Release 2026-09-20 (i4 4d9b6cfd)

.file_mapping.json

@@ -1,7 +1,7 @@
 {
-  "_source_commit": "2962eb13e76ad641803d5f498c2822555c88e79c-dirty",
-  "_dest_commit": "6c9d3916e0c5b62f8bed9694ae805d3d4c04c454",
-  "_generated_at": "2026-09-18T13:40:26Z",
+  "_source_commit": "4d9b6cfd3731fdbbca883184937d40a64d2dca52-dirty",
+  "_dest_commit": "c23e51f2f157ae3e51cfcd86ebfb5464850894f2",
+  "_generated_at": "2026-09-20T05:50:07Z",
   "files": {
     "imaginaire/__init__.py": "cosmos_framework/__init__.py",
     "imaginaire/attention/__init__.py": "cosmos_framework/model/attention/__init__.py",
@@ -292,6 +292,20 @@
     "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",
@@ -339,6 +353,7 @@
     "projects/cosmos3/cosmos3/datasets/augmentors/transfer_control_input/seg.py": "cosmos_framework/data/generator/augmentors/transfer_control_input/seg.py",
     "projects/cosmos3/cosmos3/datasets/augmentors/transfer_control_transform.py": "cosmos_framework/data/generator/augmentors/transfer_control_transform.py",
     "projects/cosmos3/cosmos3/datasets/augmentors/transfer_control_transform_test.py": "cosmos_framework/data/generator/augmentors/transfer_control_transform_test.py",
+    "projects/cosmos3/cosmos3/datasets/augmentors/transfer_target_length_test.py": "cosmos_framework/data/generator/augmentors/transfer_target_length_test.py",
     "projects/cosmos3/cosmos3/datasets/augmentors/video_editing_transform.py": "cosmos_framework/data/generator/augmentors/video_editing_transform.py",
     "projects/cosmos3/cosmos3/datasets/augmentors/video_editing_transform_test.py": "cosmos_framework/data/generator/augmentors/video_editing_transform_test.py",
     "projects/cosmos3/cosmos3/datasets/augmentors/video_parsing.py": "cosmos_framework/data/generator/augmentors/video_parsing.py",
@@ -371,6 +386,8 @@
     "projects/cosmos3/cosmos3/models/mot/__init__.py": "cosmos_framework/model/generator/mot/__init__.py",
     "projects/cosmos3/cosmos3/models/mot/action_io_projector.py": "cosmos_framework/model/generator/mot/action_io_projector.py",
     "projects/cosmos3/cosmos3/models/mot/action_io_projector_test.py": "cosmos_framework/model/generator/mot/action_io_projector_test.py",
+    "projects/cosmos3/cosmos3/models/mot/activation_marks.py": "cosmos_framework/model/generator/mot/activation_marks.py",
+    "projects/cosmos3/cosmos3/models/mot/activation_marks_test.py": "cosmos_framework/model/generator/mot/activation_marks_test.py",
     "projects/cosmos3/cosmos3/models/mot/attention.py": "cosmos_framework/model/generator/mot/attention.py",
     "projects/cosmos3/cosmos3/models/mot/attention_test.py": "cosmos_framework/model/generator/mot/attention_test.py",
     "projects/cosmos3/cosmos3/models/mot/cfgp_ar_test.py": "cosmos_framework/model/generator/mot/cfgp_ar_test.py",
@@ -460,12 +477,15 @@
     "projects/cosmos3/cosmos3/models/reasoner/qwen3_vl_moe/shared_expert_test.py": "cosmos_framework/model/generator/reasoner/qwen3_vl_moe/shared_expert_test.py",
