sglang 0.4.6.post1__py3-none-any.whl → 0.4.6.post3__py3-none-any.whl

sglang/srt/layers/attention/flashattention_backend.py

@@ -338,7 +338,7 @@ class FlashAttentionBackend(AttentionBackend):
         """Initialize forward metadata hence all layers in the forward pass can reuse it."""
         metadata = FlashAttentionMetadata()
         seqlens_in_batch = forward_batch.seq_lens
-        batch_size = len(seqlens_in_batch)
+        batch_size = forward_batch.batch_size
         device = seqlens_in_batch.device
 
         if forward_batch.forward_mode.is_decode_or_idle():
@@ -913,8 +913,10 @@ class FlashAttentionBackend(AttentionBackend):
         # Use precomputed metadata across all layers
         metadata = self.forward_metadata
         local_attn_metadata = getattr(metadata, "local_attn_metadata", None)
-        use_local_attention = (
-            self.attention_chunk_size is not None and local_attn_metadata is not None
+        use_local_attn = (
+            self.attention_chunk_size is not None
+            and local_attn_metadata is not None
+            and (hasattr(layer, "use_irope") and layer.use_irope)
         )
         # We do cascade attention for Draft Decode with topk > 1
         use_cascade_attn = self.topk > 1
@@ -970,7 +972,7 @@ class FlashAttentionBackend(AttentionBackend):
                     k_descale=k_descale,
                     v_descale=v_descale,
                 )
-            elif use_local_attention:
+            elif use_local_attn:
                 # Use chunked (local) attention batching for self-attention
                 o = flash_attn_with_kvcache(
                     q=q.contiguous().view(-1, layer.tp_q_head_num, layer.head_dim),
@@ -979,7 +981,7 @@ class FlashAttentionBackend(AttentionBackend):
                     page_table=local_attn_metadata.local_block_table,
                     cache_seqlens=local_attn_metadata.local_seqused_k,
                     cu_seqlens_q=local_attn_metadata.local_query_start_loc,
-                    cu_seqlens_k_new=metadata.cu_seqlens_k,
+                    cu_seqlens_k_new=None,
                     max_seqlen_q=local_attn_metadata.local_max_query_len,
                     softmax_scale=layer.scaling,
                     causal=True,
@@ -1127,7 +1129,6 @@ class FlashAttentionBackend(AttentionBackend):
         This creates fixed-size tensors that will be reused during CUDA graph replay
         to avoid memory allocations.
         """
-
         # This is being used by normal decode and draft decode when topk == 1
         self.decode_cuda_graph_metadata = {
             "cache_seqlens": torch.zeros(max_bs, dtype=torch.int32, device=self.device),
@@ -1154,6 +1155,34 @@ class FlashAttentionBackend(AttentionBackend):
             ),
         }
 
+        # Only allocate local attention buffers if local attention is enabled
+        # This prevents OOM errors when local attention is not being used
+        if self.attention_chunk_size is not None:
+            # Estimate maximum sizes for local attention metadata
+            max_seq_len = self.max_context_len
+            page_size = self.page_size or 1
+            attn_chunk_size = self.attention_chunk_size
+            max_virtual_batches = max_bs * (
+                (max_seq_len + attn_chunk_size - 1) // attn_chunk_size
+            )
+            max_blocks_per_seq = (max_seq_len + attn_chunk_size - 1) // attn_chunk_size
+            max_pages_per_block = (attn_chunk_size + page_size - 1) // page_size
+
+            self.decode_cuda_graph_local_attn_metadata = {
+                "local_query_start_loc": torch.zeros(
+                    max_virtual_batches + 1, dtype=torch.int32, device=self.device
+                ),
+                "local_seqused_k": torch.zeros(
+                    max_virtual_batches, dtype=torch.int32, device=self.device
+                ),
+                "local_block_table": torch.zeros(
+                    max_virtual_batches,
+                    max_blocks_per_seq * max_pages_per_block,
+                    dtype=torch.int32,
+                    device=self.device,
+                ),
+            }
+
         # This is used by draft decode's first half of metadata when topk > 1
         if self.topk > 1:
             self.draft_decode_metadata_topk_normal = {
@@ -1405,6 +1434,21 @@ class FlashAttentionBackend(AttentionBackend):
                 )
                 self.decode_cuda_graph_metadata[bs] = metadata
 
