Release 2026-09-23 (i4 461431ba)

.file_mapping.json

@@ -1,7 +1,7 @@
 {
-  "_source_commit": "4d9b6cfd3731fdbbca883184937d40a64d2dca52-dirty",
-  "_dest_commit": "c23e51f2f157ae3e51cfcd86ebfb5464850894f2",
-  "_generated_at": "2026-09-20T05:50:07Z",
+  "_source_commit": "461431bae40d4a8ccf27d812fc9d1757fe8d3b96-dirty",
+  "_dest_commit": "0460be81f16883aa380e716dc6f58c1189481172",
+  "_generated_at": "2026-09-23T04:47:38Z",
   "files": {
     "imaginaire/__init__.py": "cosmos_framework/__init__.py",
     "imaginaire/attention/__init__.py": "cosmos_framework/model/attention/__init__.py",
@@ -169,6 +169,7 @@
     "imaginaire/utils/one_logger/one_logger_global_vars.py": "cosmos_framework/utils/one_logger/one_logger_global_vars.py",
     "imaginaire/utils/one_logger/one_logger_override_utils.py": "cosmos_framework/utils/one_logger/one_logger_override_utils.py",
     "imaginaire/utils/one_logger/one_logger_utils.py": "cosmos_framework/utils/one_logger/one_logger_utils.py",
+    "imaginaire/utils/one_logger/one_logger_utils_test.py": "cosmos_framework/utils/one_logger/one_logger_utils_test.py",
     "imaginaire/utils/optim_instantiate.py": "cosmos_framework/utils/optim_instantiate.py",
     "imaginaire/utils/profiling.py": "cosmos_framework/utils/profiling.py",
     "imaginaire/utils/progress_bar.py": "cosmos_framework/utils/progress_bar.py",
@@ -292,20 +293,6 @@
     "projects/cosmos3/cosmos3/datasets/augmentors/cropping.py": "cosmos_framework/data/generator/augmentors/cropping.py",
     "projects/cosmos3/cosmos3/datasets/augmentors/duration_fps_text_timestamps.py": "cosmos_framework/data/generator/augmentors/duration_fps_text_timestamps.py",
     "projects/cosmos3/cosmos3/datasets/augmentors/duration_fps_text_timestamps_test.py": "cosmos_framework/data/generator/augmentors/duration_fps_text_timestamps_test.py",
-    "projects/cosmos3/cosmos3/datasets/augmentors/hr_lr_degradation/__init__.py": "cosmos_framework/data/generator/augmentors/hr_lr_degradation/__init__.py",
-    "projects/cosmos3/cosmos3/datasets/augmentors/hr_lr_degradation/augmentor.py": "cosmos_framework/data/generator/augmentors/hr_lr_degradation/augmentor.py",
-    "projects/cosmos3/cosmos3/datasets/augmentors/hr_lr_degradation/augmentor_test.py": "cosmos_framework/data/generator/augmentors/hr_lr_degradation/augmentor_test.py",
-    "projects/cosmos3/cosmos3/datasets/augmentors/hr_lr_degradation/bench.py": "cosmos_framework/data/generator/augmentors/hr_lr_degradation/bench.py",
-    "projects/cosmos3/cosmos3/datasets/augmentors/hr_lr_degradation/codec.py": "cosmos_framework/data/generator/augmentors/hr_lr_degradation/codec.py",
-    "projects/cosmos3/cosmos3/datasets/augmentors/hr_lr_degradation/contact_sheet.py": "cosmos_framework/data/generator/augmentors/hr_lr_degradation/contact_sheet.py",
-    "projects/cosmos3/cosmos3/datasets/augmentors/hr_lr_degradation/contact_sheet_test.py": "cosmos_framework/data/generator/augmentors/hr_lr_degradation/contact_sheet_test.py",
-    "projects/cosmos3/cosmos3/datasets/augmentors/hr_lr_degradation/degrade.py": "cosmos_framework/data/generator/augmentors/hr_lr_degradation/degrade.py",
-    "projects/cosmos3/cosmos3/datasets/augmentors/hr_lr_degradation/degrade_test.py": "cosmos_framework/data/generator/augmentors/hr_lr_degradation/degrade_test.py",
