`vllm.model_executor.models.isaac` ¶

Classes:

IsaacForConditionalGeneration –
IsaacImagePixelInputs –

Schema for validating Isaac image inputs.
Siglip2VisionTransformer –

Functions:

create_cumulative_seq_lengths –

Create cumulative sequence lengths for variable-length attention.
create_pixel_shuffle_index_map –

Build a gather-index map that tells us, for every output token after
pixel_shuffle_varlen –

Apply pixel shuffle to a packed vision sequence without unpacking per image.

`IsaacForConditionalGeneration` ¶

Bases: Module, SupportsMultiModal, SupportsLoRA, SupportsPP, SupportsMRoPE

Methods:

get_mm_mapping –

Get the module prefix in multimodal models

Source code in vllm/model_executor/models/isaac.py

@MULTIMODAL_REGISTRY.register_processor(
    IsaacMultiModalProcessor,
    info=IsaacProcessingInfo,
    dummy_inputs=IsaacDummyInputsBuilder,
)
class IsaacForConditionalGeneration(
    nn.Module, SupportsMultiModal, SupportsLoRA, SupportsPP, SupportsMRoPE
):
    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
        "gate_up_proj": [
            "gate_proj",
            "up_proj",
        ],
    }

    supports_encoder_tp_data = True

    # To ensure correct weight loading and mapping.
    hf_to_vllm_mapper = WeightsMapper(
        orig_to_new_prefix={
            "lm_head.": "language_model.lm_head.",
            "model.text_model.lm_head.": "language_model.lm_head.",
            "model.text_model.": "language_model.model.",
            "model.vision_embedding.0": "vision_embedding.transformer",
            "model.vision_embedding.1": "vision_embedding.linear_fc1",
            "model.vision_embedding.2": "vision_embedding.act",
            "model.vision_embedding.3": "vision_embedding.linear_fc2",
            "model.vision_embedding.": "vision_embedding.",
            "model.lm_head.": "language_model.lm_head.",
            "model.": "language_model.model.",
        }
    )

    @classmethod
    def get_placeholder_str(cls, modality: str, i: int) -> str | None:
        if modality.startswith("image"):
            return "<image>"

        raise ValueError("Only image modality is supported")

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = "model"):
        super().__init__()
        config: IsaacConfig = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
        self.config = config

        head_dim = config.head_dim
        calculated_mrope_section = [
            head_dim // 4,  # 2x more for temporal dim
            head_dim // 8,
            head_dim // 8,
        ]

        self.vision_token_id = _resolve_vision_token_id(
            vllm_config.model_config, config.vision_token
        )
        config.image_token_id = self.vision_token_id

        text_cfg = getattr(config, "text_config", None)
        target_cfg = (
            text_cfg
            if text_cfg is not None and not isinstance(text_cfg, dict)
            else config
        )

        rope_scaling = getattr(target_cfg, "rope_scaling", None)
        if rope_scaling is None and target_cfg is config:
            rope_scaling = getattr(config, "_rope_scaling", None)

        patch_rope_parameters(target_cfg)
        rope_parameters = target_cfg.rope_parameters
        rope_parameters["mrope_section"] = calculated_mrope_section
        if rope_scaling is not None and "mrope_interleaved" in rope_scaling:
            rope_parameters.setdefault(
                "mrope_interleaved", rope_scaling["mrope_interleaved"]
            )
        target_cfg.rope_parameters = rope_parameters

        with self._mark_language_model(vllm_config):
            self.language_model = init_vllm_registered_model(
                vllm_config=vllm_config,
                architectures=["Qwen3ForCausalLM"],
                prefix=maybe_prefix(prefix, "language_model"),
            )

        self.make_empty_intermediate_tensors = (
            self.language_model.make_empty_intermediate_tensors
        )

        vision_cfg = config.vision_config
        if vision_cfg is None:
            raise ValueError("IsaacConfig should always have vision_config")
        attn_impl = (
            config.vision_attn_implementation
            if config.vision_attn_implementation is not None
            else getattr(config, "_attn_implementation", None)
        )
        if attn_impl is not None:
            vision_cfg._attn_implementation = attn_impl