+                if self.attention_chunk_size is not None:
+                    metadata.local_attn_metadata = FlashAttentionMetadata.LocalAttentionMetadata(
+                        local_query_start_loc=self.decode_cuda_graph_local_attn_metadata[
+                            "local_query_start_loc"
+                        ],
+                        local_seqused_k=self.decode_cuda_graph_local_attn_metadata[
+                            "local_seqused_k"
+                        ],
+                        local_block_table=self.decode_cuda_graph_local_attn_metadata[
+                            "local_block_table"
+                        ],
+                        local_max_query_len=1,
+                        local_max_seq_len=1,
+                    )
+
         elif forward_mode.is_target_verify():
             if self.topk <= 1:
                 metadata.cache_seqlens_int32 = self.target_verify_metadata[
@@ -1525,12 +1569,9 @@ class FlashAttentionBackend(AttentionBackend):
                     metadata.max_seq_len_k = seq_lens_cpu.max().item() + (
                         self.speculative_step_id + 1
                     )
-                    metadata.cu_seqlens_k.copy_(
-                        torch.nn.functional.pad(
-                            torch.cumsum(
-                                metadata.cache_seqlens_int32, dim=0, dtype=torch.int32
-                            ),
-                            (1, 0),
+                    metadata.cu_seqlens_k[1:].copy_(
+                        torch.cumsum(
+                            metadata.cache_seqlens_int32, dim=0, dtype=torch.int32
                         )
                     )
 
@@ -1554,12 +1595,9 @@ class FlashAttentionBackend(AttentionBackend):
                     # metadata.max_seq_len_q = self.topk, already set in capture
                     metadata.max_seq_len_k = seq_lens_cpu.max().item()
                     # metadata.cu_seqlens_q already set in capture
-                    metadata.cu_seqlens_k.copy_(
-                        torch.nn.functional.pad(
-                            torch.cumsum(
-                                metadata.cache_seqlens_int32, dim=0, dtype=torch.int32
-                            ),
-                            (1, 0),
+                    metadata.cu_seqlens_k[1:].copy_(
+                        torch.cumsum(
+                            metadata.cache_seqlens_int32, dim=0, dtype=torch.int32
                         )
                     )
 
@@ -1578,8 +1616,7 @@ class FlashAttentionBackend(AttentionBackend):
                     metadata_expand.page_table[: cache_loc.shape[0]].copy_(
                         cache_loc[:, :decode_length].contiguous().to(torch.int32)
                     )
-                # TODO: we need to test this part for llama 4 eagle case
-                self._init_local_attn_metadata(metadata, device)
+                # TODO: Handle local attention metadata for draft decode when llama4 eagle is supported
             else:
                 metadata = self.decode_cuda_graph_metadata[bs]
                 # Normal Decode
@@ -1587,8 +1624,9 @@ class FlashAttentionBackend(AttentionBackend):
                 metadata.max_seq_len_k = max_len
 
                 metadata.cache_seqlens_int32 = seq_lens.to(torch.int32)
-                metadata.cu_seqlens_k = torch.nn.functional.pad(
-                    torch.cumsum(seq_lens, dim=0, dtype=torch.int32), (1, 0)
+                # Optimize cumulative sequence length calculation
+                metadata.cu_seqlens_k[1:].copy_(
+                    torch.cumsum(seq_lens, dim=0, dtype=torch.int32)
                 )
 
                 max_seq_pages = (
@@ -1604,7 +1642,7 @@ class FlashAttentionBackend(AttentionBackend):
                 metadata.page_table[:, :max_seq_pages].copy_(page_indices)
                 metadata.page_table[:, max_seq_pages:].fill_(0)
 