     "projects/cosmos3/cosmos3/models/utils/__init__.py": "cosmos_framework/model/generator/utils/__init__.py",
     "projects/cosmos3/cosmos3/models/utils/data_and_condition.py": "cosmos_framework/model/generator/utils/data_and_condition.py",
+    "projects/cosmos3/cosmos3/models/utils/data_and_condition_test.py": "cosmos_framework/model/generator/utils/data_and_condition_test.py",
     "projects/cosmos3/cosmos3/models/utils/load_balancing_stats.py": "cosmos_framework/model/generator/utils/load_balancing_stats.py",
     "projects/cosmos3/cosmos3/models/utils/memory.py": "cosmos_framework/model/generator/utils/memory.py",
     "projects/cosmos3/cosmos3/models/utils/moe_utils.py": "cosmos_framework/model/generator/utils/moe_utils.py",
     "projects/cosmos3/cosmos3/models/utils/safetensors_loader.py": "cosmos_framework/model/generator/utils/safetensors_loader.py",
     "projects/cosmos3/cosmos3/models/utils/safetensors_loader_indexed_test.py": "cosmos_framework/model/generator/utils/safetensors_loader_indexed_test.py",
     "projects/cosmos3/cosmos3/models/utils/safetensors_loader_test.py": "cosmos_framework/model/generator/utils/safetensors_loader_test.py",
+    "projects/cosmos3/cosmos3/models/utils/sr_latent_noise.py": "cosmos_framework/model/generator/utils/sr_latent_noise.py",
+    "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/__init__.py": "cosmos_framework/data/generator/processors/__init__.py",
@@ -492,6 +512,7 @@
     "projects/cosmos3/cosmos3/sequence_packing/natten.py": "cosmos_framework/data/generator/sequence_packing/natten.py",
     "projects/cosmos3/cosmos3/sequence_packing/packers.py": "cosmos_framework/data/generator/sequence_packing/packers.py",
     "projects/cosmos3/cosmos3/sequence_packing/runtime.py": "cosmos_framework/data/generator/sequence_packing/runtime.py",
+    "projects/cosmos3/cosmos3/sequence_packing/runtime_test.py": "cosmos_framework/data/generator/sequence_packing/runtime_test.py",
     "projects/cosmos3/cosmos3/sequence_packing/sequence.py": "cosmos_framework/data/generator/sequence_packing/sequence.py",
     "projects/cosmos3/cosmos3/sequence_packing/temporal_causal.py": "cosmos_framework/data/generator/sequence_packing/temporal_causal.py",
     "projects/cosmos3/cosmos3/tokenizers/audio/__init__.py": "cosmos_framework/model/generator/tokenizers/audio/__init__.py",
@@ -584,6 +605,26 @@
     "projects/cosmos3/interactive/models/utils/nvfp4_test.py": "cosmos_framework/model/generator/utils/nvfp4_test.py",
     "projects/cosmos3/interactive/sequence_packing.py": "cosmos_framework/data/generator/sequence_packing/autoregressive.py",
     "projects/cosmos3/interactive/utils/data_batch.py": "cosmos_framework/utils/generator/data_batch.py",
+    "projects/cosmos3/tokenizer/lidar_tokenizer/__init__.py": "cosmos_framework/model/generator/tokenizers/lidar/__init__.py",
+    "projects/cosmos3/tokenizer/lidar_tokenizer/checkpoint.py": "cosmos_framework/model/generator/tokenizers/lidar/checkpoint.py",
