tpu-inference 0.0.1rc1__py3-none-any.whl

tpu_inference/models/jax/qwen2_5_vl.py

@@ -0,0 +1,1218 @@
+import math
+from functools import partial
+from typing import (Callable, List, Literal, NamedTuple, Optional, TypedDict,
+                    Union)
+
+import jax
+import jax.numpy as jnp
+import numpy as np
+from flax import nnx
+from jax.sharding import Mesh
+from transformers import modeling_flax_utils
+from transformers.models.qwen2_5_vl.configuration_qwen2_5_vl import (
+    Qwen2_5_VLConfig, Qwen2_5_VLVisionConfig)
+from vllm.config import VllmConfig
+
+from tpu_inference import utils as utils
+from tpu_inference.layers.common.attention_interface import \
+    sharded_flash_attention
+from tpu_inference.layers.common.attention_metadata import AttentionMetadata
+from tpu_inference.logger import init_logger
+from tpu_inference.models.jax.qwen2 import Qwen2ForCausalLM
+# from vllm.model_executor.models.interfaces import MultiModalEmbeddings
+from tpu_inference.models.jax.utils.multi_modal_utils import (
+    MultiModalEmbeddings, merge_multimodal_embeddings)
+from tpu_inference.models.jax.utils.weight_utils import (get_default_maps,
+                                                         load_hf_weights)
+
+logger = init_logger(__name__)
+
+init_fn = nnx.initializers.uniform()
+
+DEFAULT_BLOCK_K_MAJOR = 128
+
+
+class SegmentIds(NamedTuple):
+    """SegmentIds for Q and KV sequences.
+
+  SegmentIds are used to generate segment mask, which prevents attention between
+  different segments in the input sequence. Each array is a list of ids
+  (integers).
+  Only the token with the same id can attend to each other.
+
+  Attributes:
+    q: segment ids along the Q sequence.
+    kv: segment ids along the KV sequence.
+  """
+
+    q: jax.Array  # [batch_size, q_seq_len]
+    kv: jax.Array  # [batch_size, kv_seq_len]
+
+
+class Qwen2_5_VLImagePixelInputs(TypedDict):
+    type: Literal["pixel_values"]
+    pixel_values: jax.Array
+    """Shape:
+    `(num_patches, num_channels * patch_size * patch_size)`
+    """
+
+    image_grid_thw: tuple[tuple[int, int, int], ...]
+    """Shape: `(num_images, 3)`
+    This should be in `(grid_t, grid_h, grid_w)` format.
+    """
+
+
+# NOTE: We are not supporting embedding inputs for now
+# The code here makes the struture consistent and
+# makes iteasier for future implementation
+class Qwen2_5_VLImageEmbeddingInputs(TypedDict):
+    type: Literal["image_embeds"]
+    image_embeds: jax.Array
+    """Supported types:
+    - list[`jax.Array`]: A list of tensors holding all images' features.
+        Each tensor holds an image's features.
+    - `jax.Array`: A tensor holding all images' features (concatenation of
+        all images' feature tensors).
+
+    Tensor shape: `(num_image_features, hidden_size)`
+    - `num_image_features` varies based on
+        the number and resolution of the images.
+    - `hidden_size` must match the hidden size of language model backbone.
+    """
+
+    image_grid_thw: jax.Array
+    """Shape: `(num_images, 3)`
+    This should be in `(grid_t, grid_h, grid_w)` format.
+    """
+
+
+Qwen2_5_VLImageInputs = Union[Qwen2_5_VLImagePixelInputs,
+                              Qwen2_5_VLImageEmbeddingInputs]
+
+
+class Qwen2_5_VisionMLP(nnx.Module):
+
+    def __init__(self, config: Qwen2_5_VLVisionConfig, dtype: jnp.dtype,
+                 rngs: nnx.Rngs):
+        in_features = config.hidden_size
+        hidden_features = config.intermediate_size
+        act_fn = modeling_flax_utils.ACT2FN[config.hidden_act]
+        self.gate_proj = nnx.Linear(
+            in_features,
+            hidden_features,
+            use_bias=True,
+            param_dtype=dtype,
+            kernel_init=nnx.with_partitioning(init_fn, (None, "model")),
+            bias_init=nnx.with_partitioning(init_fn, ("model", )),
+            rngs=rngs,
+        )
+        self.up_proj = nnx.Linear(
+            in_features,
+            hidden_features,
+            use_bias=True,
+            param_dtype=dtype,
+            kernel_init=nnx.with_partitioning(init_fn, (None, "model")),
+            bias_init=nnx.with_partitioning(init_fn, ("model", )),
+            rngs=rngs,
+        )
+        self.down_proj = nnx.Linear(
+            hidden_features,
+            in_features,
+            use_bias=True,
+            param_dtype=dtype,
+            kernel_init=nnx.with_partitioning(init_fn, ("model", None)),
+            bias_init=nnx.with_partitioning(init_fn, (None, )),
+            rngs=rngs,
+        )
+        self.act_fn = act_fn
+
+    def __call__(self, x: jax.Array) -> jax.Array:
+        gate = self.act_fn(self.gate_proj(x))
+        up = self.up_proj(x)
+        fuse = gate * up
+        result = self.down_proj(fuse)
+        return result
+
+
+def apply_rotary_pos_emb_vision(x: jax.Array,
+                                rotary_pos_emb: jax.Array) -> jax.Array:
+    # x: [B, T, N, H]
+    # rotary_pos_emb: [T, H//2]
+    _, _, _, H = x.shape
+    half_dim = H // 2
+
+    # [B, T, N, H//2]
+    x_real = x[..., :half_dim]
+    x_imag = x[..., half_dim:]
+
+    # [T, H//2]
+    cos_emb = jnp.cos(rotary_pos_emb)
+    sin_emb = jnp.sin(rotary_pos_emb)
+
+    # [1, T, 1, H//2]
+    cos_emb = cos_emb[None, :, None, :]
+    sin_emb = sin_emb[None, :, None, :]
+
+    # [B, T, N, H//2]
+    x_rotated_real = x_real * cos_emb - x_imag * sin_emb
+    x_rotated_imag = x_real * sin_emb + x_imag * cos_emb
+
+    # [B, T, N, H]
+    x_rotated = jnp.concatenate([x_rotated_real, x_rotated_imag], axis=-1)
+
+    return x_rotated
+
+
+def generate_window_segment_ids(cu_seqlens: jax.Array, seq_len: int,
+                                padded_seq_len: int) -> SegmentIds:
+    """Generates segment IDs for windowed attention
+
+    Args:
+        cu_seqlens: A 1D array of cumulative sequence lengths for each window.
+            e.g., [0, len_win0, len_win0+len_win1, ...]