-                self._init_local_attn_metadata(metadata, device)
+                self._update_local_attn_metadata_for_replay(metadata, bs)
         elif forward_mode.is_target_verify():
             if self.topk <= 1:
                 metadata = self.target_verify_metadata[bs]
@@ -1615,13 +1653,8 @@ class FlashAttentionBackend(AttentionBackend):
                 metadata.max_seq_len_k = (
                     seq_lens_cpu.max().item() + self.speculative_num_draft_tokens
                 )
-                metadata.cu_seqlens_k.copy_(
-                    torch.nn.functional.pad(
-                        torch.cumsum(
-                            metadata.cache_seqlens_int32, dim=0, dtype=torch.int32
-                        ),
-                        (1, 0),
-                    )
+                metadata.cu_seqlens_k[1:].copy_(
+                    torch.cumsum(metadata.cache_seqlens_int32, dim=0, dtype=torch.int32)
                 )
                 max_seq_pages = (
                     metadata.max_seq_len_k + self.page_size - 1
@@ -1640,13 +1673,8 @@ class FlashAttentionBackend(AttentionBackend):
                 # metadata.max_seq_len_q = self.speculative_num_draft_tokens, already set in capture
                 metadata.max_seq_len_k = seq_lens_cpu.max().item()
                 # metadata.cu_seqlens_q already set in capture
-                metadata.cu_seqlens_k.copy_(
-                    torch.nn.functional.pad(
-                        torch.cumsum(
-                            metadata.cache_seqlens_int32, dim=0, dtype=torch.int32
-                        ),
-                        (1, 0),
-                    )
+                metadata.cu_seqlens_k[1:].copy_(
+                    torch.cumsum(metadata.cache_seqlens_int32, dim=0, dtype=torch.int32)
                 )
                 page_table = self.req_to_token[
                     req_pool_indices, : metadata.max_seq_len_k
@@ -1704,14 +1732,11 @@ class FlashAttentionBackend(AttentionBackend):
                 metadata_expand.cache_seqlens_int32.copy_(
                     mask.sum(dim=1).to(torch.int32)
                 )
-                metadata_expand.cu_seqlens_k.copy_(
-                    torch.nn.functional.pad(
-                        torch.cumsum(
-                            metadata_expand.cache_seqlens_int32,
-                            dim=0,
-                            dtype=torch.int32,
-                        ),
-                        (1, 0),
+                metadata_expand.cu_seqlens_k[1:].copy_(
+                    torch.cumsum(
+                        metadata_expand.cache_seqlens_int32,
+                        dim=0,
+                        dtype=torch.int32,
                     )
                 )
                 metadata_expand.max_seq_len_k = (
@@ -1722,11 +1747,8 @@ class FlashAttentionBackend(AttentionBackend):
             # Only support encoder size 1 for now
             metadata.encoder_max_seq_len_k = encoder_lens[0]
             metadata.encoder_lens_int32.copy_(encoder_lens[:1])
-            metadata.encoder_cu_seqlens_k.copy_(
-                torch.nn.functional.pad(
-                    torch.cumsum(metadata.encoder_lens_int32, dim=0, dtype=torch.int32),
-                    (1, 0),
-                )
+            metadata.encoder_cu_seqlens_k[1:].copy_(
+                torch.cumsum(metadata.encoder_lens_int32, dim=0, dtype=torch.int32)
             )
 
             metadata.encoder_page_table[:, : metadata.encoder_max_seq_len_k].copy_(
@@ -1776,6 +1798,7 @@ class FlashAttentionBackend(AttentionBackend):
             page_table,
             self.page_size,
         )
+
         local_metadata = FlashAttentionMetadata.LocalAttentionMetadata(
             local_query_start_loc=torch.from_numpy(cu_seqlens_q_local_np).to(device),
             local_seqused_k=torch.from_numpy(seqlens_k_local_np).to(device),
@@ -1785,6 +1808,79 @@ class FlashAttentionBackend(AttentionBackend):
         )
         metadata.local_attn_metadata = local_metadata
 
