Fix long seq bug

vllm/attention/backends/abstract.py

@@ -103,6 +103,10 @@ class AttentionMetadata:
     # in block 0, and 1st slot in block 1, respectively.
     slot_mapping: torch.Tensor
 
+    # Number of original input tokens (without any decoding).
+    # Some model (phi3-) need this info to decide model settings
+    num_orig_input_tokens_tensor: torch.Tensor
+
     @property
     @abstractmethod
     def prefill_metadata(self) -> Optional["AttentionMetadata"]:
@@ -184,7 +188,8 @@ def __init__(self, input_builder: "ModelRunnerInputBuilderBase") -> None:
 
     @abstractmethod
     def build(self, seq_lens: List[int], query_lens: List[int],
-              cuda_graph_pad_size: int, batch_size: int) -> T:
+              num_orig_input_tokens_list: List[int], cuda_graph_pad_size: int,
+              batch_size: int) -> T:
         """Build attention metadata with on-device tensors."""
         raise NotImplementedError
 

vllm/attention/backends/blocksparse_attn.py

@@ -220,6 +220,9 @@ def prefill_metadata(
             query_start_loc=self.query_start_loc[:self.num_prefills + 1],
             seq_start_loc=self.seq_start_loc[:self.num_prefills + 1],
             context_lens_tensor=self.context_lens_tensor[:self.num_prefills],
+            num_orig_input_tokens_tensor=(
+                None if self.num_orig_input_tokens_tensor is None else
+                self.num_orig_input_tokens_tensor[:self.num_prefills]),
             block_tables=self.block_tables[:self.num_prefills],
             use_cuda_graph=False,
         )
@@ -248,6 +251,9 @@ def decode_metadata(self) -> Optional["BlocksparseFlashAttentionMetadata"]:
             query_start_loc=None,
             seq_start_loc=None,
             context_lens_tensor=None,
+            num_orig_input_tokens_tensor=(
+                None if self.num_orig_input_tokens_tensor is None else
+                self.num_orig_input_tokens_tensor[:self.num_prefills]),
             block_tables=self.block_tables[self.num_prefills:],
             use_cuda_graph=self.use_cuda_graph,
         )

vllm/attention/backends/flash_attn.py

@@ -263,6 +263,8 @@ def prefill_metadata(self) -> Optional["FlashAttentionMetadata"]:
             slot_mapping=self.slot_mapping[:self.num_prefill_tokens],
             seq_lens=self.seq_lens[:self.num_prefills],
             seq_lens_tensor=self.seq_lens_tensor[:self.num_prefills],
+            num_orig_input_tokens_tensor=self.
+            num_orig_input_tokens_tensor[:self.num_prefills],
             max_query_len=self.max_query_len,
             max_prefill_seq_len=self.max_prefill_seq_len,
             max_decode_seq_len=0,
@@ -291,6 +293,8 @@ def decode_metadata(self) -> Optional["FlashAttentionMetadata"]:
             slot_mapping=self.slot_mapping[self.num_prefill_tokens:],
             seq_lens=None,
             seq_lens_tensor=self.seq_lens_tensor[self.num_prefills:],
+            num_orig_input_tokens_tensor=self.
+            num_orig_input_tokens_tensor[:self.num_prefills],
             max_query_len=None,
             max_prefill_seq_len=0,
             max_decode_seq_len=self.max_decode_seq_len,
@@ -427,7 +431,8 @@ def _add_seq_group(
                                  self.block_size, inter_data.block_tables)
 
     def build(self, seq_lens: List[int], query_lens: List[int],
-              cuda_graph_pad_size: int, batch_size: int):
+              num_orig_input_tokens_list: List[int], cuda_graph_pad_size: int,
+              batch_size: int):
         """Build attention metadata with on-device tensors.
 
         Args:
@@ -499,6 +504,9 @@ def build(self, seq_lens: List[int], query_lens: List[int],
                      dim=0,
                      dtype=query_start_loc.dtype,
                      out=query_start_loc[1:])
+        num_orig_input_tokens_tensor = torch.tensor(num_orig_input_tokens_list,
+                                                    dtype=torch.long,
+                                                    device=device)
 
         return FlashAttentionMetadata(
             num_prefills=self.num_prefills,
@@ -507,6 +515,7 @@ def build(self, seq_lens: List[int], query_lens: List[int],
             num_decode_tokens=num_decode_tokens,
             seq_lens=seq_lens,
             seq_lens_tensor=seq_lens_tensor,
+            num_orig_input_tokens_tensor=num_orig_input_tokens_tensor,
             max_query_len=max_query_len,
             max_prefill_seq_len=max_prefill_seq_len,
             max_decode_seq_len=max_decode_seq_len,