+    "projects/cosmos3/tokenizer/lidar_tokenizer/checkpoint_test.py": "cosmos_framework/model/generator/tokenizers/lidar/checkpoint_test.py",
+    "projects/cosmos3/tokenizer/lidar_tokenizer/dtypes.py": "cosmos_framework/model/generator/tokenizers/lidar/dtypes.py",
+    "projects/cosmos3/tokenizer/lidar_tokenizer/dtypes_test.py": "cosmos_framework/model/generator/tokenizers/lidar/dtypes_test.py",
+    "projects/cosmos3/tokenizer/lidar_tokenizer/lidar_tokenizer_v1.py": "cosmos_framework/model/generator/tokenizers/lidar/lidar_tokenizer_v1.py",
+    "projects/cosmos3/tokenizer/lidar_tokenizer/network/__init__.py": "cosmos_framework/model/generator/tokenizers/lidar/network/__init__.py",
+    "projects/cosmos3/tokenizer/lidar_tokenizer/network/distributions.py": "cosmos_framework/model/generator/tokenizers/lidar/network/distributions.py",
+    "projects/cosmos3/tokenizer/lidar_tokenizer/network/encoding.py": "cosmos_framework/model/generator/tokenizers/lidar/network/encoding.py",
+    "projects/cosmos3/tokenizer/lidar_tokenizer/network/ops.py": "cosmos_framework/model/generator/tokenizers/lidar/network/ops.py",
+    "projects/cosmos3/tokenizer/lidar_tokenizer/network/rope3d.py": "cosmos_framework/model/generator/tokenizers/lidar/network/rope3d.py",
+    "projects/cosmos3/tokenizer/lidar_tokenizer/network/transformer_vae.py": "cosmos_framework/model/generator/tokenizers/lidar/network/transformer_vae.py",
+    "projects/cosmos3/tokenizer/lidar_tokenizer/normalization.py": "cosmos_framework/model/generator/tokenizers/lidar/normalization.py",
+    "projects/cosmos3/tokenizer/lidar_tokenizer/normalization_test.py": "cosmos_framework/model/generator/tokenizers/lidar/normalization_test.py",
+    "projects/cosmos3/tokenizer/lidar_tokenizer/paths.py": "cosmos_framework/model/generator/tokenizers/lidar/paths.py",
+    "projects/cosmos3/tokenizer/lidar_tokenizer/postprocessing.py": "cosmos_framework/model/generator/tokenizers/lidar/postprocessing.py",
+    "projects/cosmos3/tokenizer/lidar_tokenizer/postprocessing_test.py": "cosmos_framework/model/generator/tokenizers/lidar/postprocessing_test.py",
+    "projects/cosmos3/tokenizer/lidar_tokenizer/preprocessing.py": "cosmos_framework/model/generator/tokenizers/lidar/preprocessing.py",
+    "projects/cosmos3/tokenizer/lidar_tokenizer/range_projection.py": "cosmos_framework/model/generator/tokenizers/lidar/range_projection.py",
+    "projects/cosmos3/tokenizer/lidar_tokenizer/range_projection_test.py": "cosmos_framework/model/generator/tokenizers/lidar/range_projection_test.py",
     "projects/cosmos3/tokenizer/uniae/checkpoint_identity.py": "cosmos_framework/model/tokenizer/checkpoint_identity.py",
     "projects/cosmos3/tokenizer/uniae/checkpoint_io.py": "cosmos_framework/model/tokenizer/checkpoint_io.py",
     "projects/cosmos3/tokenizer/uniae/evaluation/lpips_cache.py": "cosmos_framework/model/tokenizer/evaluation/lpips_cache.py",

