[Bugfix][VLM] Add fallback to SDPA for ViT model running on CPU backend (#8061)

vllm/model_executor/models/blip.py

@@ -7,7 +7,7 @@ import torch
 import torch.nn as nn
 from PIL import Image
 from transformers import Blip2VisionConfig, BlipVisionConfig
-from xformers import ops as xops
+from transformers.models.blip.modeling_blip import BlipAttention
 
 from vllm.config import ModelConfig
 from vllm.distributed import divide, get_tensor_model_parallel_world_size
@@ -21,6 +21,12 @@ from vllm.multimodal.utils import (cached_get_tokenizer,
                                    repeat_and_pad_placeholder_tokens)
 from vllm.sequence import VLLM_TOKEN_ID_ARRAY_TYPE, SequenceData
 
+try:
+    from xformers import ops as xops
+    USE_XFORMERS_OPS = True
+except ImportError:
+    USE_XFORMERS_OPS = False
+
 
 def get_blip_patch_grid_length(*, image_size: int, patch_size: int) -> int:
     assert image_size % patch_size == 0
@@ -156,7 +162,7 @@ class BlipVisionEmbeddings(nn.Module):
         return embeddings
 
 
-class BlipAttention(nn.Module):
+class BlipParallelAttention(nn.Module):
     """Multi-headed attention from 'Attention Is All You Need' paper"""
 
     def __init__(
@@ -224,7 +230,7 @@ class BlipAttention(nn.Module):
         out = out.view(bsz, tgt_len, -1)
         attn_output, _ = self.projection(out)
 
-        return attn_output
+        return attn_output, None
 
 
 class BlipMLP(nn.Module):
@@ -261,7 +267,16 @@ class BlipEncoderLayer(nn.Module):
                  quant_config: Optional[QuantizationConfig] = None):
         super().__init__()
 
-        self.self_attn = BlipAttention(config, quant_config=quant_config)
+        # fallback to sdpa attention if tp unavailable
+        num_heads = config.num_attention_heads
+        tp_size = get_tensor_model_parallel_world_size()
+        if USE_XFORMERS_OPS and num_heads % tp_size == 0:
+            self.self_attn = BlipParallelAttention(config,
+                                                   quant_config=quant_config)
+        else:
+            # Blip doesn't have SDPA attention implemented in transformers
+            # use eager attention instead for cpu backend
+            self.self_attn = BlipAttention(config)
         self.layer_norm1 = nn.LayerNorm(config.hidden_size,
                                         eps=config.layer_norm_eps)
         self.mlp = BlipMLP(config, quant_config=quant_config)
@@ -272,7 +287,7 @@ 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/clip.py

@@ -7,7 +7,7 @@ import torch
 import torch.nn as nn
 from PIL import Image
 from transformers import CLIPVisionConfig
-from xformers import ops as xops
+from transformers.models.clip.modeling_clip import CLIPSdpaAttention
 
 from vllm.config import ModelConfig
 from vllm.distributed import divide, get_tensor_model_parallel_world_size
@@ -22,6 +22,12 @@ from vllm.multimodal.utils import (cached_get_tokenizer,
                                    repeat_and_pad_placeholder_tokens)
 from vllm.sequence import VLLM_TOKEN_ID_ARRAY_TYPE, SequenceData
 
+try:
+    from xformers import ops as xops
+    USE_XFORMERS_OPS = True
+except ImportError:
+    USE_XFORMERS_OPS = False
+
 
 def get_clip_patch_grid_length(*, image_size: int, patch_size: int) -> int:
     assert image_size % patch_size == 0
@@ -162,7 +168,7 @@ class CLIPVisionEmbeddings(nn.Module):
         return embeddings
 
 
-class CLIPAttention(nn.Module):
+class CLIPParallelAttention(nn.Module):
     """Multi-headed attention from 'Attention Is All You Need' paper"""
 
     def __init__(
@@ -231,7 +237,7 @@ class CLIPAttention(nn.Module):
         out = out.view(bsz, tgt_len, -1)
         attn_output, _ = self.out_proj(out)
 
-        return attn_output
+        return attn_output, None
 
 
 class CLIPMLP(nn.Module):
@@ -266,7 +272,13 @@ class CLIPEncoderLayer(nn.Module):
                  quant_config: Optional[QuantizationConfig] = None):
         super().__init__()
 
