Skip to content

support cosmos3 nano policy and isaac env #1243

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
helloyongyang wants to merge 8 commits into
base: main
Choose a base branch
from
Open

support cosmos3 nano policy and isaac env #1243

Show file tree
Hide file tree
Changes from 6 commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
257fc5d
update cosmos3 nano
helloyongyang Jul 12, 2026
a8c77f4
support policy
helloyongyang Jul 12, 2026
b54c245
Merge remote-tracking branch 'origin/main' into cos
helloyongyang Jul 13, 2026
21a7e35
fix libero
helloyongyang Jul 13, 2026
1ccee3c
fix curobo
helloyongyang Jul 14, 2026
ebff026
update ros
helloyongyang Jul 14, 2026
aec13d0
update
helloyongyang Jul 14, 2026
c5544ba
Merge remote-tracking branch 'origin/main' into cos
helloyongyang Jul 15, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
28 changes: 28 additions & 0 deletions configs/cosmos3/cosmos3_nano_omni_action_fd_agibotworld.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,28 @@
{
"infer_steps": 30,
"sample_guide_scale": 1.0,
"sample_shift": 10.0,
"target_height": 720,
"target_width": 640,
"target_video_length": 17,
"target_fps": 10.0,
"enable_cfg": true,
"action_mode": "forward_dynamics",
"domain_name": "agibotworld",
"view_point": "concat_view",
"action_chunk_size": 16,
"action_chunk_index": 0,
"feature_caching": "NoCaching",
"rms_norm_type": "one-pass",
"attn_rms_norm_type": "one-pass",
"rope_type": "triton",
"self_attn_type": "flash_attn3",
"causal_self_attn_type": "flash_attn3",
"add_resolution_template": false,
"add_duration_template": false,
"use_system_prompt": false,
"cosmos3_meta_init": true,
"vae_cpu_offload": false,
"cpu_offload": false,
"offload_granularity": "block"
}
30 changes: 30 additions & 0 deletions configs/cosmos3/cosmos3_nano_omni_action_fd_agibotworld_multichunk.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,30 @@
{
"infer_steps": 30,
"sample_guide_scale": 1.0,
"sample_shift": 10.0,
"target_height": 720,
"target_width": 640,
"target_video_length": 17,
"target_fps": 10.0,
"enable_cfg": true,
"action_mode": "forward_dynamics",
"domain_name": "agibotworld",
"view_point": "concat_view",
"action_chunk_size": 16,
"action_chunk_index": 0,
"action_multichunk": true,
"action_num_chunks": 4,
"feature_caching": "NoCaching",
"rms_norm_type": "one-pass",
"attn_rms_norm_type": "one-pass",
"rope_type": "triton",
"self_attn_type": "flash_attn3",
"causal_self_attn_type": "flash_attn3",
"add_resolution_template": false,
"add_duration_template": false,
"use_system_prompt": false,
"cosmos3_meta_init": true,
"vae_cpu_offload": false,
"cpu_offload": false,
"offload_granularity": "block"
}
27 changes: 27 additions & 0 deletions configs/cosmos3/cosmos3_nano_omni_action_id_av.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,27 @@
{
"infer_steps": 30,
"sample_guide_scale": 1.0,
"sample_shift": 10.0,
"target_height": 480,
"target_width": 832,
"target_video_length": 61,
"target_fps": 10.0,
"enable_cfg": true,
"action_mode": "inverse_dynamics",
"domain_name": "av",
"view_point": "ego_view",
"action_chunk_size": 60,
"feature_caching": "NoCaching",
"rms_norm_type": "one-pass",
"attn_rms_norm_type": "one-pass",
"rope_type": "triton",
"self_attn_type": "flash_attn3",
"causal_self_attn_type": "flash_attn3",
"add_resolution_template": false,
"add_duration_template": false,
"use_system_prompt": false,
"cosmos3_meta_init": true,
"vae_cpu_offload": false,
"cpu_offload": false,
"offload_granularity": "block"
}
23 changes: 23 additions & 0 deletions configs/cosmos3/cosmos3_nano_omni_i2av.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,23 @@
{
"infer_steps": 35,
"sample_guide_scale": 6.0,
"sample_shift": 10.0,
"target_height": 720,
"target_width": 1280,
"target_video_length": 189,
"target_fps": 24.0,
"enable_cfg": true,
"enable_sound": true,
"feature_caching": "NoCaching",
"rms_norm_type": "one-pass",
"attn_rms_norm_type": "one-pass",
"rope_type": "triton",
"self_attn_type": "flash_attn3",
"causal_self_attn_type": "flash_attn3",
"add_resolution_template": false,
"add_duration_template": false,
"cosmos3_meta_init": true,
"vae_cpu_offload": false,
"cpu_offload": false,
"offload_granularity": "block"
}
22 changes: 22 additions & 0 deletions configs/cosmos3/cosmos3_nano_omni_i2v.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,22 @@
{
"infer_steps": 35,
