[Lingbot World] Bound image condition storage and extend temporal RoPE (#6838)

Signed-off-by: wtz2333 <2955110911@qq.com>

examples/offline_inference/diffusion/ar_diffusion_streaming_decode.py

@@ -24,10 +24,9 @@ Examples::
 
 ``--action-script`` is a JSON list with one camera action list per AR block,
 for example ``[[["w"], ["w"], ["w"]], [["a"], [], []]]`` for a checkpoint that
-generates three latent frames per block. The frames per block, the temporal
-compression and the number of blocks that fit the condition horizon are read
-off the checkpoint rather than assumed, so the script can be repointed with
-``--model``.
+generates three latent frames per block. The frames per block and temporal
+compression are read off the checkpoint rather than assumed, so the script
+can be repointed with ``--model``.
 """
 
 from __future__ import annotations
@@ -49,9 +48,6 @@ if TYPE_CHECKING:
     import numpy as np
 
 _MODEL = "robbyant/lingbot-world-v2-14b-causal-fast-diffusers"
-# Mirrors _MAX_RAW_FRAMES in vllm_omni/diffusion/models/lingbot_world/pipeline.py:
-# the image-condition horizon is the pipeline's constant, not the checkpoint's.
-_MAX_RAW_FRAMES = 117
 
 
 def parse_args(argv: Sequence[str] | None = None) -> argparse.Namespace:
@@ -67,8 +63,8 @@ def parse_args(argv: Sequence[str] | None = None) -> argparse.Namespace:
         "--action-script",
         required=True,
         help=(
-            "JSON file holding one camera action list per AR block. Both the per-block frame count "
-            "and the number of blocks that fit are read off the checkpoint, not assumed."
+            "JSON file holding one camera action list per AR block. "
+            "The per-block frame count is read off the checkpoint, not assumed."
         ),
     )
     parser.add_argument("--output-dir", required=True, help="Directory for decoded frames and metadata.")
@@ -154,12 +150,6 @@ def _validate_against_checkpoint(
                 f"--action-script block {block_index} holds {len(block)} frame action lists, but this "
                 f"checkpoint generates {frames_per_block} latent frames per AR block."
             )
-    max_chunks = ((_MAX_RAW_FRAMES - 1) // temporal_compression + 1) // frames_per_block
-    if len(script) > max_chunks:
-        raise ValueError(
-            f"--action-script holds {len(script)} blocks, but this checkpoint's "
-            f"{_MAX_RAW_FRAMES}-frame image-condition horizon fits at most {max_chunks}."
-        )
     return (len(script) * frames_per_block - 1) * temporal_compression + 1
 
 

examples/offline_inference/diffusion/lingbot_world_v2.py

@@ -32,7 +32,6 @@ import numpy as np
 _MODEL = "robbyant/lingbot-world-v2-14b-causal-fast-diffusers"
 _MAX_SEQUENCE_LENGTH = 512
 _MAX_PIXEL_AREA = 480 * 832
-_MAX_RAW_FRAMES = 117
 _MAX_SOURCE_CAMERA_FRAMES = 4096
 _TEMPORAL_COMPRESSION = 4
 _LATENT_FRAMES_PER_BLOCK = 3
@@ -191,8 +190,6 @@ def build_request(
         raise ValueError("--height * --width must not exceed 480 * 832 pixels.")
     if args.num_frames <= 0:
         raise ValueError("--num-frames must be a positive integer.")
-    if args.num_frames > _MAX_RAW_FRAMES:
-        raise ValueError("--num-frames must not exceed 117 raw frames.")
     if args.num_frames > paths.camera_frames:
         raise ValueError("Camera arrays must contain at least --num-frames frames.")
     if (args.num_frames - 1) % _TEMPORAL_COMPRESSION:

examples/offline_inference/diffusion/lingbot_world_v2_realtime.py

@@ -1,13 +1,12 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """Run LingBot-World 2.0 as an in-process realtime AR-Diffusion session.
 
 This example exercises the internal session API. It is not a public HTTP or
 WebSocket protocol. Each JSONL event advances the world by one three-latent
-frame AR block and writes the latent plus its identity metadata. The current
-scoped implementation supports at most ten ticks per generation epoch because
-its image condition is bounded to 117 pixel frames; reset or create a session
-to start a new world.
+frame AR block and writes the latent plus its identity metadata. Session-owned
+causal VAE encoder state produces each condition block without retaining or
+repeating a full image-condition horizon.
 """
 
 from __future__ import annotations
@@ -24,8 +23,6 @@ from typing import Any
 _MODEL = "robbyant/lingbot-world-v2-14b-causal-fast-diffusers"
 _CAMERA_ACTION_SCHEMA = "lingbot.camera_actions.v1"
 _FRAMES_PER_BLOCK = 3
-# ((117 pixel frames - 1) / VAE temporal factor 4 + 1) / 3 latent frames.
-_MAX_REALTIME_TICKS = 10
 
 
 def _camera_event_data(frames: list[list[str]]) -> dict[str, Any]:
@@ -41,7 +38,7 @@ def parse_args(argv: Sequence[str] | None = None) -> argparse.Namespace:
     parser.add_argument(
         "--events",
         required=True,
-        help="JSONL file with one prompt/action event per AR block; at most 10 events.",
+        help="JSONL file with one prompt/action event per AR block.",
     )
     parser.add_argument("--output-dir", required=True, help="Directory for chunk latents and metadata.")
     parser.add_argument("--session-id", default="lingbot-world", help="Persistent world session identifier.")
@@ -84,12 +81,6 @@ def _load_events(path: Path) -> list[dict[str, Any]]:
         events.append({"event_id": event_id, "prompt": prompt, "frames": frames})
     if not events:
         raise ValueError("events file must contain at least one event.")
-    if len(events) > _MAX_REALTIME_TICKS:
-        raise ValueError(
-            "events file must contain at most "
-            f"{_MAX_REALTIME_TICKS} events because the current LingBot realtime "
-            "image-condition horizon is 117 pixel frames."
-        )
     event_ids = [event["event_id"] for event in events]
     if event_ids != sorted(set(event_ids)):
         raise ValueError("event_id values must be unique and strictly increasing.")