-    "projects/cosmos3/cosmos3/datasets/augmentors/hr_lr_degradation/diffjpeg.py": "cosmos_framework/data/generator/augmentors/hr_lr_degradation/diffjpeg.py",
-    "projects/cosmos3/cosmos3/datasets/augmentors/hr_lr_degradation/kernels.py": "cosmos_framework/data/generator/augmentors/hr_lr_degradation/kernels.py",
-    "projects/cosmos3/cosmos3/datasets/augmentors/hr_lr_degradation/ops.py": "cosmos_framework/data/generator/augmentors/hr_lr_degradation/ops.py",
-    "projects/cosmos3/cosmos3/datasets/augmentors/hr_lr_degradation/packing_test.py": "cosmos_framework/data/generator/augmentors/hr_lr_degradation/packing_test.py",
-    "projects/cosmos3/cosmos3/datasets/augmentors/hr_lr_degradation/profiles.py": "cosmos_framework/data/generator/augmentors/hr_lr_degradation/profiles.py",
     "projects/cosmos3/cosmos3/datasets/augmentors/idle_frames_text_info.py": "cosmos_framework/data/generator/augmentors/idle_frames_text_info.py",
     "projects/cosmos3/cosmos3/datasets/augmentors/image_editing_transform.py": "cosmos_framework/data/generator/augmentors/image_editing_transform.py",
     "projects/cosmos3/cosmos3/datasets/augmentors/image_editing_transform_test.py": "cosmos_framework/data/generator/augmentors/image_editing_transform_test.py",
@@ -326,7 +313,9 @@
     "projects/cosmos3/cosmos3/datasets/augmentors/reasoner/format_hot_fixes.py": "cosmos_framework/data/generator/augmentors/reasoner/format_hot_fixes.py",
     "projects/cosmos3/cosmos3/datasets/augmentors/reasoner/nvlm_data_to_conversation.py": "cosmos_framework/data/generator/augmentors/reasoner/nvlm_data_to_conversation.py",
     "projects/cosmos3/cosmos3/datasets/augmentors/reasoner/prompt_format.py": "cosmos_framework/data/generator/augmentors/reasoner/prompt_format.py",
+    "projects/cosmos3/cosmos3/datasets/augmentors/reasoner/prompt_format_test.py": "cosmos_framework/data/generator/augmentors/reasoner/prompt_format_test.py",
     "projects/cosmos3/cosmos3/datasets/augmentors/reasoner/shuffle_text_media_order.py": "cosmos_framework/data/generator/augmentors/reasoner/shuffle_text_media_order.py",
+    "projects/cosmos3/cosmos3/datasets/augmentors/reasoner/source_timestamps_test.py": "cosmos_framework/data/generator/augmentors/reasoner/source_timestamps_test.py",
     "projects/cosmos3/cosmos3/datasets/augmentors/reasoner/timestamp.py": "cosmos_framework/data/generator/augmentors/reasoner/timestamp.py",
     "projects/cosmos3/cosmos3/datasets/augmentors/reasoner/timestamp_test.py": "cosmos_framework/data/generator/augmentors/reasoner/timestamp_test.py",
     "projects/cosmos3/cosmos3/datasets/augmentors/reasoner/timestamp_with_subject_tracking.py": "cosmos_framework/data/generator/augmentors/reasoner/timestamp_with_subject_tracking.py",
@@ -488,9 +477,11 @@
     "projects/cosmos3/cosmos3/models/utils/sr_latent_noise_test.py": "cosmos_framework/model/generator/utils/sr_latent_noise_test.py",
     "projects/cosmos3/cosmos3/models/vision_encoder.py": "cosmos_framework/model/generator/vision_encoder.py",
     "projects/cosmos3/cosmos3/models/vlm_model.py": "cosmos_framework/model/generator/vlm_model.py",
+    "projects/cosmos3/cosmos3/processors/VIDEO_TIMESTAMPS.md": "cosmos_framework/data/generator/processors/VIDEO_TIMESTAMPS.md",
