[Metax][Optimization] 优化 PaddleOCR-VL 在 Metax GPU 上的视觉路径开销

fastdeploy/model_executor/models/paddleocr_vl/projector.py

@@ -17,6 +17,7 @@
 import math
 from typing import Optional
 
+import numpy as np
 import paddle
 import paddle.nn as nn
 
@@ -63,30 +64,41 @@ def __init__(self, text_config, vision_config, prefix=""):
         self.linear_2 = nn.Linear(self.hidden_size, self.text_config.hidden_size)
         self.linear_2.weight.weight_loader = self.weight_loader
 
-    def forward(self, image_features, image_grid_thw):
+    def _build_merge_permutation(self, image_grid_thw):
         m1, m2 = self.merge_kernel_size
+        if isinstance(image_grid_thw, paddle.Tensor):
+            image_grid_thw = image_grid_thw.cpu().numpy()
+
+        merge_indices = []
+        merge_lengths = []
+        start = 0
+        for image_grid in image_grid_thw:
+            t, h, w = map(int, image_grid)
+            assert h % m1 == 0 and w % m2 == 0, (image_grid, self.merge_kernel_size)
+            local = np.arange(t * h * w, dtype=np.int64).reshape((t, h // m1, m1, w // m2, m2))
+            local = local.transpose((0, 1, 3, 2, 4)).reshape(-1)
+            merge_indices.append(local + start)
+            merge_lengths.append(t * (h // m1) * (w // m2))
+            start += t * h * w
+
+        if len(merge_indices) == 0:
+            return np.empty((0,), dtype=np.int64), merge_lengths
+        return np.concatenate(merge_indices, axis=0), merge_lengths
+
+    def forward(self, image_features, image_grid_thw, return_packed: bool = False):
         if isinstance(image_features, (list, tuple)):
-            processed_features = list()
-            for image_feature, image_grid in zip(image_features, image_grid_thw):
-                image_feature = self.pre_norm(image_feature)  # shape: (T*H*W, D)
-                t, h, w = image_grid
-                from einops import rearrange
-
-                image_feature = rearrange(
-                    image_feature,
-                    "(t h p1 w p2) d -> (t h w) (p1 p2 d)",
-                    t=int(t),
-                    h=int(h // m1),
-                    p1=int(m1),
-                    w=int(w // m2),
-                    p2=int(m2),
-                )
-                hidden_states = self.linear_1(image_feature)
-                hidden_states = self.act(hidden_states)
-                hidden_states = self.linear_2(hidden_states)
-                processed_features.append(hidden_states)
-
-            return processed_features
+            packed_image_features = image_features[0] if len(image_features) == 1 else paddle.concat(image_features, axis=0)
+            packed_image_features = self.pre_norm(packed_image_features)
+            merge_indices, merge_lengths = self._build_merge_permutation(image_grid_thw)
+            merge_indices = paddle.to_tensor(merge_indices, dtype="int64", place=packed_image_features.place)
+            packed_image_features = paddle.index_select(packed_image_features, merge_indices, axis=0)
+            hidden_states = paddle.reshape(packed_image_features, [-1, self.hidden_size])
+            hidden_states = self.linear_1(hidden_states)
+            hidden_states = self.act(hidden_states)
+            hidden_states = self.linear_2(hidden_states)
+            if return_packed:
+                return hidden_states
+            return list(paddle.split(hidden_states, merge_lengths, axis=0))
 
         dim = image_features.shape[-1]
         image_features = paddle.reshape(image_features, [-1, dim])

fastdeploy/model_executor/models/paddleocr_vl/siglip.py

@@ -127,7 +127,11 @@ def forward(
         cos_emb: Optional[paddle.Tensor] = None,  # (cos, sin)
         sin_emb: Optional[paddle.Tensor] = None,  # (cos, sin)
     ):
-        B, seq_length, D = hidden_states.shape
+        if hidden_states.dim() == 3:
+            assert hidden_states.shape[0] == 1, f"SiglipAttention only supports batch=1, got {hidden_states.shape}"
+            hidden_states = hidden_states[0]
+
+        seq_length, D = hidden_states.shape
         qkv = self.qkv_proj(hidden_states)
         q, k, v = neox_rope_embedding(qkv, cos_emb, sin_emb, self.num_heads, self.head_dim)
         attn_output = self.flash_attn_func(
@@ -255,25 +259,26 @@ def forward(
                 flatten_image_grid_thw = self.flatten_list(image_grid_thw)
                 flatten_image_grid_thw = np.array(flatten_image_grid_thw)
                 assert batch_size == 1
-                start = 0
 
