Release: vfm → generator, vlm → reasoner dest rename (moves + import …

.agents/skills/cosmos3-codebase-nav/SKILL.md

@@ -22,7 +22,7 @@ All paths below are relative to this file's location (`.agents/skills/cosmos3-co
 
 - `cosmos_framework/` — main training package (data, model, trainer, callbacks, checkpoint, utils, …).
 - `cosmos_framework/configs/base/experiment/` — vfm (generator) experiment SKUs referenced by `[train.train_policy].experiment` in the recipe TOMLs.
-- `cosmos_framework/configs/base/vlm/experiment/` — vlm (reasoner) experiment SKUs.
+- `cosmos_framework/configs/base/reasoner/experiment/` — vlm (reasoner) experiment SKUs.
 - `cosmos_framework/inference/` — inference subpackage (args, model, inference engine, defaults, Ray serving, common helpers).
 - `cosmos_framework/scripts/` — top-level entry-point scripts (train, inference, eval, export_model, convert_model_to_dcp, upsample_prompts, caption_from_video, captions_to_sft_jsonl, action_policy_server, …). Invoked as `python -m cosmos_framework.scripts.<name>`.
 - `examples/toml/sft_config/<recipe>.toml` + `examples/launch_sft_<recipe>.sh` — paired SFT recipes (training entry-point input). The shell sources `examples/_sft_launcher_common.sh`, which forwards into `cosmos_framework.scripts.train --sft-toml=...`.
@@ -45,7 +45,7 @@ All paths below are relative to this file's location (`.agents/skills/cosmos3-co
 | SFT recipe TOMLs (paired with `examples/launch_sft_*.sh`) | `../../../examples/toml/sft_config/<recipe>.toml`                                                                                 |
 | SFT pydantic schema (validates the recipe TOML)           | `../../../cosmos_framework/configs/toml_config/sft_config.py`                                                                     |
 | Training experiment SKUs (vfm)                            | `../../../cosmos_framework/configs/base/experiment/`                                                                              |
-| Training experiment SKUs (vlm / reasoner)                 | `../../../cosmos_framework/configs/base/vlm/experiment/`                                                                          |
+| Training experiment SKUs (vlm / reasoner)                 | `../../../cosmos_framework/configs/base/reasoner/experiment/`                                                                     |
 | Example inputs                                            | `../../../inputs/omni/t2i.json`, `../../../inputs/omni/t2v.json`, `../../../inputs/omni/i2v.json`, …                              |
 
 Available modality modes for defaults: `text2image`, `text2video`, `image2video`, `image2image`, `video2video`, `forward_dynamics`, `inverse_dynamics`, `policy`.

.claude/skills/cosmos3-codebase-nav/SKILL.md

@@ -22,7 +22,7 @@ All paths below are relative to this file's location (`.claude/skills/cosmos3-co
 
 - `cosmos_framework/` — main training package (data, model, trainer, callbacks, checkpoint, utils, …).
 - `cosmos_framework/configs/base/experiment/` — vfm (generator) experiment SKUs referenced by `[train.train_policy].experiment` in the recipe TOMLs.
-- `cosmos_framework/configs/base/vlm/experiment/` — vlm (reasoner) experiment SKUs.
+- `cosmos_framework/configs/base/reasoner/experiment/` — vlm (reasoner) experiment SKUs.
 - `cosmos_framework/inference/` — inference subpackage (args, model, inference engine, defaults, Ray serving, common helpers).
 - `cosmos_framework/scripts/` — top-level entry-point scripts (train, inference, eval, export_model, convert_model_to_dcp, upsample_prompts, caption_from_video, captions_to_sft_jsonl, action_policy_server, …). Invoked as `python -m cosmos_framework.scripts.<name>`.
 - `examples/toml/sft_config/<recipe>.toml` + `examples/launch_sft_<recipe>.sh` — paired SFT recipes (training entry-point input). The shell sources `examples/_sft_launcher_common.sh`, which forwards into `cosmos_framework.scripts.train --sft-toml=...`.
@@ -45,7 +45,7 @@ All paths below are relative to this file's location (`.claude/skills/cosmos3-co
 | SFT recipe TOMLs (paired with `examples/launch_sft_*.sh`) | `../../../examples/toml/sft_config/<recipe>.toml`                                                                                 |
 | SFT pydantic schema (validates the recipe TOML)           | `../../../cosmos_framework/configs/toml_config/sft_config.py`                                                                     |
 | Training experiment SKUs (vfm)                            | `../../../cosmos_framework/configs/base/experiment/`                                                                              |
-| Training experiment SKUs (vlm / reasoner)                 | `../../../cosmos_framework/configs/base/vlm/experiment/`                                                                          |
+| Training experiment SKUs (vlm / reasoner)                 | `../../../cosmos_framework/configs/base/reasoner/experiment/`                                                                     |
 | Example inputs                                            | `../../../inputs/omni/t2i.json`, `../../../inputs/omni/t2v.json`, `../../../inputs/omni/i2v.json`, …                              |
 
 Available modality modes for defaults: `text2image`, `text2video`, `image2video`, `image2image`, `video2video`, `forward_dynamics`, `inverse_dynamics`, `policy`.