+
+    Returns:
+        A SegmentIds object for flash_attention.
+    """
+    indices = jnp.arange(seq_len, dtype=jnp.int32)
+    segment_ids = jnp.searchsorted(cu_seqlens[1:], indices, side='right') + 1
+    padding_segment_ids = jnp.zeros(padded_seq_len - seq_len, dtype=jnp.int32)
+    segment_ids = jnp.concatenate([segment_ids, padding_segment_ids])
+    segment_ids = segment_ids.reshape(1, -1)
+
+    return SegmentIds(q=segment_ids, kv=segment_ids)
+
+
+class Qwen2_5_VisionAttention(nnx.Module):
+
+    def __init__(self, config: Qwen2_5_VLConfig, dtype: jnp.dtype,
+                 rngs: nnx.Rngs, mesh: Mesh):
+        vision_config = config.vision_config
+        self.hidden_size = vision_config.hidden_size
+        self.num_heads = vision_config.num_heads
+        self.num_kv_heads = self.num_heads
+        self.rope_theta = config.rope_theta
+        self.rope_scaling = getattr(config, "rope_scaling", None)
+        self.head_dim_original = self.hidden_size // self.num_heads
+
+        sharding_size = mesh.shape["model"]
+        self.num_heads = utils.get_padded_num_heads(self.num_heads,
+                                                    sharding_size)
+        self.num_kv_heads = utils.get_padded_num_heads(self.num_kv_heads,
+                                                       sharding_size)
+        self.head_dim = utils.get_padded_head_dim(self.head_dim_original)
+
+        # TODO: Wenlong: Do not consider padding for now
+        self.head_dim = self.head_dim_original
+
+        self.mesh = mesh
+
+        self.qkv_proj = nnx.Linear(
+            self.hidden_size,
+            3 * self.hidden_size,
+            use_bias=True,
+            param_dtype=dtype,
+            kernel_init=nnx.with_partitioning(init_fn, (None, "model")),
+            bias_init=nnx.with_partitioning(init_fn, ("model", )),
+            rngs=rngs,
+        )
+
+        self.proj = nnx.Linear(
+            self.hidden_size,
+            self.hidden_size,
+            use_bias=True,
+            param_dtype=dtype,
+            kernel_init=nnx.with_partitioning(init_fn, ("model", None)),
+            bias_init=nnx.with_partitioning(init_fn, (None, )),
+            rngs=rngs,
+        )
+        self.flash_attention = sharded_flash_attention(
+            mesh=mesh,
+            causal=False,
+            sm_scale=1.0 / math.sqrt(self.head_dim),
+            vmem_limit_bytes=128 * 1024 * 1024,
+        )
+
+    def __call__(
+        self,
+        x: jax.Array,
+        rotary_pos_emb: jax.Array,
+        cu_window_seqlens: Optional[jax.Array] = None,
+        use_fullattn: bool = True,
+    ) -> jax.Array:
+        T, B, D = x.shape
+        assert B == 1, "Vision attention currently only supports batch size 1"
+        # [T, B, D] -> [T, B, 3 * D]
+        qkv = self.qkv_proj(x)
+
+        # Split into Q, K, V.
+        # NOTE: simplified from vLLM's split_qkv,
+        # may need to revisit for tp>1
+        # [T, B, 3 * D] -> 3 *[T, B, D]
+        q, k, v = jnp.split(qkv, 3, axis=-1)
+
+        # [T, B, N, H]
+        q = q.reshape(T, B, self.num_heads, self.head_dim)
+        k = k.reshape(T, B, self.num_heads, self.head_dim)
+        v = v.reshape(T, B, self.num_heads, self.head_dim)
+
+        # [T, B, N, H] -> [B, T, N, H]
+        q = jnp.transpose(q, (1, 0, 2, 3))
+        k = jnp.transpose(k, (1, 0, 2, 3))
+        v = jnp.transpose(v, (1, 0, 2, 3))
+
+        # rotary_pos_emb shape: (T, H)
+        q = apply_rotary_pos_emb_vision(q, rotary_pos_emb)
+        k = apply_rotary_pos_emb_vision(k, rotary_pos_emb)
+
+        # NOTE: an extra transpose because we need to
+        # align the correctness with vLLM's design.
+        # Might be able to remove one once implemented.
+        # [B, T, N, H] -> [B, N, T, H]
+        q = jnp.transpose(q, (0, 2, 1, 3))
+        k = jnp.transpose(k, (0, 2, 1, 3))
+        v = jnp.transpose(v, (0, 2, 1, 3))
+
+        # Pad the sequence length to be a multiple of 128 for flash_attention
+        block_k_major = DEFAULT_BLOCK_K_MAJOR
+        T_attn = q.shape[2]
+        padded_T = (T_attn + block_k_major -
+                    1) // block_k_major * block_k_major
+        pad_width = ((0, 0), (0, 0), (0, padded_T - T_attn), (0, 0))
+
+        q = jnp.pad(q, pad_width, 'constant')
+        k = jnp.pad(k, pad_width, 'constant')
+        v = jnp.pad(v, pad_width, 'constant')
+
+        segment_ids = generate_window_segment_ids(cu_window_seqlens, T_attn,
+                                                  padded_T)
+
+        # TODO (jacobplatin): add support for quantized KV cache?