-        self.self_attn = CLIPAttention(config, quant_config=quant_config)
+        num_heads = config.num_attention_heads
+        tp_size = get_tensor_model_parallel_world_size()
+        if USE_XFORMERS_OPS and num_heads % tp_size == 0:
+            self.self_attn = CLIPParallelAttention(config,
+                                                   quant_config=quant_config)
+        else:
+            self.self_attn = CLIPSdpaAttention(config)
         self.layer_norm1 = nn.LayerNorm(config.hidden_size,
                                         eps=config.layer_norm_eps)
         self.mlp = CLIPMLP(config, quant_config=quant_config)
@@ -278,7 +290,7 @@ 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
@@ -365,6 +377,10 @@ class CLIPVisionModel(nn.Module):
                  quant_config: Optional[QuantizationConfig] = None,
                  num_hidden_layers_override: Optional[int] = None):
         super().__init__()
+        tp_size = get_tensor_model_parallel_world_size()
+        num_heads = config.num_attention_heads
+        self.shard_weight = USE_XFORMERS_OPS and num_heads % tp_size == 0
+
         self.vision_model = CLIPVisionTransformer(
             config=config,
             quant_config=quant_config,
@@ -386,7 +402,7 @@ class CLIPVisionModel(nn.Module):
             ("qkv_proj", "q_proj", "q"),
             ("qkv_proj", "k_proj", "k"),
             ("qkv_proj", "v_proj", "v"),
-        ]
+        ] if self.shard_weight else []
         params_dict = dict(self.named_parameters())
         layer_count = len(self.vision_model.encoder.layers)
 

vllm/model_executor/models/intern_vit.py

@@ -10,7 +10,6 @@ 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
@@ -21,6 +20,12 @@ from vllm.model_executor.layers.linear import (ColumnParallelLinear,
 from vllm.model_executor.layers.quantization import QuantizationConfig
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
+try:
+    from xformers import ops as xops
+    USE_XFORMERS_OPS = True
+except ImportError:
+    USE_XFORMERS_OPS = False
+
 NORM2FN = {
     'rms_norm': RMSNorm,
     'layer_norm': nn.LayerNorm,
@@ -81,7 +86,7 @@ class InternVisionEmbeddings(nn.Module):
         return embeddings
 
 
-class InternAttention(nn.Module):
+class InternParallelAttention(nn.Module):
     """Multi-headed attention from 'Attention Is All You Need' paper"""
 
     def __init__(
@@ -140,18 +145,67 @@ class InternAttention(nn.Module):
             k = self.k_norm.forward_native(k.flatten(-2,
                                                      -1)).view(B_, N_, H_, D_)
 
-        x = xops.memory_efficient_attention_forward(
-            q,
-            k,
-            v,
-            scale=self.scale,
-        )
+        x = xops.memory_efficient_attention_forward(q, k, v, scale=self.scale)
         x = x.view(B, N, -1)
 
         x, _ = self.proj(x)
         return x
 
 
+class InternSdpaAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: PretrainedConfig):
+        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(
+                f'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.qkv = nn.Linear(self.embed_dim,
+                             3 * self.embed_dim,
+                             bias=config.qkv_bias)
+
+        self.qk_normalization = config.qk_normalization
+
+        if self.qk_normalization:
+            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)
+
+    def forward(self, x):
+        B, N, C = x.shape
+        qkv = self.qkv(x)
+        q, k, v = qkv.chunk(3, dim=-1)
+
+        q = q.view(B, N, self.num_heads, self.head_dim)
+        k = k.view(B, N, self.num_heads, self.head_dim)
+        v = v.view(B, N, self.num_heads, self.head_dim)
+
+        if self.qk_normalization:
+            B_, N_, H_, D_ = q.shape
+            q = self.q_norm.forward_native(q.flatten(-2,
+                                                     -1)).view(B_, N_, H_, D_)
+            k = self.k_norm.forward_native(k.flatten(-2,
+                                                     -1)).view(B_, N_, H_, D_)
+        q = q.transpose(1, 2)
+        k = k.transpose(1, 2)
+        v = v.transpose(1, 2)
+
+        x = F.scaled_dot_product_attention(q, k, v, scale=self.scale)
+        x = x.transpose(1, 2).view(B, N, -1)
+
+        x = self.proj(x)
+        return x
+
+
 class InternMLP(nn.Module):
 
     def __init__(self,
@@ -187,7 +241,14 @@ class InternVisionEncoderLayer(nn.Module):
         self.intermediate_size = config.intermediate_size
         self.norm_type = config.norm_type
 