cosmos_framework/configs/base/defaults/activation_checkpointing.py

@@ -17,6 +17,20 @@
 
 import attrs
 
+# The forward attention op of every backend the frontend can dispatch, by the name
+# the checkpoint policy matches against.
+#
+# * natten: fmha_forward
+# * flash2: _flash_attn_varlen_forward (varlen), _flash_attn_forward (dense)
+# * flash3: _flash_attn_forward (dense, varlen)
+#
+# cuDNN is deliberately absent since it does not support varlen attention.
+ATTENTION_FORWARD_OPS_REGEX = [
+    "fmha_forward",
+    "_flash_attn_varlen_forward",
+    "_flash_attn_forward",
+]
+
 
 @attrs.define(slots=False)
 class ActivationCheckpointingConfig:
@@ -45,10 +59,34 @@ class ActivationCheckpointingConfig:
 
     # Regex patterns for ops to save when using selective AC. Ignored if
     # mode is "full" or "none".
+    #
+    # Defaults to attention on whichever backend the frontend picks, which is what
+    # every config asking for selective AC wants and is not what a name like "fmha"
+    # delivers: that covers NATTEN alone, and NATTEN is only the *selected* backend
+    # where the others refuse the call. On sm100 cuDNN and flash2 both reject varlen
+    # so NATTEN wins and "fmha" matches; on sm90 flash3 is ranked first and takes it,
+    # leaving nothing in the region named "fmha", so selective AC there kept nothing
+    # and silently recomputed every attention it was configured to save.
+    #
+    # Copied because attrs hands this list to the instance, and a config mutating it
+    # would edit the module constant for every other config in the process.
     save_ops_regex: list[str] = attrs.field(
-        factory=lambda: ["fmha"],
+        factory=lambda: list(ATTENTION_FORWARD_OPS_REGEX),
     )
 
+    # Narrow ``save_ops_regex`` to the call sites that asked to be kept.
+    #
+    # The regex matches dispatched op names, which cannot separate calls running
+    # the same kernel. The decomposed multiview attention runs four FMHA calls per
+    # layer worth very different amounts to keep -- the same-view fold is ~96% of
+    # forward attention time and ~94% of backward, the rest are cheap to recompute.
+    #
+    # With this on, a matching op is kept only where the model marked it with
+    # ``activation_marks.mark_next_activation``. The regex still says which ops are
+    # eligible; the mark says which of them are worth it. Off by default, which is
+    # the behaviour every existing config already has.
+    save_only_marked_ops: bool = False
+
     # Stash and restore RNG state across recompute boundaries. Required for
     # deterministic output vs. non-checkpointed passes; slower otherwise.
     preserve_rng_state: bool = True

cosmos_framework/configs/base/defaults/compile.py

@@ -39,6 +39,18 @@ class CompileConfig:
     # Whether to use CUDA graphs for faster inference. This option does not work during training.
     use_cuda_graphs: bool = False
 
+    # Granularity of CUDA-graph capture for AR inference (only with ``enabled`` and
+    # ``use_cuda_graphs``).  "block": every decoder block is compiled with
+    # ``torch.compile(mode="reduce-overhead")`` and replays its own CUDA-graph tree, so a
+    # forward still pays one graph launch per block plus the Python between blocks.
+    # "forward": blocks are compiled without CUDA-graph trees and the AR loop captures one
+    # explicit graph per whole forward (denoise / KV-refresh × CFG branch) on the
+    # static-shape KV path from cache index 1 onward.
+    cuda_graph_scope: Literal["block", "forward"] = attrs.field(
+        default="block",
+        validator=attrs.validators.in_({"block", "forward"}),
+    )
+
     # AR-inference-specific behavior once the rolling KV window saturates.
     # "default" uses the global compile settings for the entire generation.
     # "static-compile" keeps the normal pre-saturation path, then uses dedicated

cosmos_framework/configs/base/defaults/model_config.py

@@ -16,6 +16,7 @@
 from cosmos_framework.configs.base.defaults.reasoner import VLMConfig
 from cosmos_framework.model.generator.mot.action_io_projector import ACTION_IO_PROJECTOR_TYPES
 from cosmos_framework.model.generator.utils.load_balancing_stats import LBLConfig
+from cosmos_framework.model.generator.utils.sr_latent_noise import SRLatentConditionNoiseConfig
 
 # Mirrors ``cosmos3.common.args.AttentionIOLayout``. Defined locally on purpose: importing
 # the ``cosmos3`` workspace package at module scope makes the whole cosmos3 config tree
@@ -351,6 +352,10 @@ class OmniMoTModelConfig:
     sound_dim: int | None = None  # Sound latent channel size (e.g., 64 for AVAE 48kHz)
     sound_latent_fps: int = 25  # Sound tokenizer's latent rate (e.g., 48kHz / 1920 hop = 25 Hz)
 
+    # Super-resolution: Gaussian noise on the LR conditioning latent of SR samples during training (L1).
+    # None disables it. See cosmos_framework/model/generator/utils/sr_latent_noise.py.
+    sr_latent_condition_noise: SRLatentConditionNoiseConfig | None = None
+
     # When False, removes bias from vae2llm, sound2llm, and the two Linear layers inside
     # time_embedder.  These biases seem to inject token-constant DC offsets that dominate
     # the MoE router input and create prompt invariant routing.  This is observed empirically