.github/workflows/gpu-tests.yml

@@ -256,13 +256,13 @@ jobs:
         run: |
           export LD_LIBRARY_PATH=
           uv run --all-extras --group=cu128-train torchrun --nproc_per_node=2 -m pytest -v \
-            cosmos_framework/model/vfm/mot/cfgp_ar_test.py -o addopts=
+            cosmos_framework/model/generator/mot/cfgp_ar_test.py -o addopts=
 
       - name: Distributed unit tests - context_parallel (torchrun, 4 ranks)
         run: |
           export LD_LIBRARY_PATH=
           uv run --all-extras --group=cu128-train torchrun --nproc_per_node=4 -m pytest -v \
-            cosmos_framework/model/vfm/mot/context_parallel_test.py -o addopts=
+            cosmos_framework/model/generator/mot/context_parallel_test.py -o addopts=
 
       # Clear everything the suite writes into the working tree (all gitignored
       # scratch): pytest tmp dirs (DCP checkpoint, logs), the script-test

NOTICE

@@ -15,18 +15,18 @@ Notices:
   Copyright 2025 The Qwen Team and The HuggingFace Inc. team.
   Copyright 2025 HuggingFace Inc. team.
 Files:
-  cosmos_framework/model/vfm/vlm/qwen3_vl/configuration_qwen3_vl.py
-  cosmos_framework/model/vfm/vlm/qwen3_vl/qwen3_vl.py
-  cosmos_framework/model/vfm/vlm/qwen3_vl/utils.py
-  cosmos_framework/model/vfm/vlm/qwen3_vl/video_processing_qwen3_vl.py
+  cosmos_framework/model/generator/vlm/qwen3_vl/configuration_qwen3_vl.py
+  cosmos_framework/model/generator/vlm/qwen3_vl/qwen3_vl.py
+  cosmos_framework/model/generator/vlm/qwen3_vl/utils.py
+  cosmos_framework/model/generator/vlm/qwen3_vl/video_processing_qwen3_vl.py
 
 ByteDance-Seed Bagel / Qwen2 tokenizer
 Source: https://github.com/ByteDance-Seed/Bagel
 License: Apache-2.0
 Notice:
   Copyright 2024 The Qwen Team and The HuggingFace Inc. team.
 File:
-  cosmos_framework/model/vfm/tokenizers/tokenization_qwen2.py
+  cosmos_framework/model/generator/tokenizers/tokenization_qwen2.py
 
 HuggingFace Diffusers UniPC scheduler
 Source: https://github.com/huggingface/diffusers
@@ -35,7 +35,7 @@ Notices:
   Copyright 2024 TSAIL Team and The HuggingFace Team. All rights reserved.
   Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved.
 File:
-  cosmos_framework/model/vfm/diffusion/samplers/fm_solvers_unipc.py
+  cosmos_framework/model/generator/diffusion/samplers/fm_solvers_unipc.py
 
 Detectron2 lazy config helpers
 Source: https://github.com/facebookresearch/detectron2
@@ -51,4 +51,4 @@ License: GPL-3.0
 Notice:
   Upstream source files do not contain an explicit copyright notice.
 Files:
-  cosmos_framework/model/vfm/utils/taylorseer.py
+  cosmos_framework/model/generator/utils/taylorseer.py

cosmos_framework/callbacks/compile_tokenizer.py

@@ -4,7 +4,7 @@
 """Training callback that defers AOT compilation of the VAE tokenizer.
 
