# SPDX-License-Identifier: Apache-2.0 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project """Lance transformer pieces. The Lance LLM is BAGEL's Qwen2-MoT transformer verbatim — the released ``Lance_3B`` checkpoint uses the identical ``*_moe_gen`` / ``q_norm`` / ``vae2llm`` / ``llm2vae`` / ``time_embedder`` / ``latent_pos_embed`` layout, so ``Bagel`` / ``Qwen2MoTForCausalLM`` / ``Qwen2MoTConfig`` / ``NaiveCache`` are re-exported unchanged. Only the understanding ViT differs (Qwen2.5-VL vision instead of SigLIP) and the video path adds a 3D latent position embedding. This module also provides :class:`LanceBagel`, a thin :class:`Bagel` subclass that overrides the two ViT entry points to consume Qwen2.5-VL's packed ``pixel_values`` + ``image_grid_thw`` layout directly (BAGEL itself assumes SigLIP-style ``(C, H, W)`` tensors that get patchified inside the model). NOTE ON mRoPE: Lance's backbone is Qwen2.5-VL and its understanding / video paths use multimodal RoPE (``mrope_section=[16,24,24]``). BAGEL's ``BagelRotaryEmbedding`` is plain 1-D RoPE on scalar position ids. For the text2img generation path Lance assigns scalar positions (the gen latent block shares a single rope position, same as BAGEL), so the reused rotary is correct there. Full mRoPE for the x2t / video understanding path is a follow-up — see ``LancePositionEmbedding3D``. """ from __future__ import annotations import numpy as np import torch from torch import nn # Lance LLM == BAGEL Qwen2-MoT: reuse verbatim (checkpoint weight names match). from vllm_omni.diffusion.models.bagel.bagel_transformer import ( # noqa: F401 Bagel, BaseNavitOutputWithPast, MLPconnector, NaiveCache, PackedAttentionMoT, PositionEmbedding, Qwen2MoTConfig, Qwen2MoTDecoderLayer, Qwen2MoTForCausalLM, Qwen2MoTModel, TimestepEmbedder, get_1d_sincos_pos_embed_from_grid, patchify, ) __all__ = [ "Bagel", "BaseNavitOutputWithPast", "LanceBagel", "LanceIdentityConnector", "LancePositionEmbedding3D", "LanceQwen2_5_VLNaViTWrapper", "LanceZeroVitPosEmbed", "MLPconnector", "NaiveCache", "PackedAttentionMoT", "PositionEmbedding", "Qwen2MoTConfig", "Qwen2MoTDecoderLayer", "Qwen2MoTForCausalLM", "Qwen2MoTModel", "TimestepEmbedder", "patchify", ] # mRoPE temporal scaling constants for Lance (Qwen2.5-VL backbone). # Upstream Lance computes vision rope positions as # ``t_index = frame_idx * tokens_per_second * second_per_grid_t`` so adjacent # latent frames sit at well-separated rope coordinates. Lance bundles a # ``Qwen2.5-VL-ViT/config.json`` with ``tokens_per_second: 2`` and uses # ``second_per_grid_ts = 1.0`` in its inference script. LANCE_TOKENS_PER_SECOND = 2 LANCE_SECONDS_PER_GRID = 1.0 # Upstream Lance ViT preprocessing (see lance-upstream # ``data/datasets_custom/validation_dataset.py`` + ``data/video/transforms/``): # ``VideoTransform(resolution=resolution_vit, mode='bucket', # divisible_crop_size=28, stride_spatial=16, aspect_ratios=[...])``. # For ``image_768res`` (Lance's default for 768px image gen), resolution_vit=672. # Resizes to the nearest aspect-ratio bucket (max_area=672², stride=16), then # center-crops to dims divisible by 28 (=patch_size * spatial_merge_size). # Standard Qwen2VLImageProcessor's smart-resize is NOT used by upstream — the # image enters the ViT at the bucket-cropped dimensions. LANCE_VIT_BUCKET_RESOLUTION = 672 LANCE_VIT_BUCKET_STRIDE = 16 LANCE_VIT_DIVISIBLE_CROP = 28 # patch_size (14) * spatial_merge_size (2) LANCE_VIT_ASPECT_RATIOS = ("21:9", "16:9", "4:3", "1:1", "3:4", "9:16") # Video-specific bucket parameters (resolution=video_480p — Lance's default for # t2v / video_edit at 12 fps). See lance-upstream # ``data/datasets_custom/validation_dataset.py`` lines 128-156: # VAE side: resolution_vae=640, divisible_crop_size=16, mean=0.5, std=0.5 # ViT side: resolution_vit=616, divisible_crop_size=28, # mean=Qwen2.5-VL CLIP-norm, std=Qwen2.5-VL CLIP-norm LANCE_VIDEO_VAE_RESOLUTION = 640 LANCE_VIDEO_VAE_DIVISIBLE_CROP = 16 LANCE_VIDEO_VIT_RESOLUTION = 616 LANCE_VIDEO_VIT_DIVISIBLE_CROP = 28 LANCE_VIDEO_BUCKET_STRIDE = 16 LANCE_VIDEO_TEMPORAL_STRIDE = 4 # Wan2.2 VAE temporal downsample LANCE_VIDEO_SAMPLE_FPS = 12 # upstream MultiClipsFrameSampler default LANCE_VIDEO_MAX_DURATION = 6.0 # upstream max_duration (seconds) LANCE_VIT_PATCH_SIZE = 14 # Qwen2.5-VL ViT spatial patch LANCE_VIT_TEMPORAL_PATCH_SIZE = 2 # Qwen2.5-VL ViT temporal patch LANCE_VIT_SPATIAL_MERGE = 2 # Qwen2.5-VL spatial merge factor LANCE_VIT_NORM_MEAN = (0.48145466, 0.4578275, 0.40821073) LANCE_VIT_NORM_STD = (0.26862954, 0.26130258, 0.27577711) def get_3d_sincos_pos_embed_from_grid(embed_dim: int, grid: np.ndarray) -> np.ndarray: """3-D sin-cos positional embedding (t, h, w), matching the upstream Lance ``modeling/lance/modeling_utils.py`` dimension split exactly.""" assert embed_dim % 2 == 0, "Embedding dimension must be even for 3D embeddings" d = embed_dim // 3 d = d if d % 2 == 0 else d - 1 dim_t, dim_h = d, d dim_w = embed_dim - 2 * d assert dim_w % 2 == 0 emb_t = get_1d_sincos_pos_embed_from_grid(dim_t, grid[0]) emb_h = get_1d_sincos_pos_embed_from_grid(dim_h, grid[1]) emb_w = get_1d_sincos_pos_embed_from_grid(dim_w, grid[2]) return np.concatenate([emb_t, emb_h, emb_w], axis=1) def get_3d_sincos_pos_embed(embed_dim: int, t: int, h: int, w: int) -> np.ndarray: grid_t = np.arange(t, dtype=np.float32) grid_h = np.arange(h, dtype=np.float32) grid_w = np.arange(w, dtype=np.float32) tt, hh, ww = np.meshgrid(grid_t, grid_h, grid_w, indexing="ij") grid = np.stack([tt, hh, ww], axis=0) return get_3d_sincos_pos_embed_from_grid(embed_dim, grid) class LancePositionEmbedding3D(nn.Module): """Frozen 3-D sin-cos latent position embedding for the video path. BAGEL only ships a 2-D ``PositionEmbedding`` (image latents). Lance's ``Lance_3B_Video`` checkpoint adds a temporal axis; this mirrors upstream ``modeling/lance/modeling_utils.py::PositionEmbedding3D``. The image path uses ``t=1`` and is numerically equivalent to the 2-D embedding. """ def __init__(self, max_num_frames: int, max_num_patch_per_side: int, hidden_size: int): super().__init__() self.max_num_frames = max_num_frames self.max_num_patch_per_side = max_num_patch_per_side self.hidden_size = hidden_size n = max_num_frames * max_num_patch_per_side * max_num_patch_per_side self.pos_embed = nn.Parameter(torch.zeros(n, hidden_size), requires_grad=False) self._init_weights() def _init_weights(self) -> None: pe = get_3d_sincos_pos_embed( self.hidden_size, self.max_num_frames, self.max_num_patch_per_side, self.max_num_patch_per_side, ) self.pos_embed.data.copy_(torch.from_numpy(pe).float()) def forward(self, position_ids: torch.Tensor) -> torch.Tensor: return self.pos_embed[position_ids] class LanceIdentityConnector(nn.Module): """No-op connector for Lance. BAGEL's ``connector`` projects the ViT hidden size to the LLM hidden size. Qwen2.5-VL's vision tower (which Lance uses) already projects to the LLM hidden size internally via ``merger`` (``out_hidden_size = hidden_size``), and the released Lance safetensors carry no ``connector.*`` weights. We therefore plug in an ``Identity`` connector so ``forward_cache_update_vit`` keeps its existing call site without a separate code path. """ def forward(self, x: torch.Tensor) -> torch.Tensor: # noqa: D401 return x class LanceZeroVitPosEmbed(nn.Module): """No-op positional embedding for Lance's ViT tokens. BAGEL adds an extra 2-D sin-cos ``vit_pos_embed`` on top of the ViT output. Qwen2.5-VL's vision tower already carries its own (rotary) positional encoding, and the released Lance safetensors carry no ``vit_pos_embed.