DDIM loop 内小张量分配优化，attention mask 缓存到 GPU，加速30s左右

scripts/evaluation/world_model_interaction.py

@@ -13,7 +13,7 @@ import time
 import json
 from contextlib import contextmanager, nullcontext
 from dataclasses import dataclass, field, asdict
-from typing import Optional, Dict, List, Any
+from typing import Optional, Dict, List, Any, Mapping
 
 from pytorch_lightning import seed_everything
 from omegaconf import OmegaConf
@@ -718,6 +718,17 @@ def preprocess_observation(
     return return_observations
 
 
+def _move_to_device(batch: Mapping[str, Any],
+                    device: torch.device) -> dict[str, Any]:
+    moved = {}
+    for key, value in batch.items():
+        if isinstance(value, torch.Tensor) and value.device != device:
+            moved[key] = value.to(device, non_blocking=True)
+        else:
+            moved[key] = value
+    return moved
+
+
 def image_guided_synthesis_sim_mode(
         model: torch.nn.Module,
         prompts: list[str],
@@ -768,7 +779,10 @@ def image_guided_synthesis_sim_mode(
     profiler = get_profiler()
 
     b, _, t, _, _ = noise_shape
+    ddim_sampler = getattr(model, "_ddim_sampler", None)
+    if ddim_sampler is None:
         ddim_sampler = DDIMSampler(model)
+        model._ddim_sampler = ddim_sampler
     batch_size = noise_shape[0]
 
     fs = torch.tensor([fs] * batch_size, dtype=torch.long, device=model.device)
@@ -1077,10 +1091,7 @@ def run_inference(args: argparse.Namespace, gpu_num: int, gpu_no: int) -> None:
                     'action':
                     torch.zeros_like(batch['action'][-1]).unsqueeze(0)
                 }
-                observation = {
+                observation = _move_to_device(observation, device)
-                    key: observation[key].to(device, non_blocking=True)
-                    for key in observation
-                }
             # Update observation queues
             cond_obs_queues = populate_queues(cond_obs_queues, observation)
 
@@ -1094,6 +1105,8 @@ def run_inference(args: argparse.Namespace, gpu_num: int, gpu_no: int) -> None:
             # Multi-round interaction with the world-model
             with pytorch_prof_ctx:
                 for itr in tqdm(range(args.n_iter)):
+                    log_every = max(1, args.step_log_every)
+                    log_step = (itr % log_every == 0)
                     profiler.current_iteration = itr
                     profiler.record_memory(f"iter_{itr}_start")
 
@@ -1111,12 +1124,10 @@ def run_inference(args: argparse.Namespace, gpu_num: int, gpu_no: int) -> None:
                                 'action':
                                 torch.stack(list(cond_obs_queues['action']), dim=1),
                             }
-                            observation = {
+                            observation = _move_to_device(observation, device)
-                                key: observation[key].to(device, non_blocking=True)
-                                for key in observation
-                            }
 
                         # Use world-model in policy to generate action
+                        if log_step:
                             print(f'>>> Step {itr}: generating actions ...')
                         with profiler.profile_section("action_generation"):
                             pred_videos_0, pred_actions, _ = image_guided_synthesis_sim_mode(
@@ -1156,12 +1167,10 @@ def run_inference(args: argparse.Namespace, gpu_num: int, gpu_no: int) -> None:
                                 'action':
                                 torch.stack(list(cond_obs_queues['action']), dim=1),
                             }
-                            observation = {
+                            observation = _move_to_device(observation, device)
-                                key: observation[key].to(device, non_blocking=True)
-                                for key in observation
-                            }
 
