sglang 0.5.3rc2__py3-none-any.whl → 0.5.4__py3-none-any.whl

sglang/srt/mem_cache/memory_pool.py

@@ -15,9 +15,12 @@ limitations under the License.
 
 from __future__ import annotations
 
+from dataclasses import dataclass
+
+from sglang.srt.configs.mamba_utils import Mamba2CacheParams
 from sglang.srt.layers.attention.nsa import index_buf_accessor
 from sglang.srt.layers.attention.nsa.quant_k_cache import quantize_k_cache
-from sglang.srt.torch_memory_saver_adapter import TorchMemorySaverAdapter
+from sglang.srt.utils.torch_memory_saver_adapter import TorchMemorySaverAdapter
 
 """
 Memory pool.
@@ -44,6 +47,8 @@ from sglang.srt.utils import get_bool_env_var, is_cuda, is_npu, next_power_of_2
 
 if TYPE_CHECKING:
     from sglang.srt.managers.cache_controller import LayerDoneCounter
+    from sglang.srt.managers.schedule_batch import Req
+
 
 logger = logging.getLogger(__name__)
 
@@ -109,92 +114,135 @@ class ReqToTokenPool:
 
 
 class MambaPool:
+    @dataclass(frozen=True, kw_only=True)
+    class State:
+        conv: torch.Tensor
+        temporal: torch.Tensor
+
+        def at_layer_idx(self, layer: int):
+            return type(self)(**{k: v[layer] for k, v in vars(self).items()})
+
+        def mem_usage_bytes(self):
+            return sum(get_tensor_size_bytes(t) for t in vars(self).values())
+
+    @dataclass(frozen=True, kw_only=True)
+    class SpeculativeState(State):
+        intermediate_ssm: torch.Tensor
+        intermediate_conv_window: torch.Tensor
+
     def __init__(
         self,
+        *,
         size: int,
-        conv_dtype: torch.dtype,
-        ssm_dtype: torch.dtype,
-        num_mamba_layers: int,
-        conv_state_shape: Tuple[int, int],
-        temporal_state_shape: Tuple[int, int],
+        cache_params: "Mamba2CacheParams",
         device: str,
         speculative_num_draft_tokens: Optional[int] = None,
     ):
-        conv_state = torch.zeros(
-            size=(num_mamba_layers, size + 1) + conv_state_shape,
-            dtype=conv_dtype,
-            device=device,
-        )
-        temporal_state = torch.zeros(
-            size=(num_mamba_layers, size + 1) + temporal_state_shape,
-            dtype=ssm_dtype,
-            device=device,
+        conv_state_shape = cache_params.shape.conv
+        temporal_state_shape = cache_params.shape.temporal
+        conv_dtype = cache_params.dtype.conv
+        ssm_dtype = cache_params.dtype.temporal
+        num_mamba_layers = len(cache_params.layers)
+
+        # for disagg with nvlink
+        self.enable_custom_mem_pool = get_bool_env_var(
+            "SGLANG_MOONCAKE_CUSTOM_MEM_POOL", "false"
         )
-        if speculative_num_draft_tokens is not None:
-            # Cache intermediate SSM states per draft token during target verify
-            # Shape: [num_layers, size + 1, speculative_num_draft_tokens, HV, K, V]
-            intermediate_ssm_state_cache = torch.zeros(
-                size=(
-                    num_mamba_layers,
-                    size + 1,
-                    speculative_num_draft_tokens,
-                    temporal_state_shape[0],
-                    temporal_state_shape[1],
-                    temporal_state_shape[2],
-                ),
-                dtype=ssm_dtype,
-                device="cuda",
-            )
-            # Cache intermediate conv windows (last K-1 inputs) per draft token during target verify
-            # Shape: [num_layers, size + 1, speculative_num_draft_tokens, dim, K-1]
-            intermediate_conv_window_cache = torch.zeros(
-                size=(
-                    num_mamba_layers,
-                    size + 1,
-                    speculative_num_draft_tokens,
-                    conv_state_shape[0],
-                    conv_state_shape[1],
-                ),
+        if self.enable_custom_mem_pool:
+            # TODO(shangming): abstract custom allocator class for more backends
+            from mooncake.allocator import NVLinkAllocator
+
+            allocator = NVLinkAllocator.get_allocator(self.device)
+            self.custom_mem_pool = torch.cuda.MemPool(allocator.allocator())
+        else:
+            self.custom_mem_pool = None
+
+        with (
+            torch.cuda.use_mem_pool(self.custom_mem_pool)
+            if self.enable_custom_mem_pool
+            else nullcontext()
+        ):
+            # assume conv_state = (dim, state_len)
+            assert conv_state_shape[0] > conv_state_shape[1]
+            conv_state = torch.zeros(