+    def _update_local_attn_metadata_for_replay(
+        self, metadata: FlashAttentionMetadata, bs: int
+    ):
+        """Update preallocated local attention metadata in-place before CUDA graph replay."""
+        if self.attention_chunk_size is None:
+            return
+
+        # Access preallocated buffers
+        local_q_buf = self.decode_cuda_graph_local_attn_metadata[
+            "local_query_start_loc"
+        ]
+        local_k_buf = self.decode_cuda_graph_local_attn_metadata["local_seqused_k"]
+        local_block_buf = self.decode_cuda_graph_local_attn_metadata[
+            "local_block_table"
+        ]
+        cu_seqlens_q = self.decode_cuda_graph_metadata["cu_seqlens_q"]
+
+        # Create a modified version for local attention that only processes the last token
+        # This mimics the normal decode pattern
+        cu_seqlens_q = torch.arange(
+            bs + 1, device=cu_seqlens_q.device, dtype=cu_seqlens_q.dtype
+        )
+        seqlens = metadata.cache_seqlens_int32[:bs]
+        # Slice the page_table to match the batch size and actual sequence length
+        # This serves three important purposes:
+        # 1. Ensures we only process the actual batch size (bs) and not the maximum batch size
+        # 2. Limits the sequence length to prevent processing padding tokens or garbage values
+        # 3. Prevents zeros in the block table which can cause garbage output during replay
+        #
+        # Without this slicing, the pre-allocated page_table may contain zeros or invalid indices
+        # beyond the actual sequence length, leading to incorrect attention calculations
+        max_seq_len = int(seqlens.max().item())
+        sliced_page_table = metadata.page_table[:bs, :max_seq_len]
+
+        cu_seqlens_q_np = cu_seqlens_q.cpu().numpy()
+        seqlens_np = seqlens.cpu().numpy()
+        (
+            seqlens_q_local_np,
+            cu_seqlens_q_local_np,
+            seqlens_k_local_np,
+            block_table_local,
+        ) = make_local_attention_virtual_batches(
+            self.attention_chunk_size,
+            cu_seqlens_q_np,
+            seqlens_np,
+            sliced_page_table,
+            self.page_size,
+        )
+
+        # Convert back to tensors
+        device = local_q_buf.device
+        cu_seqlens_q_local = torch.from_numpy(cu_seqlens_q_local_np).to(device)
+        seqlens_k_local = torch.from_numpy(seqlens_k_local_np).to(device)
+        block_table_local = block_table_local.to(device)
+        # Get sizes
+        q_len = cu_seqlens_q_local.shape[0]
+        k_len = seqlens_k_local.shape[0]
+        b0, b1 = block_table_local.shape
+
+        # In-place updates into preallocated tensors and zero out the unused space
+        local_q_buf[:q_len].copy_(cu_seqlens_q_local)
+        local_q_buf[q_len:].fill_(0)
+        local_k_buf[:k_len].copy_(seqlens_k_local)
+        local_k_buf[k_len:].fill_(0)
+        local_block_buf[:b0, :b1].copy_(block_table_local)
+        local_block_buf[b0:, :].fill_(0)
+        local_block_buf[:b0, b1:].fill_(0)
+
+        if metadata.local_attn_metadata is not None:
+            lam = metadata.local_attn_metadata
+            lam.local_max_query_len = int(seqlens_q_local_np.max())
+            lam.local_max_seq_len = int(seqlens_k_local_np.max())
+
 
 class FlashAttentionMultiStepBackend:
 

sglang/srt/layers/attention/flashinfer_backend.py

@@ -15,6 +15,12 @@ from typing import TYPE_CHECKING, Callable, List, Optional, Union
 
 import torch
 
+if os.environ["SGLANG_ENABLE_TORCH_COMPILE"] == "1":
+    import logging
+
+    torch._logging.set_logs(dynamo=logging.ERROR)
+    torch._dynamo.config.suppress_errors = True
+
 from sglang.global_config import global_config
 from sglang.srt.layers.attention.base_attn_backend import AttentionBackend
 from sglang.srt.layers.attention.utils import create_flashinfer_kv_indices_triton
@@ -82,8 +88,6 @@ class FlashInferAttnBackend(AttentionBackend):
         self.max_context_len = model_runner.model_config.context_len
         self.skip_prefill = skip_prefill
         self.is_multimodal = model_runner.model_config.is_multimodal
-        self.kv_cache_dtype = model_runner.kv_cache_dtype
-        self.kv_cache_dtype_str = model_runner.server_args.kv_cache_dtype
 
         assert not (
             model_runner.sliding_window_size is not None
@@ -104,6 +108,7 @@ class FlashInferAttnBackend(AttentionBackend):
         if (
             "Qwen2ForCausalLM" in model_runner.model_config.hf_config.architectures
             or "Qwen3ForCausalLM" in model_runner.model_config.hf_config.architectures
+            or "MiMoForCausalLM" in model_runner.model_config.hf_config.architectures
         ):
             global_config.flashinfer_workspace_size = 512 * 1024 * 1024
 
