Resolved alibi bias issue due to porting flat PA pr

Signed-off-by: Tanner Voas <tanner.voas@intel.com>
HabanaAI · Nov 5, 2024 · f8a0a13 · f8a0a13
1 parent 1bb808a
commit f8a0a13
Show file tree

Hide file tree

Showing 2 changed files with 29 additions and 6 deletions.
diff --git a/vllm/attention/backends/hpu_attn.py b/vllm/attention/backends/hpu_attn.py
@@ -118,6 +118,7 @@ def __init__(
             alibi_slopes_tensor = torch.tensor(alibi_slopes,
                                                dtype=torch.bfloat16)
             self.alibi_slopes = alibi_slopes_tensor
+        self.max_seq_len = max_seq_len
         assert self.num_heads % self.num_kv_heads == 0
         self.num_queries_per_kv = self.num_heads // self.num_kv_heads
 
@@ -235,6 +236,14 @@ def forward(
             output = out.reshape(batch_size, seq_len, hidden_size)
         else:
             # Decoding run.
+            self.position_bias = None
+            attn_bias = attn_metadata.attn_bias
+            if self.alibi_slopes is not None:
+                self.position_bias = _make_alibi_bias(self.alibi_slopes,
+                                                    self.num_kv_heads,
+                                                    attn_bias.dtype,
+                                                    self.max_seq_len if self.max_seq_len is not None else attn_bias.shape[-1])
+
             output = HPUPagedAttention.forward_decode(
                 query=query,
                 key_cache=key_cache,
@@ -245,10 +254,12 @@ def forward(
                 block_scales=attn_metadata.block_scales,
                 block_groups=attn_metadata.block_groups,
                 scale=self.scale,
+                alibi_slopes=self.position_bias,
                 matmul_qk_op=self.matmul_qk,
                 matmul_av_op=self.matmul_av,
                 keys_fetch_func=self.k_cache.fetch_from_cache,
-                values_fetch_func=self.v_cache.fetch_from_cache)
+                values_fetch_func=self.v_cache.fetch_from_cache,
+            )
         # Reshape the output tensor.
         return output.view(batch_size, seq_len, hidden_size)
 

diff --git a/vllm/worker/hpu_model_runner.py b/vllm/worker/hpu_model_runner.py
@@ -1183,6 +1183,10 @@ def _prepare_decode(
                                     dtype=self.model_config.dtype,
                                     device=self.device)
 
+        seq_lens_tensor = torch.tensor(seq_lens,
+                                       dtype=torch.int,
+                                       device=self.device)
+
         attn_metadata = self.attn_backend.make_metadata(
             is_prompt=False,
             block_list=block_list,
@@ -1193,7 +1197,7 @@ def _prepare_decode(
             block_scales=block_scales,
             block_groups=block_groups,
             attn_bias=None,
-            seq_lens_tensor=None,
+            seq_lens_tensor=seq_lens_tensor,
             context_lens_tensor=None,
             num_prefills=0,
             num_prefill_tokens=0,
@@ -1402,10 +1406,18 @@ def trim_attn_metadata(self, metadata: AttentionMetadata) -> object:
         # input_hash(123) != input_hash(321)
         # input_hash("abc") != input_hash("cba")
         attention_metadata = subtuple(metadata, 'TrimmedAttentionMetadata', [
-            'attn_bias', 'seq_lens_tensor', 'context_lens_tensor',
-            'block_list', 'block_mapping', 'block_usage', 'slot_mapping',
-            'is_prompt', 'block_indices', 'block_offsets', 'block_scales',
-            'block_groups'
+            'attn_bias',
+            'seq_lens_tensor',
+            'context_lens_tensor',
+            'block_list',
+            'block_mapping',
+            'block_usage',
+            'slot_mapping',
+            'is_prompt',
+            'block_indices',
+            'block_offsets',
+            'block_scales',
+            'block_groups',
         ])
         return attention_metadata