+                size=(num_mamba_layers, size + 1) + conv_state_shape,
                 dtype=conv_dtype,
-                device="cuda",
-            )
-            self.mamba_cache = (
-                conv_state,
-                temporal_state,
-                intermediate_ssm_state_cache,
-                intermediate_conv_window_cache,
+                device=device,
             )
-            logger.info(
-                f"Mamba Cache is allocated. "
-                f"conv_state size: {get_tensor_size_bytes(conv_state) / GB:.2f}GB, "
-                f"ssm_state size: {get_tensor_size_bytes(temporal_state) / GB:.2f}GB "
-                f"intermediate_ssm_state_cache size: {get_tensor_size_bytes(intermediate_ssm_state_cache) / GB:.2f}GB "
-                f"intermediate_conv_window_cache size: {get_tensor_size_bytes(intermediate_conv_window_cache) / GB:.2f}GB "
+            temporal_state = torch.zeros(
+                size=(num_mamba_layers, size + 1) + temporal_state_shape,
+                dtype=ssm_dtype,
+                device=device,
             )
-        else:
-            self.mamba_cache = (conv_state, temporal_state)
-            logger.info(
-                f"Mamba Cache is allocated. "
-                f"conv_state size: {get_tensor_size_bytes(conv_state) / GB:.2f}GB, "
-                f"ssm_state size: {get_tensor_size_bytes(temporal_state) / GB:.2f}GB "
+            if speculative_num_draft_tokens is not None:
+                # Cache intermediate SSM states per draft token during target verify
+                # Shape: [num_layers, size + 1, speculative_num_draft_tokens, HV, K, V]
+                intermediate_ssm_state_cache = torch.zeros(
+                    size=(
+                        num_mamba_layers,
+                        size + 1,
+                        speculative_num_draft_tokens,
+                        temporal_state_shape[0],
+                        temporal_state_shape[1],
+                        temporal_state_shape[2],
+                    ),
+                    dtype=ssm_dtype,
+                    device="cuda",
+                )
+                # Cache intermediate conv windows (last K-1 inputs) per draft token during target verify
+                # Shape: [num_layers, size + 1, speculative_num_draft_tokens, dim, K-1]
+                intermediate_conv_window_cache = torch.zeros(
+                    size=(
+                        num_mamba_layers,
+                        size + 1,
+                        speculative_num_draft_tokens,
+                        conv_state_shape[0],
+                        conv_state_shape[1],
+                    ),
+                    dtype=conv_dtype,
+                    device="cuda",
+                )
+                self.mamba_cache = self.SpeculativeState(
+                    conv=conv_state,
+                    temporal=temporal_state,
+                    intermediate_ssm=intermediate_ssm_state_cache,
+                    intermediate_conv_window=intermediate_conv_window_cache,
+                )
+                logger.info(
+                    f"Mamba Cache is allocated. "
+                    f"max_mamba_cache_size: {size}, "
+                    f"conv_state size: {get_tensor_size_bytes(conv_state) / GB:.2f}GB, "
+                    f"ssm_state size: {get_tensor_size_bytes(temporal_state) / GB:.2f}GB "
+                    f"intermediate_ssm_state_cache size: {get_tensor_size_bytes(intermediate_ssm_state_cache) / GB:.2f}GB "
+                    f"intermediate_conv_window_cache size: {get_tensor_size_bytes(intermediate_conv_window_cache) / GB:.2f}GB "
+                )
+            else:
+                self.mamba_cache = self.State(conv=conv_state, temporal=temporal_state)
+                logger.info(
+                    f"Mamba Cache is allocated. "
+                    f"max_mamba_cache_size: {size}, "
+                    f"conv_state size: {get_tensor_size_bytes(conv_state) / GB:.2f}GB, "
+                    f"ssm_state size: {get_tensor_size_bytes(temporal_state) / GB:.2f}GB "
+                )
+            self.size = size
+            self.device = device
+            self.free_slots = torch.arange(
+                self.size, dtype=torch.int64, device=self.device
             )
-        self.size = size
-        self.free_slots = list(range(size))
-        self.mem_usage = self.get_mamba_size() / GB
+            self.mem_usage = self.mamba_cache.mem_usage_bytes() / GB
+            self.num_mamba_layers = num_mamba_layers
 