*`` weights. This module returns a broadcast-friendly zero so the addition in ``forward_cache_update_vit`` is a no-op without requiring a code-path branch. """ def forward(self, position_ids: torch.Tensor) -> torch.Tensor: # noqa: D401 return torch.zeros((), device=position_ids.device) class LanceQwen2_5_VLNaViTWrapper(nn.Module): """Packed (NaViT-style) wrapper around the Qwen2.5-VL vision tower. Bridges BAGEL's ``vit(packed_pixel_values, ...) -> [num_tokens, vit_hidden]`` surface to the HF ``Qwen2_5_VisionTransformerPretrainedModel`` which consumes ``(hidden_states, grid_thw)``. The packed call additionally needs a per-image ``image_grid_thw`` so non-square images (and the spatial-merge token count) line up — :class:`LanceBagel` stashes the grid on the wrapper before invoking the ViT. """ def __init__(self, vision_model: nn.Module, spatial_merge_size: int = 2): super().__init__() # Accept either the full Qwen2_5_VLForConditionalGeneration.visual or # the bare vision transformer. self.vision_model = getattr(vision_model, "visual", vision_model) self.spatial_merge_size = spatial_merge_size # Set by ``LanceBagel.forward_cache_update_vit`` before each call so # the wrapper can pass true per-image (T, H, W) to the HF ViT. self._pending_grid_thw: torch.Tensor | None = None @property def config(self): # parity with SiglipNaViTWrapper.vision_model.config access return self.vision_model.config def set_pending_grid_thw(self, grid_thw: torch.Tensor) -> None: self._pending_grid_thw = grid_thw def forward( self, packed_pixel_values: torch.Tensor, packed_flattened_position_ids: torch.Tensor, cu_seqlens: torch.Tensor, max_seqlen: int, ) -> torch.Tensor: if self._pending_grid_thw is not None: grid_thw_t = self._pending_grid_thw.to(packed_pixel_values.device) self._pending_grid_thw = None else: # Fallback for square images when grid_thw was not pre-stashed. cu = cu_seqlens.tolist() grid_thw = [] for i in range(len(cu) - 1): n = cu[i + 1] - cu[i] side = int(round(float(n) ** 0.5)) grid_thw.append([1, side, side]) grid_thw_t = torch.tensor(grid_thw, dtype=torch.long, device=packed_pixel_values.device) hidden = packed_pixel_values if hasattr(self.vision_model, "dtype"): hidden = hidden.to(self.vision_model.dtype) out = self.vision_model(hidden_states=hidden, grid_thw=grid_thw_t) # Qwen2.5-VL returns ``BaseModelOutputWithPooling(last_hidden_state=..., # pooler_output=merged)``; the merged sequence (post spatial-merger, # shape ``[sum_tokens_after_merge, out_hidden_size]``) is what the LLM # consumes. if hasattr(out, "pooler_output") and out.pooler_output is not None: return out.pooler_output if isinstance(out, tuple): return out[0] return out class LanceBagel(Bagel): """Bagel subclass with Lance-specific ViT handling. The released Lance checkpoint pairs BAGEL's Qwen2-MoT trunk with the Qwen2.5-VL vision tower (whose ``merger`` already projects to LLM ``hidden_size`` and which carries its own rotary positional encoding). Two BAGEL assumptions therefore break for Lance: 1. ``prepare_vit_images`` calls ``transforms(image) -> (C, H, W)`` and then ``patchify(...)``; Lance's image processor returns ``(num_patches_flat, patch_features) + image_grid_thw`` already. 2. ``forward_cache_update_vit`` adds a ``connector`` projection plus a 2-D ``vit_pos_embed`` that has no checkpoint weights and would also double-count the ViT's own positional encoding. This subclass overrides exactly those two methods. Everything else — LLM trunk, VAE flow, generation loop, ``forward_cache_update_text`` / ``forward_cache_update_vae`` — is reused unchanged. """ @staticmethod def _lance_compute_vit_buckets(): """Replicate upstream Lance's ``BucketResize.init_buckets`` for the ViT path. Returns a list of ``((bucket_h, bucket_w), bucket_ratio)`` pairs in the same order as :data:`LANCE_VIT_ASPECT_RATIOS`. Stays in sync with upstream lance ``data/video/transforms/bucket_resize.py``. """ import math max_area = LANCE_VIT_BUCKET_RESOLUTION * LANCE_VIT_BUCKET_RESOLUTION stride = LANCE_VIT_BUCKET_STRIDE buckets: list[tuple[tuple[int, int], float]] = [] for name in LANCE_VIT_ASPECT_RATIOS: w_s, h_s = (int(v) for v in name.split(":")) aspect = w_s / h_s # Option 1: solve for width first. rw1 = math.sqrt(max_area * aspect) bw1 = round(rw1 / stride) * stride bh1 = round(bw1 / aspect / stride) * stride br1 = bw1 / bh1 # Option 2: solve for height first. rh2 = math.sqrt(max_area / aspect) bh2 = round(rh2 / stride) * stride bw2 = round(bh2 * aspect / stride) * stride br2 = bw2 / bh2 if abs(br1 - aspect) < abs(br2 - aspect): bh, bw = bh1, bw1 elif abs(br1 - aspect) > abs(br2 - aspect): bh, bw = bh2, bw2 else: area1 = bh1 * bw1 area2 = bh2 * bw2 if abs(area1 - max_area) <= abs(area2 - max_area): bh, bw = bh1, bw1 else: bh, bw = bh2, bw2 buckets.append(((bh, bw), bw / bh)) return buckets @staticmethod def _lance_bucket_resize_pil(image): """Apply upstream Lance's ViT preprocessing to a PIL image. Pipeline (matches lance-upstream's ``VideoTransform`` for the ViT path): 1. ``BucketResize``: pick the aspect-ratio bucket nearest to the input and resize via ``RandomResizedCrop(scale=(1,1), ratio=fixed)``, which for fixed ratio is a deterministic center-crop + resize. 2. ``DivisibleCrop(28)``: center-crop to dims divisible by ``patch_size * spatial_merge_size`` so the resulting ViT grid is clean. Returns the post-pipeline PIL image (still RGB, just resized). Doing this BEFORE the Qwen2VLImageProcessor means smart-resize is a no-op and the ViT sees exactly upstream Lance's input. """ import numpy as _np from torchvision.transforms import InterpolationMode, RandomResizedCrop if image.mode != "RGB": image = image.convert("RGB") w, h = image.size img_ratio = w / h buckets = LanceBagel._lance_compute_vit_buckets() # nearest bucket by absolute ratio difference diffs = _np.array([abs(img_ratio - br) for _, br in buckets]) bh, bw = buckets[int(diffs.argmin())][0] bratio = bw / bh # RandomResizedCrop(scale=(1,1), ratio=(bratio, bratio)) is deterministic: # picks the largest sub-image of the target aspect, then resizes to # (bh, bw). We replicate this directly to avoid the torchvision # transform's random-state dependency and to keep the result a PIL. # IMPORTANT: upstream Lance wires BucketResize through ``NaResize`` # which defaults to BICUBIC interpolation (NOT LANCZOS — BucketResize's # own default is LANCZOS but ``NaResize`` overrides it). Using # LANCZOS here gave byte-different ViT input pixels (max_diff~0.13) # which fed into the noise-query attention as wrong K/V — symptom: # vllm-omni preserved the input image's hat when upstream removed it. rrc = RandomResizedCrop( size=(bh, bw), scale=(1.0, 1.0), ratio=(bratio, bratio), interpolation=InterpolationMode.BICUBIC, ) out = rrc(image) # DivisibleCrop(28): center-crop to dims divisible by 28. ow, oh = out.size cropped_w = ow - (ow % LANCE_VIT_DIVISIBLE_CROP) cropped_h = oh - (oh % LANCE_VIT_DIVISIBLE_CROP) if cropped_w != ow or cropped_h != oh: left = (ow - cropped_w) // 2 top = (oh - cropped_h) // 2 out = out.crop((left, top, left + cropped_w, top + cropped_h)) return out # ------------------------------------------------------------------ # # Video preprocessing — frame sampler + bucket resize for VAE / ViT # ------------------------------------------------------------------ # @staticmethod def _lance_compute_video_buckets(resolution: int, stride: int): """Generic version of :meth:`_lance_compute_vit_buckets` for arbitrary ``resolution`` / ``stride`` (the video VAE uses 640/16, the video ViT uses 616/16, both with the same 6 aspect-ratio table).""" import math max_area = resolution * resolution buckets: list[tuple[tuple[int, int], float]] = [] for name in LANCE_VIT_ASPECT_RATIOS: w_s, h_s = (int(v) for v in name.split(":")) aspect = w_s / h_s rw1 = math.sqrt(max_area * aspect) bw1 = round(rw1 / stride) * stride bh1 = round(bw1 / aspect / stride) * stride br1 = bw1 / bh1 rh2 = math.sqrt(max_area / aspect) bh2 = round(rh2 / stride) * stride bw2 = round(bh2 * aspect / stride) * stride br2 = bw2 / bh2 if abs(br1 - aspect) < abs(br2 - aspect): bh, bw = bh1, bw1 elif abs(br1 - aspect) > abs(br2 - aspect): bh, bw = bh2, bw2 else: area1 = bh1 * bw1 area2 = bh2 * bw2 bh, bw = (bh1, bw1) if abs(area1 - max_area) <= abs(area2 - max_area) else (bh2, bw2) buckets.append(((bh, bw), bw / bh)) return buckets @staticmethod def _lance_sample_frame_indices(num_input_frames: int, origin_fps: float) -> list[int]: """Replicate upstream's ``MultiClipsFrameSampler`` (single-clip path). Upstream config (validation_dataset.py:110): temporal=4, sample_fps=12, truncate=False, max_duration=6.0, length_type="kn+1", assert_seconds=False. For a single clip covering the full input video this reduces to: duration = min(num_input_frames / origin_fps, max_duration) n_frames = round(duration * sample_fps) # round to k*temporal + 1 (or -temporal+1 if already a multiple) if n_frames % temporal != 0: n_frames = (n_frames // temporal) * temporal + 1 else: n_frames = (n_frames // temporal) * temporal + 1 - temporal indices = np.linspace(0, num_input_frames - 1, n_frames, dtype=int) Verified against upstream for car (93 frames @ 24fps → 45 sampled) and woman (121 frames @ 24fps → 57 sampled). """ import numpy as _np duration = num_input_frames / origin_fps duration = min(duration, LANCE_VIDEO_MAX_DURATION) n_frames = int(round(duration * LANCE_VIDEO_SAMPLE_FPS)) t = LANCE_VIDEO_TEMPORAL_STRIDE if n_frames % t != 0: n_frames = (n_frames // t) * t + 1 else: n_frames = (n_frames // t) * t + 1 - t n_frames = max(n_frames, 1) return _np.linspace(0, num_input_frames - 1, n_frames, dtype=int).tolist() @staticmethod def _lance_bucket_resize_video(frames_thwc, resolution: int, stride: int, divisible_crop: int, mean, std): """Apply upstream Lance's VideoTransform pipeline to a (T,H,W,C) uint8 ndarray. Returns ``(C, T, H', W')`` float tensor. Pipeline (mirrors lance-upstream's ``data/transforms.py::VideoTransform``): 1. BucketResize per PIL frame: RandomResizedCrop(scale=(1,1), ratio=fixed-bucket-ratio, BICUBIC) → resized PIL → to_tensor. Per-frame PIL path matches upstream's ``BucketResize.__call__(list[PIL.Image])`` exactly — using a batched tensor RRC instead drifted pixel values by ~0.13 (in [0,1] range) because torchvision's tensor-BICUBIC kernel differs subtly from PIL's BICUBIC kernel. 2. DivisibleCrop(divisible_crop) — center-crop to dims divisible by patch_size * spatial_merge_size (28) for ViT, or by VAE downsample (16) for VAE. 3. Normalize(mean, std) — VAE uses 0.5/0.5; ViT uses Qwen CLIP. 4. Rearrange "t c h w -> c t h w" """ import numpy as _np from PIL import Image as _Image from torchvision.transforms import InterpolationMode, RandomResizedCrop from torchvision.transforms.functional import center_crop as _ccrop from torchvision.transforms.functional import to_tensor as _to_tensor if not isinstance(frames_thwc, _np.ndarray): raise ValueError(f"Expected (T,H,W,C) uint8 ndarray; got {type(frames_thwc)}") if frames_thwc.dtype != _np.uint8: frames_thwc = frames_thwc.astype(_np.uint8) if frames_thwc.ndim != 4 or frames_thwc.shape[-1] != 3: raise ValueError(f"Expected (T,H,W,3) uint8; got {frames_thwc.shape}") T_in, H_in, W_in, _ = frames_thwc.shape aspect = W_in / H_in buckets = LanceBagel._lance_compute_video_buckets(resolution, stride) diffs = _np.array([abs(aspect - br) for _, br in buckets]) bh, bw = buckets[int(diffs.argmin())][0] bratio = bw / bh rrc = RandomResizedCrop( size=(bh, bw), scale=(1.0, 1.0), ratio=(bratio, bratio), interpolation=InterpolationMode.BICUBIC, ) # Per-frame PIL → RRC → to_tensor → stack — matches upstream's # ``BucketResize.__call__`` for ``list[PIL.Image]`` exactly. per_frame = [] for t in range(T_in): img = _Image.fromarray(frames_thwc[t]) img = rrc(img) per_frame.append(_to_tensor(img)) # (C, bh, bw) float [0,1] frames = torch.stack(per_frame, dim=0) # (T, C, bh, bw) # DivisibleCrop(divisible_crop) — center-crop along H/W. _, C, Hb, Wb = frames.shape Hc = Hb - (Hb % divisible_crop) Wc = Wb - (Wb % divisible_crop) if Hc != Hb or Wc != Wb: frames = _ccrop(frames, [Hc, Wc]) # Normalize. Accept scalar or per-channel mean/std. if isinstance(mean, (int, float)): mean_t = torch.tensor([float(mean)] * C, dtype=frames.dtype).view(1, C, 1, 1) else: mean_t = torch.tensor(list(mean), dtype=frames.dtype).view(1, C, 1, 1) if isinstance(std, (int, float)): std_t = torch.tensor([float(std)] * C, dtype=frames.dtype).view(1, C, 1, 1) else: std_t = torch.tensor(list(std), dtype=frames.dtype).view(1, C, 1, 1) frames = (frames - mean_t) / std_t # Rearrange "t c h w -> c t h w" frames = frames.permute(1, 0, 2, 3).contiguous() return frames @staticmethod def _lance_patchify_vit_video(video_cthw: torch.Tensor): """Replicate upstream ``patchify_video_with_merge``: (C, T, H, W) → (num_patches, patch_dim). ``patch_dim = C * tp * p² = 3 * 2 * 14² = 1176``. Returns ``(packed_pixels, grid_thw)`` where ``grid_thw`` is the pre-merge grid ``(T // tp, H // p, W // p)`` — same layout the Qwen2.5-VL ViT expects (matches ``Qwen2VLImageProcessor.video_processor`` output up to the actual pixel preprocessing). """ p = LANCE_VIT_PATCH_SIZE tp = LANCE_VIT_TEMPORAL_PATCH_SIZE ms = LANCE_VIT_SPATIAL_MERGE # (C, T, H, W) -> (T, C, H, W) video = video_cthw.permute(1, 0, 2, 3).contiguous() T, C, H, W = video.shape if T % tp != 0: raise ValueError(f"ViT T={T} must be divisible by temporal_patch={tp}") if H % p != 0 or W % p != 0: raise ValueError(f"ViT H,W=({H},{W}) must be divisible by patch={p}") gt, gh, gw = T // tp, H // p, W // p if gh % ms != 0 or gw % ms != 0: raise ValueError(f"ViT grid ({gh},{gw}) must be divisible by merge={ms}") video = video.reshape(gt, tp, C, gh // ms, ms, p, gw // ms, ms, p) # permute order matches upstream: (0,3,6,4,7,2,1,5,8) video = video.permute(0, 3, 6, 4, 7, 2, 1, 5, 8) patches = video.reshape(gt * gh * gw, C * tp * p * p) grid_thw = torch.tensor([[gt, gh, gw]], dtype=torch.long) return patches, grid_thw @classmethod def _lance_video_preprocess(cls, frames_thwc, origin_fps: float): """Full upstream-compatible video preprocessing for video_edit. Pipeline (mirrors lance-upstream ``validation_dataset.py`` for ``--resolution video_480p``): 1. Frame-sample at 12 fps with max_duration=6s (k*temporal+1 rule). 2. Bucket-resize the sampled frames TWICE: - VAE branch (resolution=640, divisible=16, norm=0.5) - ViT branch (resolution=616, divisible=28, norm=Qwen CLIP) 3. ViT branch is patchified to (num_patches, 1176) for direct consumption by Qwen2.5-VL. Returns: vae_video — (C, T_lat_in, H_vae, W_vae) float tensor in model-friendly [-1, 1] (mean=0.5, std=0.5). vit_pixels — (num_patches, 1176) float tensor. vit_grid_thw — (1, 3) long tensor with pre-merge grid (gt, gh, gw). sampled_T — number of RGB frames after the temporal sampler. """ import numpy as _np if isinstance(frames_thwc, torch.Tensor): frames_thwc = frames_thwc.detach().cpu().numpy() if frames_thwc.dtype != _np.uint8: # Allow [-1,1] / [0,1] float arrays; convert back to uint8 so the # bucket transforms operate on the same domain as upstream. arr = frames_thwc if arr.max() <= 1.5: arr = (arr * 255.0).clip(0, 255).astype(_np.uint8) else: arr = arr.astype(_np.uint8) frames_thwc = arr T_in = frames_thwc.shape[0] indices = cls._lance_sample_frame_indices(T_in, origin_fps) sampled = frames_thwc[indices] # (T_sampled, H, W, 3) vae_video = cls._lance_bucket_resize_video( sampled, resolution=LANCE_VIDEO_VAE_RESOLUTION, stride=LANCE_VIDEO_BUCKET_STRIDE, divisible_crop=LANCE_VIDEO_VAE_DIVISIBLE_CROP, mean=0.5, std=0.5, ) vit_video = cls._lance_bucket_resize_video( sampled, resolution=LANCE_VIDEO_VIT_RESOLUTION, stride=LANCE_VIDEO_BUCKET_STRIDE, divisible_crop=LANCE_VIDEO_VIT_DIVISIBLE_CROP, mean=LANCE_VIT_NORM_MEAN, std=LANCE_VIT_NORM_STD, ) # Qwen2.5-VL ViT's temporal patch size is 2 → T must be even. # Upstream pads by repeating the last frame for