when using VLLM_FLASH_ATTN_V1, set block_size to 64

vllm/attention/layer.py

@@ -103,7 +103,7 @@ class Attention(nn.Module):
             calculate_kv_scales = cache_config.calculate_kv_scales
         else:
             kv_cache_dtype = "auto"
-            block_size = 16 if not envs.VLLM_USE_FLASH_ATTN_PA else 64
+            block_size = 16 if not envs.VLLM_USE_FLASH_ATTN_PA or not envs.VLLM_FLASH_ATTN_V1 else 64
             is_attention_free = False
             calculate_kv_scales = False
         if num_kv_heads is None:
@@ -335,7 +335,7 @@ class MultiHeadAttention(nn.Module):
         attn_backend = get_attn_backend(head_size,
                                         dtype,
                                         kv_cache_dtype=None,
-                                        block_size=16 if not envs.VLLM_USE_FLASH_ATTN_PA else 64,
+                                        block_size=16 if not envs.VLLM_USE_FLASH_ATTN_PA or not envs.VLLM_FLASH_ATTN_V1 else 64,
                                         is_attention_free=False)
         backend = backend_name_to_enum(attn_backend.get_name())
         if current_platform.is_rocm():

vllm/config.py

@@ -1630,7 +1630,7 @@ PrefixCachingHashAlgo = Literal["builtin", "sha256", "sha256_cbor_64bit"]
 class CacheConfig:
     """Configuration for the KV cache."""
 
-    block_size: BlockSize = 16 if not envs.VLLM_USE_FLASH_ATTN_PA else 64  # type: ignore
+    block_size: BlockSize = 16 if not envs.VLLM_USE_FLASH_ATTN_PA or not envs.VLLM_FLASH_ATTN_V1 else 64  # type: ignore
     """Size of a contiguous cache block in number of tokens. This is ignored on
     neuron devices and set to `--max-model-len`. On CUDA devices, only block
     sizes up to 32 are supported. On HPU devices, block size defaults to 128.