-    def get_mamba_params_all_layers(self):
-        return [self.mamba_cache[i] for i in range(len(self.mamba_cache))]
+    def get_speculative_mamba2_params_all_layers(self) -> SpeculativeState:
+        assert isinstance(self.mamba_cache, self.SpeculativeState)
+        return self.mamba_cache
 
-    def get_mamba_params(self, layer_id: int):
-        return [self.mamba_cache[i][layer_id] for i in range(len(self.mamba_cache))]
-
-    def get_mamba_size(self):
-        return sum(get_tensor_size_bytes(t) for t in self.mamba_cache)
+    def mamba2_layer_cache(self, layer_id: int):
+        return self.mamba_cache.at_layer_idx(layer_id)
 
     def available_size(self):
         return len(self.free_slots)
 
-    def alloc(self, need_size: int) -> Optional[List[int]]:
+    def alloc(self, need_size: int) -> Optional[torch.Tensor]:
         if need_size > len(self.free_slots):
             return None
 
@@ -203,15 +251,46 @@ class MambaPool:
 
         return select_index
 
-    def free(self, free_index: Union[int, List[int]]):
-        if isinstance(free_index, (int,)):
-            self.free_slots.append(free_index)
-        else:
-            self.free_slots.extend(free_index)
-        self.mamba_cache[0][:, free_index] = self.mamba_cache[1][:, free_index] = 0
+    def free(self, free_index: torch.Tensor):
+        if free_index.numel() == 0:
+            return
+        self.free_slots = torch.cat((self.free_slots, free_index))
+        self.mamba_cache.conv[:, free_index] = self.mamba_cache.temporal[
+            :, free_index
+        ] = 0
 
     def clear(self):
-        self.free_slots = list(range(self.size))
+        self.free_slots = torch.arange(self.size, dtype=torch.int64, device=self.device)
+
+    def copy_from(self, src_index: torch.Tensor, dst_index: torch.Tensor):
+        self.mamba_cache.conv[:, dst_index] = self.mamba_cache.conv[:, src_index]
+        self.mamba_cache.temporal[:, dst_index] = self.mamba_cache.temporal[
+            :, src_index
+        ]
+        return
+
+    def fork_from(self, src_index: torch.Tensor) -> Optional[torch.Tensor]:
+        dst_index = self.alloc(1)
+        if dst_index == None:
+            return None
+        self.copy_from(src_index, dst_index)
+        return dst_index
+
+    def get_contiguous_buf_infos(self):
+        state_tensors = [
+            getattr(self.mamba_cache, field) for field in vars(self.mamba_cache)
+        ]
+        data_ptrs, data_lens, item_lens = [], [], []
+
+        for _, state_tensor in enumerate(state_tensors):
+            data_ptrs += [
+                state_tensor[i].data_ptr() for i in range(self.num_mamba_layers)
+            ]
+            data_lens += [state_tensor[i].nbytes for i in range(self.num_mamba_layers)]
+            item_lens += [
+                state_tensor[i][0].nbytes for i in range(self.num_mamba_layers)
+            ]
+        return data_ptrs, data_lens, item_lens
 
