Fix for breaking changes in xformers 0.0.21 (#834)

requirements.txt

@@ -5,7 +5,7 @@ sentencepiece  # Required for LLaMA tokenizer.
 numpy
 torch >= 2.0.0
 transformers >= 4.31.0  # Required for LLaMA-2.
-xformers >= 0.0.19
+xformers >= 0.0.21
 fastapi
 uvicorn
 pydantic < 2  # Required for OpenAI server.

vllm/model_executor/layers/attention.py

@@ -357,11 +357,12 @@ class PagedAttentionWithALiBi(PagedAttention):
             # be sliced from a tensor whose length is a multiple of 8.
             padded_len = (prompt_len + 7) // 8 * 8
             bias = torch.empty(
+                1,  # batch_size
                 self.num_heads,
-                padded_len,
+                prompt_len,
                 padded_len,
                 device=self.alibi_slopes.device,
-            )[:, :prompt_len, :prompt_len].copy_(bias)
+            )[:, :, :, :prompt_len].copy_(bias)
             bias.mul_(self.alibi_slopes[:, None, None])
             attn_bias = LowerTriangularMaskWithTensorBias(bias)
             input_metadata.attn_bias.append(attn_bias)