[VLM][Model] TP support for ViTs (#7186)

Co-authored-by: Roger Wang <136131678+ywang96@users.noreply.github.com>
Co-authored-by: Roger Wang <ywang@roblox.com>

tests/models/test_intern_vit.py

@@ -6,8 +6,6 @@ import torch.nn as nn
 from huggingface_hub import snapshot_download
 from transformers import AutoConfig, AutoModel, CLIPImageProcessor
 
-from vllm.model_executor.models.intern_vit import InternVisionModel
-
 from ..conftest import _ImageAssets, cleanup
 
 pytestmark = pytest.mark.vlm
@@ -49,6 +47,7 @@ def run_intern_vit_test(
         for pixel_value in pixel_values
     ]
 
+    from vllm.model_executor.models.intern_vit import InternVisionModel
     vllm_model = InternVisionModel(config)
     vllm_model.load_weights(hf_model.state_dict().items())
 

tests/models/test_internvl.py

@@ -6,9 +6,6 @@ import torch
 from PIL.Image import Image
 from transformers import AutoConfig
 
-from vllm.model_executor.models.internvl import (IMG_CONTEXT, IMG_END,
-                                                 IMG_START,
-                                                 image_to_pixel_values)
 from vllm.multimodal.utils import rescale_image_size
 from vllm.utils import is_cpu
 
@@ -33,35 +30,6 @@ models = [
 ]
 
 
-class InternVLProcessor:
-    """A simple processor for InternVL2 HF model which misses a processor."""
-
-    def __init__(self, hf_runner: HfRunner):
-        self.num_image_token = hf_runner.model.num_image_token
-        self.tokenizer = hf_runner.tokenizer
-        self.dtype = hf_runner.model.dtype
-
-        self.config = AutoConfig.from_pretrained(hf_runner.model_name)
-        self.vision_config = self.config.vision_config
-        self.use_thumbnail = self.config.use_thumbnail
-        self.min_num = self.config.min_dynamic_patch
-        self.max_num = self.config.max_dynamic_patch
-        self.image_size = self.vision_config.image_size
-
-    def __call__(self, text: str, images: Image, **kwargs):
-        pixel_values = image_to_pixel_values(images, self.image_size,
-                                             self.min_num, self.max_num,
-                                             self.use_thumbnail).to(self.dtype)
-        num_patches_list = [pixel_values.shape[0]]
-        for num_patches in num_patches_list:
-            context_tokens = IMG_CONTEXT * self.num_image_token * num_patches
-            image_tokens = IMG_START + context_tokens + IMG_END
-            text = text.replace('<image>', image_tokens, 1)
-        prompt = self.tokenizer(text, return_tensors="pt")
-        prompt.update({"pixel_values": pixel_values})
-        return prompt
-
-
 # adapted from https://huggingface.co/OpenGVLab/InternVL2-1B/blob/main/modeling_internvl_chat.py
 def generate(
     self,
@@ -127,6 +95,37 @@ def run_test(
     # if we run HF first, the cuda initialization will be done and it
     # will hurt multiprocessing backend with fork method (the default method).
 
+    class InternVLProcessor:
+        """A simple processor for InternVL2 which misses a processor."""
+
+        def __init__(self, hf_runner: HfRunner):
+            self.num_image_token = hf_runner.model.num_image_token
+            self.tokenizer = hf_runner.tokenizer
+            self.dtype = hf_runner.model.dtype
+
+            self.config = AutoConfig.from_pretrained(hf_runner.model_name)
+            self.vision_config = self.config.vision_config
+            self.use_thumbnail = self.config.use_thumbnail
+            self.min_num = self.config.min_dynamic_patch
+            self.max_num = self.config.max_dynamic_patch
+            self.image_size = self.vision_config.image_size
+
+        def __call__(self, text: str, images: Image, **kwargs):
+            from vllm.model_executor.models.internvl import (
+                IMG_CONTEXT, IMG_END, IMG_START, image_to_pixel_values)
+            pixel_values = image_to_pixel_values(
+                images, self.image_size, self.min_num, self.max_num,
+                self.use_thumbnail).to(self.dtype)
+            num_patches_list = [pixel_values.shape[0]]
+            for num_patches in num_patches_list:
+                context_tokens = IMG_CONTEXT * self.num_image_token \
+                    * num_patches
+                image_tokens = IMG_START + context_tokens + IMG_END
+                text = text.replace('<image>', image_tokens, 1)
+            prompt = self.tokenizer(text, return_tensors="pt")
+            prompt.update({"pixel_values": pixel_values})
+            return prompt
+
     # max_model_len should be greater than image_feature_size
     with vllm_runner(model,
                      max_model_len=4096,