vllm/attention/backends/flashinfer.py

@@ -134,6 +134,8 @@ def graph_capture(self, max_batch_size: int):
                                           device=self.runner.device)
         self._graph_block_tables = torch.from_numpy(
             self.runner.graph_block_tables).to(device=self.runner.device)
+        self._num_orig_input_tokens_tensor = torch.zeros(
+            max_batch_size, dtype=torch.int32, device=self.runner.device)
         self._graph_decode_workspace_buffer = self._get_workspace_buffer()
         self._graph_indices_buffer = torch.empty(
             max_batch_size * self.runner.cache_config.num_gpu_blocks,
@@ -154,6 +156,7 @@ def graph_capture(self, max_batch_size: int):
         del self._graph_indptr_buffer
         del self._graph_last_page_len_buffer
         del self._graph_decode_wrapper
+        del self._num_orig_input_tokens_tensor
 
     def graph_clone(self, batch_size: int):
         assert self._is_graph_capturing
@@ -200,6 +203,8 @@ def graph_capture_get_metadata_for_batch(self, batch_size: int):
             slot_mapping=self._graph_slot_mapping[:batch_size],
             num_prefill_tokens=0,
             num_decode_tokens=batch_size,
+            num_orig_input_tokens_tensor=self.
+            _num_orig_input_tokens_tensor[:batch_size],
             max_prefill_seq_len=0,
             block_tables=self._graph_block_tables,
             paged_kv_indptr=paged_kv_indptr_tensor_host,
@@ -221,10 +226,15 @@ def graph_capture_get_metadata_for_batch(self, batch_size: int):
 
     def get_graph_input_buffers(self, attn_metadata):
         return {
-            "slot_mapping": attn_metadata.slot_mapping,
+            "slot_mapping":
+            attn_metadata.slot_mapping,
+            "num_orig_input_tokens_tensor":
+            attn_metadata.num_orig_input_tokens_tensor,
         }
 
     def prepare_graph_input_buffers(self, input_buffers, attn_metadata):
+        input_buffers["num_orig_input_tokens_tensor"].copy_(
+            attn_metadata.num_orig_input_tokens_tensor, non_blocking=True)
         return
 
     def begin_forward(self, model_input):
@@ -494,7 +504,8 @@ def _update_paged_kv_tensors(self, block_table: List[int], seq_len: int):
         self.paged_kv_last_page_len.append(last_page_len)
 
     def build(self, seq_lens: List[int], query_lens: List[int],
-              cuda_graph_pad_size: int, batch_size: int):
+              num_orig_input_tokens_list: List[int], cuda_graph_pad_size: int,
+              batch_size: int):
         """Build attention metadata with on-device tensors.
 
         Args:
@@ -564,6 +575,10 @@ def build(self, seq_lens: List[int], query_lens: List[int],
                      dtype=query_start_loc.dtype,
                      out=query_start_loc[1:])
 
+        num_orig_input_tokens_tensor = torch.tensor(num_orig_input_tokens_list,
+                                                    dtype=torch.long,
+                                                    device=device)
+
         if len(self.paged_kv_indptr) > 0:
             paged_kv_indices_tensor = torch.tensor(self.paged_kv_indices,
                                                    device="cpu",
@@ -585,6 +600,7 @@ def build(self, seq_lens: List[int], query_lens: List[int],
             slot_mapping=slot_mapping_tensor,
             num_prefill_tokens=self.num_prefill_tokens,
             num_decode_tokens=num_decode_tokens,
+            num_orig_input_tokens_tensor=num_orig_input_tokens_tensor,
             max_prefill_seq_len=max_prefill_seq_len,
             block_tables=block_tables,
             paged_kv_indptr=paged_kv_indptr_tensor,