                         # Interaction with the world-model
+                        if log_step:
                             print(f'>>> Step {itr}: interacting with world model ...')
                         with profiler.profile_section("world_model_interaction"):
                             pred_videos_1, _, pred_states = image_guided_synthesis_sim_mode(
@@ -1364,6 +1373,12 @@ def get_parser():
         default="fp32",
         help="Dtype for VAE/first_stage_model weights and forward autocast."
     )
+    parser.add_argument(
+        "--step_log_every",
+        type=int,
+        default=1,
+        help="Print per-iteration step logs every N iterations."
+    )
     parser.add_argument(
         "--n_action_steps",
         type=int,

src/unifolm_wma/models/samplers/ddim.py

@@ -28,6 +28,11 @@ class DDIMSampler(object):
                       ddim_discretize="uniform",
                       ddim_eta=0.,
                       verbose=True):
+        device = self.model.betas.device
+        cache_key = (ddim_num_steps, ddim_discretize, float(ddim_eta),
+                     str(device))
+        if getattr(self, "_schedule_cache", None) == cache_key:
+            return
         self.ddim_timesteps = make_ddim_timesteps(
             ddim_discr_method=ddim_discretize,
             num_ddim_timesteps=ddim_num_steps,
@@ -67,16 +72,26 @@ class DDIMSampler(object):
             ddim_timesteps=self.ddim_timesteps,
             eta=ddim_eta,
             verbose=verbose)
+        ddim_sigmas = torch.as_tensor(ddim_sigmas,
+                                      device=self.model.device,
+                                      dtype=torch.float32)
+        ddim_alphas = torch.as_tensor(ddim_alphas,
+                                      device=self.model.device,
+                                      dtype=torch.float32)
+        ddim_alphas_prev = torch.as_tensor(ddim_alphas_prev,
+                                           device=self.model.device,
+                                           dtype=torch.float32)
         self.register_buffer('ddim_sigmas', ddim_sigmas)
         self.register_buffer('ddim_alphas', ddim_alphas)
         self.register_buffer('ddim_alphas_prev', ddim_alphas_prev)
         self.register_buffer('ddim_sqrt_one_minus_alphas',
-                             np.sqrt(1. - ddim_alphas))
+                             torch.sqrt(1. - ddim_alphas))
         sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt(
             (1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) *
             (1 - self.alphas_cumprod / self.alphas_cumprod_prev))
         self.register_buffer('ddim_sigmas_for_original_num_steps',
                              sigmas_for_original_sampling_steps)
+        self._schedule_cache = cache_key
 
     @torch.no_grad()
     def sample(
@@ -228,10 +243,14 @@ class DDIMSampler(object):
             'x_inter_state': [state],
             'pred_x0_state': [state],
         }
-        time_range = reversed(range(
+        if ddim_use_original_steps:
-            0, timesteps)) if ddim_use_original_steps else np.flip(timesteps)
+            time_range = np.arange(timesteps - 1, -1, -1)
-        total_steps = timesteps if ddim_use_original_steps else timesteps.shape[
+        else:
-            0]
+            time_range = np.flip(timesteps)
+        time_range = np.ascontiguousarray(time_range)
+        total_steps = int(time_range.shape[0])
+        t_seq = torch.as_tensor(time_range, device=device, dtype=torch.long)
+        ts_batch = t_seq.unsqueeze(1).expand(total_steps, b).contiguous()
         if verbose:
             iterator = tqdm(time_range, desc='DDIM Sampler', total=total_steps)
         else:
@@ -243,7 +262,7 @@ class DDIMSampler(object):
         dp_ddim_scheduler_state.set_timesteps(len(timesteps))
         for i, step in enumerate(iterator):
             index = total_steps - i - 1
-            ts = torch.full((b, ), step, device=device, dtype=torch.long)
+            ts = ts_batch[i]
 
             # Use mask to blend noised original latent (img_orig) & new sampled latent (img)
             if mask is not None:
@@ -378,16 +397,14 @@ class DDIMSampler(object):
         sigmas = self.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas
 
         if is_video:
-            size = (b, 1, 1, 1, 1)
+            size = (1, 1, 1, 1, 1)
         else:
-            size = (b, 1, 1, 1)
+            size = (1, 1, 1, 1)
 