        hidden_dim = vision_cfg.hidden_size * (vision_cfg.pixel_shuffle_scale_factor**2)

        with self._mark_tower_model(vllm_config, "image"):
            self.vision_embedding = IsaacVisionEmbedding(
                vision_cfg=vision_cfg,
                hidden_dim=hidden_dim,
                output_dim=config.hidden_size,
                quant_config=quant_config,
                prefix=maybe_prefix(prefix, "vision_embedding"),
            )

    def iter_mm_grid_hw(
        self, input_tokens: list[int], mm_features: list[MultiModalFeatureSpec]
    ) -> Iterator[tuple[int, int, int]]:
        spatial_merge_size = self.config.vision_config.pixel_shuffle_scale_factor
        for mm_feature in sorted(mm_features, key=lambda f: f.mm_position.offset):
            offset = mm_feature.mm_position.offset
            if mm_feature.modality == "image":
                t, h, w = mm_feature.data["image_grid_thw"].data.tolist()
                assert t == 1, f"Image must have 1 frame, got {t}"
                yield offset, h // spatial_merge_size, w // spatial_merge_size
            else:
                raise ValueError(f"Unsupported modality: {mm_feature.modality}")

    def get_mrope_input_positions(
        self,
        input_tokens: list[int],
        mm_features: list[MultiModalFeatureSpec],
    ) -> tuple[torch.Tensor, int]:
        llm_pos_ids_list = []
        st = 0
        for offset, llm_grid_h, llm_grid_w in self.iter_mm_grid_hw(
            input_tokens, mm_features
        ):
            text_len = offset - st
            st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
            llm_pos_ids_list.append(
                np.broadcast_to(np.arange(text_len), (3, text_len)) + st_idx
            )

            grid_indices = np.indices((1, llm_grid_h, llm_grid_w)).reshape(3, -1)
            grid_indices[0, :] = grid_indices[0, :] + text_len + st_idx
            llm_pos_ids_list.append(grid_indices)
            st = offset + llm_grid_h * llm_grid_w

        if st < len(input_tokens):
            st_idx = llm_pos_ids_list[-1][0, -1] + 1 if len(llm_pos_ids_list) > 0 else 0
            text_len = len(input_tokens) - st
            llm_pos_ids_list.append(
                np.broadcast_to(np.arange(text_len), (3, text_len)) + st_idx
            )

        llm_positions = np.concatenate(llm_pos_ids_list, axis=1).reshape(3, -1)
        mrope_position_delta = (llm_positions.max() + 1 - len(input_tokens)).item()

        return torch.from_numpy(llm_positions), mrope_position_delta

    def _parse_and_validate_image_input(
        self, **kwargs: object
    ) -> IsaacImagePixelInputs | None:
        pixel_values = kwargs.get("pixel_values")
        image_grid_thw = kwargs.get("image_grid_thw")
        if pixel_values is None or image_grid_thw is None:
            return None

        # TensorSchema will automatically validate shapes on initialization
        return IsaacImagePixelInputs(
            pixel_values=pixel_values,
            image_grid_thw=image_grid_thw,
        )

    def _process_image_input(
        self,
        image_input: IsaacImagePixelInputs,
    ) -> tuple[torch.Tensor, ...]:
        pixel_values = image_input["pixel_values"]
        image_grid_thw = image_input["image_grid_thw"]
        if pixel_values.numel() == 0:
            return ()

        device = next(self.language_model.parameters()).device
        dtype = self.vision_embedding.linear_fc1.weight.dtype
        pixel_values = pixel_values.to(device=device, dtype=dtype)
        spatial_grids = image_grid_thw[:, 1:3].to(device, dtype=torch.int32)

        vision_embeddings = self.vision_embedding((pixel_values, spatial_grids))
        merge_size = self.config.vision_config.pixel_shuffle_scale_factor
        sizes = spatial_grids.prod(-1) // (merge_size * merge_size)
        return tuple(vision_embeddings.split(sizes.tolist()))