 The actual compilation logic lives in
-:meth:`~cosmos_framework.model.vfm.tokenizers.wan2pt2_vae_4x16x16.Wan2pt2VAEInterface.compile_encode`.
+:meth:`~cosmos_framework.model.generator.tokenizers.wan2pt2_vae_4x16x16.Wan2pt2VAEInterface.compile_encode`.
 This module provides a :class:`CompileTokenizer` callback that invokes it
 at the right point during training (after ``compile_after_iterations``
 steps, to avoid NCCL timeouts during CUDA/cuDNN warm-up).
@@ -27,7 +27,7 @@
 
 from cosmos_framework.utils import log
 from cosmos_framework.utils.callback import Callback
-from cosmos_framework.model.vfm.omni_mot_model import OmniMoTModel
+from cosmos_framework.model.generator.omni_mot_model import OmniMoTModel
 
 
 class CompileTokenizer(Callback):

cosmos_framework/callbacks/every_n_draw_sample.py

@@ -21,7 +21,7 @@
 from cosmos_framework.utils import distributed, log, misc
 from cosmos_framework.utils.easy_io import easy_io
 from cosmos_framework.tools.visualize.video import save_img_or_video
-from cosmos_framework.utils.vfm.data_utils import slice_data_batch
+from cosmos_framework.utils.generator.data_utils import slice_data_batch
 
 
 def resize_image(image: torch.Tensor, size: int = 1024) -> torch.Tensor:

cosmos_framework/callbacks/expert_heatmap.py

@@ -10,7 +10,7 @@
 from cosmos_framework.model._base import ImaginaireModel
 from cosmos_framework.trainer import ImaginaireTrainer
 from cosmos_framework.utils import distributed
-from cosmos_framework.model.vfm.reasoner.qwen3_vl_moe.qwen3_vl_moe import Qwen3VLMoeTextSparseMoeBlock
+from cosmos_framework.model.generator.reasoner.qwen3_vl_moe.qwen3_vl_moe import Qwen3VLMoeTextSparseMoeBlock
 
 
 def compute_expert_heatmap(vfm: torch.nn.Module) -> dict[str, torch.Tensor]:

cosmos_framework/callbacks/hf_export.py

@@ -133,7 +133,7 @@ def on_save_checkpoint(self, model: Any, state_dict: dict[str, Any]) -> None:
             return
 
         # Deferred import to avoid circular dependency at module load time.
-        from cosmos_framework.model.vfm.vlm_model import VLMModel
+        from cosmos_framework.model.generator.vlm_model import VLMModel
 
         if not isinstance(model, VLMModel):
             # The legacy vlm/train.py path passes model_parts: list[nn.Module] (raw HF

cosmos_framework/callbacks/moe_specialization_callback.py

@@ -39,7 +39,7 @@
 from cosmos_framework.model._base import ImaginaireModel
 from cosmos_framework.trainer import ImaginaireTrainer
 from cosmos_framework.utils import distributed
-from cosmos_framework.model.vfm.reasoner.qwen3_vl_moe.qwen3_vl_moe import Qwen3VLMoeTextSparseMoeBlock
+from cosmos_framework.model.generator.reasoner.qwen3_vl_moe.qwen3_vl_moe import Qwen3VLMoeTextSparseMoeBlock
 
 
 def _get_device_mesh(vfm: torch.nn.Module):

cosmos_framework/callbacks/moe_stability_callback.py

@@ -107,7 +107,7 @@
 from cosmos_framework.model._base import ImaginaireModel
 from cosmos_framework.trainer import ImaginaireTrainer
 from cosmos_framework.utils import distributed
-from cosmos_framework.model.vfm.reasoner.qwen3_vl_moe.qwen3_vl_moe import Qwen3VLMoeTextSparseMoeBlock
+from cosmos_framework.model.generator.reasoner.qwen3_vl_moe.qwen3_vl_moe import Qwen3VLMoeTextSparseMoeBlock
 
 
 def _effective_experts(

cosmos_framework/callbacks/norm_monitor.py

@@ -14,7 +14,7 @@
 from cosmos_framework.utils import distributed, log, misc
 from cosmos_framework.utils.callback import Callback
 from cosmos_framework.utils.easy_io import easy_io
-from cosmos_framework.data.vfm.sequence_packing.runtime import get_gen_seq
+from cosmos_framework.data.generator.sequence_packing.runtime import get_gen_seq
 
 try:
     from apex.contrib.layer_norm import FastLayerNorm

cosmos_framework/callbacks/sequence_packing_padding.py

@@ -7,7 +7,7 @@
 from cosmos_framework.callbacks.every_n import EveryN
 from cosmos_framework.model._base import ImaginaireModel
 from cosmos_framework.trainer import ImaginaireTrainer
-from cosmos_framework.data.vfm.sequence_packing.runtime import get_padding_stats
+from cosmos_framework.data.generator.sequence_packing.runtime import get_padding_stats
 
 
 class SequencePackingPadding(EveryN):