+        output = self.flash_attention(q, k, v, segment_ids)
+
+        # Unpad the output
+        output = output[:, :, :T_attn, :]
+
+        # [B, N, T, H] -> [T, B, N, H]
+        output = jnp.transpose(output, (2, 0, 1, 3))
+
+        output = output.reshape(T, B, D)
+
+        output = self.proj(output)
+
+        return output
+
+
+class Qwen2_5_VisionBlock(nnx.Module):
+
+    def __init__(self, config: Qwen2_5_VLConfig, dtype: jnp.dtype,
+                 rngs: nnx.Rngs, mesh: Mesh):
+        vision_config = config.vision_config
+        dim = vision_config.hidden_size
+        norm_layer = partial(nnx.RMSNorm,
+                             epsilon=config.rms_norm_eps,
+                             scale_init=nnx.with_partitioning(
+                                 init_fn, (None, )))
+
+        self.norm1 = norm_layer(dim, dtype=dtype, rngs=rngs)
+        self.norm2 = norm_layer(dim, dtype=dtype, rngs=rngs)
+        self.attn = Qwen2_5_VisionAttention(config=config,
+                                            dtype=dtype,
+                                            rngs=rngs,
+                                            mesh=mesh)
+        self.mlp = Qwen2_5_VisionMLP(config=vision_config,
+                                     dtype=dtype,
+                                     rngs=rngs)
+
+    def __call__(self,
+                 x: jax.Array,
+                 rotary_pos_emb: jax.Array,
+                 cu_window_seqlens: Optional[jax.Array] = None,
+                 use_fullattn: bool = True) -> jax.Array:
+
+        x = x + self.attn(self.norm1(x), rotary_pos_emb, cu_window_seqlens,
+                          use_fullattn)
+        x = x + self.mlp(self.norm2(x))
+
+        return x
+
+
+class Qwen2_5_VisionPatchEmbed(nnx.Module):
+
+    def __init__(
+        self,
+        rngs: nnx.Rngs,
+        patch_size: int = 14,
+        temporal_patch_size: int = 2,
+        in_channels: int = 3,
+        hidden_size: int = 1152,
+        dtype: jnp.dtype = jnp.bfloat16,
+    ) -> None:
+        self.patch_size = patch_size
+        self.temporal_patch_size = temporal_patch_size
+        self.hidden_size = hidden_size
+        kernel_size = (temporal_patch_size, patch_size, patch_size)
+        self.proj = nnx.Conv(in_features=in_channels,
+                             out_features=hidden_size,
+                             kernel_size=kernel_size,
+                             strides=kernel_size,
+                             use_bias=False,
+                             param_dtype=dtype,
+                             kernel_init=nnx.with_partitioning(
+                                 init_fn, (None, None, None, None, "model")),
+                             rngs=rngs)
+
+    def __call__(self, x: jax.Array) -> jax.Array:
+        # x is (L, C * T * H * W)
+        L, dim = x.shape
+        C = dim // (self.temporal_patch_size * self.patch_size *
+                    self.patch_size)
+        # Reshape to (L, T, H, W, C) for Conv3D with channels_last
+        x = x.reshape(L, C, self.temporal_patch_size, self.patch_size,
+                      self.patch_size)
+        # L,T,H,W,C
+        x = jnp.transpose(x, (0, 2, 3, 4, 1))
+        x = self.proj(x)
+        # After conv, shape is (L, T_out, H_out, W_out, C_out)
+        # With stride=kernel_size, T_out=H_out=W_out=1.
+        # So shape is (L, 1, 1, 1, hidden_size)
+        x = x.reshape(L, self.hidden_size)
+        return x
+
+
+class Qwen2_5_VisionPatchMerger(nnx.Module):
+
+    def __init__(self, d_model: int, context_dim: int, norm_layer: Callable,
+                 spatial_merge_size: int, dtype: jnp.dtype, rngs: nnx.Rngs):
+        self.hidden_size = context_dim * (spatial_merge_size**2)
+        self.ln_q = norm_layer(context_dim,
+                               dtype=dtype,
+                               rngs=rngs,
+                               scale_init=nnx.with_partitioning(
+                                   init_fn, (None, )))
+        self.mlp_fc1 = nnx.Linear(
+            self.hidden_size,
+            self.hidden_size,
+            use_bias=True,
+            param_dtype=dtype,
+            kernel_init=nnx.with_partitioning(init_fn, (None, "model")),
+            bias_init=nnx.with_partitioning(init_fn, ("model", )),
+            rngs=rngs)
+        self.mlp_act = modeling_flax_utils.ACT2FN["gelu"]
+        self.mlp_fc2 = nnx.Linear(
+            self.hidden_size,
+            d_model,
+            use_bias=True,
+            param_dtype=dtype,
+            kernel_init=nnx.with_partitioning(init_fn, ("model", None)),
+            bias_init=nnx.with_partitioning(init_fn, (None, )),
+            rngs=rngs)
+
+    def __call__(self, x: jax.Array) -> jax.Array:
+        x = self.ln_q(x)
+        x = x.reshape(-1, self.hidden_size)
+        x = self.mlp_fc1(x)
+        x = self.mlp_act(x)
+        x = self.mlp_fc2(x)
+        return x
+
+
+class Qwen2_5_VisionRotaryEmbedding(nnx.Module):
+
+    def __init__(self, dim: int, theta: float = 10000.0):
+        self.dim = dim
+        self.theta = theta
+
+    def __call__(self, seqlen: int) -> jax.Array:
+        inv_freq = 1.0 / (self.theta**(
+            jnp.arange(0, self.dim, 2, dtype=jnp.float32) / self.dim))
+        seq = jnp.arange(seqlen, dtype=jnp.float32)
+        freqs = jnp.outer(seq, inv_freq)
+        return freqs.astype(jnp.bfloat16)
+
+
+class Qwen2_5_VisionTransformer(nnx.Module):
+
+    def __init__(self,
+                 vllm_config: VllmConfig,
+                 rngs: nnx.Rngs,
+                 mesh: Mesh,
+                 norm_eps: float = 1e-6):
+        model_config = vllm_config.model_config
+        hf_config = model_config.hf_config
+        vision_config = hf_config.vision_config
+        dtype = model_config.dtype
+
+        self.config = vision_config
+        self.dtype = dtype
+
+        patch_size = vision_config.patch_size
+        temporal_patch_size = vision_config.temporal_patch_size
+        in_channels = vision_config.in_channels
+        self.hidden_size = vision_config.hidden_size
+        self.num_heads = vision_config.num_heads
+
+        # args for get_window_index_thw
+        self.window_size = vision_config.window_size
+        self.patch_size = vision_config.patch_size
+        self.spatial_merge_size = vision_config.spatial_merge_size
+        self.fullatt_block_indexes = vision_config.fullatt_block_indexes
+        self.spatial_merge_unit = self.spatial_merge_size**2
+
+        self.patch_embed = Qwen2_5_VisionPatchEmbed(
+            patch_size=patch_size,