@@ -268,6 +273,12 @@ class FlashInferAttnBackend(AttentionBackend):
             cuda_graph_kv_indices.clone() for _ in range(self.num_wrappers - 1)
         ]
 
+        # Ensure tensors are properly allocated
+        for i in range(self.num_wrappers):
+            # Force allocation by performing a small operation
+            if len(self.cuda_graph_kv_indices[i]) > 0:
+                self.cuda_graph_kv_indices[i][0] = 0
+
         if not self.skip_prefill:
             self.cuda_graph_custom_mask = torch.zeros(
                 (max_bs * self.max_context_len),
@@ -396,8 +407,6 @@ class FlashInferAttnBackend(AttentionBackend):
         forward_batch: ForwardBatch,
         save_kv_cache=True,
     ):
-        k_scale = layer.k_scale_float if self.kv_cache_dtype_str != "auto" else None
-        v_scale = layer.v_scale_float if self.kv_cache_dtype_str != "auto" else None
         prefill_wrapper_paged = self.forward_metadata.prefill_wrappers[
             self._get_wrapper_idx(layer)
         ]
@@ -409,39 +418,47 @@ class FlashInferAttnBackend(AttentionBackend):
 
         logits_soft_cap = layer.logit_cap
 
+        q = q.contiguous()
         if not self.forward_metadata.use_ragged:
             if k is not None:
                 assert v is not None
                 if save_kv_cache:
                     forward_batch.token_to_kv_pool.set_kv_buffer(
-                        layer, cache_loc, k, v, k_scale, v_scale
+                        layer, cache_loc, k, v, layer.k_scale, layer.v_scale
                     )
 
             o = prefill_wrapper_paged.forward(
-                q.contiguous().view(-1, layer.tp_q_head_num, layer.head_dim),
+                q.view(-1, layer.tp_q_head_num, layer.head_dim),
                 forward_batch.token_to_kv_pool.get_kv_buffer(layer.layer_id),
                 causal=not layer.is_cross_attention,
                 sm_scale=layer.scaling,
                 window_left=layer.sliding_window_size,
                 logits_soft_cap=logits_soft_cap,
-                k_scale=k_scale,
-                v_scale=v_scale,
+                k_scale=layer.k_scale,
+                v_scale=layer.v_scale,
             )
         else:
-            o1, s1 = self.prefill_wrapper_ragged.forward_return_lse(
-                q.view(-1, layer.tp_q_head_num, layer.head_dim),
-                k.view(-1, layer.tp_k_head_num, layer.head_dim),
-                v.view(-1, layer.tp_v_head_num, layer.head_dim),
-                causal=True,
-                sm_scale=layer.scaling,
-                logits_soft_cap=logits_soft_cap,
-            )
-
             if self.forward_metadata.extend_no_prefix:
-                o = o1
+                o = self.prefill_wrapper_ragged.forward(
+                    q.view(-1, layer.tp_q_head_num, layer.head_dim),
+                    k.view(-1, layer.tp_k_head_num, layer.head_dim),
+                    v.view(-1, layer.tp_v_head_num, layer.head_dim),
+                    causal=True,
+                    sm_scale=layer.scaling,
+                    logits_soft_cap=logits_soft_cap,
+                )
+
             else:
+                o1, s1 = self.prefill_wrapper_ragged.forward_return_lse(
+                    q.view(-1, layer.tp_q_head_num, layer.head_dim),
+                    k.view(-1, layer.tp_k_head_num, layer.head_dim),
+                    v.view(-1, layer.tp_v_head_num, layer.head_dim),
+                    causal=True,
+                    sm_scale=layer.scaling,
+                    logits_soft_cap=logits_soft_cap,
+                )
                 o2, s2 = prefill_wrapper_paged.forward_return_lse(
-                    q.contiguous().view(-1, layer.tp_q_head_num, layer.head_dim),
+                    q.view(-1, layer.tp_q_head_num, layer.head_dim),
                     forward_batch.token_to_kv_pool.get_kv_buffer(layer.layer_id),
                     causal=False,
                     sm_scale=layer.scaling,
@@ -452,7 +469,7 @@ class FlashInferAttnBackend(AttentionBackend):
 
