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): 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): 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