[Bugfix] Remove hardcoded head_size=256 for Deepseek v2 and v3 (#12067)

Signed-off-by: Isotr0py <2037008807@qq.com>

tests/kernels/test_attention.py

@@ -31,9 +31,9 @@ NUM_GEN_SEQS = [7]  # Arbitrary values for testing
 NUM_PREFILL_SEQS = [3]  # Arbitrary values for testing
 NUM_HEADS = [(40, 40), (64, 8)]  # Arbitrary values for testing
 
-# FlashAttention forward only supports head dimension at most 128
+# This should be sync with get_supported_head_sizes() in
-# https://github.com/ROCmSoftwarePlatform/flash-attention/blob/3d2b6f5d037782cc2c906909a46fb7e2e1b48b25/csrc/flash_attn_rocm/flash_api.cpp#L62
+# vllm.attention.ops.paged_attn.PagedAttention
-HEAD_SIZES = [64, 80, 120, 256]
+HEAD_SIZES = [32, 64, 80, 96, 112, 120, 128, 192, 256]
 
 BLOCK_SIZES = [16, 32]
 USE_ALIBI = [False, True]

vllm/config.py

@@ -733,9 +733,12 @@ class ModelConfig:
         if hasattr(self.hf_text_config,
                    "model_type") and (self.hf_text_config.model_type
                                       in ('deepseek_v2', 'deepseek_v3')):
-            # FlashAttention supports only head_size 32, 64, 128, 256,
+            qk_rope_head_dim = getattr(self.hf_text_config, "qk_rope_head_dim",
-            # we need to pad head_size 192 to 256
+                                       0)
-            return 256
+            qk_nope_head_dim = getattr(self.hf_text_config, "qk_nope_head_dim",
+                                       0)
+            if qk_rope_head_dim and qk_nope_head_dim:
+                return qk_rope_head_dim + qk_nope_head_dim
 
         if self.is_attention_free:
             return 0

vllm/model_executor/models/deepseek_v2.py

@@ -262,14 +262,8 @@ class DeepseekV2Attention(nn.Module):
             mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
             self.scaling = self.scaling * mscale * mscale
 
-        # self.attn = Attention(self.num_heads,
-        #                       self.qk_head_dim,
-        #                       self.scaling,
-        #                       num_kv_heads=self.num_heads)
-
-        # TODO, support head_size 192
         self.attn = Attention(self.num_local_heads,
-                              256,
+                              self.qk_head_dim,
                               self.scaling,
                               num_kv_heads=self.num_local_heads,
                               cache_config=cache_config,
@@ -319,18 +313,14 @@ class DeepseekV2Attention(nn.Module):
         k = torch.empty_like(q)
         k[..., :self.qk_nope_head_dim] = k_nope
         k[..., self.qk_nope_head_dim:] = k_pe
-        q = torch.nn.functional.pad(q, [0, 256 - self.qk_head_dim],
+        # padding value to qk_head_dim for alignment
-                                    value=0).view(-1,
+        v = torch.nn.functional.pad(
-                                                  self.num_local_heads * 256)
+            v, [0, self.qk_head_dim - self.v_head_dim],
-        k = torch.nn.functional.pad(k, [0, 256 - self.qk_head_dim],
+            value=0).view(-1, self.num_local_heads * self.qk_head_dim)
-                                    value=0).view(-1,
-                                                  self.num_local_heads * 256)
-        v = torch.nn.functional.pad(v, [0, 256 - self.v_head_dim],
-                                    value=0).view(-1,
-                                                  self.num_local_heads * 256)
         attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
         attn_output = attn_output.view(
-            -1, self.num_local_heads, 256)[..., :self.v_head_dim].reshape(
+            -1, self.num_local_heads,
+            self.qk_head_dim)[..., :self.v_head_dim].reshape(
                 -1, self.num_local_heads * self.v_head_dim)
         output, _ = self.o_proj(attn_output)
         return output

vllm/model_executor/models/deepseek_v3.py

@@ -269,14 +269,8 @@ class DeepseekV3Attention(nn.Module):
             mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
             self.scaling = self.scaling * mscale * mscale
 
-        # self.attn = Attention(self.num_heads,
-        #                       self.qk_head_dim,
-        #                       self.scaling,
-        #                       num_kv_heads=self.num_heads)
-
-        # TODO, support head_size 192
         self.attn = Attention(self.num_local_heads,
-                              256,
+                              self.qk_head_dim,
                               self.scaling,
                               num_kv_heads=self.num_local_heads,
                               cache_config=cache_config,
@@ -326,18 +320,14 @@ class DeepseekV3Attention(nn.Module):
         k = torch.empty_like(q)
         k[..., :self.qk_nope_head_dim] = k_nope
         k[..., self.qk_nope_head_dim:] = k_pe
-        q = torch.nn.functional.pad(q, [0, 256 - self.qk_head_dim],
+        # padding value to qk_head_dim for alignment
-                                    value=0).view(-1,
+        v = torch.nn.functional.pad(
-                                                  self.num_local_heads * 256)
+            v, [0, self.qk_head_dim - self.v_head_dim],
-        k = torch.nn.functional.pad(k, [0, 256 - self.qk_head_dim],
+            value=0).view(-1, self.num_local_heads * self.qk_head_dim)
-                                    value=0).view(-1,
-                                                  self.num_local_heads * 256)
-        v = torch.nn.functional.pad(v, [0, 256 - self.v_head_dim],
-                                    value=0).view(-1,
-                                                  self.num_local_heads * 256)
         attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
         attn_output = attn_output.view(
-            -1, self.num_local_heads, 256)[..., :self.v_head_dim].reshape(
+            -1, self.num_local_heads,
+            self.qk_head_dim)[..., :self.v_head_dim].reshape(
                 -1, self.num_local_heads * self.v_head_dim)
         output, _ = self.o_proj(attn_output)
         return output