vllm/model_executor/models/blip.py

@@ -7,12 +7,14 @@ import torch
 import torch.nn as nn
 from PIL import Image
 from transformers import Blip2VisionConfig, BlipVisionConfig
-from transformers.models.blip.modeling_blip import BlipAttention
+from xformers import ops as xops
 
 from vllm.config import ModelConfig
+from vllm.distributed import divide, get_tensor_model_parallel_world_size
 from vllm.inputs import LLMInputs
 from vllm.model_executor.layers.activation import get_act_fn
 from vllm.model_executor.layers.linear import (ColumnParallelLinear,
+                                               QKVParallelLinear,
                                                RowParallelLinear)
 from vllm.model_executor.layers.quantization import QuantizationConfig
 from vllm.multimodal.utils import (cached_get_tokenizer,
@@ -154,6 +156,77 @@ class BlipVisionEmbeddings(nn.Module):
         return embeddings
 
 
+class BlipAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(
+        self,
+        config: BlipVisionConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+    ):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_heads
+        if self.head_dim * self.num_heads != self.embed_dim:
+            raise ValueError(
+                "embed_dim must be divisible by num_heads "
+                f"(got `embed_dim`: {self.embed_dim} and `num_heads`:"
+                f" {self.num_heads}).")
+        self.scale = self.head_dim**-0.5
+        self.dropout = config.attention_dropout
+
+        self.qkv = QKVParallelLinear(
+            self.embed_dim,
+            self.head_dim,
+            self.num_heads,
+            bias=config.qkv_bias,
+            quant_config=quant_config,
+        )
+        self.projection = RowParallelLinear(
+            self.embed_dim,
+            self.embed_dim,
+            quant_config=quant_config,
+        )
+
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.num_heads_per_partition = divide(self.num_heads, self.tp_size)
+
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads,
+                           self.head_dim).transpose(1, 2).contiguous()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+    ):
+        """Input shape: Batch x Time x Channel"""
+        bsz, tgt_len, _ = hidden_states.size()
+
+        qkv_states, _ = self.qkv(hidden_states)
+        query_states, key_states, value_states = qkv_states.chunk(3, dim=-1)
+        query_states = query_states.view(bsz, tgt_len,
+                                         self.num_heads_per_partition,
+                                         self.head_dim)
+        key_states = key_states.view(bsz, tgt_len,
+                                     self.num_heads_per_partition,
+                                     self.head_dim)
+        value_states = value_states.view(bsz, tgt_len,
+                                         self.num_heads_per_partition,
+                                         self.head_dim)
+
+        out = xops.memory_efficient_attention_forward(query_states,
+                                                      key_states,
+                                                      value_states,
+                                                      p=self.dropout,
+                                                      scale=self.scale)
+        out = out.view(bsz, tgt_len, -1)
+        attn_output, _ = self.projection(out)
+
+        return attn_output
+
+
 class BlipMLP(nn.Module):
 
     def __init__(self,
@@ -188,7 +261,7 @@ class BlipEncoderLayer(nn.Module):
                  quant_config: Optional[QuantizationConfig] = None):
         super().__init__()
 
-        self.self_attn = BlipAttention(config)
+        self.self_attn = BlipAttention(config, quant_config=quant_config)
         self.layer_norm1 = nn.LayerNorm(config.hidden_size,
                                         eps=config.layer_norm_eps)
         self.mlp = BlipMLP(config, quant_config=quant_config)
@@ -199,7 +272,7 @@ class BlipEncoderLayer(nn.Module):
         residual = hidden_states
 
         hidden_states = self.layer_norm1(hidden_states)
-        hidden_states, _ = self.self_attn(hidden_states=hidden_states)
+        hidden_states = self.self_attn(hidden_states=hidden_states)
         hidden_states = residual + hidden_states
 
         residual = hidden_states

vllm/model_executor/models/blip2.py

@@ -714,8 +714,7 @@ class Blip2ForConditionalGeneration(nn.Module, SupportsMultiModal):
             use_default_weight_loading = False
             if "vision" in name:
                 if self.vision_model is not None:
-                    # We only do sharding for language model and
+                    # BlipVisionModel does not need sharding
-                    # not vision model for now.
                     use_default_weight_loading = True
             else:
                 for (param_name, weight_name,