             if save_kv_cache:
                 forward_batch.token_to_kv_pool.set_kv_buffer(
-                    layer, cache_loc, k, v, k_scale, v_scale
+                    layer, cache_loc, k, v, layer.k_scale, layer.v_scale
                 )
 
         return o.view(-1, layer.tp_q_head_num * layer.head_dim)
@@ -466,8 +483,6 @@ class FlashInferAttnBackend(AttentionBackend):
         forward_batch: ForwardBatch,
         save_kv_cache=True,
     ):
-        k_scale = layer.k_scale_float if self.kv_cache_dtype_str != "auto" else None
-        v_scale = layer.v_scale_float if self.kv_cache_dtype_str != "auto" else None
         decode_wrapper = self.forward_metadata.decode_wrappers[
             self._get_wrapper_idx(layer)
         ]
@@ -481,16 +496,17 @@ class FlashInferAttnBackend(AttentionBackend):
             assert v is not None
             if save_kv_cache:
                 forward_batch.token_to_kv_pool.set_kv_buffer(
-                    layer, cache_loc, k, v, k_scale, v_scale
+                    layer, cache_loc, k, v, layer.k_scale, layer.v_scale
                 )
 
+        # Call the wrapped function
         o = decode_wrapper.forward(
             q.contiguous().view(-1, layer.tp_q_head_num, layer.head_dim),
             forward_batch.token_to_kv_pool.get_kv_buffer(layer.layer_id),
             sm_scale=layer.scaling,
             logits_soft_cap=layer.logit_cap,
-            k_scale=k_scale,
-            v_scale=v_scale,
+            k_scale=layer.k_scale,
+            v_scale=layer.v_scale,
         )
 
         return o.view(-1, layer.tp_q_head_num * layer.head_dim)
@@ -1146,8 +1162,9 @@ def fast_decode_plan(
     pos_encoding_mode: str = "NONE",
     window_left: int = -1,
     logits_soft_cap: Optional[float] = None,
-    data_type: Union[str, torch.dtype] = "float16",
     q_data_type: Optional[Union[str, torch.dtype]] = None,
+    kv_data_type: Optional[Union[str, torch.dtype]] = None,
+    data_type: Optional[Union[str, torch.dtype]] = None,
     sm_scale: Optional[float] = None,
     rope_scale: Optional[float] = None,
     rope_theta: Optional[float] = None,
@@ -1163,6 +1180,18 @@ def fast_decode_plan(
     if logits_soft_cap is None:
         logits_soft_cap = 0.0
 
+    # Handle data types consistently
+    if data_type is not None:
+        if q_data_type is None:
+            q_data_type = data_type
+        if kv_data_type is None:
+            kv_data_type = data_type
+    elif q_data_type is None:
+        q_data_type = "float16"
+
+    if kv_data_type is None:
+        kv_data_type = q_data_type
+
     if self.use_tensor_cores:
         qo_indptr_host = _get_range_buf(batch_size + 1, "cpu")
 
@@ -1178,36 +1207,33 @@ def fast_decode_plan(
             raise ValueError(
                 "The size of indices should be less than or equal to the allocated buffer"
             )
-        # Skip these copies because we directly write to them during prepartion
-        # self._paged_kv_indptr_buf.copy_(indptr)
-        # self._paged_kv_indices_buf[: len(indices)] = indices
-        # self._paged_kv_last_page_len_buf.copy_(last_page_len)
     else:
         self._paged_kv_indptr_buf = indptr
         self._paged_kv_indices_buf = indices
         self._paged_kv_last_page_len_buf = last_page_len
-        self._qo_indptr_buf = qo_indptr_host.to(self.device, non_blocking=non_blocking)
-
-    # NOTE(Zihao): the following tensors acts as placeholder to pass dtype info
-    if not q_data_type:
-        q_data_type = data_type
-
-    if not hasattr(self, "empty_q_data"):
-        self.empty_q_data = torch.empty(
-            0,
-            dtype=(
-                getattr(torch, q_data_type)
-                if isinstance(q_data_type, str)
-                else q_data_type
-            ),
-        )
-        self.empty_kv_cache = torch.empty(
-            0,
-            dtype=(
-                getattr(torch, data_type) if isinstance(data_type, str) else data_type
-            ),
-        )
-        self.last_page_len = torch.ones(32768, dtype=torch.int32)
+        if self.use_tensor_cores:
+            self._qo_indptr_buf = qo_indptr_host.to(
+                self.device, non_blocking=non_blocking
+            )
+
+    # Create empty tensors for dtype info if needed
+    empty_q_data = torch.empty(
+        0,
+        dtype=(
+            getattr(torch, q_data_type) if isinstance(q_data_type, str) else q_data_type
+        ),
+        device=self.device,
+    )
+
+    empty_kv_cache = torch.empty(
+        0,
+        dtype=(
+            getattr(torch, kv_data_type)
+            if isinstance(kv_data_type, str)
+            else kv_data_type
+        ),
+        device=self.device,
+    )
 