                 assert sum([np.prod(x) for x in flatten_image_grid_thw]) == embeddings.shape[1], (
                     flatten_image_grid_thw,
                     embeddings.shape,
                 )
                 embeddings = embeddings.squeeze(0)
-                tmp_embeddings = list()
-                for image_grid in image_grid_thw:
-                    t, h, w = image_grid
-                    end = start + t * h * w
-                    image_embeddings = embeddings[int(start) : int(end), :]
-                    position_embedding = (
-                        self.interpolate_pos_encoding(image_embeddings, h, w, True).squeeze(0).tile((t, 1))
-                    ).astype(image_embeddings.dtype)
-                    image_embeddings = image_embeddings + position_embedding
-                    tmp_embeddings.append(image_embeddings)
-                    start = end
-                embeddings = paddle.concat(tmp_embeddings, axis=0).unsqueeze(0)
+                packed_position_embeddings = []
+                for t, h, w in flatten_image_grid_thw:
+                    t, h, w = map(int, (t, h, w))
+                    position_embedding = self.fetch_position_embedding_lfu_cache(embeddings, h, w).squeeze(0)
+                    if t > 1:
+                        position_embedding = position_embedding.tile((t, 1))
+                    if position_embedding.dtype != embeddings.dtype:
+                        position_embedding = position_embedding.astype(embeddings.dtype)
+                    packed_position_embeddings.append(position_embedding)
+                if len(packed_position_embeddings) == 1:
+                    packed_position_embeddings = packed_position_embeddings[0]
+                else:
+                    packed_position_embeddings = paddle.concat(packed_position_embeddings, axis=0)
+                embeddings = (embeddings + packed_position_embeddings).unsqueeze(0)
             else:
                 embeddings = embeddings + self.packing_position_embedding(position_ids)
             return embeddings
@@ -307,7 +312,7 @@ def weight_loader(self, param, loaded_weight, loaded_shard_id: Optional[str] = N
 
     def forward(self, hidden_states: paddle.Tensor) -> paddle.Tensor:
         hidden_states = self.fc1(hidden_states)
-        hidden_states = get_activation_fn(self.config.hidden_act)(hidden_states[0])
+        hidden_states = get_activation_fn(self.config.hidden_act)(hidden_states)
         hidden_states = self.fc2(hidden_states)
         return hidden_states
 
@@ -331,6 +336,32 @@ def forward(
         cos_emb=None,
         sin_emb=None,
     ):
+        if hidden_states.dim() == 3 and hidden_states.shape[0] == 1:
+            hidden_states = hidden_states[0]
+
+            residual = hidden_states
+            ln1_out = self.layer_norm1(hidden_states)
+
+            x = self.self_attn(
+                hidden_states=ln1_out,
+                attention_mask=attention_mask,
+                output_attentions=output_attentions,
+                cu_seqlens=cu_seqlens,
+                max_seqlen=max_seqlen,
+                cos_emb=cos_emb,
+                sin_emb=sin_emb,
+            )
+
+            hs_post_attn = residual + x
+
+            residual = hs_post_attn
+            ln2_out = self.layer_norm2(residual)
+
+            mlp_out = self.mlp(ln2_out)
+
+            hidden_states_out = residual + mlp_out
+
+            return (hidden_states_out.unsqueeze(0),)
 
         residual = hidden_states
         ############################
@@ -677,7 +708,6 @@ def forward(
             end = cu_seqlens[i + 1]
             tensor = last_hidden_state[:, start:end, :].squeeze(0)
             sample_hidden_state.append(tensor)
-
         return sample_hidden_state
 
 

fastdeploy/model_executor/models/paddleocr_vl/siglip_ops.py

@@ -37,27 +37,33 @@ def rotate_half(x):
 
 
 def apply_rotary_pos_emb_vision(x, cos, sin):
-    orig_dtype = x.dtype
-    x = x.astype("float32")
     x_embed = (x * cos) + (rotate_half(x) * sin)
-    return x_embed.astype(orig_dtype)
+    return x_embed
 
 
 def native_neox_rope_embedding(qkv, cos, sin, num_heads):
-    B, seq_length, D = qkv.shape
-    if seq_length == -1:
-        _, seq_length, _ = paddle.shape(qkv)
-    qkv = qkv.reshape(
-        [
-            seq_length,
-            3,
-            num_heads,
-            -1,
-        ]
-    ).transpose(perm=[1, 0, 2, 3])
-    q, k, v = qkv.unbind(axis=0)
+    if qkv.dim() == 3:
+        B, seq_length, D = qkv.shape
+        if seq_length == -1:
+            _, seq_length, _ = paddle.shape(qkv)
+        token_count = B * seq_length
+    else:
+        token_count, D = qkv.shape
+        if token_count == -1:
+            token_count, _ = paddle.shape(qkv)
+    qkv = qkv.reshape([token_count, 3, num_heads, -1])
+    q_dtype = qkv.dtype
+    if q_dtype != paddle.float32:
+        qk = qkv[:, :2].astype("float32")
+        q, k = qk[:, 0], qk[:, 1]
+    else:
+        q, k = qkv[:, 0], qkv[:, 1]
+    v = qkv[:, 2]
     q = apply_rotary_pos_emb_vision(q, cos, sin)
     k = apply_rotary_pos_emb_vision(k, cos, sin)
+    if q.dtype != q_dtype:
+        q = q.astype(q_dtype)
+        k = k.astype(q_dtype)
     return q, k, v