vllm.model_executor.models.isaac ¶

@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",
        )

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),
    ]

class IsaacImageProcessor:
    patch_size = 16
    max_num_patches = 6144
    min_num_patches = 256
    pixel_shuffle_scale = 2

    valid_kwargs = IsaacImageProcessorKwargs
    model_input_names = ["pixel_values", "image_grid_thw"]

    def __init__(self, kwargs):
        self.patch_size = kwargs.pop("patch_size", self.patch_size)
        self.vision_max_num_patches = kwargs.pop(
            "vision_max_num_patches", self.max_num_patches
        )
        self.vision_min_num_patches = kwargs.pop(
            "vision_min_num_patches", self.min_num_patches
        )
        self.pixel_shuffle_scale = kwargs.pop("pixel_shuffle_scale", 2)

    def preprocess(
        self,
        images: list[torch.Tensor],
        return_tensors: str | TensorType | None,
        **kwargs: Unpack[IsaacImageProcessorKwargs],
    ) -> BatchFeature:
        """Preprocess images into format compatibile with vLLM input processing."""

        all_pixel_values: list[torch.Tensor] = []
        all_image_grids: list[torch.Tensor] = []

        for image in images:
            image_tensor = extract_image_pil(image)

            patches, dims_virtual = process_vision_for_patches(
                image_tensor,
                patch_size=self.patch_size,
                max_num_patches=self.vision_max_num_patches,
                min_num_patches=self.vision_min_num_patches,
                pixel_shuffle_scale=self.pixel_shuffle_scale,
            )

            # Isaac packs a dummy temporal dim for images
            patches = patches.unsqueeze(1)  # [N, T=1, Hp, Wp, D]

            hp, wp, dim = patches.shape[-3], patches.shape[-2], patches.shape[-1]
            current_num_patches = hp * wp
            pixel_values = patches.reshape(current_num_patches, dim)  # [N_tokens, D]

            # Use real patch dimensions for image_grid_thw, not virtual dimensions
            # This ensures the vision model receives correct grid info for pixel shuffle
            dims_real = [1, hp, wp]  # Real patch dimensions
            image_grid_thw = torch.tensor(dims_real).unsqueeze(0)

            all_pixel_values.append(pixel_values)
            all_image_grids.append(image_grid_thw)

        if all_pixel_values:
            final_pixel_values = torch.cat(all_pixel_values, dim=0)
            final_image_grids = torch.cat(all_image_grids, dim=0)
        else:
            final_pixel_values = torch.empty(0, 0)
            final_image_grids = torch.empty(0, 3)

        return BatchFeature(
            data={
                "pixel_values": final_pixel_values,
                "image_grid_thw": final_image_grids,
            },
            tensor_type=return_tensors,
        )

class IsaacProcessor:
    """Processor wrapper (tokenizer + IsaacImageProcessor)."""

    def __init__(self, image_processor=None, tokenizer=None, **kwargs):
        self.image_token = kwargs.pop("image_token", "<image>")
        self.image_processor = image_processor or IsaacImageProcessor(kwargs)
        self.tokenizer = tokenizer

    def __call__(self, text=None, images=None, **kwargs) -> BatchFeature:
        result = {}

        if images is not None:
            image_inputs = self.image_processor.preprocess(images, **kwargs)
            image_grid_thw = image_inputs["image_grid_thw"]
            result.update(image_inputs)

            if text is not None:
                if not isinstance(text, list):
                    text = [text]

                text = text.copy()  # below lines change text in-place
                merge_length = self.image_processor.pixel_shuffle_scale**2
                index = 0
                for i in range(len(text)):
                    while self.image_token in text[i]:
                        num_image_tokens = image_grid_thw[index].prod() // merge_length
                        text[i] = text[i].replace(
                            self.image_token, "<|placeholder|>" * num_image_tokens, 1
                        )
                        index += 1
                    text[i] = text[i].replace("<|placeholder|>", "<|image_pad|>")

        if text is not None:
            result.update(self.tokenizer(text, **kwargs))

        return BatchFeature(result)

    def apply_chat_template(
        self,
        messages: list[dict[str, Any]],
        tokenize: bool = False,
        add_generation_prompt: bool = False,
        **kwargs,
    ) -> Any:
        # Convert mixed content messages to simple text format
        processed_messages = []

        for message in messages:
            if "content" in message and isinstance(message["content"], list):
                # Handle mixed content (text + image)
                text_parts = []
                for content_item in message["content"]:
                    if content_item.get("type") == "text":
                        text_parts.append(content_item.get("text", ""))
                    elif content_item.get("type") == "image":
                        # Replace image with vision token
                        text_parts.append(self.image_token)

                processed_message = {
                    "role": message.get("role", "user"),
                    "content": "".join(text_parts),
                }
                processed_messages.append(processed_message)
            else:
                # Regular text message
                processed_messages.append(message)

        kwargs["return_dict"] = False
        return self.tokenizer.apply_chat_template(
            processed_messages,
            tokenize=tokenize,
            add_generation_prompt=add_generation_prompt,
            **kwargs,
        )

class Siglip2VisionTransformer(nn.Module):
    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]:
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("qkv_proj", "q_proj", "q"),
            ("qkv_proj", "k_proj", "k"),
            ("qkv_proj", "v_proj", "v"),
        ]
        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()

        for name, loaded_weight in weights:
            for param_name, weight_name, shard_id in stacked_params_mapping:
                if weight_name not in name:
                    continue
                name = name.replace(weight_name, param_name)

                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
                param = params_dict[name]
                weight_loader = getattr(param, "weight_loader", default_weight_loader)
                weight_loader(param, loaded_weight)
            loaded_params.add(name)
        return loaded_params