[VLM] Add MLA with pure RoPE support for deepseek-vl2 models (#12729)

vllm/attention/backends/mla/utils.py

@@ -26,7 +26,8 @@ from vllm.model_executor.layers.quantization.utils.fp8_utils import (
     apply_fp8_linear_generic, current_platform_fp8_dtype, is_fp8)
 from vllm.model_executor.layers.quantization.utils.quant_utils import (
     scaled_dequantize, scaled_quantize)
-from vllm.model_executor.layers.rotary_embedding import RotaryEmbedding
+from vllm.model_executor.layers.rotary_embedding import (
+    DeepseekScalingRotaryEmbedding, RotaryEmbedding)
 
 try:
     from vllm.vllm_flash_attn import flash_attn_varlen_func
@@ -174,6 +175,8 @@ class MLACommonImpl(MLAAttentionImpl[T], Generic[T]):
         self.v_head_dim = v_head_dim
 
         self.rotary_emb = rotary_emb
+        self.use_yarn_rope = isinstance(rotary_emb,
+                                        DeepseekScalingRotaryEmbedding)
         self.q_proj = q_proj
         self.kv_b_proj = kv_b_proj
         self.o_proj = o_proj
@@ -420,6 +423,24 @@ class MLACommonImpl(MLAAttentionImpl[T], Generic[T]):
     ) -> torch.Tensor:
         raise NotImplementedError
 
+    def apply_pure_rope(
+        self,
+        input_positions: torch.Tensor,
+        q_pe: torch.Tensor,
+        k_pe: torch.Tensor,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        seq_len = input_positions.size(0)
+        ori_q_pe_shape, ori_k_pe_shape = q_pe.shape, k_pe.shape
+
+        q_pe, k_pe = self.rotary_emb(
+            input_positions,
+            q_pe.reshape(seq_len, -1),
+            k_pe.reshape(seq_len, -1),
+        )
+        q_pe, k_pe = q_pe.view(ori_q_pe_shape), k_pe.view(ori_k_pe_shape)
+
+        return q_pe, k_pe
+
     def forward(
         self,
         layer: AttentionLayer,
@@ -444,13 +465,14 @@ class MLACommonImpl(MLAAttentionImpl[T], Generic[T]):
         # Restore head dim (for rotary embedding)
         k_pe = k_pe.unsqueeze(1)
         assert hasattr(attn_metadata, "input_positions")
+        rope_fn = (self.rotary_emb
+                   if self.use_yarn_rope else self.apply_pure_rope)
 
         if is_decode:
             q_nope = self._q_proj_and_k_up_proj(hidden_states_or_q_c)
             q_pe = torch.matmul(hidden_states_or_q_c, self.W_QR)\
                 .view(-1, self.num_heads, self.qk_rope_head_dim)
-            q_pe, k_pe = \
-                self.rotary_emb(attn_metadata.input_positions, q_pe, k_pe)
+            q_pe, k_pe = rope_fn(attn_metadata.input_positions, q_pe, k_pe)
         else:
             assert is_prefill
             q = self.q_proj(hidden_states_or_q_c)[0]\
@@ -458,7 +480,7 @@ class MLACommonImpl(MLAAttentionImpl[T], Generic[T]):
 
             # TODO(lucas): there must be a nicer way to write this line
             q[..., self.qk_nope_head_dim:], k_pe = \
-                self.rotary_emb(
+                rope_fn(
                     attn_metadata.input_positions,
                     q[..., self.qk_nope_head_dim:], k_pe)
 

vllm/model_executor/models/deepseek_v2.py

@@ -414,7 +414,8 @@ class DeepseekV2MLAAttention(nn.Module):
                                         quant_config=quant_config,
                                         prefix=f"{prefix}.o_proj")
 
-        rope_scaling["rope_type"] = 'deepseek_yarn'
+        if rope_scaling:
+            rope_scaling["rope_type"] = 'deepseek_yarn'
         self.rotary_emb = get_rope(qk_rope_head_dim,
                                    rotary_dim=qk_rope_head_dim,
                                    max_position=max_position_embeddings,

vllm/model_executor/models/deepseek_v3.py

@@ -422,7 +422,8 @@ class DeepseekV3MLAAttention(nn.Module):
                                         quant_config=quant_config,
                                         prefix=f"{prefix}.o_proj")
 
-        rope_scaling["rope_type"] = 'deepseek_yarn'
+        if rope_scaling:
+            rope_scaling["rope_type"] = 'deepseek_yarn'
         self.rotary_emb = get_rope(qk_rope_head_dim,
                                    rotary_dim=qk_rope_head_dim,
                                    max_position=max_position_embeddings,