vllm/model_executor/models/clip.py

@@ -7,12 +7,14 @@ import torch
 import torch.nn as nn
 from PIL import Image
 from transformers import CLIPVisionConfig
-from transformers.models.clip.modeling_clip import CLIPAttention
+from xformers import ops as xops
 
 from vllm.config import ModelConfig
+from vllm.distributed import divide, get_tensor_model_parallel_world_size
 from vllm.inputs import LLMInputs
 from vllm.model_executor.layers.activation import get_act_fn
 from vllm.model_executor.layers.linear import (ColumnParallelLinear,
+                                               QKVParallelLinear,
                                                RowParallelLinear)
 from vllm.model_executor.layers.quantization import QuantizationConfig
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
@@ -160,6 +162,78 @@ class CLIPVisionEmbeddings(nn.Module):
         return embeddings
 
 
+class CLIPAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(
+        self,
+        config: CLIPVisionConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+    ):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_heads
+        if self.head_dim * self.num_heads != self.embed_dim:
+            raise ValueError(
+                "embed_dim must be divisible by num_heads "
+                f"(got `embed_dim`: {self.embed_dim} and `num_heads`:"
+                f" {self.num_heads}).")
+        self.scale = self.head_dim**-0.5
+        self.dropout = config.attention_dropout
+
+        self.qkv_proj = QKVParallelLinear(
+            hidden_size=self.embed_dim,
+            head_size=self.head_dim,
+            total_num_heads=self.num_heads,
+            quant_config=quant_config,
+        )
+
+        self.out_proj = RowParallelLinear(
+            input_size=self.embed_dim,
+            output_size=self.embed_dim,
+            quant_config=quant_config,
+        )
+
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.num_heads_per_partition = divide(self.num_heads, self.tp_size)
+
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads,
+                           self.head_dim).transpose(1, 2).contiguous()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+    ):
+        """Input shape: Batch x Time x Channel"""
+        bsz, tgt_len, _ = hidden_states.size()
+
+        qkv_states, _ = self.qkv_proj(hidden_states)
+        query_states, key_states, value_states = qkv_states.chunk(3, dim=-1)
+
+        query_states = query_states.view(bsz, tgt_len,
+                                         self.num_heads_per_partition,
+                                         self.head_dim)
+        key_states = key_states.view(bsz, tgt_len,
+                                     self.num_heads_per_partition,
+                                     self.head_dim)
+        value_states = value_states.view(bsz, tgt_len,
+                                         self.num_heads_per_partition,
+                                         self.head_dim)
+
+        out = xops.memory_efficient_attention_forward(query_states,
+                                                      key_states,
+                                                      value_states,
+                                                      p=self.dropout,
+                                                      scale=self.scale)
+        out = out.view(bsz, tgt_len, -1)
+        attn_output, _ = self.out_proj(out)
+
+        return attn_output
+
+
 class CLIPMLP(nn.Module):
 
     def __init__(self,
@@ -192,7 +266,7 @@ class CLIPEncoderLayer(nn.Module):
                  quant_config: Optional[QuantizationConfig] = None):
         super().__init__()
 
-        self.self_attn = CLIPAttention(config)
+        self.self_attn = CLIPAttention(config, quant_config=quant_config)
         self.layer_norm1 = nn.LayerNorm(config.hidden_size,
                                         eps=config.layer_norm_eps)
         self.mlp = CLIPMLP(config, quant_config=quant_config)
@@ -204,7 +278,7 @@ class CLIPEncoderLayer(nn.Module):
         residual = hidden_states
 
         hidden_states = self.layer_norm1(hidden_states)
-        hidden_states, _ = self.self_attn(hidden_states=hidden_states)
+        hidden_states = self.self_attn(hidden_states=hidden_states)
         hidden_states = residual + hidden_states
 
         residual = hidden_states
@@ -304,7 +378,15 @@ class CLIPVisionModel(nn.Module):
     def device(self):
         return next(self.parameters()).device
 
+    # (TODO) Add prefix argument for filtering out weights to be loaded
+    #        ref: https://github.com/vllm-project/vllm/pull/7186#discussion_r1734163986
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+        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())
         layer_count = len(self.vision_model.encoder.layers)
 