cosmos_framework/callbacks/tokens_per_sec.py

@@ -11,7 +11,7 @@
 ``MFUCallback`` is written for the OmniMoT/VFM network -- it reads
 ``model.net.language_model.config`` and per-modality token counts
 (``output_batch["und_token_length"]`` etc.). ``VLMModel``
-(``cosmos_framework/model/vfm/vlm_model.py``) is a plain HF wrapper: it exposes
+(``cosmos_framework/model/generator/vlm_model.py``) is a plain HF wrapper: it exposes
 ``self.model`` (no ``.net``) and its ``training_step`` returns only
 ``{"loss", "loss_avg", "labels"}``. ``MFUCallback`` would therefore silently
 no-op (token length is ``None``) or fail on ``model.net``. This callback instead

cosmos_framework/callbacks/training_stats.py

@@ -11,7 +11,7 @@
 from cosmos_framework.utils import distributed
 from cosmos_framework.utils.callback import Callback
 from cosmos_framework.callbacks.wandb_log import _LossRecord
-from cosmos_framework.data.vfm.action.domain_utils import EMBODIMENT_TO_DOMAIN_ID
+from cosmos_framework.data.generator.action.domain_utils import EMBODIMENT_TO_DOMAIN_ID
 
 # Build inverse mapping: domain_id -> embodiment_type. First occurrence wins when multiple embodiment names share the
 # same domain id.

cosmos_framework/checkpoint/dcp.py

@@ -75,7 +75,7 @@
 from cosmos_framework.model._base import ImaginaireModel
 from cosmos_framework.utils import callback, distributed, log, misc
 from cosmos_framework.utils.easy_io import easy_io
-from cosmos_framework.utils.vfm.rand_state import get_rand_state_dict, set_rand_state_dict
+from cosmos_framework.utils.generator.rand_state import get_rand_state_dict, set_rand_state_dict
 
 
 class ModelWrapper(Stateful):

cosmos_framework/configs/base/defaults/activation_checkpointing.py

@@ -35,9 +35,9 @@ class ActivationCheckpointingConfig:
     Read sites:
 
     - MoT path consumes every field — see
-      cosmos_framework/model/vfm/mot/parallelize_unified_mot.py.
+      cosmos_framework/model/generator/mot/parallelize_unified_mot.py.
     - VLM path consumes only ``mode`` (and only ``"full"`` enables
-      checkpointing) — see cosmos_framework/model/vfm/vlm_model.py.
+      checkpointing) — see cosmos_framework/model/generator/vlm_model.py.
     """
 
     # AC mode:

cosmos_framework/configs/base/defaults/model.py

@@ -6,7 +6,7 @@
 from cosmos_framework.utils.lazy_config import LazyCall as L
 from cosmos_framework.configs.base.defaults.model_config import OmniMoTModelConfig
 from cosmos_framework.configs.base.defaults.parallelism import ParallelismConfig
-from cosmos_framework.model.vfm.omni_mot_model import OmniMoTModel
+from cosmos_framework.model.generator.omni_mot_model import OmniMoTModel
 
 MOT_DDP_CONFIG = dict(
     trainer=dict(

cosmos_framework/configs/base/defaults/open_source_dataloader.py

@@ -21,17 +21,17 @@
 
 Original YAML reference target paths use the ``cosmos3._src.vfm.*`` namespace
 (the OSS-release form of ``projects.cosmos3.vfm.*``); inside this released
-tree the same modules live under ``cosmos_framework.data.vfm.*``.
+tree the same modules live under ``cosmos_framework.data.generator.*``.
 """
 
 from hydra.core.config_store import ConfigStore
 
 from cosmos_framework.configs.base.defaults.reasoner import create_qwen2_tokenizer_with_download
-from cosmos_framework.data.vfm.joint_dataloader import (
+from cosmos_framework.data.generator.joint_dataloader import (
     PackingDataLoader,
     RankPartitionedDataLoader,
 )
-from cosmos_framework.data.vfm.local_datasets.sft_dataset import get_sft_dataset
+from cosmos_framework.data.generator.local_datasets.sft_dataset import get_sft_dataset
 from cosmos_framework.utils.lazy_config import LazyCall as L
 
 # ---------------------------------------------------------------------------

cosmos_framework/configs/base/defaults/optimizer.py

@@ -8,7 +8,7 @@
 from cosmos_framework.utils.lazy_config import PLACEHOLDER
 from cosmos_framework.utils.lazy_config import LazyCall as L
 from cosmos_framework.utils.config_helper import ConfigStore
-from cosmos_framework.utils.vfm.optimizer import build_lr_scheduler, build_optimizer
+from cosmos_framework.utils.generator.optimizer import build_lr_scheduler, build_optimizer
 
 OPTIMIZER_KWARGS: dict[str, Any] = dict(
     # Learning rate for the optimizer.