-        a_t = torch.full(size, alphas[index], device=device)
+        a_t = alphas[index].view(size)
-        a_prev = torch.full(size, alphas_prev[index], device=device)
+        a_prev = alphas_prev[index].view(size)
-        sigma_t = torch.full(size, sigmas[index], device=device)
+        sigma_t = sigmas[index].view(size)
-        sqrt_one_minus_at = torch.full(size,
+        sqrt_one_minus_at = sqrt_one_minus_alphas[index].view(size)
-                                       sqrt_one_minus_alphas[index],
-                                       device=device)
 
         if self.model.parameterization != "v":
             pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
@@ -395,12 +412,8 @@ class DDIMSampler(object):
             pred_x0 = self.model.predict_start_from_z_and_v(x, t, model_output)
 
         if self.model.use_dynamic_rescale:
-            scale_t = torch.full(size,
+            scale_t = self.ddim_scale_arr[index].view(size)
-                                 self.ddim_scale_arr[index],
+            prev_scale_t = self.ddim_scale_arr_prev[index].view(size)
-                                 device=device)
-            prev_scale_t = torch.full(size,
-                                      self.ddim_scale_arr_prev[index],
-                                      device=device)
             rescale = (prev_scale_t / scale_t)
             pred_x0 *= rescale
 

src/unifolm_wma/modules/attention.py

@@ -99,6 +99,7 @@ class CrossAttention(nn.Module):
         self.agent_state_context_len = agent_state_context_len
         self.agent_action_context_len = agent_action_context_len
         self.cross_attention_scale_learnable = cross_attention_scale_learnable
+        self._attn_mask_cache = {}
         if self.image_cross_attention:
             self.to_k_ip = nn.Linear(context_dim, inner_dim, bias=False)
             self.to_v_ip = nn.Linear(context_dim, inner_dim, bias=False)
@@ -275,7 +276,8 @@ class CrossAttention(nn.Module):
                 attn_mask_aa = self._get_attn_mask_aa(x.shape[0],
                                                       q.shape[1],
                                                       k_aa.shape[1],
-                                                      block_size=16).to(k_aa.device)
+                                                      block_size=16,
+                                                      device=k_aa.device)
         else:
             if not spatial_self_attn:
                 assert 1 > 2, ">>> ERROR: you should never go into here ..."
@@ -386,14 +388,26 @@ class CrossAttention(nn.Module):
 
         return self.to_out(out)
 
-    def _get_attn_mask_aa(self, b, l1, l2, block_size=16):
+    def _get_attn_mask_aa(self,
+                          b,
+                          l1,
+                          l2,
+                          block_size=16,
+                          device=None):
+        if device is None:
+            device = self.to_q.weight.device
+        cache_key = (b, l1, l2, block_size, str(device))
+        if cache_key in self._attn_mask_cache:
+            return self._attn_mask_cache[cache_key]
         num_token = l2 // block_size
-        start_positions = ((torch.arange(b) % block_size) + 1) * num_token
+        start_positions = ((torch.arange(b, device=device) % block_size) +
-        col_indices = torch.arange(l2)
+                           1) * num_token
+        col_indices = torch.arange(l2, device=device)
         mask_2d = col_indices.unsqueeze(0) >= start_positions.unsqueeze(1)
         mask = mask_2d.unsqueeze(1).expand(b, l1, l2)
-        attn_mask = torch.zeros_like(mask, dtype=torch.float)
+        attn_mask = torch.zeros_like(mask, dtype=torch.float32)
         attn_mask[mask] = float('-inf')
+        self._attn_mask_cache[cache_key] = attn_mask
         return attn_mask
 
 

useful.sh

@@ -107,3 +107,15 @@ embedder：
   1. 新增 --encoder_mode {fp32, autocast, bf16_full}
   2. bf16_full = 权重 BF16 + 前向 BF16
   3. autocast = 权重 FP32 + 仅主干 autocast（现在的实现）
+
+
+
+    1. DDIM loop 内小张量分配优化（已完成）
+
+  - 每步 torch.full(...) 改成预先构造/广播，减少 loop 内分配
+  - 位置：src/unifolm_wma/models/samplers/ddim.py
+
+  2. attention mask 缓存到 GPU（已完成）
+
+  - _get_attn_mask_aa 现在直接在目标 device 构造并缓存，避免每步 CPU→GPU 拷贝
+  - 位置：src/unifolm_wma/modules/attention.py