@@ -318,7 +400,16 @@ class CLIPVisionModel(nn.Module):
                 if layer_idx >= layer_count:
                     continue
 
-            param = params_dict[name]
+            for (param_name, weight_name, shard_id) in stacked_params_mapping:
-            weight_loader = getattr(param, "weight_loader",
+                if weight_name not in name:
-                                    default_weight_loader)
+                    continue
-            weight_loader(param, loaded_weight)
+
+                param = params_dict[name.replace(weight_name, param_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)

vllm/model_executor/models/intern_vit.py

@@ -10,10 +10,13 @@ import torch
 import torch.nn as nn
 import torch.nn.functional as F
 from transformers import PretrainedConfig
+from xformers import ops as xops
 
+from vllm.distributed import divide, get_tensor_model_parallel_world_size
 from vllm.model_executor.layers.activation import get_act_fn
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.linear import (ColumnParallelLinear,
+                                               QKVParallelLinear,
                                                RowParallelLinear)
 from vllm.model_executor.layers.quantization import QuantizationConfig
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
@@ -81,7 +84,11 @@ class InternVisionEmbeddings(nn.Module):
 class InternAttention(nn.Module):
     """Multi-headed attention from 'Attention Is All You Need' paper"""
 
-    def __init__(self, config: PretrainedConfig):
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+    ):
         super().__init__()
         self.config = config
         self.embed_dim = config.hidden_size
@@ -94,9 +101,13 @@ class InternAttention(nn.Module):
                 f' {self.num_heads}).')
 
         self.scale = self.head_dim**-0.5
-        self.qkv = nn.Linear(self.embed_dim,
+        self.qkv = QKVParallelLinear(
-                             3 * self.embed_dim,
+            self.embed_dim,
-                             bias=config.qkv_bias)
+            self.head_dim,
+            self.num_heads,
+            bias=config.qkv_bias,
+            quant_config=quant_config,
+        )
 
         self.qk_normalization = config.qk_normalization
 
@@ -104,25 +115,40 @@ class InternAttention(nn.Module):
             self.q_norm = RMSNorm(self.embed_dim, eps=config.layer_norm_eps)
             self.k_norm = RMSNorm(self.embed_dim, eps=config.layer_norm_eps)
 
-        self.proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.proj = RowParallelLinear(
+            self.embed_dim,
+            self.embed_dim,
+            quant_config=quant_config,
+        )
+
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.num_heads_per_partition = divide(self.num_heads, self.tp_size)
 
     def forward(self, x):
         B, N, C = x.shape
-        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads,
+        qkv, _ = self.qkv(x)
-                                  C // self.num_heads).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv.chunk(3, dim=-1)
-        q, k, v = qkv.unbind(0)
+
+        q = q.view(B, N, self.num_heads_per_partition, self.head_dim)
+        k = k.view(B, N, self.num_heads_per_partition, self.head_dim)
+        v = v.view(B, N, self.num_heads_per_partition, self.head_dim)
 
         if self.qk_normalization:
-            B_, H_, N_, D_ = q.shape
+            B_, N_, H_, D_ = q.shape
-            q = self.q_norm.forward_native(q.transpose(1, 2).flatten(
+            q = self.q_norm.forward_native(q.flatten(-2,
-                -2, -1)).view(B_, N_, H_, D_).transpose(1, 2)
+                                                     -1)).view(B_, N_, H_, D_)
-            k = self.k_norm.forward_native(k.transpose(1, 2).flatten(
+            k = self.k_norm.forward_native(k.flatten(-2,
-                -2, -1)).view(B_, N_, H_, D_).transpose(1, 2)
+                                                     -1)).view(B_, N_, H_, D_)
 
-        x = F.scaled_dot_product_attention(q, k, v, scale=self.scale)
+        x = xops.memory_efficient_attention_forward(
-        x = x.transpose(1, 2).reshape(B, N, C)
+            q,
+            k,
+            v,
+            scale=self.scale,
+        )
+        x = x.view(B, N, -1)
 
-        x = self.proj(x)
+        x, _ = self.proj(x)
         return x
 
 
@@ -161,7 +187,7 @@ class InternVisionEncoderLayer(nn.Module):
         self.intermediate_size = config.intermediate_size
         self.norm_type = config.norm_type
 