+            temporal_patch_size=temporal_patch_size,
+            in_channels=in_channels,
+            hidden_size=self.hidden_size,
+            dtype=dtype,
+            rngs=rngs)
+
+        head_dim = vision_config.hidden_size // vision_config.num_heads
+        self.rotary_pos_emb = Qwen2_5_VisionRotaryEmbedding(head_dim // 2)
+
+        self.blocks = [
+            Qwen2_5_VisionBlock(
+                config=hf_config,
+                dtype=dtype,
+                rngs=rngs,
+                mesh=mesh,
+            ) for _ in range(vision_config.depth)
+        ]
+        self.merger = Qwen2_5_VisionPatchMerger(
+            d_model=vision_config.out_hidden_size,
+            context_dim=vision_config.hidden_size,
+            norm_layer=partial(nnx.RMSNorm, epsilon=norm_eps),
+            spatial_merge_size=vision_config.spatial_merge_size,
+            dtype=dtype,
+            rngs=rngs)
+
+        additional_config = getattr(vllm_config, "additional_config",
+                                    None) or {}
+        self.enable_dynamic_image_sizes = additional_config.get(
+            "enable_dynamic_image_sizes", False)
+
+    def rotary_pos_emb_thw(self, t, h, w):
+        hpos_ids, wpos_ids = jnp.indices((h, w))
+        hpos_ids = hpos_ids.reshape(
+            h // self.spatial_merge_size,
+            self.spatial_merge_size,
+            w // self.spatial_merge_size,
+            self.spatial_merge_size,
+        ).transpose(0, 2, 1, 3).flatten()
+        wpos_ids = wpos_ids.reshape(
+            h // self.spatial_merge_size,
+            self.spatial_merge_size,
+            w // self.spatial_merge_size,
+            self.spatial_merge_size,
+        ).transpose(0, 2, 1, 3).flatten()
+        pos_ids = jnp.stack([hpos_ids, wpos_ids], axis=-1)
+        pos_ids = jnp.tile(pos_ids, (t, 1))
+
+        max_size = max(h, w)
+        rotary_pos_emb_full = self.rotary_pos_emb(max_size)
+
+        rotary_pos_emb = rotary_pos_emb_full[pos_ids].reshape(
+            pos_ids.shape[0], -1)
+        rotary_pos_emb = rotary_pos_emb.reshape(
+            rotary_pos_emb.shape[0] // self.spatial_merge_unit,
+            self.spatial_merge_unit, -1)
+
+        return rotary_pos_emb
+
+    def get_window_index_thw(self, grid_t, grid_h, grid_w):
+        vit_merger_window_size = (self.window_size //
+                                  self.spatial_merge_size // self.patch_size)
+
+        llm_grid_h = grid_h // self.spatial_merge_size
+        llm_grid_w = grid_w // self.spatial_merge_size
+
+        index = jnp.arange(grid_t * llm_grid_h * llm_grid_w).reshape(
+            grid_t, llm_grid_h, llm_grid_w)
+
+        pad_h = vit_merger_window_size - llm_grid_h % vit_merger_window_size
+        pad_w = vit_merger_window_size - llm_grid_w % vit_merger_window_size
+        num_windows_h = (llm_grid_h + pad_h) // vit_merger_window_size
+        num_windows_w = (llm_grid_w + pad_w) // vit_merger_window_size
+
+        index_padded = jnp.pad(index, ((0, 0), (0, pad_h), (0, pad_w)),
+                               constant_values=-100)
+        index_padded = index_padded.reshape(grid_t, num_windows_h,
+                                            vit_merger_window_size,
+                                            num_windows_w,
+                                            vit_merger_window_size)
+        index_padded = jnp.transpose(index_padded, (0, 1, 3, 2, 4)).reshape(
+            grid_t, num_windows_h * num_windows_w, vit_merger_window_size,
+            vit_merger_window_size)
+        seqlens = (index_padded != -100).sum([2, 3]).reshape(-1)
+        index_padded = index_padded.reshape(-1)
+        # The number of valid indices is static because grid_t, grid_h, grid_w
+        # are static.
+        num_valid_indices = grid_t * llm_grid_h * llm_grid_w
+        valid_indices = jnp.nonzero(index_padded != -100,
+                                    size=num_valid_indices)[0]
+        index_new = index_padded[valid_indices]
+        cu_seqlens_tmp = jnp.cumsum(seqlens) * self.spatial_merge_unit
+        cu_seqlens_tmp = cu_seqlens_tmp.astype(jnp.int32)
+
+        # NOTE (wenlong): Pytorch code uses this to reduce replication,
+        # but I don't think there is a need here, plus it would cause problem in JIT
+        # Please refer here if there is a problem down-stream
+        # cu_seqlens_tmp = jnp.unique(cu_seqlens_tmp)
+
+        return index_new, cu_seqlens_tmp
+
+    def get_rope_by_thw(self, t, h, w):
+        window_index_thw, cu_seqlens_window_thw = self.get_window_index_thw(
+            t, h, w)
+
+        rotary_pos_emb_thw = self.rotary_pos_emb_thw(t, h, w)
+
+        rotary_pos_emb_thw = rotary_pos_emb_thw[window_index_thw, :, :]
+        rotary_pos_emb_thw = rotary_pos_emb_thw.reshape(
+            -1, rotary_pos_emb_thw.shape[-1])
+        cu_seqlens_thw = jnp.full(t, h * w, dtype=jnp.int32)
+
+        return (rotary_pos_emb_thw, window_index_thw, cu_seqlens_window_thw,
+                cu_seqlens_thw)
+
+    def compute_attn_mask_seqlen(
+        self,
+        cu_seqlens: jax.Array,
+    ) -> tuple[Optional[int], Optional[list[int]]]:
+        max_seqlen, seqlens = None
+        max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max().item()
+        seqlens = (cu_seqlens[1:] - cu_seqlens[:-1]).tolist()
+        return max_seqlen, seqlens
+
+    def compute_aux_arrays(self, grid_thw: tuple[tuple[int, int, int]]):
+        # num of images/videoes
+        num_grids = len(grid_thw)
+
+        rotary_pos_emb = []
+        window_index: list = []
+        cu_window_seqlens: list = [jnp.array([0], dtype=jnp.int32)]
+        cu_seqlens: list = []
+
+        window_index_id = 0
+        cu_window_seqlens_last = 0
+        for i in range(num_grids):
+            t, h, w = grid_thw[i]
+
+            llm_h = h // self.spatial_merge_size
+            llm_w = w // self.spatial_merge_size
+
+            (
+                rotary_pos_emb_thw,
+                window_index_thw,
+                cu_seqlens_window_thw,
+                cu_seqlens_thw,
+            ) = self.get_rope_by_thw(t, h, w)
+
+            window_index.append(window_index_thw + window_index_id)
+            window_index_id += (t * llm_h * llm_w)
+
+            cu_seqlens_window_thw = (cu_seqlens_window_thw +
+                                     cu_window_seqlens_last)
+            cu_window_seqlens_last = cu_seqlens_window_thw[-1]
+            cu_window_seqlens.append(cu_seqlens_window_thw)
+
+            rotary_pos_emb.append(rotary_pos_emb_thw)
+
+            cu_seqlens.append(cu_seqlens_thw)
