[Bugfix] Fix FP8 KV cache support (#4869)

vllm/attention/backends/flash_attn.py

@@ -200,15 +200,15 @@ class FlashAttentionImpl(AttentionImpl):
         num_heads: int,
         head_size: int,
         scale: float,
-        num_kv_heads: Optional[int] = None,
-        alibi_slopes: Optional[List[float]] = None,
-        sliding_window: Optional[int] = None,
-        kv_cache_dtype: str = "auto",
+        num_kv_heads: int,
+        alibi_slopes: Optional[List[float]],
+        sliding_window: Optional[int],
+        kv_cache_dtype: str,
     ) -> None:
         self.num_heads = num_heads
         self.head_size = head_size
         self.scale = float(scale)
-        self.num_kv_heads = num_heads if num_kv_heads is None else num_kv_heads
+        self.num_kv_heads = num_kv_heads
         if alibi_slopes is not None:
             alibi_slopes = torch.tensor(alibi_slopes, dtype=torch.float32)
         self.alibi_slopes = alibi_slopes

vllm/attention/backends/flashinfer.py

@@ -164,15 +164,15 @@ class FlashInferImpl(AttentionImpl):
         num_heads: int,
         head_size: int,
         scale: float,
-        num_kv_heads: Optional[int] = None,
-        alibi_slopes: Optional[List[float]] = None,
-        sliding_window: Optional[int] = None,
-        kv_cache_dtype: str = "auto",
+        num_kv_heads: int,
+        alibi_slopes: Optional[List[float]],
+        sliding_window: Optional[int],
+        kv_cache_dtype: str,
     ) -> None:
         self.num_heads = num_heads
         self.head_size = head_size
         self.scale = float(scale)
-        self.num_kv_heads = num_heads if num_kv_heads is None else num_kv_heads
+        self.num_kv_heads = num_kv_heads
         if alibi_slopes is not None:
             alibi_slopes = torch.tensor(alibi_slopes, dtype=torch.float32)
         self.alibi_slopes = alibi_slopes

vllm/attention/backends/rocm_flash_attn.py

@@ -197,15 +197,15 @@ class ROCmFlashAttentionImpl(AttentionImpl):
         num_heads: int,
         head_size: int,
         scale: float,
-        num_kv_heads: Optional[int] = None,
-        alibi_slopes: Optional[List[float]] = None,
-        sliding_window: Optional[int] = None,
-        kv_cache_dtype: str = "auto",
+        num_kv_heads: int,
+        alibi_slopes: Optional[List[float]],
+        sliding_window: Optional[int],
+        kv_cache_dtype: str,
     ) -> None:
         self.num_heads = num_heads
         self.head_size = head_size
         self.scale = float(scale)
-        self.num_kv_heads = num_heads if num_kv_heads is None else num_kv_heads
+        self.num_kv_heads = num_kv_heads
         if alibi_slopes is not None:
             alibi_slopes = torch.tensor(alibi_slopes, dtype=torch.float32)
         self.alibi_slopes = alibi_slopes

vllm/attention/backends/torch_sdpa.py

@@ -96,15 +96,15 @@ class TorchSDPABackendImpl(AttentionImpl[TorchSDPAMetadata]):
         num_heads: int,
         head_size: int,
         scale: float,
-        num_kv_heads: Optional[int] = None,
-        alibi_slopes: Optional[List[float]] = None,
-        sliding_window: Optional[int] = None,
-        kv_cache_dtype: str = "auto",
+        num_kv_heads: int,
+        alibi_slopes: Optional[List[float]],
+        sliding_window: Optional[int],
+        kv_cache_dtype: str,
     ) -> None:
         self.num_heads = num_heads
         self.head_size = head_size
         self.scale = float(scale)
-        self.num_kv_heads = num_heads if num_kv_heads is None else num_kv_heads
+        self.num_kv_heads = num_kv_heads
         if alibi_slopes is not None:
             alibi_slopes = torch.tensor(alibi_slopes, dtype=torch.float32)
         self.alibi_slopes = alibi_slopes

vllm/attention/backends/xformers.py

@@ -208,15 +208,15 @@ class XFormersImpl(AttentionImpl[XFormersMetadata]):
         num_heads: int,
         head_size: int,
         scale: float,
-        num_kv_heads: Optional[int] = None,
-        alibi_slopes: Optional[List[float]] = None,
-        sliding_window: Optional[int] = None,
-        kv_cache_dtype: str = "auto",
+        num_kv_heads: int,
+        alibi_slopes: Optional[List[float]],
+        sliding_window: Optional[int],
+        kv_cache_dtype: str,
     ) -> None:
         self.num_heads = num_heads
         self.head_size = head_size
         self.scale = float(scale)
-        self.num_kv_heads = num_heads if num_kv_heads is None else num_kv_heads
+        self.num_kv_heads = num_kv_heads
         if alibi_slopes is not None:
             alibi_slopes = torch.tensor(alibi_slopes, dtype=torch.float32)
         self.alibi_slopes = alibi_slopes

vllm/attention/layer.py

@@ -48,7 +48,7 @@ class Attention(nn.Module):
                                         block_size)
         impl_cls = attn_backend.get_impl_cls()
         self.impl = impl_cls(num_heads, head_size, scale, num_kv_heads,
-                             alibi_slopes, sliding_window)
+                             alibi_slopes, sliding_window, kv_cache_dtype)
 
     def forward(
         self,