    def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings | None:
        image_input = self._parse_and_validate_image_input(**kwargs)
        if image_input is None:
            return ()
        return self._process_image_input(image_input)

    def forward(
        self,
        input_ids: torch.Tensor | None,
        positions: torch.Tensor,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
        **kwargs: object,
    ) -> torch.Tensor | IntermediateTensors:
        return self.language_model(
            input_ids=input_ids,
            positions=positions,
            intermediate_tensors=intermediate_tensors,
            inputs_embeds=inputs_embeds,
            **kwargs,
        )

    def compute_logits(self, hidden_states: torch.Tensor) -> torch.Tensor | None:
        return self.language_model.compute_logits(hidden_states)

    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
        loader = AutoWeightsLoader(self)
        return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)

    def get_mm_mapping(self) -> MultiModelKeys:
        """
        Get the module prefix in multimodal models
        """
        return MultiModelKeys.from_string_field(
            language_model="language_model",
            connector="vision_embedding.linear_fc2",  # The final linear layer
            tower_model="vision_embedding",
        )

`get_mm_mapping()` ¶

Get the module prefix in multimodal models

Source code in vllm/model_executor/models/isaac.py

def get_mm_mapping(self) -> MultiModelKeys:
    """
    Get the module prefix in multimodal models
    """
    return MultiModelKeys.from_string_field(
        language_model="language_model",
        connector="vision_embedding.linear_fc2",  # The final linear layer
        tower_model="vision_embedding",
    )

`IsaacImagePixelInputs` ¶

Bases: TensorSchema

Schema for validating Isaac image inputs.

Dimensions

np: Number of patches
d: Patch dimension
ni: Number of images

The schema enforces

pixel_values must be 2D: (num_patches, patch_dim)
image_grid_thw must be 2D: (num_images, 3) where 3 represents [T, H, W]

Source code in vllm/model_executor/models/isaac.py

class IsaacImagePixelInputs(TensorSchema):
    """
    Schema for validating Isaac image inputs.

    Dimensions:
        - np: Number of patches
        - d: Patch dimension
        - ni: Number of images

    The schema enforces:
        - pixel_values must be 2D: (num_patches, patch_dim)
        - image_grid_thw must be 2D: (num_images, 3)
          where 3 represents [T, H, W]
    """

    pixel_values: Annotated[
        torch.Tensor,
        TensorShape("np", "d"),
    ]

    image_grid_thw: Annotated[
        torch.Tensor,
        TensorShape("ni", 3),
    ]

`Siglip2VisionTransformer` ¶

Bases: Module

Methods:

forward –

spatial_shapes (torch.LongTensor of shape (batch_size, 2)):

Source code in vllm/model_executor/models/isaac.py

class Siglip2VisionTransformer(nn.Module):
    hf_to_vllm_mapper = WeightsMapper(
        orig_to_new_stacked={
            ".q_proj": (".qkv_proj", "q"),
            ".k_proj": (".qkv_proj", "k"),
            ".v_proj": (".qkv_proj", "v"),
        }
    )

    def __init__(
        self,
        config: PixelShuffleSiglip2VisionConfig,
        quant_config: QuantizationConfig | None = None,
        prefix: str = "",
    ):
        super().__init__()
        self.config = config
        self.quant_config = quant_config
        embed_dim = config.hidden_size

        self.embeddings = Siglip2VariableSequenceEmbeddings(config)
        self.pixel_shuffle_scale_factor = config.pixel_shuffle_scale_factor
        self.encoder = Siglip2Encoder(
            config,
            quant_config=quant_config,
            prefix=f"{prefix}.encoder",
        )
        self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)

    def forward(
        self,
        packed_seq_patches: tuple[torch.Tensor, torch.Tensor],
    ) -> torch.Tensor:
        r"""
        spatial_shapes (`torch.LongTensor` of shape `(batch_size, 2)`):
            Tensor containing the spatial dimensions (height, width)
            of the input images.
        """

        seq_patches, token_grids = packed_seq_patches
        seq_sizes = torch.prod(token_grids, dim=-1)

