Release: scrub docstring URLs; drop cluster/unittest from migration

- Loosened REPLACE-NEXT semantics in the rewriter pick up the next *matching*
  line; lets a directive placed above a docstring scrub URLs inside it.
  Applied to avae.py and utils/misc.py to scrub two internal
  gitlab-master.nvidia.com URLs that previously survived in module/class
  docstrings.
- cluster.py and unittest.py removed from cosmos_framework/configs/base/defaults/
  and excluded from the release mapping (CF-owned going forward).
- Other small updates picked up from current i4 source.

No dangling imports.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Author: yy-code-nv

cosmos_framework/auxiliary/guardrail/common/presets.py

@@ -7,9 +7,6 @@
 from cosmos_framework.auxiliary.guardrail.common.core import GuardrailRunner
 from cosmos_framework.auxiliary.guardrail.face_blur_filter.face_blur_filter import RetinaFaceFilter
 from cosmos_framework.auxiliary.guardrail.qwen3guard.qwen3guard import Qwen3Guard
-from cosmos_framework.auxiliary.guardrail.video_content_safety_filter.video_content_safety_filter import (
-    VideoContentSafetyFilter,
-)
 from cosmos_framework.utils import log
 
 
@@ -27,9 +24,9 @@ def create_video_guardrail_runner(offload_model_to_cpu: bool = False) -> Guardra
     """Create the video guardrail runner."""
     return GuardrailRunner(
         safety_models=[
-            #VideoContentSafetyFilter(offload_model_to_cpu=offload_model_to_cpu) 
+            # VideoContentSafetyFilter(offload_model_to_cpu=offload_model_to_cpu)
             # Too many false positives, add back when fixed
-            ],
+        ],
         postprocessors=[RetinaFaceFilter(offload_model_to_cpu=offload_model_to_cpu)],
     )
 

cosmos_framework/configs/base/config.py

@@ -38,7 +38,6 @@ class Config(config.Config):
             {"ema": "power"},
             {"tokenizer": "wan2pt2_tokenizer"},
             {"sound_tokenizer": None},  # Optional: for audio-video generation
-            {"cluster": "default"},
             {"vlm_config": None},
             {"ckpt_type": "dcp"},
             {"experiment": None},
@@ -72,7 +71,6 @@ def make_config() -> Config:
 
     from cosmos_framework.configs.base.defaults.callbacks import register_callbacks
     from cosmos_framework.configs.base.defaults.checkpointer import register_checkpoint, register_ckpt_type
-    from cosmos_framework.configs.base.defaults.cluster import register_cluster
     from cosmos_framework.configs.base.defaults.ema import register_ema
 
     # from cosmos_framework.configs.base.defaults.data import register_data
@@ -92,7 +90,6 @@ def make_config() -> Config:
     register_tokenizer()
     register_sound_tokenizer()
     register_ema()
-    register_cluster()
     register_vlm()
 
     # Register shipped experiments explicitly. (vision_sft_nano also defines

cosmos_framework/configs/base/defaults/cluster.py

@@ -136,7 +136,7 @@ class VLMConfig:
     # HuggingFace model identifier or local path. Drives AutoConfig + AutoModel selection.
     model_name: str = ""
 
-    # Safetensor path for model for load a safetensor from different folder 
+    # Safetensor path for model for load a safetensor from different folder
     safetensors_path: str = ""
 
     # Optional pretrained-weights overlay (separate from the AutoModel structural

cosmos_framework/configs/base/defaults/unittest.py

@@ -142,6 +142,7 @@ def build_processor(
     else:
         raise ValueError(f"Tokenizer type {tokenizer_type} not supported")
 
+
 def build_processor_lazy(
     *args,
     repository: Optional[str] = None,
@@ -178,4 +179,4 @@ def build_processor_lazy(
         if subdir:
             local_path = os.path.join(local_path, subdir)
         return sys.modules[__name__].build_processor(local_path, **kwargs)
-    return sys.modules[__name__].build_processor(*args, **kwargs)
+    return sys.modules[__name__].build_processor(*args, **kwargs)

cosmos_framework/configs/base/defaults/vlm.py

@@ -2435,12 +2435,8 @@ def _single_velocity_fn(tokens: list[list[int]], skip_text_tokens: bool):
             # always-on globally) would silently ignore the per-timestep
             # ``guidance_interval`` gate.
             if _dp_shard_group is not None:
-                _cfg_t = torch.tensor(
-                    [1 if needs_cfg else 0], device=_align_device, dtype=torch.int32
-                )
-                torch.distributed.all_reduce(
-                    _cfg_t, op=torch.distributed.ReduceOp.MAX, group=_dp_shard_group
-                )
+                _cfg_t = torch.tensor([1 if needs_cfg else 0], device=_align_device, dtype=torch.int32)
+                torch.distributed.all_reduce(_cfg_t, op=torch.distributed.ReduceOp.MAX, group=_dp_shard_group)
                 _any_needs_cfg = bool(_cfg_t.item())
             else:
                 _any_needs_cfg = needs_cfg
@@ -2480,9 +2476,7 @@ def _single_velocity_fn(tokens: list[list[int]], skip_text_tokens: bool):
         # to pad their FSDP allgather sequence.
         if _dp_shard_group is not None:
             _local_steps_t = torch.tensor([num_steps], device=_align_device, dtype=torch.int32)
-            torch.distributed.all_reduce(
-                _local_steps_t, op=torch.distributed.ReduceOp.MAX, group=_dp_shard_group
-            )
+            torch.distributed.all_reduce(_local_steps_t, op=torch.distributed.ReduceOp.MAX, group=_dp_shard_group)
             _max_num_steps = int(_local_steps_t.item())
         else:
             _max_num_steps = num_steps

cosmos_framework/data/vfm/processors/__init__.py

@@ -3,7 +3,7 @@
 
 """
 AVAE (Audio Variational AutoEncoder) Tokenizer for Imaginaire4
-ported from https://gitlab-master.nvidia.com/ADLR/bigvgan
+ported from https://invalid_url
 commit hash: 80fbd8cfecb1867cc864e6d4fe0a474d8403a474
 """
 

cosmos_framework/model/vfm/omni_mot_model.py

@@ -29,6 +29,7 @@ def save_video(grid, video_name, fps=30):
         macro_block_size=1,
     )
 
+
 def save_img_or_video(
     sample: Tensor,  # [C,T,H,W] in [0,1] range
     save_fp_wo_ext: Union[str, IO[Any]],

cosmos_framework/model/vfm/tokenizers/audio/avae.py

@@ -542,7 +542,7 @@ def __exit__(self, exc_type, exc_val, exc_tb):
 
 class StragglerDetectorV2:
     """StragglerDetectorV2 is a class that allows you to easily integrate "straggler" tool:
-    https://gitlab-master.nvidia.com/dl/gwe/fault_tolerance_related/straggler/-/tree/cupti?ref_type=heads.
+    https://invalid_url.
 
     This tool detects stragglers using low-level CUPTI tool, which can gather kernel execution time with very low overhead.
     The execution times are compared across different ranks, as well as to the execution time of the exact same kernels in the past.