-        self.attn = InternAttention(config)
+        self.attn = InternAttention(config, quant_config=quant_config)
         self.mlp = InternMLP(config, quant_config=quant_config)
         self.norm1 = NORM2FN[self.norm_type](self.embed_dim,
                                              eps=config.layer_norm_eps)

vllm/model_executor/models/paligemma.py

@@ -145,7 +145,6 @@ class PaliGemmaForConditionalGeneration(nn.Module, SupportsMultiModal):
         self.config = config
         self.multimodal_config = multimodal_config
 
-        # TODO(ywang96): Port over SiglipVisionModel & TP
         self.vision_tower = SiglipVisionModel(config.vision_config)
         self.multi_modal_projector = PaliGemmaMultiModalProjector(
             vision_hidden_size=config.vision_config.hidden_size,
@@ -308,34 +307,27 @@ class PaliGemmaForConditionalGeneration(nn.Module, SupportsMultiModal):
                 if key_to_modify in name:
                     name = name.replace(key_to_modify, new_key)
             use_default_weight_loading = False
-            if "vision" in name:
+            for (param_name, shard_name, shard_id) in stacked_params_mapping:
-                if self.vision_tower is not None:
+                if shard_name not in name:
-                    # We only do sharding for language model and
+                    continue
-                    # not vision model for now.
+                name = name.replace(shard_name, param_name)
-                    use_default_weight_loading = True
+                # Skip loading extra bias for GPTQ models.
+                if name.endswith(".bias") and name not in params_dict:
+                    continue
+                param = params_dict[name]
+                weight_loader = param.weight_loader
+                weight_loader(param, loaded_weight, shard_id)
+                break
             else:
-                for (param_name, shard_name,
+                # lm_head is not used in vllm as it is tied with
-                     shard_id) in stacked_params_mapping:
+                # embed_token. To prevent errors, skip loading
-                    if shard_name not in name:
+                # lm_head.weight.
-                        continue
+                if "lm_head.weight" in name:
-                    name = name.replace(shard_name, param_name)
+                    continue
-                    # Skip loading extra bias for GPTQ models.
+                # Skip loading extra bias for GPTQ models.
-                    if name.endswith(".bias") and name not in params_dict:
+                if name.endswith(".bias") and name not in params_dict:
-                        continue
+                    continue
-                    param = params_dict[name]
+                use_default_weight_loading = True
-                    weight_loader = param.weight_loader
-                    weight_loader(param, loaded_weight, shard_id)
-                    break
-                else:
-                    # lm_head is not used in vllm as it is tied with
-                    # embed_token. To prevent errors, skip loading
-                    # lm_head.weight.
-                    if "lm_head.weight" in name:
-                        continue
-                    # Skip loading extra bias for GPTQ models.
-                    if name.endswith(".bias") and name not in params_dict:
-                        continue
-                    use_default_weight_loading = True
 
             if use_default_weight_loading:
                 param = params_dict[name]

vllm/model_executor/models/phi3v.py

@@ -71,6 +71,23 @@ CLIP_VIT_LARGE_PATCH14_336_CONFIG = CLIPVisionConfig(dropout=0.0,
                                                      projection_dim=768)
 
 
+def _init_img_processor(hf_config: PretrainedConfig):
+    clip_config = CLIP_VIT_LARGE_PATCH14_336_CONFIG
+    layer_idx = hf_config.img_processor.get('layer_idx', -2)
+
+    # Initialize the CLIP only up to the required feature layer
+    if layer_idx < 0:
+        num_hidden_layers = clip_config.num_hidden_layers + \
+            layer_idx + 1
+    else:
+        num_hidden_layers = layer_idx + 1
+
+    img_processor = CLIPVisionModel(
+        clip_config, num_hidden_layers_override=num_hidden_layers)
+
+    return img_processor
+
+
 class Phi3VImagePixelInputs(TypedDict):
     type: Literal["pixel_values"]
     data: Union[torch.Tensor, List[torch.Tensor]]
@@ -139,18 +156,8 @@ class Phi3HDImageEmbedding(Phi3ImageEmbeddingBase):
         hidden_size = config.n_embd if hasattr(
             config, 'n_embd') else config.hidden_size
 
-        clip_config = CLIP_VIT_LARGE_PATCH14_336_CONFIG
+        self.img_processor = _init_img_processor(config)
-        self.layer_idx = config.img_processor.get('layer_idx', -2)
 