odd T (see # ``validation_dataset.py:266``): # ``video_tensor = torch.cat([video_tensor, last_frame], dim=1)`` if vit_video.shape[1] % 2 == 1: vit_video = torch.cat([vit_video, vit_video[:, -1:]], dim=1) vit_pixels, vit_grid_thw = cls._lance_patchify_vit_video(vit_video) return vae_video, vit_pixels, vit_grid_thw, len(indices) @staticmethod def _qwen_vl_processor_call(processor, image): """Invoke the Qwen2-VL image processor and return ``(pixel_values, grid_thw)``. Applies upstream Lance's BucketResize+DivisibleCrop BEFORE calling the standard Qwen2VLImageProcessor. After the pre-resize the input dims are already divisible by ``patch_size * spatial_merge_size`` and within the processor's pixel bounds, so the processor's own smart-resize is a no-op — the ViT sees the exact bucket-cropped pixels upstream uses. """ prepped = LanceBagel._lance_bucket_resize_pil(image) proc_out = processor(images=prepped, return_tensors="pt") pixel_values = proc_out["pixel_values"] grid_thw = proc_out["image_grid_thw"] return pixel_values, grid_thw # ------------------------------------------------------------------ # # 3-D mRoPE plumbing # ------------------------------------------------------------------ # @staticmethod def _mrope_broadcast(position_ids: torch.Tensor) -> torch.Tensor: """Convert scalar ``(S,)`` position ids to mrope ``(3, S)`` by broadcasting the scalar value to all three (t, h, w) axes. Lance/Qwen2.5-VL's mRoPE expects 3-D positions per token; for text tokens (and image-edit context tokens) all three axes share the same scalar position id. Vision-generation paths override this with per-token ``(t, h, w)`` indices. """ if position_ids is None: return position_ids if position_ids.ndim == 1: return position_ids.unsqueeze(0).expand(3, -1).contiguous() return position_ids def prepare_prompts(self, *args, **kwargs): gen_input, newlens, new_rope = super().prepare_prompts(*args, **kwargs) if "packed_text_position_ids" in gen_input: gen_input["packed_text_position_ids"] = self._mrope_broadcast(gen_input["packed_text_position_ids"]) return gen_input, newlens, new_rope def _per_token_mrope_for_vae_latent(self, image_sizes, curr_rope) -> torch.Tensor: """Build per-token Qwen2.5-VL mRoPE positions for the gen latent block. Matches upstream Lance (``Lance.validation_gen_KVcache`` → ``get_rope_index``): start_of_image -> (P-1, P-1, P-1) latent (hi, wi) -> (P, P+hi, P+wi) end_of_image -> (P+max(h,w), ..., ...) where ``P = curr_position_id`` is the position counter AFTER the text prefix (BAGEL parent already passes ``gen_context['ropes']``). The ``-1`` on start_of_image matches upstream's get_rope_index layout: the start marker sits one position before the latent block (which is anchored at P), and the end marker sits ``max(h,w)`` positions past. Without per-token (h, w) variation, attention can't see the spatial layout → image PSNR collapses ~6 dB after 30 steps. """ pos_t: list[int] = [] pos_h: list[int] = [] pos_w: list[int] = [] for (H, W), P in zip(image_sizes, curr_rope): h = H // self.latent_downsample w = W // self.latent_downsample # Upstream Lance (``shift_position_ids`` + ``get_rope_index``) # layout for the gen latent block: # start_of_image -> (P, P, P) # latent (hi, wi) -> (P+1, P+1+hi, P+1+wi) # end_of_image -> (P+max+1, P+max+1, P+max+1) # where ``P = curr_position_id`` is the rope counter AFTER the # text+ViT prefix (i.e., :attr:`gen_context['ropes']` snapshot # taken before the VAE-ref prefill in ``_forward_image_edit``). # This matches :meth:`_lance_native_prepare_vae_images` so the # gen latent occupies the SAME rope range as the VAE-ref block # (modality==1 / modality==2 share rope per upstream's # ``i_sample_modality==2`` branch in ``shift_position_ids``). max_hw = max(h, w) pos_t.append(P) pos_h.append(P) pos_w.append(P) for hi in range(h): for wi in range(w): pos_t.append(P + 1) pos_h.append(P + 1 + hi) pos_w.append(P + 1 + wi) end_p = P + max_hw + 1 pos_t.append(end_p) pos_h.append(end_p) pos_w.append(end_p) return torch.stack( [ torch.tensor(pos_t, dtype=torch.long), torch.tensor(pos_h, dtype=torch.long), torch.tensor(pos_w, dtype=torch.long), ], dim=0, ) def _per_token_mrope_for_video_latent(self, video_shapes, curr_rope) -> torch.Tensor: """3-D analogue of :meth:`_per_token_mrope_for_vae_latent` for video. Matches upstream Lance's ``get_rope_index`` for ``modality==noise`` (and ``modality==ref_source`` since they share rope per ``shift_position_ids``). The temporal axis is amplified by ``LANCE_TOKENS_PER_SECOND * LANCE_SECONDS_PER_GRID`` so neighbouring latent frames are well-separated, matching the upstream Qwen2.5-VL rope convention. Verified against upstream's rope dump for video_edit at video_480p (call09 cond noise iter 0): start_of_image -> (P, P, P) latent (t,h,w) -> (P+1+t*2, P+1+h, P+1+w) end_of_image -> (P+max+1, P+max+1, P+max+1) where ``max = max(t_lat*2-1, h_lat-1, w_lat-1) + 1`` matches the end marker placement upstream uses (call09 final token at 159 for T_lat=12, H_lat=35, W_lat=47 → max(22, 34, 46)+1=47, start=111, end=158... actual upstream end=159, suggesting +2 from the body's max). We mirror :meth:`_lance_native_prepare_vae_images` exactly so the VAE_ref prefill and the gen noise share the same rope range. """ downsample_t = int(getattr(self.config.vae_config, "downsample_temporal", 4)) t_scale = LANCE_TOKENS_PER_SECOND * LANCE_SECONDS_PER_GRID pos_t: list[int] = [] pos_h: list[int] = [] pos_w: list[int] = [] for (T, H, W), P in zip(video_shapes, curr_rope): h = H // self.latent_downsample w = W // self.latent_downsample t = (T - 1) // downsample_t + 1 max_thw = max(int((t - 1) * t_scale), h - 1, w - 1) + 1 pos_t.append(P) pos_h.append(P) pos_w.append(P) for ti in range(t): for hi in range(h): for wi in range(w): pos_t.append(P + 1 + int(ti * t_scale)) pos_h.append(P + 1 + hi) pos_w.append(P + 1 + wi) end_p = P + max_thw + 1 pos_t.append(end_p) pos_h.append(end_p) pos_w.append(end_p) return torch.stack( [ torch.tensor(pos_t, dtype=torch.long), torch.tensor(pos_h, dtype=torch.long), torch.tensor(pos_w, dtype=torch.long), ], dim=0, ) def prepare_vae_latent(self, curr_kvlens, curr_rope, image_sizes, new_token_ids): gen_input = super().prepare_vae_latent(curr_kvlens, curr_rope, image_sizes, new_token_ids) gen_input["packed_position_ids"] = self._per_token_mrope_for_vae_latent(image_sizes, curr_rope) return gen_input def prepare_vae_latent_cfg(self, curr_kvlens, curr_rope, image_sizes): gen_input = super().prepare_vae_latent_cfg(curr_kvlens, curr_rope, image_sizes) gen_input["cfg_packed_position_ids"] = self._per_token_mrope_for_vae_latent(image_sizes, curr_rope) return gen_input def prepare_start_tokens(self, *args, **kwargs): gen_input = super().prepare_start_tokens(*args, **kwargs) if isinstance(gen_input, dict): for k in ( "packed_query_position_ids", "packed_text_position_ids", "packed_position_ids", ): if k in gen_input and torch.is_tensor(gen_input[k]): gen_input[k] = self._mrope_broadcast(gen_input[k]) return gen_input def prepare_vit_videos( self, curr_kvlens, curr_rope, videos, new_token_ids, precomputed_vit=None, ): """Multi-frame ViT prefill for the ``x2t_video`` / ``video_edit`` paths. ``videos`` is a list of per-request video tensors / numpy arrays of shape ``(T, H, W, 3)``. By default the Qwen2-VL video processor is used to convert each video to ``(pixel_values_videos, video_grid_thw)``. For ``video_edit`` precision matching, the pipeline may pre-compute the upstream-style BucketResize output and pass it via ``precomputed_vit`` — a list of ``(pixel_values, grid_thw)`` per video, in which case the processor call is skipped. """ processor = getattr(self, "_lance_video_processor", None) if processor is None and precomputed_vit is None: raise RuntimeError( "LanceBagel.prepare_vit_videos requires either ``precomputed_vit`` " "or ``bagel._lance_video_processor`` (a Qwen2-VL-compatible video processor)." ) if precomputed_vit is not None and len(precomputed_vit) != len(videos): raise ValueError( f"precomputed_vit length ({len(precomputed_vit)}) must match number of videos ({len(videos)})" ) packed_vit_token_indexes: list[int] = [] vit_token_seqlens: list[int] = [] packed_vit_tokens: list[torch.Tensor] = [] packed_vit_position_ids: list[torch.Tensor] = [] packed_text_ids: list[int] = [] packed_text_indexes: list[int] = [] packed_seqlens: list[int] = [] packed_indexes: list[int] = [] packed_key_value_indexes: list[int] = [] grid_thw_list: list[torch.Tensor] = [] per_axis_pos: list[tuple[int, int, int]] = [] merge_size = int(getattr(self.vit_model, "spatial_merge_size", 2)) merge_factor = merge_size * merge_size _curr = curr = 0 newlens: list[int] = [] new_rope: list[int] = [] for vi, (video, curr_kvlen, curr_position_id) in enumerate(zip(videos, curr_kvlens, curr_rope)): packed_key_value_indexes.extend(range(curr, curr + curr_kvlen)) curr += curr_kvlen packed_text_ids.append(new_token_ids["start_of_image"]) packed_text_indexes.append(_curr) packed_indexes.append(curr) curr += 1 _curr += 1 per_axis_pos.append((curr_position_id, curr_position_id, curr_position_id)) if precomputed_vit is not None: pixel_values, grid_thw = precomputed_vit[vi] if not torch.is_tensor(grid_thw): grid_thw = torch.as_tensor(grid_thw, dtype=torch.long) if grid_thw.dim() == 1: grid_thw = grid_thw.unsqueeze(0) else: proc_out = processor(videos=video, return_tensors="pt") # Qwen2-VL video processor: pixel_values_videos is packed # (num_patches_flat, patch_features) with a temporal axis # ``T_lat`` in video_grid_thw. pixel_values = proc_out["pixel_values_videos"] grid_thw = proc_out["video_grid_thw"] T_lat, H, W = (int(v) for v in grid_thw[0].tolist()) num_patches_pre_merge = T_lat * H * W assert num_patches_pre_merge == pixel_values.shape[0], ( f"pixel_values rows ({pixel_values.shape[0]}) != T_lat*H*W ({num_patches_pre_merge}); " f"video_grid_thw={grid_thw.tolist()}" ) num_vit_tokens = num_patches_pre_merge // merge_factor packed_vit_tokens.append(pixel_values) vit_token_seqlens.append(num_patches_pre_merge) packed_vit_position_ids.append(torch.arange(num_patches_pre_merge, dtype=torch.long)) grid_thw_list.append(grid_thw[0].to(torch.long)) packed_vit_token_indexes.extend(range(_curr, _curr + num_vit_tokens)) packed_indexes.extend(range(curr, curr + num_vit_tokens)) curr += num_vit_tokens _curr += num_vit_tokens # Upstream Lance ViT layout (verified against video_edit rope dump # for car @ video_480p — call2 t=(1000,1046), h,w=(64,110)): # start_of_image -> (P, P, P) # body (t,h,w) -> (P+1+t*ts, P+1+h, P+1+w) # end_of_image -> (P+max+2, P+max+2, P+max+2) # where max = max(ts*(T-1), h-1, w-1) # then the entire ViT block's t-channel is shifted to 1000+ via # ``shift_position_ids(pos_shift=1000)`` for modality=ref_vit. h_merged = H // merge_size w_merged = W // merge_size t_scale = LANCE_TOKENS_PER_SECOND * LANCE_SECONDS_PER_GRID for ti in range(T_lat): for hi in range(h_merged): for wi in range(w_merged): per_axis_pos.append( ( curr_position_id + 1 + int(ti * t_scale), curr_position_id + 1 + hi, curr_position_id + 1 + wi, ) ) packed_text_ids.append(new_token_ids["end_of_image"]) packed_text_indexes.append(_curr) packed_indexes.append(curr) curr += 1 _curr += 1 max_offset = max(int((T_lat - 1) * t_scale), h_merged - 1, w_merged - 1) end_p = curr_position_id + max_offset + 2 per_axis_pos.append((end_p, end_p, end_p)) packed_seqlens.append(num_vit_tokens + 2) newlens.append(curr_kvlen + num_vit_tokens + 2) new_rope.append(end_p + 1) pos_t = torch.tensor([p[0] for p in per_axis_pos], dtype=torch.long) pos_h = torch.tensor([p[1] for p in per_axis_pos], dtype=torch.long) pos_w = torch.tensor([p[2] for p in per_axis_pos], dtype=torch.long) # Apply upstream's ViT t-channel shift (modality=ref_vit=4 → +1000). # Without this, the ViT block's K/V at the t-channel collides with the # VAE_ref / noise block in mRoPE space (same low-100s range), and # noise-query attention picks up cross-modal leakage. pos_t = pos_t - pos_t[0] + 1000 packed_position_ids_3d = torch.stack([pos_t, pos_h, pos_w], dim=0) generation_input = { "packed_text_ids": torch.tensor(packed_text_ids, dtype=torch.long), "packed_text_indexes": torch.tensor(packed_text_indexes, dtype=torch.long), "vit_token_seqlens": torch.tensor(vit_token_seqlens, dtype=torch.int), "packed_vit_tokens": torch.cat(packed_vit_tokens, dim=0), "packed_vit_position_ids": torch.cat(packed_vit_position_ids, dim=0), "packed_vit_token_indexes": torch.tensor(packed_vit_token_indexes, dtype=torch.long), "packed_position_ids": packed_position_ids_3d, "packed_seqlens": torch.tensor(packed_seqlens, dtype=torch.int), "packed_indexes": torch.tensor(packed_indexes, dtype=torch.long), "packed_key_value_indexes": torch.tensor(packed_key_value_indexes, dtype=torch.long), "key_values_lens": torch.tensor(curr_kvlens, dtype=torch.int), "_lance_grid_thw": torch.stack(grid_thw_list, dim=0), } return generation_input, newlens, new_rope def prepare_vit_images(self, curr_kvlens, curr_rope, images, transforms, new_token_ids): # ``transforms`` is set up by ``BagelPipeline.forward`` to be # ``processor(images=img, return_tensors='pt').pixel_values[0]``, which # discards the ``image_grid_thw``. For Lance we need both, so we ignore # the supplied lambda and re-call the processor here directly. Lance's # pipeline always sets ``self.image_processor`` to ``Qwen2VLImageProcessor``. if not hasattr(self, "_lance_image_processor"): raise RuntimeError( "LanceBagel.prepare_vit_images requires the pipeline to set " "``bagel._lance_image_processor`` (a Qwen2-VL-compatible processor)." ) processor = self._lance_image_processor packed_vit_token_indexes: list[int] = [] vit_token_seqlens: list[int] = [] packed_vit_tokens: list[torch.Tensor] = [] packed_vit_position_ids: list[torch.Tensor] = [] packed_text_ids: list[int] = [] packed_text_indexes: list[int] = [] packed_seqlens: list[int] = [] packed_indexes: list[int] = [] packed_key_value_indexes: list[int] = [] grid_thw_list: list[torch.Tensor] = [] merge_size = int(getattr(self.vit_model, "spatial_merge_size", 2)) merge_factor = merge_size * merge_size _curr = curr = 0 newlens: list[int] = [] new_rope: list[int] = [] # 3-D position ids per query token (mRoPE) — text framing tokens use # scalar broadcast, image patches use per-token ``(P + t, P + h_merged, # P + w_merged)`` so the Qwen2.5-VL backbone sees the spatial layout # along the (h, w) axes of mRoPE. per_axis_pos: list[tuple[int, int, int]] = [] for image, curr_kvlen, curr_position_id in zip(images, curr_kvlens, curr_rope): packed_key_value_indexes.extend(range(curr, curr + curr_kvlen)) curr += curr_kvlen packed_text_ids.append(new_token_ids["start_of_image"]) packed_text_indexes.append(_curr) packed_indexes.append(curr) curr += 1 _curr += 1 per_axis_pos.append((curr_position_id, curr_position_id, curr_position_id)) pixel_values, grid_thw = self._qwen_vl_processor_call(processor, image) # pixel_values: (num_patches_flat, patch_features). grid_thw: (1, 3) with (T, H, W) in patch units. T, H, W = (int(v) for v in grid_thw[0].tolist()) num_patches_pre_merge = T * H * W assert num_patches_pre_merge == pixel_values.shape[0], ( f"pixel_values rows ({pixel_values.shape[0]}) != T*H*W ({num_patches_pre_merge}); " f"image_grid_thw={grid_thw.tolist()}" ) num_img_tokens = num_patches_pre_merge // merge_factor packed_vit_tokens.append(pixel_values) vit_token_seqlens.append(num_patches_pre_merge) packed_vit_position_ids.append(torch.arange(num_patches_pre_merge, dtype=torch.long)) grid_thw_list.append(grid_thw[0].to(torch.long)) packed_vit_token_indexes.extend(range(_curr, _curr + num_img_tokens)) packed_indexes.extend(range(curr, curr + num_img_tokens)) curr += num_img_tokens _curr += num_img_tokens # 3-D mRoPE positions for image patches. Upstream Lance # (``modeling/lance/lance.py::get_rope_index`` + ``shift_position_ids``) # places ViT body tokens at ``(P+1+t, P+1+h, P+1+w)`` and the # surrounding markers at ``(P, ...)