        # Get embeddings from packed sequence
        hidden_states = self.embeddings((seq_patches, seq_sizes, token_grids))

        # Add a pseudo batch dimension for the encoder
        hidden_states = hidden_states.unsqueeze(0)

        cu_seqlens, max_seqlen = create_cumulative_seq_lengths(
            seq_sizes, hidden_states.device
        )

        hidden_states = self.encoder(
            inputs_embeds=hidden_states,
            cu_seqlens=cu_seqlens,
            max_seqlen=max_seqlen,
        )
        hidden_states = self.post_layernorm(hidden_states)

        if self.pixel_shuffle_scale_factor > 1:
            hidden_states = pixel_shuffle_varlen(
                x=hidden_states,
                token_grids=token_grids,
                scale_factor=self.pixel_shuffle_scale_factor,
            )
        # Remove the pseudo batch dimension we added earlier
        hidden_states = hidden_states.squeeze(0)

        # return last_hidden_state
        return hidden_states

    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
        loader = AutoWeightsLoader(self)
        return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)

`forward(packed_seq_patches)` ¶

spatial_shapes (torch.LongTensor of shape (batch_size, 2)): Tensor containing the spatial dimensions (height, width) of the input images.

Source code in vllm/model_executor/models/isaac.py

def forward(
    self,
    packed_seq_patches: tuple[torch.Tensor, torch.Tensor],
) -> torch.Tensor:
    r"""
    spatial_shapes (`torch.LongTensor` of shape `(batch_size, 2)`):
        Tensor containing the spatial dimensions (height, width)
        of the input images.
    """

    seq_patches, token_grids = packed_seq_patches
    seq_sizes = torch.prod(token_grids, dim=-1)

    # Get embeddings from packed sequence
    hidden_states = self.embeddings((seq_patches, seq_sizes, token_grids))

    # Add a pseudo batch dimension for the encoder
    hidden_states = hidden_states.unsqueeze(0)

    cu_seqlens, max_seqlen = create_cumulative_seq_lengths(
        seq_sizes, hidden_states.device
    )

    hidden_states = self.encoder(
        inputs_embeds=hidden_states,
        cu_seqlens=cu_seqlens,
        max_seqlen=max_seqlen,
    )
    hidden_states = self.post_layernorm(hidden_states)

    if self.pixel_shuffle_scale_factor > 1:
        hidden_states = pixel_shuffle_varlen(
            x=hidden_states,
            token_grids=token_grids,
            scale_factor=self.pixel_shuffle_scale_factor,
        )
    # Remove the pseudo batch dimension we added earlier
    hidden_states = hidden_states.squeeze(0)

    # return last_hidden_state
    return hidden_states

`create_cumulative_seq_lengths(seq_sizes, device)` ¶

Create cumulative sequence lengths for variable-length attention.

Source code in vllm/model_executor/models/isaac.py

def create_cumulative_seq_lengths(
    seq_sizes: torch.Tensor, device: torch.device
) -> tuple[torch.Tensor, torch.Tensor]:
    """Create cumulative sequence lengths for variable-length attention."""
    cu_seqlens = torch.zeros(len(seq_sizes) + 1, dtype=torch.int32, device=device)
    cu_seqlens[1:] = seq_sizes.cumsum(0)
    max_seqlen = (
        seq_sizes.max()
        if len(seq_sizes) > 0
        else torch.tensor(0, dtype=torch.int32, device=device)
    )
    return cu_seqlens, max_seqlen

`create_pixel_shuffle_index_map(seq_sizes, token_grids, scale_factor=1, device=None)` ¶

Build a gather-index map that tells us, for every output token after pixel-shuffle, which scale_factor**2 input tokens are being merged.

Args¶

seq_sizes : (num_images,) - #patches in each image (row-major order) token_grids : (num_images,2) - (height, width) for every image scale_factor : spatial down-scale factor (≥2) device : (optional) overrides seq_sizes.device

Returns¶

gather_idx : (new_total_seq_len, scale_factor2) int64 tensor. gather_idx[i, j] is the flat index into the original packed sequence for the j-th sub-patch that forms the i-th output token.

