dtype cos = freqs_cis[0] sin = freqs_cis[1] q_embed = (xq * cos) + (rotate_half(xq) * sin) k_embed = (xk * cos) + (rotate_half(xk) * sin) - return q_embed, k_embed + return q_embed.to(org_dtype), k_embed.to(org_dtype) class Attention(nn.Module): From 103a12cb668303f197b22f52bb2981bb1539beea Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Mon, 8 Sep 2025 14:30:26 -0700 Subject: [PATCH 7/7] Support qwen inpaint controlnet. (#9772) --- comfy/controlnet.py | 16 +++++++++++++--- 1 file changed, 13 insertions(+), 3 deletions(-) diff --git a/comfy/controlnet.py b/comfy/controlnet.py index e3dfedf55..f08ff4b36 100644 --- a/comfy/controlnet.py +++ b/comfy/controlnet.py @@ -253,7 +253,10 @@ class ControlNet(ControlBase): to_concat = [] for c in self.extra_concat_orig: c = c.to(self.cond_hint.device) - c = comfy.utils.common_upscale(c, self.cond_hint.shape[3], self.cond_hint.shape[2], self.upscale_algorithm, "center") + c = comfy.utils.common_upscale(c, self.cond_hint.shape[-1], self.cond_hint.shape[-2], self.upscale_algorithm, "center") + if c.ndim < self.cond_hint.ndim: + c = c.unsqueeze(2) + c = comfy.utils.repeat_to_batch_size(c, self.cond_hint.shape[2], dim=2) to_concat.append(comfy.utils.repeat_to_batch_size(c, self.cond_hint.shape[0])) self.cond_hint = torch.cat([self.cond_hint] + to_concat, dim=1) @@ -585,11 +588,18 @@ def load_controlnet_flux_instantx(sd, model_options={}): def load_controlnet_qwen_instantx(sd, model_options={}): model_config, operations, load_device, unet_dtype, manual_cast_dtype, offload_device = controlnet_config(sd, model_options=model_options) - control_model = comfy.ldm.qwen_image.controlnet.QwenImageControlNetModel(operations=operations, device=offload_device, dtype=unet_dtype, **model_config.unet_config) + control_latent_channels = sd.get("controlnet_x_embedder.weight").shape[1] + + extra_condition_channels = 0 + concat_mask = False + if control_latent_channels == 68: #inpaint controlnet + extra_condition_channels = control_latent_channels - 64 + concat_mask = True + control_model = comfy.ldm.qwen_image.controlnet.QwenImageControlNetModel(extra_condition_channels=extra_condition_channels, operations=operations, device=offload_device, dtype=unet_dtype, **model_config.unet_config) control_model = controlnet_load_state_dict(control_model, sd) latent_format = comfy.latent_formats.Wan21() extra_conds = [] - control = ControlNet(control_model, compression_ratio=1, latent_format=latent_format, load_device=load_device, manual_cast_dtype=manual_cast_dtype, extra_conds=extra_conds) + control = ControlNet(control_model, compression_ratio=1, latent_format=latent_format, concat_mask=concat_mask, load_device=load_device, manual_cast_dtype=manual_cast_dtype, extra_conds=extra_conds) return control def convert_mistoline(sd):