+
+        rotary_pos_emb = jnp.concatenate(rotary_pos_emb, axis=0)
+        window_index = jnp.concatenate(window_index, axis=0)
+        cu_window_seqlens = jnp.concatenate(cu_window_seqlens, axis=0)
+
+        cu_seqlens = jnp.concatenate(cu_seqlens, axis=0)
+        cu_seqlens = jnp.cumsum(cu_seqlens, axis=0, dtype=jnp.int32)
+        cu_seqlens = jnp.pad(cu_seqlens, ((1, 0), ),
+                             mode='constant',
+                             constant_values=0)
+        return window_index, rotary_pos_emb, cu_seqlens, cu_window_seqlens
+
+    def pad_inputs(self, x, window_index, rotary_pos_emb, cu_seqlens,
+                   cu_window_seqlens):
+        # padding
+        num_patches = int(rotary_pos_emb.shape[0])
+        bucket_num_patches = 1 << (num_patches - 1).bit_length()
+        num_tokens = window_index.shape[0]
+        bucket_num_tokens = bucket_num_patches // self.spatial_merge_unit
+        vit_merger_window_size = (self.window_size //
+                                  self.spatial_merge_size // self.patch_size)
+        max_windows = (bucket_num_tokens // vit_merger_window_size) + 2
+
+        rotary_pos_emb = jnp.pad(rotary_pos_emb,
+                                 ((0, bucket_num_patches - num_patches),
+                                  (0, 0)))
+        window_index = jnp.concatenate([
+            window_index,
+            jnp.arange(num_tokens, bucket_num_tokens, dtype=jnp.int32)
+        ])
+        cu_window_seqlens = jnp.append(cu_window_seqlens, bucket_num_patches)
+        pad_w = max(0, max_windows + 1 - cu_window_seqlens.shape[0])
+        cu_window_seqlens = jnp.pad(cu_window_seqlens, (0, pad_w), mode='edge')
+        cu_seqlens = jnp.append(cu_seqlens, bucket_num_patches)
+
+        x_padded = jnp.pad(x, ((0, bucket_num_patches - x.shape[0]), (0, 0)))
+
+        return x_padded, window_index, rotary_pos_emb, cu_seqlens, cu_window_seqlens, num_tokens
+
+    def compute_hidden_states(self, x: jax.Array, window_index: jax.Array,
+                              rotary_pos_emb: jax.Array, cu_seqlens: jax.Array,
+                              cu_window_seqlens: jax.Array) -> jax.Array:
+        hidden_states = self.patch_embed(x)
+
+        # num of patches
+        seq_len = x.shape[0]
+
+        hidden_states = hidden_states.reshape(
+            seq_len // self.spatial_merge_unit, self.spatial_merge_unit, -1)
+        hidden_states = hidden_states[window_index, :, :]
+        hidden_states = hidden_states.reshape(seq_len, -1)
+
+        hidden_states = jnp.expand_dims(hidden_states, axis=1)
+
+        for layer_num, blk in enumerate(self.blocks):
+            if layer_num in self.fullatt_block_indexes:
+                hidden_states = blk(hidden_states,
+                                    rotary_pos_emb=rotary_pos_emb,
+                                    cu_window_seqlens=cu_seqlens,
+                                    use_fullattn=True)
+            else:
+                hidden_states = blk(hidden_states,
+                                    rotary_pos_emb=rotary_pos_emb,
+                                    cu_window_seqlens=cu_window_seqlens,
+                                    use_fullattn=False)
+
+        # adapter
+        hidden_states = self.merger(hidden_states)
+        reverse_indices = jnp.argsort(window_index)
+        hidden_states = hidden_states[reverse_indices, :]
+        return hidden_states
+
+    @jax.jit
+    def encode_padded_jit(self, x_padded, window_index, rotary_pos_emb,
+                          cu_seqlens, cu_window_seqlens):
+        return self.compute_hidden_states(x_padded, window_index,
+                                          rotary_pos_emb, cu_seqlens,
+                                          cu_window_seqlens)
+
+    @partial(
+        jax.jit,
+        static_argnames=("grid_thw", ),
+    )
+    def encode_jit(self, x, grid_thw):
+        window_index, rotary_pos_emb, cu_seqlens, cu_window_seqlens = self.compute_aux_arrays(
+            grid_thw)
+        return self.compute_hidden_states(x, window_index, rotary_pos_emb,
+                                          cu_seqlens, cu_window_seqlens)
+
+    def __call__(self, x: jax.Array, grid_thw: tuple[tuple[int, int,
+                                                           int]]) -> jax.Array:
+        # x: pixel_values: jax.Array
+        # """Shape:
+        # `(num_patches, num_channels * patch_size * patch_size)`
+        # """
+
+        # grid_thw: image_grid_thw: jax.Array
+        # """Shape: `(num_images, 3)`
+        # This should be in `(grid_t, grid_h, grid_w)` format.
+        # """
+        if self.enable_dynamic_image_sizes:
+            window_index, rotary_pos_emb, cu_seqlens, cu_window_seqlens = self.compute_aux_arrays(
+                grid_thw)
+            x_padded, window_index, rotary_pos_emb, cu_seqlens, cu_window_seqlens, num_tokens = self.pad_inputs(
+                x, window_index, rotary_pos_emb, cu_seqlens, cu_window_seqlens)
+
+            hidden_states = self.encode_padded_jit(x_padded, window_index,
+                                                   rotary_pos_emb, cu_seqlens,
+                                                   cu_window_seqlens)
+            return hidden_states[:num_tokens]
+
+        else:
+            return self.encode_jit(x, grid_thw)
+
+
+class Qwen2_5_VLForConditionalGeneration(nnx.Module):
+
+    def __init__(self, vllm_config: VllmConfig, rng_key: jax.Array,
+                 mesh: Mesh) -> None:
+        config: Qwen2_5_VLConfig = vllm_config.model_config.hf_config
+        multimodal_config = vllm_config.model_config.multimodal_config
+
+        self.vllm_config = vllm_config
+        self.rng = nnx.Rngs(rng_key)
+        self.mesh = mesh
+
+        self.config = config
+        self.multimodal_config = multimodal_config
+
+        self.visual = Qwen2_5_VisionTransformer(
+            vllm_config=vllm_config,
+            rngs=self.rng,
+            mesh=mesh,
+            norm_eps=getattr(config, "rms_norm_eps", 1e-6),
+        )
+        self.language_model = Qwen2ForCausalLM(vllm_config, rng_key, mesh)
+
+    def get_mrope_input_positions(
+        self,
+        input_tokens: list[int],
+        hf_config,
+        image_grid_thw,
+        video_grid_thw,
+        second_per_grid_ts: list[float],
+        context_len: int = 0,
+        seq_len: int | None = None,
+        audio_feature_lengths=None,
+        use_audio_in_video: bool = False,
+    ) -> tuple[jax.Array, int]:
+        """Get mrope input positions and delta value."""