vllm/attention/backends/rocm_flash_attn.py

@@ -136,6 +136,8 @@ def prefill_metadata(self) -> Optional["ROCmFlashAttentionMetadata"]:
             slot_mapping=self.slot_mapping[:self.num_prefill_tokens],
             seq_lens=self.seq_lens[:self.num_prefills],
             seq_lens_tensor=self.seq_lens_tensor[:self.num_prefills],
+            num_orig_input_tokens_tensor=self.
+            num_orig_input_tokens_tensor[:self.num_prefills],
             max_query_len=self.max_query_len,
             max_prefill_seq_len=self.max_prefill_seq_len,
             max_decode_seq_len=0,
@@ -164,6 +166,8 @@ def decode_metadata(self) -> Optional["ROCmFlashAttentionMetadata"]:
             slot_mapping=self.slot_mapping[self.num_prefill_tokens:],
             seq_lens=None,
             seq_lens_tensor=self.seq_lens_tensor[self.num_prefills:],
+            num_orig_input_tokens_tensor=self.
+            num_orig_input_tokens_tensor[:self.num_prefills],
             max_query_len=None,
             max_prefill_seq_len=0,
             max_decode_seq_len=self.max_decode_seq_len,

vllm/attention/backends/utils.py

@@ -190,7 +190,8 @@ def _add_seq_group(
                                  self.block_size, inter_data.block_tables)
 
     def build(self, seq_lens: List[int], query_lens: List[int],
-              cuda_graph_pad_size: int, batch_size: int):
+              num_orig_input_tokens_list: List[int], cuda_graph_pad_size: int,
+              batch_size: int):
         """Build attention metadata with on-device tensors.
 
         Args:
@@ -258,13 +259,18 @@ def build(self, seq_lens: List[int], query_lens: List[int],
                      dtype=query_start_loc.dtype,
                      out=query_start_loc[1:])
 
+        num_orig_input_tokens_tensor = torch.tensor(num_orig_input_tokens_list,
+                                                    dtype=torch.long,
+                                                    device=device)
+
         return self._metadata_cls(  # type: ignore
             num_prefills=self.num_prefills,
             slot_mapping=slot_mapping_tensor,
             num_prefill_tokens=self.num_prefill_tokens,
             num_decode_tokens=num_decode_tokens,
             seq_lens=seq_lens,
             seq_lens_tensor=seq_lens_tensor,
+            num_orig_input_tokens_tensor=num_orig_input_tokens_tensor,
             max_query_len=max_query_len,
             max_prefill_seq_len=max_prefill_seq_len,
             max_decode_seq_len=max_decode_seq_len,
@@ -294,11 +300,16 @@ def graph_capture(self, max_batch_size: int):
                                           device=self.runner.device)
         self._graph_block_tables = torch.from_numpy(
             self.runner.graph_block_tables).to(device=self.runner.device)
+
+        self._num_orig_input_tokens_tensor = torch.zeros(
+            max_batch_size, dtype=torch.int32, device=self.runner.device)
+
         yield
         self._is_graph_capturing = False
         del self._graph_slot_mapping
         del self._graph_seq_lens
         del self._graph_block_tables
+        del self._num_orig_input_tokens_tensor
 
     def graph_clone(self, batch_size: int) -> "CommonAttentionState":
         assert self._is_graph_capturing
@@ -313,6 +324,8 @@ def graph_capture_get_metadata_for_batch(self, batch_size: int):
             slot_mapping=self._graph_slot_mapping[:batch_size],
             seq_lens=None,
             seq_lens_tensor=self._graph_seq_lens[:batch_size],
+            num_orig_input_tokens_tensor=self.
+            _num_orig_input_tokens_tensor[:batch_size],
             max_query_len=None,
             max_prefill_seq_len=0,
             max_decode_seq_len=self.runner.max_seq_len_to_capture,
@@ -326,9 +339,14 @@ def graph_capture_get_metadata_for_batch(self, batch_size: int):
 
     def get_graph_input_buffers(self, attn_metadata) -> Dict[str, Any]:
         return {
-            "slot_mapping": attn_metadata.slot_mapping,
-            "seq_lens_tensor": attn_metadata.decode_metadata.seq_lens_tensor,
-            "block_tables": attn_metadata.decode_metadata.block_tables,
+            "slot_mapping":
+            attn_metadata.slot_mapping,
+            "seq_lens_tensor":
+            attn_metadata.decode_metadata.seq_lens_tensor,
+            "block_tables":
+            attn_metadata.decode_metadata.block_tables,
+            "num_orig_input_tokens_tensor":
+            attn_metadata.num_orig_input_tokens_tensor,
         }
 
     def prepare_graph_input_buffers(self, input_buffers,
@@ -337,6 +355,8 @@ def prepare_graph_input_buffers(self, input_buffers,
             attn_metadata.decode_metadata.seq_lens_tensor, non_blocking=True)
         input_buffers["block_tables"].copy_(
             attn_metadata.decode_metadata.block_tables, non_blocking=True)
+        input_buffers["num_orig_input_tokens_tensor"].copy_(
+            attn_metadata.num_orig_input_tokens_tensor, non_blocking=True)
 