Source code in vllm/model_executor/models/isaac.py

def create_pixel_shuffle_index_map(
    seq_sizes: torch.Tensor,
    token_grids: torch.Tensor,
    scale_factor: int = 1,
    device: torch.device | None = None,
) -> torch.Tensor:
    """
    Build a gather-index map that tells us, for every *output* token after
    pixel-shuffle, which `scale_factor**2` *input* tokens are being merged.

    Args
    ----
    seq_sizes     : (num_images,)  - #patches in each image (row-major order)
    token_grids   : (num_images,2) - (height, width) for every image
    scale_factor  : spatial down-scale factor (≥2)
    device        : (optional) overrides `seq_sizes.device`

    Returns
    -------
    gather_idx : (new_total_seq_len, scale_factor**2) int64 tensor.
                 gather_idx[i, j] is the *flat* index into the *original*
                 packed sequence for the j-th sub-patch that forms the
                 i-th output token.
    """
    if device is None:
        device = seq_sizes.device

    r = int(scale_factor)
    if r < 2:
        raise ValueError("`scale_factor` must be ≥ 2")

    # Safety: all spatial dims must be divisible by r
    # Cannot run under torch compile fullgraph mode hence
    if not torch.compiler.is_compiling() and not (
        (token_grids[:, 0] % r == 0).all() and (token_grids[:, 1] % r == 0).all()
    ):
        raise AssertionError(
            "Every (H,W) in `token_grids` must be divisible by "
            f"scale_factor={r}, got {token_grids.tolist()}"
        )

    gather_chunks: list[torch.Tensor] = []
    tok_offset = 0

    for seq_len, (h, w) in zip(seq_sizes.tolist(), token_grids.tolist(), strict=False):
        # Build the (H, W) grid of flat indices for this image
        grid = torch.arange(seq_len, device=device, dtype=torch.int64) + tok_offset
        grid = grid.view(h, w)  # (H, W)

        # -------- identical ordering to your fixed-res routine --------
        # Step 1: split width into blocks of r
        grid = grid.view(h, w // r, r)  # (H, W/r, r)
        # Step 2: now split height into blocks of r
        grid = grid.view(h // r, r, w // r, r)  # (H/r, r, W/r, r)
        # Step 3: final permutation to (H/r, W/r, r, r)
        grid = grid.permute(0, 2, 1, 3).contiguous()  # (H/r, W/r, r, r)
        # Step 4: each (r, r) block forms one output token
        gather_chunks.append(grid.reshape(-1, r * r))  # (H*W / r², r²)

        tok_offset += seq_len

    # Concatenate over all images in the packed batch
    gather_idx = torch.cat(gather_chunks, dim=0)  # (Σ_i HᵢWᵢ/r², r²)
    return gather_idx

`pixel_shuffle_varlen(x, token_grids, scale_factor=1)` ¶

Apply pixel shuffle to a packed vision sequence without unpacking per image.

Parameters:

x ¶
(`torch.Tensor`) –

Concatenated vision embeddings. Accepts (seq_len, hidden_size) or (1, seq_len, hidden_size) shapes produced by stacking image patches.
token_grids ¶
(`torch.Tensor`) –

Integer tensor of shape (num_images, 2) whose rows give the (height, width) patch grid sizes corresponding to each image segment inside x.
scale_factor ¶
(`int`, *optional*, defaults to 1, default: 1 ) –

Spatial down-sampling factor specific to pixel shuffle. Values greater than one merge scale_factor**2 neighboring patches into a single embedding channel-group.

Returns:

Tensor –

torch.Tensor: Pixel-shuffled embeddings with shape matching the input
convention ( Tensor ) –

(seq_len, hidden_size * scale_factor**2) when the input
Tensor –

was 2D, or (1, seq_len, hidden_size * scale_factor**2) if the
Tensor –

singleton batch dimension was present.

Raises:

ValueError –

If more than one batch item is provided.