+
+        image_token_id = hf_config.image_token_id
+        video_token_id = hf_config.video_token_id
+        vision_start_token_id = hf_config.vision_start_token_id
+        spatial_merge_size = hf_config.vision_config.spatial_merge_size
+        tokens_per_second = getattr(hf_config.vision_config,
+                                    "tokens_per_second", 1.0)
+
+        input_tokens_tensor = np.array(input_tokens)
+        vision_start_indices = np.argwhere(
+            input_tokens_tensor == vision_start_token_id).squeeze(1)
+        vision_tokens = input_tokens_tensor[vision_start_indices + 1]
+        image_nums = np.sum(vision_tokens == image_token_id)
+        video_nums = np.sum(vision_tokens == video_token_id)
+        llm_pos_ids_list: list = []
+
+        st = 0
+        remain_images, remain_videos = image_nums, video_nums
+
+        image_index, video_index = 0, 0
+        for _ in range(image_nums + video_nums):
+            video_second_per_grid_t = 0.0
+            if remain_images > 0:
+                try:
+                    ed_image = input_tokens.index(image_token_id, st)
+                except ValueError:
+                    ed_image = len(input_tokens) + 1
+            else:
+                ed_image = len(input_tokens) + 1
+            if remain_videos > 0:
+                try:
+                    ed_video = input_tokens.index(video_token_id, st)
+                except ValueError:
+                    ed_video = len(input_tokens) + 1
+            else:
+                ed_video = len(input_tokens) + 1
+            if ed_image < ed_video:
+                t, h, w = (
+                    image_grid_thw[image_index][0],
+                    image_grid_thw[image_index][1],
+                    image_grid_thw[image_index][2],
+                )
+                image_index += 1
+                remain_images -= 1
+                ed = ed_image
+            else:
+                t, h, w = (
+                    video_grid_thw[video_index][0],
+                    video_grid_thw[video_index][1],
+                    video_grid_thw[video_index][2],
+                )
+                video_second_per_grid_t = 1.0
+                if second_per_grid_ts:
+                    video_second_per_grid_t = second_per_grid_ts[video_index]
+                video_index += 1
+                remain_videos -= 1
+                ed = ed_video
+
+            llm_grid_t, llm_grid_h, llm_grid_w = (
+                t,
+                h // spatial_merge_size,
+                w // spatial_merge_size,
+            )
+            text_len = ed - st
+
+            st_idx = llm_pos_ids_list[-1].max().item() + 1 if len(
+                llm_pos_ids_list) > 0 else 0
+            llm_pos_ids_list.append(
+                jnp.broadcast_to(
+                    jnp.arange(text_len, dtype=jnp.int32).reshape(1, -1),
+                    (3, text_len)) + st_idx)
+
+            t_index = ((jnp.broadcast_to(
+                jnp.arange(llm_grid_t, dtype=jnp.int32).reshape(-1, 1),
+                (llm_grid_t, llm_grid_h * llm_grid_w)) *
+                        video_second_per_grid_t * tokens_per_second).astype(
+                            jnp.int32).flatten())
+
+            h_index = (jnp.broadcast_to(
+                jnp.arange(llm_grid_h, dtype=jnp.int32).reshape(1, -1, 1),
+                (llm_grid_t, llm_grid_h, llm_grid_w)).flatten())
+            w_index = (jnp.broadcast_to(
+                jnp.arange(llm_grid_w, dtype=jnp.int32).reshape(1, 1, -1),
+                (llm_grid_t, llm_grid_h, llm_grid_w)).flatten())
+
+            llm_pos_ids_list.append(
+                jnp.stack([t_index, h_index, w_index]) + text_len + st_idx)
+            st = ed + llm_grid_t * llm_grid_h * llm_grid_w
+
+        if st < len(input_tokens):
+            st_idx = llm_pos_ids_list[-1].max().item() + 1 if len(
+                llm_pos_ids_list) > 0 else 0
+            text_len = len(input_tokens) - st
+
+            llm_pos_ids_list.append(
+                jnp.broadcast_to(
+                    jnp.arange(text_len, dtype=jnp.int32).reshape(1, -1),
+                    (3, text_len)) + st_idx)
+
+        llm_positions = jnp.concatenate(llm_pos_ids_list,
+                                        axis=1).reshape(3, -1)
+        mrope_position_delta = (llm_positions.max() + 1 -
+                                len(input_tokens)).item()
+        llm_positions = llm_positions[:, context_len:seq_len]
+
+        return llm_positions, mrope_position_delta
+
+    def _validate_and_reshape_mm_tensor(self, mm_input: object,
+                                        name: str) -> jax.Array:
+        if isinstance(mm_input, list):
+            # Assuming it's a list of arrays (e.g., np.ndarray, torch.Tensor)
+            # that can be concatenated.
+            arrays_to_concat = [jnp.asarray(item) for item in mm_input]
+            return jnp.concatenate(arrays_to_concat, axis=0)
+
+        # Handle single array-like objects (np.ndarray, torch.Tensor, jax.Array)
+        if hasattr(mm_input, 'ndim'):
+            array_input = jnp.asarray(mm_input)
+            if array_input.ndim == 2:
+                return array_input
+            if array_input.ndim == 3:
+                # This reshapes the batched 3D tensor to a 2D tensor.
+                return array_input.reshape(-1, array_input.shape[-1])
+
+        raise ValueError(f"Incorrect type of {name}. "
+                         f"Got type: {type(mm_input)}")
+
+    def _parse_and_validate_image_input(
+            self, image_grid_thw: tuple[tuple[int, int, int], ...],
+            **kwargs: object) -> Optional[Qwen2_5_VLImageInputs]:
+        pixel_values = kwargs.pop("pixel_values", None)
+        image_embeds = kwargs.pop("image_embeds", None)
+        # image_grid_thw = kwargs.pop("image_grid_thw", None)
+
+        if pixel_values is None and image_embeds is None:
+            return None
+
+        if pixel_values is not None:
+            pixel_values = self._validate_and_reshape_mm_tensor(
+                pixel_values, "image pixel values")
+            # image_grid_thw = self._validate_and_reshape_mm_tensor(
+            #     image_grid_thw, "image grid_thw")
+
+            if not isinstance(pixel_values, jax.Array):
+                raise ValueError("Incorrect type of image pixel values. "
+                                 f"Got type: {type(pixel_values)}")
+
+            return Qwen2_5_VLImagePixelInputs(type="pixel_values",
+                                              pixel_values=pixel_values,
+                                              image_grid_thw=image_grid_thw)
+
+        # Note: comment them out for now and save for future support
+        # if image_embeds is not None:
+        #     image_embeds = self._validate_and_reshape_mm_tensor(
+        #         image_embeds, "image embeds")
+        #     image_grid_thw = self._validate_and_reshape_mm_tensor(
+        #         image_grid_thw, "image grid_thw")
+
+        #     if not isinstance(image_embeds, jax.Array):
+        #         raise ValueError("Incorrect type of image embeddings. "
+        #                          f"Got type: {type(image_embeds)}")
+        #     return Qwen2_5_VLImageEmbeddingInputs(
+        #         type="image_embeds",
+        #         image_embeds=image_embeds,
+        #         image_grid_thw=image_grid_thw)
+
+    def _parse_and_validate_multimodal_inputs(self,
+                                              image_grid_thw: tuple[tuple[int,
+                                                                          int,
+                                                                          int],
+                                                                    ...],
+                                              **kwargs: object) -> dict:
+        mm_input_by_modality = {}
+
+        # Preserve the order of modalities if there are multiple of them
+        # from the order of kwargs.
