rebase all codes

python/sglang/srt/layers/attention_backend.py

@@ -478,3 +478,185 @@ def forward_decode(self, q, k, v, layer: nn.Module, input_metadata: InputMetadat
             layer.logit_cap,
         )
         return o
+
+
+class DoubleSparseAttnBackend(AttentionBackend):
+    def __init__(self, model_runner: ModelRunner):
+        # Lazy import to avoid the initialization of cuda context
+        from sglang.srt.layers.triton_attention.decode_attention import (
+            decode_attention_fwd,
+        )
+        from sglang.srt.layers.triton_attention.extend_attention import (
+            extend_attention_fwd,
+        )
+        from sglang.srt.layers.triton_attention.sparse_decode_attention import (
+            decode_sparse_attention_fwd
+        )
+
+        super().__init__()
+
+        self.decode_attention_fwd = decode_attention_fwd
+        self.decode_sparse_attention_fwd = decode_sparse_attention_fwd
+        self.extend_attention_fwd = extend_attention_fwd
+        self.num_head = model_runner.model_config.num_attention_heads
+
+        if global_server_args_dict.get("triton_attention_reduce_in_fp32", False):
+            self.reduce_dtype = torch.float32
+        else:
+            self.reduce_dtype = torch.float16
+
+        self.forward_metadata = None
+
+        self.cuda_graph_max_seq_len = model_runner.model_config.context_len
+
+    def init_forward_metadata(
+        self, batch: ScheduleBatch, input_metadata: InputMetadata
+    ):
+        """Init auxiliary variables for triton attention backend."""
+
+        if input_metadata.forward_mode.is_decode():
+            start_loc = torch.zeros_like(input_metadata.seq_lens, dtype=torch.int32)
+            start_loc[1:] = torch.cumsum(input_metadata.seq_lens[:-1], dim=0)
+
+            total_num_tokens = torch.sum(input_metadata.seq_lens).item()
+            attn_logits = torch.empty(
+                (self.num_head, total_num_tokens),
+                dtype=self.reduce_dtype,
+                device="cuda",
+            )
+
+            max_seq_len = torch.max(input_metadata.seq_lens).item()
+            max_extend_len = None
+            #NOTE: Align sequence order with req_to_token order
+            ds_req_to_token = input_metadata.req_to_token_pool.req_to_token[input_metadata.req_pool_indices]
+        else:
+            start_loc = attn_logits = max_seq_len = None
+            prefix_lens = torch.tensor(batch.prefix_lens_cpu, device="cuda")
+            max_extend_len = torch.max(input_metadata.seq_lens - prefix_lens).item()
+            ds_req_to_token = None
+
+        self.forward_metadata = start_loc, attn_logits, max_seq_len, max_extend_len, ds_req_to_token
+
+    def init_cuda_graph_state(self, max_bs: int):
+        #TODO(Andy): Support CUDA graph for double sparse attention
+        raise ValueError("Double sparse attention does not support CUDA graph for now. Please --disable-cuda-graph")
+        self.cuda_graph_max_total_num_tokens = max_bs * self.cuda_graph_max_seq_len
+
+        self.cuda_graph_start_loc = torch.zeros(
+            (max_bs,), dtype=torch.int32, device="cuda"
+        )
+        self.cuda_graph_attn_logits = torch.empty(
+            (
+                self.num_head,
+                self.cuda_graph_max_total_num_tokens,
+            ),
+            dtype=self.reduce_dtype,
+            device="cuda",
+        )
+
+    def init_forward_metadata_capture_cuda_graph(
+        self, bs: int, req_pool_indices, seq_lens
+    ):
+        self.forward_metadata = (
+            self.cuda_graph_start_loc,
+            self.cuda_graph_attn_logits,
+            self.cuda_graph_max_seq_len,
+            None,
+        )
+
+    def init_forward_metadata_replay_cuda_graph(
+        self, bs: int, req_pool_indices, seq_lens
+    ):
+        self.cuda_graph_start_loc.zero_()
+        self.cuda_graph_start_loc[1:bs] = torch.cumsum(seq_lens[: bs - 1], dim=0)
+
+    def get_cuda_graph_seq_len_fill_value(self):
+        return 1
+
+    def forward_extend(self, q, k, v, layer: nn.Module, input_metadata: InputMetadata):
+        # TODO: reuse the buffer across layers