 
 class HybridReqToTokenPool(ReqToTokenPool):
@@ -219,16 +298,14 @@ class HybridReqToTokenPool(ReqToTokenPool):
 
     def __init__(
         self,
+        *,
         size: int,
+        mamba_size: int,
         max_context_len: int,
         device: str,
         enable_memory_saver: bool,
-        conv_dtype: torch.dtype,
-        ssm_dtype: torch.dtype,
-        mamba_layers: List[int],
-        conv_state_shape: Tuple[int, int],
-        temporal_state_shape: Tuple[int, int],
-        speculative_num_draft_tokens: int,
+        cache_params: "Mamba2CacheParams",
+        speculative_num_draft_tokens: int = None,
     ):
         super().__init__(
             size=size,
@@ -236,31 +313,37 @@ class HybridReqToTokenPool(ReqToTokenPool):
             device=device,
             enable_memory_saver=enable_memory_saver,
         )
+        self._init_mamba_pool(
+            size=mamba_size,
+            cache_params=cache_params,
+            device=device,
+            speculative_num_draft_tokens=speculative_num_draft_tokens,
+        )
 
+    def _init_mamba_pool(
+        self,
+        size: int,
+        cache_params: "Mamba2CacheParams",
+        device: str,
+        speculative_num_draft_tokens: int = None,
+    ):
         self.mamba_pool = MambaPool(
-            size,
-            conv_dtype,
-            ssm_dtype,
-            len(mamba_layers),
-            conv_state_shape,
-            temporal_state_shape,
-            device,
-            speculative_num_draft_tokens,
+            size=size,
+            cache_params=cache_params,
+            device=device,
+            speculative_num_draft_tokens=speculative_num_draft_tokens,
         )
-        self.mamba_map = {layer_id: i for i, layer_id in enumerate(mamba_layers)}
+        self.mamba_map = {layer_id: i for i, layer_id in enumerate(cache_params.layers)}
 
         self.device = device
         self.req_index_to_mamba_index_mapping: torch.Tensor = torch.zeros(
             size, dtype=torch.int32, device=self.device
         )
 
-        self.rid_to_mamba_index_mapping: Dict[str, int] = {}
-        self.mamba_index_to_rid_mapping: Dict[int, str] = {}
-
     # For chunk prefill req, we do not need to allocate mamba cache,
     # We could use allocated mamba cache instead.
     def alloc(
-        self, need_size: int, reqs: Optional[List["Req"]] = None
+        self, need_size: int, reqs: Optional[List[Req]] = None
     ) -> Optional[List[int]]:
         select_index = super().alloc(need_size)
         if select_index == None:
@@ -268,14 +351,14 @@ class HybridReqToTokenPool(ReqToTokenPool):
 
         mamba_index = []
         for req in reqs:
-            rid = req.rid
-            if rid in self.rid_to_mamba_index_mapping:
-                mid = self.rid_to_mamba_index_mapping[rid]
-            elif (mid := self.mamba_pool.alloc(1)) is not None:
-                mid = mid[0]
-                self.rid_to_mamba_index_mapping[rid] = mid
-                self.mamba_index_to_rid_mapping[mid] = rid
-            mamba_index.append(mid)
+            mid = None
+            if req.mamba_pool_idx is not None:  # for radix cache
+                mid = req.mamba_pool_idx
+            else:
+                mid = self.mamba_pool.alloc(1)[0]
+                req.mamba_pool_idx = mid
+            if mid is not None:
+                mamba_index.append(mid)
         assert len(select_index) == len(
             mamba_index
         ), f"Not enough space for mamba cache, try to increase --max-mamba-cache-size."
@@ -287,26 +370,21 @@ class HybridReqToTokenPool(ReqToTokenPool):
     def get_mamba_indices(self, req_indices: torch.Tensor) -> torch.Tensor:
         return self.req_index_to_mamba_index_mapping[req_indices]
 