`` (start) / ``(P+max+1, ...)`` # (end), then SHIFTS the t-channel of the entire ViT block to # 1000+ (modality=ref_vit=4) so ViT keys don't conflate with VAE # ref / noise keys in mRoPE space. ``new_rope = P + max + 2``. h_merged = H // merge_size w_merged = W // merge_size max_hw = max(h_merged, w_merged) for hi in range(h_merged): for wi in range(w_merged): per_axis_pos.append((curr_position_id + 1, curr_position_id + 1 + hi, curr_position_id + 1 + wi)) packed_text_ids.append(new_token_ids["end_of_image"]) packed_text_indexes.append(_curr) packed_indexes.append(curr) curr += 1 _curr += 1 end_p = curr_position_id + max_hw + 1 per_axis_pos.append((end_p, end_p, end_p)) packed_seqlens.append(num_img_tokens + 2) newlens.append(curr_kvlen + num_img_tokens + 2) new_rope.append(end_p + 1) pos_t = torch.tensor([p[0] for p in per_axis_pos], dtype=torch.long) pos_h = torch.tensor([p[1] for p in per_axis_pos], dtype=torch.long) pos_w = torch.tensor([p[2] for p in per_axis_pos], dtype=torch.long) # Upstream Lance shifts the ViT block's t-channel into a far-away rope # range (1000+) via ``shift_position_ids(pos_shift=1000)`` on # modality=ref_vit (4). This makes ViT keys/queries mRoPE-orthogonal # to noise/VAE-ref tokens along the t-channel, preventing leakage in # the time-frequency dimension. We apply the same shift here so the # ViT segment's K, V cache content matches upstream. pos_t = pos_t - pos_t[0] + 1000 packed_position_ids_3d = torch.stack([pos_t, pos_h, pos_w], dim=0) generation_input = { "packed_text_ids": torch.tensor(packed_text_ids, dtype=torch.long), "packed_text_indexes": torch.tensor(packed_text_indexes, dtype=torch.long), "vit_token_seqlens": torch.tensor(vit_token_seqlens, dtype=torch.int), "packed_vit_tokens": torch.cat(packed_vit_tokens, dim=0), "packed_vit_position_ids": torch.cat(packed_vit_position_ids, dim=0), "packed_vit_token_indexes": torch.tensor(packed_vit_token_indexes, dtype=torch.long), "packed_position_ids": packed_position_ids_3d, "packed_seqlens": torch.tensor(packed_seqlens, dtype=torch.int), "packed_indexes": torch.tensor(packed_indexes, dtype=torch.long), "packed_key_value_indexes": torch.tensor(packed_key_value_indexes, dtype=torch.long), "key_values_lens": torch.tensor(curr_kvlens, dtype=torch.int), # Lance-extra: per-image (T, H, W) so the ViT wrapper can pass the # true grid to the HF tower (non-square images would otherwise be # reconstructed incorrectly as ``sqrt(n)`` squares). "_lance_grid_thw": torch.stack(grid_thw_list, dim=0), } return generation_input, newlens, new_rope def forward_cache_update_vit(self, *args, **kwargs): # Pluck the Lance-specific extra and stash on the wrapper for the # upcoming ViT call. Everything else delegates to the parent so we # don't have to copy 60 lines of LLM-call boilerplate. grid_thw = kwargs.pop("_lance_grid_thw", None) if grid_thw is not None and hasattr(self.vit_model, "set_pending_grid_thw"): self.vit_model.set_pending_grid_thw(grid_thw) return super().forward_cache_update_vit(*args, **kwargs) def prepare_video_latent(self, curr_kvlens, curr_rope, video_shapes, new_token_ids): """3-D analogue of :meth:`prepare_vae_latent`. ``video_shapes`` is a list of ``(T, H, W)`` per request (RGB pixel space). We package one packed-init-noise tensor over ``T_lat × H_lat × W_lat`` latent tokens per video, plus 1-D indices into the 3-D position embedding table maintained by :class:`LancePositionEmbedding3D` (`bagel.latent_pos_embed`). Latent geometry: - spatial: ``H_lat = H // latent_downsample`` (``=16`` for Lance) - temporal: ``T_lat = (T - 1) // downsample_temporal + 1`` (``=4`` for Wan2.2) - channels: ``latent_channel = 48`` Position ids are flattened ``t * max_per_side² + h * max_per_side + w`` so they index directly into the ``(max_num_frames * max_per_side², hidden_size)`` table. """ downsample_t = int(getattr(self.config.vae_config, "downsample_temporal", 4)) max_per_side = int(self.max_latent_size) packed_text_ids, packed_text_indexes = [], [] packed_vae_position_ids, packed_vae_token_indexes, packed_init_noises = [], [], [] packed_seqlens, packed_indexes = [], [] packed_key_value_indexes = [] # 3-D position ids per token (mRoPE). ``per_axis_pos[i]`` is the list # of (t, h, w) for the i-th query token; we stack into ``(3, S)`` at the # end. Text framing tokens broadcast scalar to all axes; video latent # tokens carry per-token ``(P + t, P + h, P + w)``. per_axis_pos: list[tuple[int, int, int]] = [] query_curr = curr = 0 for (T, H, W), curr_kvlen, curr_position_id in zip(video_shapes, curr_kvlens, curr_rope): packed_key_value_indexes.extend(range(curr, curr + curr_kvlen)) curr += curr_kvlen packed_text_ids.append(new_token_ids["start_of_image"]) packed_text_indexes.append(query_curr) packed_indexes.append(curr) curr += 1 query_curr += 1 # Upstream Lance places gen-latent start_of_image at ``(P, P, P)`` # (NOT ``P-1`` and NOT shifted to ViT t-range), body at # ``(P+1+ti*t_scale, P+1+hi, P+1+wi)``, end at ``(P+max+1, ...)``. # Verified against upstream's video_edit rope dump for T_lat=12, # H_lat=35, W_lat=47 (call09 / cond noise iter 0). per_axis_pos.append((curr_position_id, curr_position_id, curr_position_id)) h = H // self.latent_downsample w = W // self.latent_downsample t = (T - 1) // downsample_t + 1 num_video_tokens = t * h * w # 1-D position ids into the 3-D table (frame-major, then row, then col). tt, hh, ww = torch.meshgrid(torch.arange(t), torch.arange(h), torch.arange(w), indexing="ij") tt_flat = tt.flatten().tolist() hh_flat = hh.flatten().tolist() ww_flat = ww.flatten().tolist() vae_position_ids = (tt * (max_per_side * max_per_side) + hh * max_per_side + ww).flatten() packed_vae_position_ids.append(vae_position_ids) packed_init_noises.append(torch.randn(num_video_tokens, self.latent_channel * self.latent_patch_size**2)) packed_vae_token_indexes.extend(range(query_curr, query_curr + num_video_tokens)) packed_indexes.extend(range(curr, curr + num_video_tokens)) curr += num_video_tokens query_curr += num_video_tokens # Per-token 3-D mRoPE positions, matching upstream Lance's # ``get_rope_index``: the temporal axis is amplified by # ``tokens_per_second * second_per_grid_t`` (see Qwen2.5-VL docs) # so neighbouring latent frames are well-separated in rope space. # ``second_per_grid_t = 1.0`` matches upstream's default; # ``tokens_per_second = 2`` is Lance's vit config. t_scale = LANCE_TOKENS_PER_SECOND * LANCE_SECONDS_PER_GRID for ti, hi, wi in zip(tt_flat, hh_flat, ww_flat): per_axis_pos.append( ( curr_position_id + 1 + int(ti * t_scale), curr_position_id + 1 + hi, curr_position_id + 1 + wi, ) ) packed_text_ids.append(new_token_ids["end_of_image"]) packed_text_indexes.append(query_curr) packed_indexes.append(curr) curr += 1 query_curr += 1 # end_of_image sits "past" the latent block at ``P+max+1`` on every # axis, where ``max = max(t_scale*(t-1), h-1, w-1) + 1``. Verified # against upstream's call09 final token at (159, 159, 159) for # T_lat=12 (t_scale=2 → max_t=22), H_lat=35 (max_h=34), W_lat=47 # (max_w=46), with P=111: end_p = 111 + 46 + 1 + 1 = 159. ✓ max_thw = max(int((t - 1) * t_scale), h - 1, w - 1) + 1 end_p = curr_position_id + max_thw + 1 per_axis_pos.append((end_p, end_p, end_p)) packed_seqlens.append(num_video_tokens + 2) # Stack into (3, total_query_tokens) for mRoPE. pos_t = torch.tensor([p[0] for p in per_axis_pos], dtype=torch.long) pos_h = torch.tensor([p[1] for p in per_axis_pos], dtype=torch.long) pos_w = torch.tensor([p[2] for p in per_axis_pos], dtype=torch.long) packed_position_ids_3d = torch.stack([pos_t, pos_h, pos_w], dim=0) generation_input = { "packed_text_ids": torch.tensor(packed_text_ids, dtype=torch.long), "packed_text_indexes": torch.tensor(packed_text_indexes, dtype=torch.long), "packed_init_noises": torch.cat(packed_init_noises, dim=0), "packed_vae_position_ids": torch.cat(packed_vae_position_ids, dim=0), "packed_vae_token_indexes": torch.tensor(packed_vae_token_indexes, dtype=torch.long), "packed_seqlens": torch.tensor(packed_seqlens, dtype=torch.int), "packed_position_ids": packed_position_ids_3d, "key_values_lens": torch.tensor(curr_kvlens, dtype=torch.int), "packed_indexes": torch.tensor(packed_indexes, dtype=torch.long), "packed_key_value_indexes": torch.tensor(packed_key_value_indexes, dtype=torch.long), } return generation_input def prepare_video_latent_cfg(self, curr_kvlens, curr_rope, video_shapes): """3-D analogue of :meth:`prepare_vae_latent_cfg` (CFG side). Mirrors :meth:`prepare_video_latent`'s mRoPE 3-D position layout EXACTLY, including the ``LANCE_TOKENS_PER_SECOND * LANCE_SECONDS_PER_GRID`` temporal scaling. Without that scaling, the cfg_text branch attends with different rope coordinates than the cond branch (and than upstream's ``get_rope_index``), which makes ``cfg_text_v_t`` diverge from upstream and the CFG combination amplifies the error every denoise step. """ downsample_t = int(getattr(self.config.vae_config, "downsample_temporal", 4)) packed_indexes, packed_key_value_indexes = [], [] per_axis_pos: list[tuple[int, int, int]] = [] query_curr = curr = 0 t_scale = LANCE_TOKENS_PER_SECOND * LANCE_SECONDS_PER_GRID for (T, H, W), curr_kvlen, curr_position_id in zip(video_shapes, curr_kvlens, curr_rope): packed_key_value_indexes.extend(range(curr, curr + curr_kvlen)) curr += curr_kvlen packed_indexes.append(curr) curr += 1 query_curr += 1 # Match :meth:`prepare_video_latent` convention exactly (see # comments there for the upstream rope layout this replicates). per_axis_pos.append((curr_position_id, curr_position_id, curr_position_id)) h = H // self.latent_downsample w = W // self.latent_downsample t = (T - 1) // downsample_t + 1 num_video_tokens = t * h * w packed_indexes.extend(range(curr, curr + num_video_tokens)) curr += num_video_tokens query_curr += num_video_tokens tt, hh, ww = torch.meshgrid(torch.arange(t), torch.arange(h), torch.arange(w), indexing="ij") for ti, hi, wi in zip(tt.flatten().tolist(), hh.flatten().tolist(), ww.flatten().tolist()): per_axis_pos.append( ( curr_position_id + 1 + int(ti * t_scale), curr_position_id + 1 + hi, curr_position_id + 1 + wi, ) ) packed_indexes.append(curr) curr += 1 query_curr += 1 max_thw = max(int((t - 1) * t_scale), h - 1, w - 1) + 1 end_p = curr_position_id + max_thw + 1 per_axis_pos.append((end_p, end_p, end_p)) pos_t = torch.tensor([p[0] for p in per_axis_pos], dtype=torch.long) pos_h = torch.tensor([p[1] for p in per_axis_pos], dtype=torch.long) pos_w = torch.tensor([p[2] for p in per_axis_pos], dtype=torch.long) cfg_packed_position_ids_3d = torch.stack([pos_t, pos_h, pos_w], dim=0) return { "cfg_packed_position_ids": cfg_packed_position_ids_3d, "cfg_packed_query_indexes": torch.tensor(packed_indexes, dtype=torch.long), "cfg_key_values_lens": torch.tensor(curr_kvlens, dtype=torch.int), "cfg_packed_key_value_indexes": torch.tensor(packed_key_value_indexes, dtype=torch.long), } # ------------------------------------------------------------------ # # Image-edit / video-edit VAE prefill # ------------------------------------------------------------------ # def _lance_native_prepare_vae_images( self, curr_kvlens, curr_rope, images, transforms, new_token_ids, timestep=0, is_video: bool = False, ): """Lance-native ``prepare_vae_images`` with 3-D mRoPE positions. Mirrors :meth:`Bagel.prepare_vae_images` but emits per-token mRoPE ``(curr + t, curr + h, curr + w)`` positions for VAE latent tokens (BAGEL emits scalar ``curr_position_id`` for all of them, which triggers a CUDA index-out-of-bounds gather on Lance's Qwen2.5-VL backbone). For still images ``t = 0``; for video clips ``t`` ranges over ``T_lat = (T - 1) // downsample_temporal + 1``. """ downsample_t = int(getattr(self.config.vae_config, "downsample_temporal", 4)) max_per_side = int(self.max_latent_size) packed_text_ids: list[int] = [] packed_text_indexes: list[int] = [] packed_vae_position_ids: list[torch.Tensor] = [] packed_vae_token_indexes: list[int] = [] packed_indexes: list[int] = [] packed_key_value_indexes: list[int] = [] packed_seqlens: list[int] = [] patchified_vae_latent_shapes: list[tuple] = [] vae_image_tensors: list[torch.Tensor] = [] per_axis_pos: list[tuple[int, int, int]] = [] query_curr = curr = 0 newlens: list[int] = [] new_rope: list[int] = [] for image, curr_kvlen, curr_position_id in zip(images, curr_kvlens, curr_rope): packed_key_value_indexes.extend(range(curr, curr + curr_kvlen)) curr += curr_kvlen packed_text_ids.append(new_token_ids["start_of_image"]) packed_text_indexes.append(query_curr) packed_indexes.append(curr) curr += 1 query_curr += 1 per_axis_pos.append((curr_position_id, curr_position_id, curr_position_id)) image_tensor = transforms(image) vae_image_tensors.append(image_tensor) # ``image_tensor`` is ``(C, H, W)`` for image input or ``(C, T, H, W)`` # for video input. We derive the *latent* grid from the RGB spatial # dims and our Wan2.2 VAE downsample factor. if image_tensor.dim() == 3: H, W = image_tensor.shape[1:] T = 1 elif image_tensor.dim() == 4: _, T, H, W = image_tensor.shape else: raise ValueError(f"vae transforms must return 3-D or 4-D tensor; got {image_tensor.shape}") h_lat = H // self.latent_downsample w_lat = W // self.latent_downsample t_lat = ((T - 1) // downsample_t + 1) if is_video else 1 patchified_vae_latent_shapes.append((t_lat, h_lat, w_lat) if is_video else (h_lat, w_lat)) # 1-D index into the (3-D when video) ``latent_pos_embed`` table. tt, hh, ww = torch.meshgrid(torch.arange(t_lat), torch.arange(h_lat), torch.arange(w_lat), indexing="ij") vae_pos_ids = (tt * (max_per_side * max_per_side) + hh * max_per_side + ww).flatten() packed_vae_position_ids.append(vae_pos_ids) num_image_tokens = t_lat * h_lat * w_lat packed_vae_token_indexes.extend(range(query_curr, query_curr + num_image_tokens)) packed_indexes.extend(range(curr, curr + num_image_tokens)) curr += num_image_tokens query_curr += num_image_tokens # Upstream Lance places VAE-ref body at ``(P+1+t*t_scale, P+1+h, P+1+w)`` # with the start marker at ``(P, P, P)`` and the end marker at # ``(P+max+1, ...)``; gen-latent tokens (noise) share the SAME # rope range via the modality_map==1/2 trick, so this is the # convention used by :meth:`_per_token_mrope_for_vae_latent` (2D) # and :meth:`_per_token_mrope_for_video_latent` (3D) below. # For video, ``t_scale = LANCE_TOKENS_PER_SECOND * LANCE_SECONDS_PER_GRID`` # matches upstream Qwen2.5-VL's temporal-axis scaling; for still # images we leave the t-channel scalar (T=1 → ``ti=0`` anyway). t_scale = LANCE_TOKENS_PER_SECOND * LANCE_SECONDS_PER_GRID if is_video else 1 max_thw = max(int((t_lat - 1) * t_scale), h_lat - 1, w_lat - 1) + 1 for ti, hi, wi in zip(tt.flatten().tolist(), hh.flatten().tolist(), ww.flatten().tolist()): per_axis_pos.append( ( curr_position_id + 1 + int(ti * t_scale), curr_position_id + 1 + hi, curr_position_id + 1 + wi, ) ) packed_text_ids.append(new_token_ids["end_of_image"]) packed_text_indexes.append(query_curr) packed_indexes.append(curr) curr += 1 query_curr += 1 end_p = curr_position_id + max_thw + 1 per_axis_pos.append((end_p, end_p, end_p)) packed_seqlens.append(num_image_tokens + 2) newlens.append(curr_kvlen + num_image_tokens + 2) new_rope.append(end_p + 1) image_sizes = [item.shape for item in vae_image_tensors] max_image_size = [max(item) for item in list(zip(*image_sizes))] padded_images = torch.zeros(size=(len(vae_image_tensors), *max_image_size)) for i, image_tensor in enumerate(vae_image_tensors): if image_tensor.dim() == 3: padded_images[i, :, : image_tensor.shape[1], : image_tensor.shape[2]] = image_tensor else: padded_images[i, :, : image_tensor.shape[1], : image_tensor.shape[2], : image_tensor.shape[3]] = ( image_tensor ) pos_t = torch.tensor([p[0] for p in per_axis_pos], dtype=torch.long) pos_h = torch.tensor([p[1] for p in per_axis_pos], dtype=torch.long) pos_w = torch.tensor([p[2] for p in per_axis_pos], dtype=torch.long) packed_position_ids_3d = torch.stack([pos_t, pos_h, pos_w], dim=0) generation_input = { "padded_images": padded_images, "patchified_vae_latent_shapes": patchified_vae_latent_shapes, "packed_vae_position_ids": torch.cat(packed_vae_position_ids, dim=0), "packed_timesteps": torch.tensor([timestep]), "packed_vae_token_indexes": torch.tensor(packed_vae_token_indexes, dtype=torch.long), "packed_text_ids": torch.tensor(packed_text_ids, dtype=torch.long), "packed_text_indexes": torch.tensor(packed_text_indexes, dtype=torch.long), "packed_position_ids": packed_position_ids_3d, "packed_seqlens": torch.tensor(packed_seqlens, dtype=torch.int), "packed_indexes": torch.tensor(packed_indexes, dtype=torch.long), "packed_key_value_indexes": torch.tensor(packed_key_value_indexes, dtype=torch.long), "key_values_lens": torch.tensor(curr_kvlens, dtype=torch.int), } return generation_input, newlens, new_rope def prepare_vae_images(self, curr_kvlens, curr_rope, images, transforms, new_token_ids, timestep=0): """VAE prefill router. - When ``images`` is non-empty (image_edit / video_edit path on the image side): delegate to :meth:`_lance_native_prepare_vae_images` which emits Lance's 3-D mRoPE positions and a real VAE prefill, letting BAGEL's parent ``image_edit`` flow handle the rest. - When ``images`` is empty (t2i / x2t paths): short-circuit with the no-op output, mirroring BAGEL's "no image to prefill" sentinel. """ if images: return self._lance_native_prepare_vae_images( curr_kvlens=curr_kvlens, curr_rope=curr_rope, images=images, transforms=transforms, new_token_ids=new_token_ids, timestep=timestep, is_video=False, ) generation_input = { "padded_images": torch.empty(0, 3, 0, 0), "patchified_vae_latent_shapes": [], "packed_vae_position_ids": torch.empty(0, dtype=torch.long), "packed_timesteps": torch.tensor([timestep]), "packed_vae_token_indexes": torch.empty(0, dtype=torch.long), "packed_text_ids": torch.empty(0, dtype=torch.long), "packed_text_indexes": torch.empty(0, dtype=torch.long), "packed_position_ids": torch.empty(0, dtype=torch.long), "packed_seqlens": torch.empty(0, dtype=torch.int), "packed_indexes": torch.empty(0, dtype=torch.long), "packed_key_value_indexes": torch.empty(0, dtype=torch.long), "key_values_lens": torch.tensor(curr_kvlens, dtype=torch.int), } return generation_input, list(curr_kvlens), list(curr_rope) def forward_cache_update_vae( self, vae_model, past_key_values, padded_images=None, patchified_vae_latent_shapes=None, packed_vae_position_ids=None, packed_timesteps=None, packed_vae_token_indexes=None, packed_text_ids=None, packed_text_indexes=None, packed_position_ids=None, packed_seqlens=None, packed_indexes=None, key_values_lens=None, packed_key_value_indexes=None, precomputed_latent=None, ): """Lance-native VAE prefill that *actually* scatters the encoded latents into the LLM query sequence. :meth:`Bagel.forward_cache_update_vae` in vllm-omni computes ``packed_latent = vae2llm(...) + time_embed + pos_embed`` and then passes only ``packed_text_ids`` to the LLM — the VAE embeddings never enter the query sequence (the LLM builds it from ``embed_tokens(packed_text_ids)`` which is just the 2 framing tokens). The mismatch between ``query_lens = [num_vae + 2]`` and the resulting 2-token sequence is what crashes the gather inside attention. We scatter both pieces explicitly: text framing tokens at ``packed_text_indexes`` and the VAE latent embeddings at ``packed_vae_token_indexes``, producing a full-length ``(sum(packed_seqlens), hidden)`` sequence the LLM can attend over. Empty prep data (legacy x2t / x2t_video no-op path) is short-circuited. """ if ( packed_text_ids is None or packed_text_ids.numel() == 0 or (precomputed_latent is None and (padded_images is None or padded_images.numel() == 0)) ): return past_key_values # ``vae_model.encode`` samples from the posterior via # ``mu + std * randn_like(std)`` (Wan2.2 VAE's ``reparameterize`` path # — both upstream and vllm-omni default to ``use_sample=True``). Each # call therefore returns a DIFFERENT latent. Upstream's image_edit # encodes the reference image ONCE and reuses the resulting latent # for both the cond and cfg_text branches; calling encode separately # per branch made vllm-omni's gen vs cfg VAE-ref K cache diverge by # ~6% rel_l2, which cascaded into ~33% v_t divergence at iter 0 on # the cond branch (cfg matched within bf16 noise because it sees the # SAME ref VAE K in both implementations — only gen-vs-cfg drift # matters for the qualitative edit, since the noise query attends to # both branches' caches at every iter). The pipeline now encodes # the ref once and threads the encoded latent into both prefill # calls via ``precomputed_latent``. if precomputed_latent is not None: padded_latent = precomputed_latent else: padded_latent = vae_model.encode(padded_images) p = self.latent_patch_size packed_latent_list = [] for latent, shape in zip(padded_latent, patchified_vae_latent_shapes): if isinstance(shape, tuple) and len(shape) == 3: t_lat, h_lat, w_lat = shape latent = latent[:, :t_lat, : h_lat * p, : w_lat * p].reshape( self.latent_channel, t_lat, h_lat, p, w_lat, p ) latent = torch.einsum("cthpwq->thwpqc", latent).reshape(-1, p * p * self.latent_channel) else: h_lat, w_lat = shape latent = latent[:, : h_lat * p, : w_lat * p].reshape(self.latent_channel, h_lat, p, w_lat, p) latent = torch.einsum("chpwq->hwpqc", latent).reshape(-1, p * p * self.latent_channel) packed_latent_list.append(latent) packed_latent = torch.cat(packed_latent_list, dim=0) packed_pos_embed = self.latent_pos_embed(packed_vae_position_ids) packed_timestep_embeds = self.time_embedder(packed_timesteps) packed_latent = self.vae2llm(packed_latent) + packed_timestep_embeds + packed_pos_embed # ---- Scatter text + VAE embeds into the full query sequence ---- # total_len = int(packed_seqlens.sum().item()) packed_text_embed = self.language_model.model.embed_tokens(packed_text_ids) packed_query_sequence = packed_latent.new_zeros((total_len, packed_latent.shape[-1])) packed_query_sequence[packed_text_indexes] = packed_text_embed.to(packed_query_sequence.dtype) packed_query_sequence[packed_vae_token_indexes] = packed_latent.to(packed_query_sequence.dtype) # Upstream Lance routes the VAE-ref block through MoE_GEN because # ``packed_gen_token_indexes`` in ``Lance.validation_gen_KVcache`` # is set to ``current_vae_token_indexes_local`` — i.e., it includes # BOTH ref VAE tokens (modality=2 ref_source) AND noise tokens # (modality=1). For the VAE-ref segment forward, the mask # ``packed_gen_token_indexes ∈ [current_cond_start, current_cond_end)`` # filters in the VAE-ref body tokens → ``mode_ = "gen"``. extra_inputs = {} if self.use_moe: extra_inputs = { "mode": "gen", "packed_vae_token_indexes": packed_vae_token_indexes, "packed_text_indexes": packed_text_indexes, } # ``Qwen2MoTForCausalLM.forward`` post-main-merge no longer accepts # ``packed_query_indexes`` / ``key_values_lens`` / # ``packed_key_value_indexes`` — derived internally now. output = self.language_model.forward( packed_query_sequence=packed_query_sequence, query_lens=packed_seqlens, packed_query_position_ids=packed_position_ids, past_key_values=past_key_values, update_past_key_values=True, is_causal=False, **extra_inputs, ) return output.past_key_values