-        self.attn = InternAttention(config, quant_config=quant_config)
+        # fallback to sdpa attention if tp unavailable
+        tp_size = get_tensor_model_parallel_world_size()
+        num_heads = config.num_attention_heads
+        if USE_XFORMERS_OPS and num_heads % tp_size == 0:
+            self.attn = InternParallelAttention(config,
+                                                quant_config=quant_config)
+        else:
+            self.attn = InternSdpaAttention(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

@@ -307,26 +307,30 @@ class PaliGemmaForConditionalGeneration(nn.Module, SupportsMultiModal):
                 if key_to_modify in name:
                     name = name.replace(key_to_modify, new_key)
             use_default_weight_loading = False
-            for (param_name, shard_name, shard_id) in stacked_params_mapping:
-                if shard_name not in name:
-                    continue
-                name = name.replace(shard_name, param_name)
-                # 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
+            if "vision" not in name or self.vision_tower.shard_weight:
+                for (param_name, shard_name,
+                     shard_id) in stacked_params_mapping:
+                    if shard_name not in name:
+                        continue
+                    name = name.replace(shard_name, param_name)
+                    # 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:
+                    # 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
             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:

vllm/model_executor/models/siglip.py

@@ -9,7 +9,7 @@ import torch
 from PIL import Image
 from torch import nn
 from transformers import SiglipVisionConfig
-from xformers import ops as xops
+from transformers.models.siglip.modeling_siglip import SiglipSdpaAttention
 
 from vllm.config import ModelConfig
 from vllm.distributed import divide, get_tensor_model_parallel_world_size
@@ -26,6 +26,12 @@ from vllm.multimodal.utils import (cached_get_tokenizer,
                                    repeat_and_pad_placeholder_tokens)
 from vllm.sequence import VLLM_TOKEN_ID_ARRAY_TYPE, SequenceData
 
+try:
+    from xformers import ops as xops
+    USE_XFORMERS_OPS = True
+except ImportError:
+    USE_XFORMERS_OPS = False
+
 
 def get_siglip_patch_grid_length(*, image_size: int, patch_size: int) -> int:
     # Since interpolation is applied, the image size need not be divisible
@@ -219,7 +225,7 @@ class SiglipVisionEmbeddings(nn.Module):
         return embeddings
 
 
-class SiglipAttention(nn.Module):
+class SiglipParallelAttention(nn.Module):
 
     def __init__(
         self,
@@ -282,7 +288,7 @@ class SiglipAttention(nn.Module):
         out = out.view(batch_size, q_len, -1)
         attn_output, _ = self.out_proj(out)
 
-        return attn_output
+        return attn_output, None
 
 
 class SiglipMLP(nn.Module):
@@ -327,7 +333,14 @@ class SiglipEncoderLayer(nn.Module):
         super().__init__()
         self.embed_dim = config.hidden_size
 
-        self.self_attn = SiglipAttention(config, quant_config=quant_config)
+        num_heads = config.num_attention_heads
+        tp_size = get_tensor_model_parallel_world_size()
+        if USE_XFORMERS_OPS and num_heads % tp_size == 0:
+            self.self_attn = SiglipParallelAttention(config,
+                                                     quant_config=quant_config)
+        else:
+            self.self_attn = SiglipSdpaAttention(config)
+
         self.layer_norm1 = nn.LayerNorm(self.embed_dim,
                                         eps=config.layer_norm_eps)
         self.mlp = SiglipMLP(
@@ -344,7 +357,7 @@ 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
@@ -476,6 +489,10 @@ class SiglipVisionModel(nn.Module):
         num_hidden_layers_override: Optional[int] = None,
     ):
         super().__init__()
+        num_heads = config.num_attention_heads
+        tp_size = get_tensor_model_parallel_world_size()
+        self.shard_weight = USE_XFORMERS_OPS and num_heads % tp_size == 0
+
         self.vision_model = SiglipVisionTransformer(
             config,
             quant_config,