-    def get_mamba_params(self, layer_id: int):
+    def mamba2_layer_cache(self, layer_id: int):
         assert layer_id in self.mamba_map
-        return self.mamba_pool.get_mamba_params(self.mamba_map[layer_id])
+        return self.mamba_pool.mamba2_layer_cache(self.mamba_map[layer_id])
 
-    def get_mamba_params_all_layers(self):
-        return self.mamba_pool.get_mamba_params_all_layers()
+    def get_speculative_mamba2_params_all_layers(self) -> MambaPool.SpeculativeState:
+        return self.mamba_pool.get_speculative_mamba2_params_all_layers()
 
     # For chunk prefill, we can not free mamba cache, we need use it in the future
     def free(self, free_index: Union[int, List[int]], free_mamba_cache: bool = True):
+        if isinstance(free_index, (int,)):
+            free_index = [free_index]
         super().free(free_index)
         if free_mamba_cache:
             mamba_index = self.req_index_to_mamba_index_mapping[free_index]
-            mamba_index_list = mamba_index.tolist()
-            if isinstance(mamba_index_list, int):
-                mamba_index_list = [mamba_index_list]
-            self.mamba_pool.free(mamba_index_list)
-            for mid in mamba_index_list:
-                rid = self.mamba_index_to_rid_mapping[mid]
-                self.mamba_index_to_rid_mapping.pop(mid)
-                self.rid_to_mamba_index_mapping.pop(rid)
+            self.mamba_pool.free(mamba_index)
 
     def clear(self):
         super().clear()
@@ -349,6 +427,19 @@ class KVCache(abc.ABC):
         # default state for optional layer-wise transfer control
         self.layer_transfer_counter = None
 
+        # for disagg with nvlink
+        self.enable_custom_mem_pool = get_bool_env_var(
+            "SGLANG_MOONCAKE_CUSTOM_MEM_POOL", "false"
+        )
+        if self.enable_custom_mem_pool:
+            # TODO(shangming): abstract custom allocator class for more backends
+            from mooncake.allocator import NVLinkAllocator
+
+            allocator = NVLinkAllocator.get_allocator(self.device)
+            self.custom_mem_pool = torch.cuda.MemPool(allocator.allocator())
+        else:
+            self.custom_mem_pool = None
+
     def _finalize_allocation_log(self, num_tokens: int):
         """Common logging and mem_usage computation for KV cache allocation.
         Supports both tuple (K, V) size returns and single KV size returns.
@@ -400,6 +491,9 @@ class KVCache(abc.ABC):
     def load_cpu_copy(self, kv_cache_cpu, indices):
         raise NotImplementedError()
 
+    def maybe_get_custom_mem_pool(self):
+        return self.custom_mem_pool
+
 
 class MHATokenToKVPool(KVCache):
 
@@ -415,6 +509,7 @@ class MHATokenToKVPool(KVCache):
         enable_memory_saver: bool,
         start_layer: Optional[int] = None,
         end_layer: Optional[int] = None,
+        enable_kv_cache_copy: bool = False,
     ):
         super().__init__(
             size,
@@ -429,25 +524,61 @@ class MHATokenToKVPool(KVCache):
         self.head_num = head_num
         self.head_dim = head_dim
 
