[dlinfer]change llm op interface of paged_prefill_attention. (#2977)

* [dlinfer]modify interface to support camb multi-batch-conv

* [dlinfer]change order for paged_prefill

Author: JackWeiw

lmdeploy/pytorch/backends/dlinfer/attention.py

@@ -16,6 +16,7 @@ class DlinferAttentionMetadata(AttentionMetadata):
     max_q_seq_len: int = 1
     max_kv_seq_len: int = 1
     quant_meta: Dict = None
+    cu_seq_lens_kv: Optional[Tensor] = None
 
 
 class DlinferAttentionImpl(AttentionImpl[DlinferAttentionMetadata]):
@@ -79,6 +80,8 @@ def forward(
         max_q_seq_len = attn_metadata.max_q_seq_len
         max_kv_seq_len = attn_metadata.max_kv_seq_len
         quant_bits = attn_metadata.quant_policy
+        cu_seq_lens_kv = attn_metadata.cu_seq_lens_kv
+
         if attn_metadata.quant_meta is not None:
             k_scales_zeros = [
                 next(attn_metadata.quant_meta['k_scales']),
@@ -128,6 +131,7 @@ def forward(
             q_start_loc=q_start_loc,
             q_seqlens=q_seqlens,
             kv_seqlens=kv_seqlens,
+            cu_seq_lens_kv=cu_seq_lens_kv,
             max_q_seq_len=max_q_seq_len,
             max_kv_seq_len=max_kv_seq_len,
             is_decoding=is_decoding,

lmdeploy/pytorch/kernels/dlinfer/pagedattention.py

@@ -15,7 +15,9 @@ def prefill_attention(
     q_start_loc: Tensor,
     q_seq_len: Tensor,
     kv_seq_len: Tensor,
+    cu_seq_lens_kv: Tensor,
     max_q_seq_len: int,
+    max_kv_seq_len: int,
     block_size: int,
     attn_mask: Sequence[Optional[Tensor]],
     is_unpaged_prefill: Optional[bool],
@@ -51,7 +53,9 @@ def prefill_attention(
             q_start_loc,
             q_seq_len,
             kv_seq_len,
+            cu_seq_lens_kv,
             max_q_seq_len,
+            max_kv_seq_len,
             num_q_heads,
             num_kv_heads,
             attn_mask,
@@ -105,6 +109,7 @@ def paged_attention_fwd(
     q_start_loc: Tensor,
     q_seqlens: Tensor,
     kv_seqlens: Tensor,
+    cu_seq_lens_kv: Tensor,
     max_q_seq_len: int,
     max_kv_seq_len: int,
     is_decoding: bool,
@@ -127,7 +132,9 @@ def paged_attention_fwd(
             q_start_loc,
             q_seqlens,
             kv_seqlens,
+            cu_seq_lens_kv,
             max_q_seq_len,
+            max_kv_seq_len,
             block_size,
             attn_mask,
             is_unpaged_prefill,