diff --git a/vllm_omni/diffusion/models/z_image/pipeline_z_image.py b/vllm_omni/diffusion/models/z_image/pipeline_z_image.py index 62d92900a..2c938f39b 100644 --- a/vllm_omni/diffusion/models/z_image/pipeline_z_image.py +++ b/vllm_omni/diffusion/models/z_image/pipeline_z_image.py @@ -527,6 +527,16 @@ class ZImagePipeline(nn.Module): num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) self._num_timesteps = len(timesteps) + # Precompute normalized timesteps once to avoid per-step GPU->CPU sync (.item() causes cudaStreamSynchronize) + if isinstance(timesteps, torch.Tensor): + timesteps_tensor = timesteps.to(device=device, dtype=torch.float32) + else: + timesteps_tensor = torch.as_tensor(timesteps, device=device, dtype=torch.float32) + norm_timesteps = (1000 - timesteps_tensor) / 1000 + t_norm_list = norm_timesteps.cpu().tolist() + if not isinstance(t_norm_list, list): + t_norm_list = [t_norm_list] + # 6. Denoising loop for i, t in enumerate(timesteps): if self.interrupt: @@ -535,8 +545,9 @@ class ZImagePipeline(nn.Module): # broadcast to batch dimension in a way that's compatible with ONNX/Core ML timestep = t.expand(latents.shape[0]) timestep = (1000 - timestep) / 1000 - # Normalized time for time-aware config (0 at start, 1 at end) - t_norm = timestep[0].item() + # Normalized time for time-aware config (0 at start, 1 at end); + # use precomputed to avoid .item() sync per step + t_norm = t_norm_list[i] # Handle cfg truncation current_guidance_scale = self.guidance_scale @@ -582,13 +593,17 @@ class ZImagePipeline(nn.Module): pred = pos + current_guidance_scale * (pos - neg) - # Renormalization + # Renormalization (torch.where avoids GPU->CPU sync from Python if/scalar comparison) if self._cfg_normalization and float(self._cfg_normalization) > 0.0: ori_pos_norm = torch.linalg.vector_norm(pos) new_pos_norm = torch.linalg.vector_norm(pred) max_new_norm = ori_pos_norm * float(self._cfg_normalization) - if new_pos_norm > max_new_norm: - pred = pred * (max_new_norm / new_pos_norm) + scale = torch.where( + new_pos_norm > max_new_norm, + (max_new_norm / new_pos_norm.clamp(min=1e-12)).to(pred.dtype), + pred.new_tensor(1.0), + ) + pred = pred * scale noise_pred.append(pred) diff --git a/vllm_omni/diffusion/models/z_image/z_image_transformer.py b/vllm_omni/diffusion/models/z_image/z_image_transformer.py index 2b7d4eb5b..8efb89931 100644 --- a/vllm_omni/diffusion/models/z_image/z_image_transformer.py +++ b/vllm_omni/diffusion/models/z_image/z_image_transformer.py @@ -852,7 +852,13 @@ class ZImageTransformer2DModel(CachedTransformer): # Match t_embedder output dtype to x for layerwise casting compatibility adaln_input = t.type_as(x) - x[torch.cat(x_inner_pad_mask)] = self.x_pad_token + # Use torch.where instead of x[mask]= to avoid aten::index_put_/nonzero and cudaStreamSynchronize + x_pad_mask = torch.cat(x_inner_pad_mask) + x = torch.where( + x_pad_mask.unsqueeze(1).expand_as(x), + self.x_pad_token.expand(x.shape[0], -1), + x, + ) x = list(x.split(x_item_seqlens, dim=0)) x_cos, x_sin = self.rope_embedder(torch.cat(x_pos_ids, dim=0)) x_cos = list(x_cos.split(x_item_seqlens, dim=0)) @@ -875,7 +881,13 @@ class ZImageTransformer2DModel(CachedTransformer): cap_feats = torch.cat(cap_feats, dim=0) cap_feats = self.cap_embedder(cap_feats) - cap_feats[torch.cat(cap_inner_pad_mask)] = self.cap_pad_token + # Use torch.where instead of cap_feats[mask]= to avoid aten::index_put_/nonzero and cudaStreamSynchronize + cap_pad_mask = torch.cat(cap_inner_pad_mask) + cap_feats = torch.where( + cap_pad_mask.unsqueeze(1).expand_as(cap_feats), + self.cap_pad_token.expand(cap_feats.shape[0], -1), + cap_feats, + ) cap_feats = list(cap_feats.split(cap_item_seqlens, dim=0)) cap_cos, cap_sin = self.rope_embedder(torch.cat(cap_pos_ids, dim=0)) cap_cos = list(cap_cos.split(cap_item_seqlens, dim=0))