     indptr_host = (
         global_override_indptr_cpu
@@ -1215,48 +1241,57 @@ def fast_decode_plan(
         else indptr.cpu()
     )
 
-    if self.use_tensor_cores:
-        kv_lens_arr_host = get_seq_lens(
-            indptr_host, self.last_page_len[:batch_size], page_size
-        )
-
-        self._plan_info = self._cached_module.plan(
-            self._float_workspace_buffer,
-            self._int_workspace_buffer,
-            self._pin_memory_int_workspace_buffer,
-            qo_indptr_host,
-            indptr_host,
-            kv_lens_arr_host,
-            batch_size,  # total_num_rows
-            batch_size,
-            num_qo_heads,
-            num_kv_heads,
-            page_size,
-            self.is_cuda_graph_enabled,
-            head_dim,
-            head_dim,
-            False,  # causal
-            torch.cuda.current_stream().cuda_stream,
-        )
-    else:
-        self._plan_info = self._cached_module.plan(
-            self._float_workspace_buffer,
-            self._int_workspace_buffer,
-            self._pin_memory_int_workspace_buffer,
-            indptr_host,
-            batch_size,
-            num_qo_heads,
-            num_kv_heads,
-            page_size,
-            self.is_cuda_graph_enabled,
-            window_left,
-            logits_soft_cap,
-            head_dim,
-            head_dim,
-            self.empty_q_data,
-            self.empty_kv_cache,
-            torch.cuda.current_stream().cuda_stream,
-        )
+    with torch.cuda.device(self.device):
+
+        if self.use_tensor_cores:
+            # ALSO convert last_page_len to CPU
+            last_page_len_host = last_page_len.cpu()
+
+            kv_lens_arr_host = get_seq_lens(indptr_host, last_page_len_host, page_size)
+
+            try:
+                # Make sure we pass exactly 15 arguments for tensor core version
+                self._plan_info = self._cached_module.plan(
+                    self._float_workspace_buffer,
+                    self._int_workspace_buffer,
+                    self._pin_memory_int_workspace_buffer,
+                    qo_indptr_host,
+                    indptr_host,
+                    kv_lens_arr_host,
+                    batch_size,  # total_num_rows
+                    batch_size,
+                    num_qo_heads,
+                    num_kv_heads,
+                    page_size,
+                    self.is_cuda_graph_enabled,
+                    head_dim,
+                    head_dim,
+                    False,  # causal
+                )
+            except Exception as e:
+                raise RuntimeError(f"Error in standard plan: {e}")
+        else:
+            try:
+                # Make sure we pass exactly 15 arguments for standard version
+                self._plan_info = self._cached_module.plan(
+                    self._float_workspace_buffer,
+                    self._int_workspace_buffer,
+                    self._pin_memory_int_workspace_buffer,
+                    indptr_host,
+                    batch_size,
+                    num_qo_heads,
+                    num_kv_heads,
+                    page_size,
+                    self.is_cuda_graph_enabled,
+                    window_left,
+                    logits_soft_cap,
+                    head_dim,
+                    head_dim,
+                    empty_q_data,
+                    empty_kv_cache,
+                )
+            except Exception as e:
+                raise RuntimeError(f"Error in standard plan: {e}")
 
     self._pos_encoding_mode = pos_encoding_mode
     self._window_left = window_left