Skip to content

Cosmos3: give the generation positions a contiguous tensor so a new prompt length does not recompile the denoise step #922

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

Merged
NSagan271 merged 2 commits into from
Oct 28, 2026
Merged

Cosmos3: give the generation positions a contiguous tensor so a new prompt length does not recompile the denoise step #922

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
3deacd2
cosmos3: give the generation band's position ids their own tensor
merceod Oct 27, 2026
0556690
cosmos3: say what the position-id layout test guards against
merceod Oct 27, 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
9 changes: 9 additions & 0 deletions mstar/model/cosmos3/components/packing.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -497,10 +497,18 @@ def build_static_inputs(
pos_dtype = torch.float32 if any(p.is_floating_point() for p in parts) else torch.long
position_ids = torch.cat([p.to(pos_dtype) for p in parts], dim=1)
static["position_ids"] = position_ids
static["gen_position_ids"] = _gen_band(position_ids, text["und_len"])
static["sequence_length"] = sequence_length
return static


def _gen_band(position_ids: torch.Tensor, und_len: int) -> torch.Tensor:
"""The generation band's position ids as their own tensor. A slice of
``position_ids`` has the prompt length in its stride, which the compiled
denoise step guards on, so every new prompt length would recompile it."""
return position_ids[:, und_len:].contiguous()


def build_t2i_static_inputs(
input_ids: list[int],
latent_shape: tuple[int, int, int, int, int],
Expand Down Expand Up @@ -585,5 +593,6 @@ def build_action_static_inputs(
"num_noisy_action_tokens": len(action_noisy),
"action_mode": mode,
"position_ids": position_ids,
"gen_position_ids": _gen_band(position_ids, text["und_len"]),
"sequence_length": curr + action_chunk_size,
}
26 changes: 12 additions & 14 deletions mstar/model/cosmos3/submodules.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1758,16 +1758,16 @@ def _forward_video_sound_gen(self, attn, st, latents, sound_latents, time_index,

if not cfg_active:
velocity, sound_v = self.transformer.denoise_step(
latents, vts, st["cond"]["position_ids"][:, st["cond"]["und_len"]:],
latents, vts, st["cond"]["gen_position_ids"],
st["cond"]["vision_token_shapes"], st["cond"]["vision_noisy_frame_indexes"],
st["cond"]["mse_gen_indexes"], COND_LABEL, attn,
**self._sound_kwargs(st["cond"], sound_latents, sts),
)
elif self.batched_cfg:
(cond_v, s_c), (uncond_v, s_u) = self.transformer.denoise_step_batched_cfg(
latents, vts,
st["cond"]["position_ids"][:, st["cond"]["und_len"]:],
st["uncond"]["position_ids"][:, st["uncond"]["und_len"]:],
st["cond"]["gen_position_ids"],
st["uncond"]["gen_position_ids"],
st["cond"]["vision_token_shapes"],
st["cond"]["vision_noisy_frame_indexes"],
st["cond"]["mse_gen_indexes"], CFG_BATCHED_LABEL, attn,
Expand All @@ -1777,13 +1777,13 @@ def _forward_video_sound_gen(self, attn, st, latents, sound_latents, time_index,
sound_v = s_u + st["gs"] * (s_c - s_u)
else:
cond_v, s_c = self.transformer.denoise_step(
latents, vts, st["cond"]["position_ids"][:, st["cond"]["und_len"]:],
latents, vts, st["cond"]["gen_position_ids"],
st["cond"]["vision_token_shapes"], st["cond"]["vision_noisy_frame_indexes"],
st["cond"]["mse_gen_indexes"], COND_LABEL, attn,
**self._sound_kwargs(st["cond"], sound_latents, sts),
)
uncond_v, s_u = self.transformer.denoise_step(
latents, vts, st["uncond"]["position_ids"][:, st["uncond"]["und_len"]:],
latents, vts, st["uncond"]["gen_position_ids"],
st["uncond"]["vision_token_shapes"], st["uncond"]["vision_noisy_frame_indexes"],
st["uncond"]["mse_gen_indexes"], UNCOND_LABEL, attn,
**self._sound_kwargs(st["uncond"], sound_latents, sts),
Expand All @@ -1806,11 +1806,10 @@ def _forward_video_sound_gen(self, attn, st, latents, sound_latents, time_index,
}