-        # Initialize the CLIP only up to the required feature layer
-        if self.layer_idx < 0:
-            num_hidden_layers = clip_config.num_hidden_layers + \
-                self.layer_idx + 1
-        else:
-            num_hidden_layers = self.layer_idx + 1
-
-        self.img_processor = CLIPVisionModel(
-            clip_config, num_hidden_layers_override=num_hidden_layers)
         image_dim_out = config.img_processor['image_dim_out']
         self.num_img_tokens = config.img_processor['num_img_tokens']
 
@@ -656,23 +663,27 @@ class Phi3VForCausalLM(nn.Module, SupportsMultiModal):
             (".gate_up_proj", ".gate_proj", 0),
             (".gate_up_proj", ".up_proj", 1),
         ]
+
+        # TODO(ChristopherCho): This is a temporary fix to load
+        #     the vision weights with CLIPVisionModel.load_weights()
+        vision_weights = []
         params_dict = dict(self.named_parameters())
         for name, loaded_weight in weights:
             if "rotary_emb.inv_freq" in name:
                 continue
-            # post_layernorm is not needed in CLIPVisionModel
+            # Skip loading the img_processor weights since they are
-            if "vision_model.post_layernorm" in name:
+            # loaded separately.
+            if "vision_embed_tokens.img_processor" in name:
+                vision_weights.append((name, loaded_weight))
                 continue
+
             for key_to_modify, new_key in _KEYS_TO_MODIFY_MAPPING.items():
                 if key_to_modify in name:
                     name = name.replace(key_to_modify, new_key)
             for (param_name, weight_name, shard_id) in stacked_params_mapping:
-                # We only do sharding for language model
-                # and not vision model for now.
-                if "vision_embed_tokens" in name and self.vision_embed_tokens:
-                    continue
                 if weight_name not in name:
                     continue
+
                 param = params_dict[name.replace(weight_name, param_name)]
                 weight_loader = param.weight_loader
                 weight_loader(param, loaded_weight, shard_id)
@@ -686,3 +697,11 @@ class Phi3VForCausalLM(nn.Module, SupportsMultiModal):
                     weight_loader = getattr(param, "weight_loader",
                                             default_weight_loader)
                     weight_loader(param, loaded_weight)
+
+        # We use regex to extract the sub-module name
+        # from "model.vision_embed_tokens.img_processor.*"
+        vision_weights = [
+            (re.search(r"vision_embed_tokens\.img_processor\.(.*)",
+                       n).group(1), w) for n, w in vision_weights
+        ]
+        self.vision_embed_tokens.img_processor.load_weights(vision_weights)

vllm/model_executor/models/siglip.py

@@ -9,12 +9,10 @@ import torch
 from PIL import Image
 from torch import nn
 from transformers import SiglipVisionConfig
-from transformers.models.siglip.modeling_siglip import SiglipAttention
+from xformers import ops as xops
-from vllm_flash_attn import flash_attn_func
-from xformers.ops import memory_efficient_attention
 
 from vllm.config import ModelConfig
-from vllm.distributed import get_tensor_model_parallel_world_size
+from vllm.distributed import divide, get_tensor_model_parallel_world_size
 from vllm.inputs import LLMInputs
 from vllm.model_executor.layers.activation import get_act_fn
 from vllm.model_executor.layers.linear import (ColumnParallelLinear,
@@ -221,9 +219,7 @@ class SiglipVisionEmbeddings(nn.Module):
         return embeddings
 
 
-# NOTE: Not used - kept for later when we TP the ViT
+class SiglipAttention(nn.Module):
-# TODO(ChristopherCho): Implement TP version of Attention
-class SiglipTPAttention(nn.Module):
 