+        for input_key in kwargs:
+            if input_key in ("pixel_values", "image_embeds"
+                             ) and "image" not in mm_input_by_modality:
+                mm_input_by_modality[
+                    "image"] = self._parse_and_validate_image_input(
+                        image_grid_thw, **kwargs)
+            # if input_key in ("pixel_values_videos", "video_embeds"
+            #                  ) and "video" not in mm_input_by_modality:
+            #     mm_input_by_modality[
+            #         "video"] = self._parse_and_validate_video_input(**kwargs)
+        return mm_input_by_modality
+
+    def get_single_image_embedding(self, image_pixel_values, image_grid_thw):
+        return self.visual(image_pixel_values, (image_grid_thw, ))
+
+    def _process_image_input(
+            self, image_input: Qwen2_5_VLImageInputs) -> tuple[jax.Array, ...]:
+
+        grid_thw = image_input["image_grid_thw"]
+
+        if image_input["type"] == "image_embeds":
+            image_embeds = image_input["image_embeds"].astype(
+                self.visual.dtype)
+        else:
+            pixel_values = image_input["pixel_values"]
+            image_embeds = []
+            current_idx = 0
+            for image_thw in grid_thw:
+                t, h, w = image_thw
+                image_size = t * h * w
+                end_idx = current_idx + image_size
+                image_pixel_values = pixel_values[current_idx:end_idx, :]
+                image_embeds.append(
+                    self.get_single_image_embedding(image_pixel_values,
+                                                    image_thw))
+                current_idx = end_idx
+            image_embeds = jnp.concatenate(image_embeds, axis=0)
+
+        # Split concatenated embeddings for each image item.
+        merge_size = self.visual.config.spatial_merge_size
+        sizes = np.prod(np.array(grid_thw, dtype=np.int64),
+                        axis=-1) // merge_size // merge_size
+
+        if sizes.size == 0:
+            return ()
+        if sizes.size == 1:
+            return (image_embeds, )
+
+        split_indices = np.cumsum(sizes)[:-1]
+        return tuple(jnp.split(image_embeds, split_indices))
+
+    def get_multimodal_embeddings(self, image_grid_thw: tuple[tuple[int, int,
+                                                                    int], ...],
+                                  **kwargs: object) -> MultiModalEmbeddings:
+
+        mm_input_by_modality = self._parse_and_validate_multimodal_inputs(
+            image_grid_thw, **kwargs)
+        if not mm_input_by_modality:
+            return []
+
+        # The result multimodal_embeddings is tuple of tensors, with each
+        # tensor correspoending to a multimodal data item (image or video).
+        multimodal_embeddings: tuple[jax.Array, ...] = ()
+
+        # NOTE: It is important to iterate over the keys in this dictionary
+        # to preserve the order of the modalities.
+        for modality in mm_input_by_modality:
+            multimodal_input = mm_input_by_modality[modality]
+            if modality == "image":
+                vision_embeddings = self._process_image_input(multimodal_input)
+                multimodal_embeddings += vision_embeddings
+            # if modality == "video":
+            #     video_embeddings = self._process_video_input(multimodal_input)
+            #     multimodal_embeddings += video_embeddings
+
+        return multimodal_embeddings
+
+    def get_input_embeddings(
+            self, input_ids: jax.Array,
+            multimodal_embeddings: Optional[jax.Array]) -> jax.Array:
+
+        inputs_embeds = self.language_model.model.embed(input_ids)
+
+
+        if multimodal_embeddings is not None \
+            and multimodal_embeddings.shape[0] != 0:
+            inputs_embeds = merge_multimodal_embeddings(
+                input_ids, inputs_embeds, multimodal_embeddings,
+                [self.config.image_token_id, self.config.video_token_id])
+
+        return inputs_embeds
+
+    def __call__(
+        self,
+        kv_caches: list[jax.Array],
+        input_ids: Optional[jax.Array],
+        attention_metadata: AttentionMetadata,
+        inputs_embeds: Optional[jax.Array] = None,
+        *args,
+    ) -> tuple[list[jax.Array], jax.Array, List[jax.Array]]:
+        # The logic of choosing between input_ids and inputs_embeds is
+        # handled inside self.language_model.__call__
+        kv_caches, x, [] = self.language_model(
+            kv_caches=kv_caches,
+            input_ids=input_ids,
+            attention_metadata=attention_metadata,
+            inputs_embeds=inputs_embeds,
+        )
+        return kv_caches, x, []
+
+    def compute_logits(self, hidden_states: jax.Array) -> jax.Array:
+        return self.language_model.compute_logits(hidden_states)
+
+    def load_weights(self, rng_key: jax.Array) -> None:
+        self.rng = nnx.Rngs(rng_key)
+        self.language_model.rng = self.rng
+
+        # Key: path to a HF layer weight
+        # Value: a tuple of (path to a nnx layer weight, nnx weight sharding)
+
+        mappings = {
+            "model.embed_tokens": "language_model.model.embed.embedding",
+            "model.layers.*.input_layernorm":
+            "language_model.model.layers.*.input_layernorm.scale",
+            "model.layers.*.mlp.down_proj":
+            "language_model.model.layers.*.mlp.down_proj.kernel",
+            "model.layers.*.mlp.gate_proj":
+            "language_model.model.layers.*.mlp.gate_proj.kernel",
+            "model.layers.*.mlp.up_proj":
+            "language_model.model.layers.*.mlp.up_proj.kernel",
+            "model.layers.*.post_attention_layernorm":
+            "language_model.model.layers.*.post_attention_layernorm.scale",
+            "model.layers.*.self_attn.k_proj":
+            "language_model.model.layers.*.self_attn.k_proj.kernel",
+            "model.layers.*.self_attn.o_proj":
+            "language_model.model.layers.*.self_attn.o_proj.kernel",
+            "model.layers.*.self_attn.q_proj":
+            "language_model.model.layers.*.self_attn.q_proj.kernel",
+            "model.layers.*.self_attn.v_proj":
+            "language_model.model.layers.*.self_attn.v_proj.kernel",
+            "model.layers.*.self_attn.q_proj.bias":
+            "language_model.model.layers.*.self_attn.q_proj.bias",
+            "model.layers.*.self_attn.k_proj.bias":
