[AMD] Derive AITER verify tokens-per-req from input shape (#31221)

python/sglang/srt/layers/attention/aiter_backend.py

@@ -857,6 +857,11 @@ class AiterAttnBackend(AttentionBackend):
         seq_lens_cpu = (
             forward_batch.seq_lens.cpu() if in_capture else forward_batch.seq_lens_cpu
         )
+        verify_tokens_per_req = (
+            forward_batch.input_ids.shape[0] // forward_batch.batch_size
+            if forward_batch.forward_mode.is_target_verify()
+            else None
+        )
         self._apply_cuda_graph_metadata(
             bs=forward_batch.batch_size,
             req_pool_indices=forward_batch.req_pool_indices,
@@ -866,6 +871,7 @@ class AiterAttnBackend(AttentionBackend):
             forward_mode=forward_batch.forward_mode,
             spec_info=forward_batch.spec_info,
             seq_lens_cpu=seq_lens_cpu,
+            verify_tokens_per_req=verify_tokens_per_req,
         )
 
         # Refill the SWA write-target buffer from the live out_cache_loc and
@@ -1196,8 +1202,8 @@ class AiterAttnBackend(AttentionBackend):
                     run_graph=False,
                 )
             else:
+                draft_num = forward_batch.input_ids.shape[0] // bs
                 bs = len(forward_batch.req_pool_indices)
-                draft_num = spec_info.draft_token_num
 
                 if self._use_unified_verify:
                     page_table, qo_indptr, max_q_len, swa_page_table = (
@@ -1500,6 +1506,7 @@ class AiterAttnBackend(AttentionBackend):
         forward_mode: ForwardMode,
         spec_info: Optional[SpecInput],
         seq_lens_cpu: Optional[torch.Tensor],
+        verify_tokens_per_req: Optional[int],
     ):
 
         num_kv_splits = None
@@ -1652,11 +1659,17 @@ class AiterAttnBackend(AttentionBackend):
 
         elif forward_mode.is_target_verify():
             bs = len(req_pool_indices)
+            assert verify_tokens_per_req is not None
+            # MLA uses a fixed draft length (num_draft_tokens); the non-MLA
+            # unified path derives it per batch from input_ids.
+            tokens_per_req = (
+                self.num_draft_tokens if self.use_mla else verify_tokens_per_req
+            )
             qo_indptr = self.qo_indptr[: bs + 1]
             qo_indptr[: bs + 1] = torch.arange(
                 0,
-                (1 + bs) * self.num_draft_tokens,
+                (1 + bs) * tokens_per_req,
-                step=self.num_draft_tokens,
+                step=tokens_per_req,
                 dtype=torch.int32,
                 device=self.device,
             )
@@ -1689,9 +1702,9 @@ class AiterAttnBackend(AttentionBackend):
                 self.req_to_token.stride(0),
             )
             kv_last_page_len = self.cuda_graph_kv_last_page_len[:bs]
-            max_q_len = self.num_draft_tokens
 
             if self.use_mla:
+                max_q_len = self.num_draft_tokens
                 if _use_mla_ps_kernel:
                     num_kv_splits = self.max_split_per_batch
 
@@ -1735,6 +1748,7 @@ class AiterAttnBackend(AttentionBackend):
                     num_kv_splits=num_kv_splits,
                 )
             else:
+                max_q_len = verify_tokens_per_req
                 if self._use_unified_verify:
                     max_num_blocks_per_seq = (
                         self.max_context_len + self.page_size - 1
@@ -1753,7 +1767,7 @@ class AiterAttnBackend(AttentionBackend):
                             bs,
                             seq_lens,
                             req_pool_indices,
-                            self.num_draft_tokens,
+                            verify_tokens_per_req,
                             page_table_dest=page_table,
                             swa_page_table_dest=swa_page_table,
                         )
@@ -2278,7 +2292,9 @@ class AiterAttnBackend(AttentionBackend):
                         v=v_unified,
                         out=o.view(-1, layer.tp_q_head_num, layer.v_head_dim),
                         cu_seqlens_q=self.forward_metadata.qo_indptr,
-                        seqused_k=forward_batch.seq_lens + self.num_draft_tokens,
+                        seqused_k=(
+                            forward_batch.seq_lens + self.forward_metadata.max_q_len
+                        ),
                         max_seqlen_q=self.forward_metadata.max_q_len,
                         max_seqlen_k=max_kv_len,
                         softmax_scale=layer.scaling,