     def __init__(
         self,
@@ -233,38 +229,30 @@ class SiglipTPAttention(nn.Module):
         super().__init__()
         self.config = config
         self.embed_dim = config.hidden_size
-
+        self.num_heads = config.num_attention_heads
-        tp_size = get_tensor_model_parallel_world_size()
+        self.head_dim = self.embed_dim // self.num_heads
-        self.total_num_heads = config.num_attention_heads
+        if self.head_dim * self.num_heads != self.embed_dim:
-        if self.total_num_heads % tp_size != 0:
-            raise ValueError(
-                f"Number of attention heads ({self.total_num_heads}) "
-                "must be divisible by the tensor model parallel size"
-                f" ({tp_size}).")
-
-        self.num_heads = self.total_num_heads // tp_size
-        self.head_dim = self.embed_dim // self.total_num_heads
-        if self.head_dim * self.total_num_heads != self.embed_dim:
             raise ValueError(f"embed_dim must be divisible by num_heads (got "
                              "`embed_dim`: {self.embed_dim} and `num_heads`:"
                              f" {self.num_heads}).")
-        self.qkv_size = self.num_heads * self.head_dim
+
         self.scale = self.head_dim**-0.5
         self.dropout = config.attention_dropout
-
         self.qkv_proj = QKVParallelLinear(
             hidden_size=self.embed_dim,
             head_size=self.head_dim,
-            total_num_heads=self.total_num_heads,
+            total_num_heads=self.num_heads,
             quant_config=quant_config,
         )
+
         self.out_proj = RowParallelLinear(
             input_size=self.embed_dim,
             output_size=self.embed_dim,
             quant_config=quant_config,
         )
 
-        self.attn_fn = self._basic_attention_forward
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.num_heads_per_partition = divide(self.num_heads, self.tp_size)
 
     def forward(
         self,
@@ -274,163 +262,29 @@ class SiglipTPAttention(nn.Module):
         batch_size, q_len, _ = hidden_states.size()
 
         qkv_states, _ = self.qkv_proj(hidden_states)
-        query_states, key_states, value_states = qkv_states.split(
+        query_states, key_states, value_states = qkv_states.chunk(3, dim=-1)
-            [self.qkv_size] * 3, dim=-1)
 
-        attn_output = self.attn_fn(
+        query_states = query_states.view(batch_size, q_len,
-            q=query_states,
+                                         self.num_heads_per_partition,
-            k=key_states,
+                                         self.head_dim)
-            v=value_states,
+        key_states = key_states.view(batch_size, q_len,
-            batch_size=batch_size,
+                                     self.num_heads_per_partition,
-            q_len=q_len,
+                                     self.head_dim)
-        )
+        value_states = value_states.view(batch_size, q_len,
+                                         self.num_heads_per_partition,
+                                         self.head_dim)
 
-        attn_output, _ = self.out_proj(attn_output)
+        out = xops.memory_efficient_attention_forward(query_states,
-        return attn_output
+                                                      key_states,
-
+                                                      value_states,
-    def _basic_attention_forward(self, q, k, v, batch_size, q_len):
+                                                      p=self.dropout,
-        q = q.view(batch_size, q_len, self.num_heads,
+                                                      scale=self.scale)
-                   self.head_dim).transpose(1, 2)
+        out = out.view(batch_size, q_len, -1)
-        k = k.view(batch_size, q_len, self.num_heads,
+        attn_output, _ = self.out_proj(out)
-                   self.head_dim).transpose(1, 2)
-        v = v.view(batch_size, q_len, self.num_heads,
-                   self.head_dim).transpose(1, 2)
-
-        k_v_seq_len = k.shape[-2]
-        attn_weights = torch.matmul(q, k.transpose(2, 3)) * self.scale
-
-        if attn_weights.size() != (
-                batch_size,
-                self.num_heads,
-                q_len,
-                k_v_seq_len,
-        ):
-            raise ValueError(
-                "Attention weights should be of size "
-                f"{(batch_size, self.num_heads, q_len, k_v_seq_len)}, but is"
-                f" {attn_weights.size()}")
-
-        # upcast attention to fp32
-        attn_weights = nn.functional.softmax(attn_weights,
-                                             dim=-1,
-                                             dtype=torch.float32).to(q.dtype)
-        attn_weights = nn.functional.dropout(attn_weights,
-                                             p=self.dropout,
-                                             training=self.training)
-        attn_output = torch.matmul(attn_weights, v)
-
-        if attn_output.size() != (
-                batch_size,
-                self.num_heads,
-                q_len,
-                self.head_dim,
-        ):
-            raise ValueError(
-                "`attn_output` should be of size "
-                f"{(batch_size, self.num_heads, q_len, self.head_dim)}, but is"
-                f" {attn_output.size()}")
-
-        attn_output = attn_output.transpose(1, 2).contiguous()
-        attn_output = attn_output.reshape(batch_size, q_len, self.embed_dim)
 