+        if layer.qk_head_dim != layer.v_head_dim:
+            o = q.new_empty((q.shape[0], layer.tp_q_head_num * layer.v_head_dim))
+        else:
+            o = torch.empty_like(q)
+
+        k_label = torch.gather(k, 2, input_metadata.sorted_channels[layer.layer_id].unsqueeze(0).expand(k.shape[0], -1, -1))
+
+        input_metadata.token_to_kv_pool.set_kv_buffer(
+            layer.layer_id, input_metadata.out_cache_loc, k, v, k_label
+        )
+
+        start_loc, attn_logits, max_seq_len, max_extend_len, ds_req_to_token = self.forward_metadata
+        self.extend_attention_fwd(
+            q.view(-1, layer.tp_q_head_num, layer.qk_head_dim),
+            k.contiguous(),
+            v.contiguous(),
+            o.view(-1, layer.tp_q_head_num, layer.v_head_dim),
+            input_metadata.token_to_kv_pool.get_key_buffer(layer.layer_id),
+            input_metadata.token_to_kv_pool.get_value_buffer(layer.layer_id),
+            input_metadata.req_to_token_pool.req_to_token,
+            input_metadata.req_pool_indices,
+            input_metadata.seq_lens,
+            input_metadata.extend_seq_lens,
+            input_metadata.extend_start_loc,
+            max_extend_len,
+            layer.scaling,
+            layer.logit_cap,
+        )
+        return o
+
+    def forward_decode(self, q, k, v, layer: nn.Module, input_metadata: InputMetadata):
+        # During torch.compile, there is a bug in rotary_emb that causes the
+        # output value to have a 3D tensor shape. This reshapes the output correctly.
+        q = q.reshape(-1, layer.tp_q_head_num * layer.qk_head_dim)
+
+        # TODO: reuse the buffer across layers
+        if layer.qk_head_dim != layer.v_head_dim:
+            o = q.new_empty((q.shape[0], layer.tp_q_head_num * layer.v_head_dim))
+        else:
+            o = torch.empty_like(q)
+
+        start_loc, attn_logits, max_seq_len, max_extend_len, ds_req_to_token = self.forward_metadata
+
+        k_label = torch.gather(k, 2, input_metadata.sorted_channels[layer.layer_id].unsqueeze(0).expand(k.shape[0], -1, -1))
+
+        input_metadata.token_to_kv_pool.set_kv_buffer(
+            layer.layer_id, input_metadata.out_cache_loc, k, v, k_label
+        )
+
+        # NOTE(Andy) shouldn't be used when max_len_in_batch < heavy_token_num
+        #            and set a minimum value for sparse_decode
+        if max_seq_len < input_metadata.heavy_token_num or max_seq_len < input_metadata.sparse_decode_thresold:
+            self.decode_attention_fwd(
+                q.view(-1, layer.tp_q_head_num, layer.qk_head_dim),
+                input_metadata.token_to_kv_pool.get_key_buffer(layer.layer_id),
+                input_metadata.token_to_kv_pool.get_value_buffer(layer.layer_id),
+                o.view(-1, layer.tp_q_head_num, layer.v_head_dim),
+                input_metadata.req_to_token_pool.req_to_token,
+                input_metadata.req_pool_indices,
+                start_loc,
+                input_metadata.seq_lens,
+                attn_logits,
+                max_seq_len,
+                layer.scaling,
+                layer.logit_cap,
+            )
+        else:
+            #TODO(Andy): indexing with torch.gather or torch.index_select or customized kernel
+            q_label = torch.gather(q.view(-1, layer.tp_q_head_num, layer.qk_head_dim), 2, input_metadata.sorted_channels[layer.layer_id].unsqueeze(0).expand(q.shape[0], -1, -1))
+            self.decode_sparse_attention_fwd(
+                q.view(-1, layer.tp_q_head_num, layer.qk_head_dim),
+                input_metadata.token_to_kv_pool.get_key_buffer(layer.layer_id),
+                input_metadata.token_to_kv_pool.get_value_buffer(layer.layer_id),
+                o.view(-1, layer.tp_q_head_num, layer.qk_head_dim),
+                q_label,
+                input_metadata.token_to_kv_pool.get_label_buffer(layer.layer_id),
+                ds_req_to_token,
+                input_metadata.seq_lens,
+                max_seq_len,
+                layer.scaling,
+                layer.logit_cap,
+                input_metadata.heavy_token_num,
+            )
+
+        return o