def _denoise_action(self, attn, static, latents, action_latents, vts, ats, domain, label):
und_len = static["und_len"]
return self.transformer.denoise_step(
latents,
vts,
static["position_ids"][:, und_len:],
static["gen_position_ids"],
static["vision_token_shapes"],
static["vision_noisy_frame_indexes"],
static["mse_gen_indexes"],
Expand Down Expand Up @@ -1873,8 +1872,8 @@ def _forward_action_gen(self, attn, st, latents, action_latents, time_index, **k
(video_v, action_v), (v_u, a_u) = self.transformer.denoise_step_batched_cfg(
latents,
vts,
st["cond"]["position_ids"][:, st["cond"]["und_len"]:],
st["uncond"]["position_ids"][:, st["uncond"]["und_len"]:],
st["cond"]["gen_position_ids"],
st["uncond"]["gen_position_ids"],
st["cond"]["vision_token_shapes"],
st["cond"]["vision_noisy_frame_indexes"],
st["cond"]["mse_gen_indexes"],
Expand Down Expand Up @@ -2055,8 +2054,8 @@ def _forward_batched_sound(self, engine_inputs, latents, sound_latents, time_ind
reqs.append({
"latents": lat,
"vision_timesteps": torch.full((st["cond"]["num_noisy_vision_tokens"],), t.item(), device=lat.device),
"position_ids_cond": cond["position_ids"][:, cond["und_len"]:],
"position_ids_uncond": unc["position_ids"][:, unc["und_len"]:],
"position_ids_cond": cond["gen_position_ids"],
"position_ids_uncond": unc["gen_position_ids"],
"vision_token_shapes": cond["vision_token_shapes"],
"vision_noisy_frame_indexes": cond["vision_noisy_frame_indexes"],
"vision_mse_loss_indexes": cond["mse_gen_indexes"],
Expand Down Expand Up @@ -2108,13 +2107,12 @@ def _forward_batched_action(self, engine_inputs, latents, action_latents, time_i
# step below.
t = st["scheduler"].timesteps[min(step_index, n_steps - 1)]
cond = st["cond"]
und = cond["und_len"]
req = {
"latents": lat,
"action_latents": act,
"vision_timesteps": torch.full((st["cond"]["num_noisy_vision_tokens"],), t.item(), device=lat.device),
"action_timesteps": torch.full((st["cond"]["num_noisy_action_tokens"],), t.item(), device=lat.device),
"position_ids_cond": cond["position_ids"][:, und:],
"position_ids_cond": cond["gen_position_ids"],
"vision_token_shapes": cond["vision_token_shapes"],
"vision_noisy_frame_indexes": cond["vision_noisy_frame_indexes"],
"vision_mse_loss_indexes": cond["mse_gen_indexes"],
Expand All @@ -2125,7 +2123,7 @@ def _forward_batched_action(self, engine_inputs, latents, action_latents, time_i
}
if with_cfg:
unc = st["uncond"]
req["position_ids_uncond"] = unc["position_ids"][:, unc["und_len"]:]
req["position_ids_uncond"] = unc["gen_position_ids"]
reqs.append(req)
meta.append((rid, st, lat, act, ti, t))

Expand Down
27 changes: 27 additions & 0 deletions mstar/model/cosmos3/tests/test_action.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -24,6 +24,7 @@
from mstar.model.cosmos3.components.packing import (
action_condition_frame_indexes,
build_action_static_inputs,
build_static_inputs,
get_3d_mrope_ids_action_tokens,
vision_condition_frame_indexes,
)
Expand Down Expand Up @@ -118,6 +119,32 @@ def test_action_static_layout() -> None:
assert s["action_mse_loss_indexes"].numel() == exp_act_noisy


def test_gen_position_ids_do_not_depend_on_prompt_length() -> None:
"""The denoise step is compiled with dynamic=False, which guards on sizes
and strides, so the generation band's position ids must have the same
layout for every prompt length."""
cfg = _cfg()
latent_shape = (1, cfg.latent_channel, 3, 4, 4)
builds = {
"action": lambda ids: build_action_static_inputs(
ids, latent_shape, 8, "policy", cfg, cfg.vae.scale_factor_temporal,
fps=10.0, action_fps=10.0, action_start_offset=1, device="cpu",
),
"sound": lambda ids: build_static_inputs(
ids, latent_shape, cfg, cfg.vae.scale_factor_temporal, fps=10.0, device="cpu",
sound_latent_frames=4,
),
}
for build in builds.values():
layouts = set()
for n in (4, 9, 23):
s = build(list(range(1, n + 1)))
gen = s["gen_position_ids"]
assert torch.equal(gen, s["position_ids"][:, s["und_len"]:])
layouts.add((tuple(gen.shape), gen.stride(), gen.storage_offset()))
assert len(layouts) == 1


def _run_mode(model, cfg, mode, latent_shape, action_chunk, ids):
s = build_action_static_inputs(
ids, latent_shape, action_chunk, mode, cfg, cfg.vae.scale_factor_temporal,
Expand Down
Loading