         return attn_output
 
 
-# NOTE: Not used - kept for later when we TP the ViT
-# TODO(ChristopherCho): flash_attn_func is not working properly.
-#                       It constantly throws a CUDA error.
-class SiglipFlashAttention2(SiglipTPAttention):
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        self.attn_fn = self._flash_attention_forward
-
-    # Ported from https://github.com/huggingface/transformers/blob/v4.43.3/src/transformers/models/siglip/modeling_siglip.py#L449
-    # and https://github.com/huggingface/transformers/blob/v4.43.3/src/transformers/modeling_flash_attention_utils.py#L133
-    def _flash_attention_forward(self, q, k, v, batch_size, q_len, *args,
-                                 **kwargs):
-        """Implements the multihead softmax attention.
-        Arguments
-        ---------
-            q, k, v: The tensor containing the
-                     query, key, and value. (B, S, H, D)
-        """
-
-        q = q.view(batch_size, q_len, self.num_heads, self.head_dim)
-        k = k.view(batch_size, q_len, self.num_heads, self.head_dim)
-        v = v.view(batch_size, q_len, self.num_heads, self.head_dim)
-
-        attn_output = flash_attn_func(
-            q,
-            k,
-            v,
-            dropout_p=self.dropout,
-            causal=False,
-        )
-
-        attn_output = attn_output.reshape(batch_size, q_len,
-                                          self.embed_dim).contiguous()
-
-        return attn_output
-
-
-# NOTE: Not used - kept for later when we TP the ViT
-class SiglipSdpaAttention(SiglipTPAttention):
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        self.is_causal = False
-        self.attn_fn = self._sdpa_attention_forward
-
-    def _sdpa_attention_forward(self, q, k, v, batch_size, q_len):
-        q = q.view(batch_size, q_len, self.num_heads,
-                   self.head_dim).transpose(1, 2)
-        k = k.view(batch_size, q_len, self.num_heads,
-                   self.head_dim).transpose(1, 2)
-        v = v.view(batch_size, q_len, self.num_heads,
-                   self.head_dim).transpose(1, 2)
-
-        attn_output = torch.nn.functional.scaled_dot_product_attention(
-            q, k, v, dropout_p=self.dropout, is_causal=False, scale=self.scale)
-
-        attn_output = attn_output.transpose(1, 2).contiguous()
-        attn_output = attn_output.view(batch_size, q_len, self.embed_dim)
-
-        return attn_output
-
-
-# NOTE: Not used - kept for later when we TP the ViT
-class SiglipxFormersAttention(SiglipTPAttention):
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        self.attn_fn = self._xformers_attention_forward
-
-    def _xformers_attention_forward(self, q, k, v, batch_size, q_len):
-        q = q.view(batch_size, q_len, self.num_heads, self.head_dim)
-        k = k.view(batch_size, q_len, self.num_heads, self.head_dim)
-        v = v.view(batch_size, q_len, self.num_heads, self.head_dim)
-
-        attn_output = memory_efficient_attention(q,
-                                                 k,
-                                                 v,
-                                                 p=0.0,
-                                                 scale=self.scale)
-        attn_output = attn_output.reshape(batch_size, q_len,
-                                          self.embed_dim).contiguous()
-
-        return attn_output
-
-
-# NOTE: Not used - kept for later when we TP the ViT
-SIGLIP_ATTENTION_CLASSES = {
-    "eager": SiglipTPAttention,
-    "flash_attention_2": SiglipFlashAttention2,
-    "sdpa": SiglipSdpaAttention,
-    "xformers": SiglipxFormersAttention,
-}
-
-
 class SiglipMLP(nn.Module):
 
     def __init__(
@@ -473,8 +327,7 @@ class SiglipEncoderLayer(nn.Module):
         super().__init__()
         self.embed_dim = config.hidden_size
 
-        # TODO(ChristopherCho): use TP'ed Attention block
+        self.self_attn = SiglipAttention(config, quant_config=quant_config)
-        self.self_attn = SiglipAttention(config)
         self.layer_norm1 = nn.LayerNorm(self.embed_dim,
                                         eps=config.layer_norm_eps)
         self.mlp = SiglipMLP(
@@ -491,7 +344,7 @@ class SiglipEncoderLayer(nn.Module):
         residual = hidden_states
 
         hidden_states = self.layer_norm1(hidden_states)
-        hidden_states, _ = self.self_attn(hidden_states=hidden_states)
+        hidden_states = self.self_attn(hidden_states=hidden_states)
         hidden_states = residual + hidden_states
 
         residual = hidden_states