+            "language_model.model.layers.*.self_attn.k_proj.bias",
+            "model.layers.*.self_attn.v_proj.bias":
+            "language_model.model.layers.*.self_attn.v_proj.bias",
+            "model.norm": "language_model.model.norm.scale",
+            "visual.blocks.*.attn.proj.bias": "visual.blocks.*.attn.proj.bias",
+            "visual.blocks.*.attn.proj": "visual.blocks.*.attn.proj.kernel",
+            "visual.blocks.*.attn.qkv.bias":
+            "visual.blocks.*.attn.qkv_proj.bias",
+            "visual.blocks.*.attn.qkv": "visual.blocks.*.attn.qkv_proj.kernel",
+            "visual.blocks.*.mlp.down_proj.bias":
+            "visual.blocks.*.mlp.down_proj.bias",
+            "visual.blocks.*.mlp.down_proj":
+            "visual.blocks.*.mlp.down_proj.kernel",
+            "visual.blocks.*.mlp.gate_proj.bias":
+            "visual.blocks.*.mlp.gate_proj.bias",
+            "visual.blocks.*.mlp.gate_proj":
+            "visual.blocks.*.mlp.gate_proj.kernel",
+            "visual.blocks.*.mlp.up_proj.bias":
+            "visual.blocks.*.mlp.up_proj.bias",
+            "visual.blocks.*.mlp.up_proj":
+            "visual.blocks.*.mlp.up_proj.kernel",
+            "visual.blocks.*.norm1": "visual.blocks.*.norm1.scale",
+            "visual.blocks.*.norm2": "visual.blocks.*.norm2.scale",
+            "visual.merger.ln_q": "visual.merger.ln_q.scale",
+            "visual.merger.mlp.0.bias": "visual.merger.mlp_fc1.bias",
+            "visual.merger.mlp.0": "visual.merger.mlp_fc1.kernel",
+            "visual.merger.mlp.2.bias": "visual.merger.mlp_fc2.bias",
+            "visual.merger.mlp.2": "visual.merger.mlp_fc2.kernel",
+            "visual.patch_embed.proj": "visual.patch_embed.proj.kernel",
+        }
+
+        # Add lm_head mapping only if it's not tied to embeddings
+        hf_config = self.vllm_config.model_config.hf_config
+        if not hf_config.tie_word_embeddings:
+            mappings.update({
+                "lm_head": "language_model.model.lm_head",
+            })
+
+        metadata_map = get_default_maps(self.vllm_config.model_config,
+                                        self.mesh, mappings)
+        load_hf_weights(vllm_config=self.vllm_config,
+                        model=self,
+                        metadata_map=metadata_map,
+                        mesh=self.mesh)
+
+    def precompile_vision_encoder(
+        self,
+        run_compilation_fn: Callable,
+    ) -> None:
+        vc = self.vllm_config.model_config.hf_config.vision_config
+        patch_input_dim = vc.in_channels * vc.temporal_patch_size * vc.patch_size * vc.patch_size
+        if self.visual.enable_dynamic_image_sizes:
+            spatial_merge_unit = vc.spatial_merge_size**2
+            max_num_batched_tokens = self.vllm_config.scheduler_config.max_num_batched_tokens
+            mm_kwargs = self.vllm_config.model_config.multimodal_config.mm_processor_kwargs or {}
+            limit_pixels = float(mm_kwargs.get("max_pixels", float('inf')))
+
+            max_patches = int(
+                min(max_num_batched_tokens * spatial_merge_unit,
+                    limit_pixels / (vc.patch_size**2)))
+
+            num_patches_paddings = [
+                1 << i for i in range(4, (max_patches - 1).bit_length() + 1)
+            ]
+            rotary_dim = vc.hidden_size // vc.num_heads // 2
+            vit_merger_window_size = (vc.window_size //
+                                      vc.spatial_merge_size // vc.patch_size)
+
+            for num_patches in num_patches_paddings:
+                dummy_x_padded = jnp.ones(
+                    (num_patches, patch_input_dim),
+                    dtype=self.vllm_config.model_config.dtype)
+
+                num_tokens = num_patches // spatial_merge_unit
+                dummy_window_index = jnp.arange(num_tokens, dtype=jnp.int32)
+
+                dummy_rotary_pos_emb = jnp.ones(
+                    (num_patches, rotary_dim),
+                    dtype=self.vllm_config.model_config.dtype)
+
+                dummy_cu_seqlens = jnp.array([0, num_patches, num_patches],
+                                             dtype=jnp.int32)
+
+                max_windows = (num_tokens // vit_merger_window_size) + 2
+                patches_per_window = (vit_merger_window_size**
+                                      2) * spatial_merge_unit
+                dummy_cu_window_seqlens = jnp.arange(
+                    max_windows + 1, dtype=jnp.int32) * patches_per_window
+                dummy_cu_window_seqlens = jnp.minimum(dummy_cu_window_seqlens,
+                                                      num_patches)
+
+                run_compilation_fn("vision_encoder_padded",
+                                   self.visual.encode_padded_jit,
+                                   dummy_x_padded,
+                                   dummy_window_index,
+                                   dummy_rotary_pos_emb,
+                                   dummy_cu_seqlens,
+                                   dummy_cu_window_seqlens,
+                                   num_patches=num_patches)
+        else:
+            image_shapes = []
+            if (warmup_config := self.vllm_config.additional_config.get(
+                    "vision_warmup_config")):
+                image_shapes = warmup_config.get("image_shapes")
+
+            factor = vc.patch_size * vc.spatial_merge_size
+            for input_hw in image_shapes:
+                if not isinstance(input_hw, list) or len(input_hw) != 2:
+                    logger.warning(f"Skipping invalid shape {input_hw}.")
+                    continue
+                h_input, w_input = input_hw
+                h_processed = round(h_input / factor) * factor
+                w_processed = round(w_input / factor) * factor
+                t, h, w = 1, h_processed // vc.patch_size, w_processed // vc.patch_size
+                grid_thw = (t, h, w)
+                num_patches = t * h * w
+
+                dummy_pixel_values = jnp.ones(
+                    (num_patches, patch_input_dim),
+                    self.vllm_config.model_config.dtype,
+                )
+                dummy_grid_thw = (grid_thw, )
+
+                run_compilation_fn("vision_encoder",
+                                   self.visual.encode_jit,
+                                   dummy_pixel_values,
+                                   dummy_grid_thw,
+                                   image_shape=input_hw)