"sample_guide_scale": 6.0,
"sample_shift": 10.0,
"target_height": 720,
"target_width": 1280,
"target_video_length": 189,
"target_fps": 24.0,
"enable_cfg": true,
"feature_caching": "NoCaching",
"rms_norm_type": "one-pass",
"attn_rms_norm_type": "one-pass",
"rope_type": "triton",
"self_attn_type": "flash_attn3",
"causal_self_attn_type": "flash_attn3",
"add_resolution_template": false,
"add_duration_template": false,
"cosmos3_meta_init": true,
"vae_cpu_offload": false,
"cpu_offload": false,
"offload_granularity": "block"
}
23 changes: 23 additions & 0 deletions configs/cosmos3/cosmos3_nano_omni_t2av.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,23 @@
{
"infer_steps": 35,
"sample_guide_scale": 6.0,
"sample_shift": 10.0,
"target_height": 720,
"target_width": 1280,
"target_video_length": 189,
"target_fps": 24.0,
"enable_cfg": true,
"enable_sound": true,
"feature_caching": "NoCaching",
"rms_norm_type": "one-pass",
"attn_rms_norm_type": "one-pass",
"rope_type": "triton",
"self_attn_type": "flash_attn3",
"causal_self_attn_type": "flash_attn3",
"add_resolution_template": false,
"add_duration_template": false,
"cosmos3_meta_init": true,
"vae_cpu_offload": false,
"cpu_offload": false,
"offload_granularity": "block"
}
22 changes: 22 additions & 0 deletions configs/cosmos3/cosmos3_nano_omni_t2v.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,22 @@
{
"infer_steps": 35,
"sample_guide_scale": 6.0,
"sample_shift": 10.0,
"target_height": 720,
"target_width": 1280,
"target_video_length": 189,
"target_fps": 24.0,
"enable_cfg": true,
"feature_caching": "NoCaching",
"rms_norm_type": "one-pass",
"attn_rms_norm_type": "one-pass",
"rope_type": "triton",
"self_attn_type": "flash_attn3",
"causal_self_attn_type": "flash_attn3",
"add_resolution_template": false,
"add_duration_template": false,
"cosmos3_meta_init": true,
"vae_cpu_offload": false,
"cpu_offload": false,
"offload_granularity": "block"
}
28 changes: 28 additions & 0 deletions configs/cosmos3/cosmos3_nano_omni_t2v_cfg_ulysses_8gpu.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,28 @@
{
"infer_steps": 35,
"sample_guide_scale": 6.0,
"sample_shift": 10.0,
"target_height": 720,
"target_width": 1280,
"target_video_length": 189,
"target_fps": 24.0,
"enable_cfg": true,
"feature_caching": "NoCaching",
"rms_norm_type": "one-pass",
"attn_rms_norm_type": "one-pass",
"rope_type": "triton",
"self_attn_type": "flash_attn2",
"causal_self_attn_type": "flash_attn2",
"add_resolution_template": false,
"add_duration_template": false,
"cosmos3_meta_init": true,
"unload_modules": false,
"vae_cpu_offload": true,
"cpu_offload": true,
"offload_granularity": "block",
"parallel": {
"seq_p_size": 4,
"seq_p_attn_type": "ulysses",
"cfg_p_size": 2
}
}
38 changes: 38 additions & 0 deletions configs/cosmos3/cosmos3_nano_policy_droid_cfg_ulysses_8gpu.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,38 @@
{
"infer_steps": 4,
"sample_guide_scale": 3.0,
"sample_shift": 5.0,
"target_height": 544,
"target_width": 736,
"target_video_length": 33,
"target_fps": 15.0,
"enable_cfg": true,
"action_mode": "policy",
"domain_name": "droid_lerobot",
"view_point": "concat_view",
"action_chunk_size": 32,
"raw_action_dim": 8,
"policy_use_state": true,
"policy_history_length": 1,
"policy_flip_gripper": true,
"decode_video": false,
"feature_caching": "NoCaching",
"rms_norm_type": "one-pass",
"attn_rms_norm_type": "one-pass",
"rope_type": "triton",
"self_attn_type": "flash_attn2",
"causal_self_attn_type": "flash_attn2",
"add_resolution_template": false,
"add_duration_template": false,
"use_system_prompt": false,
"cosmos3_meta_init": true,
"unload_modules": false,
"vae_cpu_offload": true,
"cpu_offload": true,
"offload_granularity": "block",
"parallel": {
"seq_p_size": 4,
"seq_p_attn_type": "ulysses",
"cfg_p_size": 2
}
}
4 changes: 2 additions & 2 deletions configs/fastwam/libero_i2va.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
{
"adapter_model_path": "/data/nvme1/yongyang/nb/FastWAM/checkpoints/fastwam_release/libero_uncond_2cam224.pt",
"dataset_stats_path": "/data/nvme1/yongyang/nb/FastWAM/checkpoints/fastwam_release/libero_uncond_2cam224_dataset_stats.json",
"adapter_model_path": "/data/nvme7/yongyang/fastwam_release/libero_uncond_2cam224.pt",
"dataset_stats_path": "/data/nvme7/yongyang/fastwam_release/libero_uncond_2cam224_dataset_stats.json",
Comment on lines +2 to +3

