[Misc] Qwen2.5-VL Optimization (#13155)

vllm/model_executor/models/qwen2_5_vl.py

@@ -45,6 +45,7 @@ from vllm.distributed import utils as dist_utils
 from vllm.logger import init_logger
 from vllm.model_executor import SamplingMetadata
 from vllm.model_executor.layers.activation import _ACTIVATION_REGISTRY
+from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.linear import (ColumnParallelLinear,
                                                RowParallelLinear)
 from vllm.model_executor.layers.quantization import QuantizationConfig
@@ -271,8 +272,13 @@ class Qwen2_5_VisionAttention(nn.Module):
         q, k, v = (rearrange(x, "s b ... -> b s ...").contiguous()
                    for x in (q, k, v))
         if rotary_pos_emb is not None:
-            q = apply_rotary_pos_emb_vision(q, rotary_pos_emb)
-            k = apply_rotary_pos_emb_vision(k, rotary_pos_emb)
+            use_flash_attn = self.attn_backend == _Backend.FLASH_ATTN
+            q = apply_rotary_pos_emb_vision(q,
+                                            rotary_pos_emb,
+                                            use_flash_attn=use_flash_attn)
+            k = apply_rotary_pos_emb_vision(k,
+                                            rotary_pos_emb,
+                                            use_flash_attn=use_flash_attn)
 
         if self.attn_backend == _Backend.FLASH_ATTN:
             # from vllm_flash_attn.flash_attn_interface import (
@@ -296,20 +302,23 @@ class Qwen2_5_VisionAttention(nn.Module):
                                       "(b s) ... -> b s ...",
                                       b=batch_size)
         elif self.attn_backend == _Backend.TORCH_SDPA:
-            seq_length = q.size(1)
-            q, k, v = (rearrange(x, "b s h d -> b h s d") for x in [q, k, v])
-            attention_mask = torch.zeros([1, seq_length, seq_length],
-                                         device=q.device,
-                                         dtype=torch.bool)
+            # Execute attention entry by entry for speed & less VRAM.
+            outputs = []
             for i in range(1, len(cu_seqlens)):
-                attention_mask[..., cu_seqlens[i - 1]:cu_seqlens[i],
-                               cu_seqlens[i - 1]:cu_seqlens[i]] = True
-            output = F.scaled_dot_product_attention(q,
-                                                    k,
-                                                    v,
-                                                    attention_mask,
+                start_idx = cu_seqlens[i - 1]
+                end_idx = cu_seqlens[i]
+                q_i = q[:, start_idx:end_idx]
+                k_i = k[:, start_idx:end_idx]
+                v_i = v[:, start_idx:end_idx]
+                q_i, k_i, v_i = (rearrange(x, "b s h d -> b h s d")
+                                 for x in [q_i, k_i, v_i])
+                output_i = F.scaled_dot_product_attention(q_i,
+                                                          k_i,
+                                                          v_i,
                                                           dropout_p=0.0)
-            context_layer = rearrange(output, "b h s d -> b s h d ")
+                output_i = rearrange(output_i, "b h s d -> b s h d ")
+                outputs.append(output_i)
+            context_layer = torch.cat(outputs, dim=1)
         elif self.attn_backend == _Backend.XFORMERS:
             from xformers import ops as xops
             from xformers.ops.fmha.attn_bias import BlockDiagonalMask
@@ -327,25 +336,6 @@ class Qwen2_5_VisionAttention(nn.Module):
         return output
 
 
-class Qwen2RMSNorm(nn.Module):
-
-    def __init__(self, hidden_size, eps=1e-6):
-        super().__init__()
-        self.weight = nn.Parameter(torch.ones(hidden_size))
-        self.variance_epsilon = eps
-
-    def forward(self, hidden_states):
-        input_dtype = hidden_states.dtype
-        hidden_states = hidden_states.to(torch.float32)
-        variance = hidden_states.pow(2).mean(-1, keepdim=True)
-        hidden_states = hidden_states * torch.rsqrt(variance +
-                                                    self.variance_epsilon)
-        return self.weight * hidden_states.to(input_dtype)
-
-    def extra_repr(self):
-        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
-
-
 class Qwen2_5_VisionBlock(nn.Module):
 
     def __init__(
@@ -516,8 +506,7 @@ class Qwen2_5_VisionTransformer(nn.Module):
             hidden_size=self.hidden_size,
         )
 
-        # NOTE: We use torch native RMSNorm here for precision purposes.
-        norm_layer = partial(Qwen2RMSNorm, eps=norm_eps)
+        norm_layer = partial(RMSNorm, eps=norm_eps)
         head_dim = self.hidden_size // self.num_heads
         self.rotary_pos_emb = Qwen2_5_VisionRotaryEmbedding(head_dim // 2)
 

vllm/model_executor/models/qwen2_vl.py

@@ -226,11 +226,15 @@ def apply_rotary_emb_torch(x: torch.Tensor,
 
 
 def apply_rotary_pos_emb_vision(t: torch.Tensor,
-                                freqs: torch.Tensor) -> torch.Tensor:
+                                freqs: torch.Tensor,
+                                use_flash_attn=False) -> torch.Tensor:
     t_ = t.float()
     cos = freqs.cos()
     sin = freqs.sin()
-    output = apply_rotary_emb_torch(t_, cos, sin).type_as(t)
+    apply_rotary_emb = apply_rotary_emb_torch
+    if use_flash_attn:
+        from flash_attn.layers.rotary import apply_rotary_emb
+    output = apply_rotary_emb(t_, cos, sin).type_as(t)
     return output
 
 
@@ -336,20 +340,23 @@ class Qwen2VisionAttention(nn.Module):
                                       "(b s) ... -> b s ...",
                                       b=batch_size)
         elif self.attn_backend == _Backend.TORCH_SDPA:
-            seq_length = q.size(1)
-            q, k, v = (rearrange(x, "b s h d -> b h s d") for x in [q, k, v])
-            attention_mask = torch.zeros([1, seq_length, seq_length],
-                                         device=q.device,
-                                         dtype=torch.bool)
+            # Execute attention entry by entry for speed & less VRAM.
+            outputs = []
             for i in range(1, len(cu_seqlens)):
-                attention_mask[..., cu_seqlens[i - 1]:cu_seqlens[i],
-                               cu_seqlens[i - 1]:cu_seqlens[i]] = True
-            output = F.scaled_dot_product_attention(q,
-                                                    k,
-                                                    v,
-                                                    attention_mask,
+                start_idx = cu_seqlens[i - 1]
+                end_idx = cu_seqlens[i]
+                q_i = q[:, start_idx:end_idx]
+                k_i = k[:, start_idx:end_idx]
+                v_i = v[:, start_idx:end_idx]
+                q_i, k_i, v_i = (rearrange(x, "b s h d -> b h s d")
+                                 for x in [q_i, k_i, v_i])
+                output_i = F.scaled_dot_product_attention(q_i,
+                                                          k_i,
+                                                          v_i,
                                                           dropout_p=0.0)
-            context_layer = rearrange(output, "b h s d -> b s h d ")
+                output_i = rearrange(output_i, "b h s d -> b s h d ")
+                outputs.append(output_i)
+            context_layer = torch.cat(outputs, dim=1)
         elif self.attn_backend == _Backend.XFORMERS:
             from xformers import ops as xops
             from xformers.ops.fmha.attn_bias import BlockDiagonalMask