     "projects/cosmos3/cosmos3/processors/__init__.py": "cosmos_framework/data/generator/processors/__init__.py",
     "projects/cosmos3/cosmos3/processors/audio_utils.py": "cosmos_framework/data/generator/processors/audio_utils.py",
     "projects/cosmos3/cosmos3/processors/base.py": "cosmos_framework/data/generator/processors/base.py",
+    "projects/cosmos3/cosmos3/processors/base_test.py": "cosmos_framework/data/generator/processors/base_test.py",
     "projects/cosmos3/cosmos3/processors/cosmos3_edge_processing.py": "cosmos_framework/data/generator/processors/cosmos3_edge_processing.py",
     "projects/cosmos3/cosmos3/processors/cosmos3_edge_processing_test.py": "cosmos_framework/data/generator/processors/cosmos3_edge_processing_test.py",
     "projects/cosmos3/cosmos3/processors/nemotron3densevl_processor.py": "cosmos_framework/data/generator/processors/nemotron3densevl_processor.py",
@@ -504,6 +495,7 @@
     "projects/cosmos3/cosmos3/processors/qwen3vl_nemo_chat_processor.py": "cosmos_framework/data/generator/processors/qwen3vl_nemo_chat_processor.py",
     "projects/cosmos3/cosmos3/processors/qwen3vl_nemo_chat_processor_test.py": "cosmos_framework/data/generator/processors/qwen3vl_nemo_chat_processor_test.py",
     "projects/cosmos3/cosmos3/processors/qwen3vl_processor.py": "cosmos_framework/data/generator/processors/qwen3vl_processor.py",
+    "projects/cosmos3/cosmos3/processors/source_video_timing_test.py": "cosmos_framework/data/generator/processors/source_video_timing_test.py",
     "projects/cosmos3/cosmos3/scripts/multiview_auto/multiview_collage.py": "cosmos_framework/scripts/multiview_collage.py",
     "projects/cosmos3/cosmos3/scripts/multiview_auto/multiview_collage_test.py": "cosmos_framework/scripts/multiview_collage_test.py",
     "projects/cosmos3/cosmos3/sequence_packing/__init__.py": "cosmos_framework/data/generator/sequence_packing/__init__.py",
@@ -576,8 +568,10 @@
     "projects/cosmos3/cosmos3/utils/reasoner/pretrained_models_downloader.py": "cosmos_framework/utils/generator/reasoner/pretrained_models_downloader.py",
     "projects/cosmos3/cosmos3/utils/reasoner/pretrained_models_downloader_test.py": "cosmos_framework/utils/generator/reasoner/pretrained_models_downloader_test.py",
     "projects/cosmos3/cosmos3/utils/reasoner/true_packing.py": "cosmos_framework/utils/generator/reasoner/true_packing.py",
+    "projects/cosmos3/cosmos3/utils/source_video_timing.py": "cosmos_framework/utils/generator/source_video_timing.py",
     "projects/cosmos3/cosmos3/utils/video_frame_sampling.py": "cosmos_framework/utils/generator/video_frame_sampling.py",
     "projects/cosmos3/cosmos3/utils/video_preprocess.py": "cosmos_framework/utils/generator/video_preprocess.py",
+    "projects/cosmos3/cosmos3/utils/video_source_metadata.py": "cosmos_framework/utils/generator/video_source_metadata.py",
     "projects/cosmos3/interactive/configs/defaults/flex_attention.py": "cosmos_framework/configs/base/defaults/causal_flex_attention.py",
     "projects/cosmos3/interactive/configs/defaults/replay_attention.py": "cosmos_framework/configs/base/defaults/replay_attention.py",
     "projects/cosmos3/interactive/models/attention_io_layout.py": "cosmos_framework/model/generator/attention_io_layout.py",