@gemini-code-assist gemini-code-assist Bot Jul 14, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

medium

Avoid hardcoding absolute paths pointing to a specific user's directory (e.g., /data/nvme7/yongyang/...) in shared configuration files. This makes the configuration non-portable and will cause failures on other machines or in production environments. Consider using relative paths or environment variables to resolve these paths dynamically.

"camera_size": 224,
"action_chunk_size": 32,
"actions_per_plan": 10,
Expand Down
1 change: 0 additions & 1 deletion configs/fastwam/robotwin_i2va.json
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,7 +12,6 @@
"robot_state_dim": 14,
"policy_profile": "robotwin",
"normalize_mode": "z-score",
"target_video_length": 9,
"binarize_gripper": false,
"gripper_postprocess": false,
"default_prompt": "A video recorded from a robot's point of view executing the following instruction: {task_prompt}"
Expand Down
24 changes: 21 additions & 3 deletions lightx2v/models/networks/cosmos3/infer/pre_infer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -145,7 +145,16 @@ def _prepare_sound_segment(self, sound_latents, text_segment, vision_segment, de
"sound_len": sound_len,
}

def _prepare_action_segment(self, action_latents, text_segment, vision_segment, sound_segment, device, condition_frame_indexes=None):
def _prepare_action_segment(
self,
action_latents,
text_segment,
vision_segment,
sound_segment,
device,
condition_frame_indexes=None,
start_frame_offset=1,
):
action_len = int(action_latents.shape[0])
condition_frame_indexes = [] if condition_frame_indexes is None else condition_frame_indexes
cond_set = {int(idx) for idx in condition_frame_indexes if 0 <= int(idx) < action_len}
Expand All @@ -162,7 +171,7 @@ def _prepare_action_segment(self, action_latents, text_segment, vision_segment,
base_fps=self.base_fps,
temporal_compression_factor=1,
base_temporal_compression_factor=self.config.get("vae_scale_factor_temporal", 4),
start_frame_offset=1,
start_frame_offset=int(start_frame_offset),
)
sequence_indexes = torch.arange(curr, curr + action_len, dtype=torch.long, device=device)
return {
Expand All @@ -186,6 +195,7 @@ def infer(
action_latents=None,
action_domain_id=None,
action_condition_frame_indexes=None,
action_start_frame_offset=1,
raw_action_dim=None,
):
device = latents.device
Expand All @@ -196,7 +206,15 @@ def infer(
if sound_latents is not None:
sound_segment = self._prepare_sound_segment(sound_latents, text_segment, vision_segment, device)
if action_latents is not None:
action_segment = self._prepare_action_segment(action_latents, text_segment, vision_segment, sound_segment, device, condition_frame_indexes=action_condition_frame_indexes)
action_segment = self._prepare_action_segment(
action_latents,
text_segment,
vision_segment,
sound_segment,
device,
condition_frame_indexes=action_condition_frame_indexes,
start_frame_offset=action_start_frame_offset,
)
sequence_length = text_segment["und_len"] + vision_segment["num_vision_tokens"] + sound_segment.get("sound_len", 0) + action_segment.get("action_len", 0)

packed_text_embedding = weights.embed_tokens.apply(text_segment["input_ids"])
Expand Down
1 change: 1 addition & 0 deletions lightx2v/models/networks/cosmos3/model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -133,6 +133,7 @@ def _infer_cond_uncond(self, input_ids):
action_latents=getattr(self.scheduler, "action_latents", None),
action_domain_id=getattr(self.scheduler, "action_domain_id", None),
action_condition_frame_indexes=getattr(self.scheduler, "action_condition_frame_indexes", None),
action_start_frame_offset=getattr(self.scheduler, "action_start_frame_offset", 1),
raw_action_dim=getattr(self.scheduler, "raw_action_dim", None),
)
if self.config["seq_parallel"]:
Expand Down
Loading
Loading