     def begin_forward(self, model_input) -> None:
         return

vllm/attention/backends/xformers.py

@@ -200,6 +200,9 @@ def prefill_metadata(self) -> Optional["XFormersMetadata"]:
                            self.seq_lens_tensor[:self.num_prefills])
         context_lens_tensor = (None if self.context_lens_tensor is None else
                                self.context_lens_tensor[:self.num_prefills])
+        num_orig_input_tokens_tensor = (
+            None if self.num_orig_input_tokens_tensor is None else
+            self.num_orig_input_tokens_tensor[:self.num_prefills])
         block_tables = (None if self.block_tables is None else
                         self.block_tables[:self.num_prefills])
 
@@ -211,6 +214,7 @@ def prefill_metadata(self) -> Optional["XFormersMetadata"]:
             slot_mapping=slot_mapping,
             seq_lens=seq_lens,
             seq_lens_tensor=seq_lens_tensor,
+            num_orig_input_tokens_tensor=num_orig_input_tokens_tensor,
             max_query_len=self.max_query_len,
             max_prefill_seq_len=self.max_prefill_seq_len,
             max_decode_seq_len=0,
@@ -245,6 +249,9 @@ def decode_metadata(self) -> Optional["XFormersMetadata"]:
                            self.seq_lens_tensor[self.num_prefills:])
         block_tables = (None if self.block_tables is None else
                         self.block_tables[self.num_prefills:])
+        num_orig_input_tokens_tensor = (
+            None if self.num_orig_input_tokens_tensor is None else
+            self.num_orig_input_tokens_tensor[:self.num_prefills])
 
         # Construct & cache decode-phase attention metadata structure
         self._cached_decode_metadata = XFormersMetadata(
@@ -253,6 +260,7 @@ def decode_metadata(self) -> Optional["XFormersMetadata"]:
             num_decode_tokens=self.num_decode_tokens,
             slot_mapping=slot_mapping,
             seq_lens_tensor=seq_lens_tensor,
+            num_orig_input_tokens_tensor=num_orig_input_tokens_tensor,
             max_prefill_seq_len=0,
             max_decode_seq_len=self.max_decode_seq_len,
             block_tables=block_tables,

vllm/model_executor/layers/rotary_embedding.py

@@ -536,8 +536,8 @@ def __init__(
                 1 + math.log(scale) /
                 math.log(self.original_max_position_embeddings))
 
-        short_cache = self._compute_cos_sin_cache(
-            original_max_position_embeddings, short_factor, short_mscale)
+        short_cache = self._compute_cos_sin_cache(max_position_embeddings,
+                                                  short_factor, short_mscale)
         short_cache = short_cache.to(dtype)
         self.register_buffer("short_cos_sin_cache",
                              short_cache,
@@ -582,11 +582,18 @@ def forward(
         query: torch.Tensor,
         key: torch.Tensor,
         offsets: Optional[torch.Tensor] = None,
+        *,
+        num_orig_input_tokens_tensor: torch.Tensor = None,
     ) -> Tuple[torch.Tensor, torch.Tensor]:
         query = query.view(*query.shape[:-1], -1, self.head_size)
         key = key.view(*key.shape[:-1], -1, self.head_size)
 
         k = self.original_max_position_embeddings
+        long_prompt_offset = torch.where(
+            num_orig_input_tokens_tensor <= k,
+            torch.zeros_like(num_orig_input_tokens_tensor),
+            torch.full_like(num_orig_input_tokens_tensor,
+                            self.max_position_embeddings))
         long_prompt_offset = (torch.any(positions > k).float() *
                               torch.full_like(positions, k)).long()
         idx = (torch.add(positions, long_prompt_offset)

vllm/model_executor/models/__init__.py

@@ -48,7 +48,7 @@
     "OrionForCausalLM": ("orion", "OrionForCausalLM"),
     "PersimmonForCausalLM": ("persimmon", "PersimmonForCausalLM"),
     "PhiForCausalLM": ("phi", "PhiForCausalLM"),
-    "Phi3ForCausalLM": ("llama", "LlamaForCausalLM"),
+    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
     "QWenLMHeadModel": ("qwen", "QWenLMHeadModel"),
     "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
     "Qwen2MoeForCausalLM": ("qwen2_moe", "Qwen2MoeForCausalLM"),