Source code in vllm/model_executor/models/isaac.py

def pixel_shuffle_varlen(
    x: torch.Tensor,
    token_grids: torch.Tensor,
    scale_factor: int = 1,
) -> torch.Tensor:
    r"""Apply pixel shuffle to a packed vision sequence without unpacking per image.

    Args:
        x (`torch.Tensor`):
            Concatenated vision embeddings. Accepts `(seq_len, hidden_size)` or
            `(1, seq_len, hidden_size)` shapes produced by stacking image
            patches.
        token_grids (`torch.Tensor`):
            Integer tensor of shape `(num_images, 2)` whose rows give the
            `(height, width)` patch grid sizes corresponding to each image
            segment inside `x`.
        scale_factor (`int`, *optional*, defaults to 1):
            Spatial down-sampling factor specific to pixel shuffle. Values
            greater than one merge `scale_factor**2` neighboring patches into a
            single embedding channel-group.

    Returns:
        `torch.Tensor`: Pixel-shuffled embeddings with shape matching the input
        convention: `(seq_len, hidden_size * scale_factor**2)` when the input
        was 2D, or `(1, seq_len, hidden_size * scale_factor**2)` if the
        singleton batch dimension was present.

    Raises:
        ValueError: If more than one batch item is provided.
    """
    keep_batch_dim = x.dim() == 3
    if keep_batch_dim:
        if x.size(0) != 1:
            raise AssertionError("Packed sequence is expected to have batch_size == 1")
        x_ = x.squeeze(0)  # (seq, embed)
    else:
        x_ = x  # (seq, embed)

    embed_dim = x_.size(-1)
    r = int(scale_factor)

    # Calculate seq_sizes from token_grids
    seq_sizes = torch.prod(token_grids, dim=-1)

    # Build index map and gather in one go
    gather_idx = create_pixel_shuffle_index_map(
        seq_sizes=seq_sizes,
        token_grids=token_grids,
        scale_factor=r,
        device=x_.device,
    )  # (new_seq, r²)

    # Gather → (new_seq, r², embed_dim)
    gathered = x_[gather_idx]  # fancy indexing keeps gradient

    # Merge the r² group dimension into channels to finish the shuffle
    out = gathered.reshape(gathered.size(0), embed_dim * r * r)

    # Restore batch dimension if needed
    if keep_batch_dim:
        out = out.unsqueeze(0)
    return out

`vllm.model_executor.models.isaac` ¶

`IsaacForConditionalGeneration` ¶

`get_mm_mapping()` ¶

`IsaacImagePixelInputs` ¶

`Siglip2VisionTransformer` ¶

`forward(packed_seq_patches)` ¶

`create_cumulative_seq_lengths(seq_sizes, device)` ¶

`create_pixel_shuffle_index_map(seq_sizes, token_grids, scale_factor=1, device=None)` ¶

Args¶

Returns¶

`pixel_shuffle_varlen(x, token_grids, scale_factor=1)` ¶

`x` ¶

`token_grids` ¶

`scale_factor` ¶

vllm.model_executor.models.isaac ¶

IsaacForConditionalGeneration ¶

get_mm_mapping() ¶

IsaacImagePixelInputs ¶

Siglip2VisionTransformer ¶

forward(packed_seq_patches) ¶

create_cumulative_seq_lengths(seq_sizes, device) ¶

create_pixel_shuffle_index_map(seq_sizes, token_grids, scale_factor=1, device=None) ¶

Args¶

Returns¶

pixel_shuffle_varlen(x, token_grids, scale_factor=1) ¶

x ¶

token_grids ¶

scale_factor ¶

`vllm.model_executor.models.isaac` ¶

`IsaacForConditionalGeneration` ¶

`get_mm_mapping()` ¶

`IsaacImagePixelInputs` ¶

`Siglip2VisionTransformer` ¶

`forward(packed_seq_patches)` ¶

`create_cumulative_seq_lengths(seq_sizes, device)` ¶

`create_pixel_shuffle_index_map(seq_sizes, token_grids, scale_factor=1, device=None)` ¶

`pixel_shuffle_varlen(x, token_grids, scale_factor=1)` ¶

`x` ¶

`token_grids` ¶

`scale_factor` ¶