cosmos_framework/configs/base/defaults/multiview_attention.py

@@ -78,8 +78,11 @@ def resolve_caption_scope(access: CaptionAccess, *, per_view_captions: bool) ->
 #   ``decomposed_temporal_window_seconds`` is set: the two streams do not share a frame
 #   index, but they do share real capture time, which the window compares instead.
 #
-# Read by the ``flex_*`` backends only. The ``"maskless"`` backend is its own attention pattern
-# and does not take a scope -- see ``BackendPreference``.
+# Read by every backend, but not the same way. A ``flex_*`` backend expresses the scope as a mask.
+# The ``"maskless"`` backend expresses ``"same_view"`` and ``"decomposed"`` as partitions of the
+# GEN stream and refuses ``"all_views"``, which is not a partition at all -- so there the scope
+# decides whether the cross-instant pass exists rather than describing one attention two ways. See
+# ``BackendPreference`` and ``models.mot.multiview_maskless_attention.MASKLESS_ATTENTION_SCOPES``.
 AttentionScope = Literal["all_views", "same_view", "decomposed"]
 
 # The scopes of ``AttentionScope`` at runtime, which the annotation itself is not.

cosmos_framework/configs/base/reasoner/defaults/augmentors.py

@@ -37,6 +37,7 @@ def create_data_augmentor_config() -> dict[str, Any]:
             max_fps_thres=60,
             target_fps="${data_setting.qwen_target_fps}",  # type: ignore
             video_temporal_mode="${data_setting.qwen_video_temporal_mode}",
+            video_timestamp_mode="${data_setting.video_timestamp_mode}",
             max_video_token_length="${data_setting.qwen_max_video_token_length}",  # type: ignore
             processor=processor,
             extract_audio="${model.config.sound_und}",
@@ -45,6 +46,7 @@ def create_data_augmentor_config() -> dict[str, Any]:
         "prompt_format": L(PromptFormat)(  # takes text_keys and output "conversation"
             input_keys=["texts"],
             text_chat_order="${data_setting.text_chat_order}",
+            strip_thinking_prob="${data_setting.strip_thinking_prob}",
         ),
         "shuffle_text_media_order": L(ShuffleTextMediaOrder)(),
         "format_hot_fixes": L(FormatHotFixes)(),
@@ -130,6 +132,7 @@ def create_data_augmentor_config() -> dict[str, Any]:
             custom_system_prompt="${data_setting.custom_system_prompt}",
             strip_original_system_prompt="${data_setting.strip_original_system_prompt}",
             video_temporal_mode="${data_setting.qwen_video_temporal_mode}",
+            video_timestamp_mode="${data_setting.video_timestamp_mode}",
             text_only=False,
             sound_und="${model.config.sound_und}",
             audio_encoder_type="${model.config.sound_und_config.audio_encoder_type}",
@@ -168,6 +171,7 @@ def create_data_augmentor_config() -> dict[str, Any]:
 
 processor = L(build_processor_lazy)(
     tokenizer_type="${model.config.policy.backbone.model_name}",
+    use_native_edge_processor="${data_setting.use_native_edge_processor}",
     credentials="${checkpoint.load_from_object_store.credentials}",
     bucket="${checkpoint.load_from_object_store.bucket}",
 )

cosmos_framework/configs/base/reasoner/defaults/config.py

@@ -17,6 +17,7 @@ class DataSetting:
         qwen_max_video_token_length: Maximum video token length.
         qwen_target_fps: Target fps for video sampling.
         text_chat_order: Order of text items in user messages.
+        strip_thinking_prob: Per-sample probability of converting thinking data into non-thinking data.
         custom_system_prompt: System prompt injected when a conversation has no leading system message.
         strip_original_system_prompt: Remove existing system messages before optional custom prompt injection.
         distributor_type: "with_replace" (WeightedShardlistBasic) or "no_replace" (NoReplaceShardlistBasic).
@@ -30,6 +31,11 @@ class DataSetting:
     qwen_max_video_token_length: int = 8192
     qwen_max_image_token_length: int = 8192
     qwen_target_fps: float = 4.0
+    use_native_edge_processor: bool = False
+    video_timestamp_mode: str = attrs.field(
+        default="qwen_index",
+        validator=attrs.validators.in_({"qwen_index", "legacy_fps", "source_pts"}),
+    )
     qwen_video_temporal_mode: str = attrs.field(
         default="native", validator=attrs.validators.in_({"native", "framewise"})
     )
@@ -38,6 +44,14 @@ class DataSetting:
         default="text_end",
         validator=attrs.validators.in_({"text_end", "text_start", "random"}),
     )
+    strip_thinking_prob: float = attrs.field(
+        default=0.0,
+        validator=attrs.validators.and_(
+            attrs.validators.instance_of((int, float)),
+            attrs.validators.ge(0.0),
+            attrs.validators.le(1.0),
+        ),
+    )
     custom_system_prompt: str | None = "You are a helpful assistant."
     strip_original_system_prompt: bool = False
     temporal_localization_output_format: str = attrs.field(

cosmos_framework/data/generator/action/utils/domain_utils.py

@@ -54,6 +54,10 @@
     # RoboCasa PandaOmron mobile manipulation (10/15/20D raw action per
     # ``use_base_action`` / ``base_encoding``); appended above the maximum.
     "robocasa": 30,
+    # embodiment_b nvidia-20260828 ingestion: a new one-shot dataset, distinct from
+    # "embodiment_b" (domain 9, an earlier unrelated sample drop with its own 30D
+    # contract).
+    "embodiment_b_20260828": 32,
 }
 
 
@@ -88,6 +92,7 @@
     "so101-bimanual-midtrain-conditional": 20,
     "geniesim3_g2a": 29,
     "geniesim3_g2a_joint": 16,
+    "embodiment_b_20260828": 50,
     # NOTE: ``libero`` (7/10/13 depending on ``rotation_space``), ``hand_pose``
     # (variable with ``keypoint_option`` and ``rotation_format``) and ``robocasa``
     # (10 arm-only, 15/20 with the mobile base, per ``use_base_action`` /