-        # for disagg with nvlink
-        self.enable_custom_mem_pool = get_bool_env_var(
-            "SGLANG_MOONCAKE_CUSTOM_MEM_POOL", "false"
-        )
-        if self.enable_custom_mem_pool:
-            # TODO(shangming): abstract custom allocator class for more backends
-            from mooncake.allocator import NVLinkAllocator
-
-            allocator = NVLinkAllocator.get_allocator(self.device)
-            self.custom_mem_pool = torch.cuda.MemPool(allocator.allocator())
-        else:
-            self.custom_mem_pool = None
-
         self._create_buffers()
 
         self.device_module = torch.get_device_module(self.device)
         self.alt_stream = self.device_module.Stream() if _is_cuda else None
+
+        if enable_kv_cache_copy:
+            self._init_kv_copy_and_warmup()
+        else:
+            self._kv_copy_config = None
+
         self._finalize_allocation_log(size)
 
+    def _init_kv_copy_and_warmup(self):
+        # Heuristics for KV copy tiling
+        _KV_COPY_STRIDE_THRESHOLD_LARGE = 8192
+        _KV_COPY_STRIDE_THRESHOLD_MEDIUM = 4096
+        _KV_COPY_TILE_SIZE_LARGE = 512
+        _KV_COPY_TILE_SIZE_MEDIUM = 256
+        _KV_COPY_TILE_SIZE_SMALL = 128
+        _KV_COPY_NUM_WARPS_LARGE_TILE = 8
+        _KV_COPY_NUM_WARPS_SMALL_TILE = 4
+
+        stride_bytes = int(self.data_strides[0].item())
+        if stride_bytes >= _KV_COPY_STRIDE_THRESHOLD_LARGE:
+            bytes_per_tile = _KV_COPY_TILE_SIZE_LARGE
+        elif stride_bytes >= _KV_COPY_STRIDE_THRESHOLD_MEDIUM:
+            bytes_per_tile = _KV_COPY_TILE_SIZE_MEDIUM
+        else:
+            bytes_per_tile = _KV_COPY_TILE_SIZE_SMALL
+
+        self._kv_copy_config = {
+            "bytes_per_tile": bytes_per_tile,
+            "byte_tiles": (stride_bytes + bytes_per_tile - 1) // bytes_per_tile,
+            "num_warps": (
+                _KV_COPY_NUM_WARPS_SMALL_TILE
+                if bytes_per_tile <= _KV_COPY_TILE_SIZE_MEDIUM
+                else _KV_COPY_NUM_WARPS_LARGE_TILE
+            ),
+        }
+
+        dummy_loc = torch.zeros(1, dtype=torch.int32, device=self.device)
+        grid = (self.data_ptrs.numel(), self._kv_copy_config["byte_tiles"])
+
+        copy_all_layer_kv_cache_tiled[grid](
+            self.data_ptrs,
+            self.data_strides,
+            dummy_loc,
+            dummy_loc,
+            1,
+            1,
+            BYTES_PER_TILE=self._kv_copy_config["bytes_per_tile"],
+            num_warps=self._kv_copy_config["num_warps"],
+            num_stages=2,
+        )
+
     def _create_buffers(self):
         with self.memory_saver_adapter.region(GPU_MEMORY_TYPE_KV_CACHE):
             with (
@@ -535,9 +666,6 @@ class MHATokenToKVPool(KVCache):
         ]
         return kv_data_ptrs, kv_data_lens, kv_item_lens
 
-    def maybe_get_custom_mem_pool(self):
-        return self.custom_mem_pool
-
     def get_cpu_copy(self, indices):
         torch.cuda.synchronize()
         kv_cache_cpu = []
@@ -642,13 +770,28 @@ class MHATokenToKVPool(KVCache):
             self.v_buffer[layer_id - self.start_layer][loc] = cache_v
 
     def move_kv_cache(self, tgt_loc: torch.Tensor, src_loc: torch.Tensor):
-        copy_all_layer_kv_cache[(len(self.data_ptrs),)](
+        N = tgt_loc.numel()
+        if N == 0:
+            return
+
+        assert (
+            self._kv_copy_config is not None
+        ), "KV copy not initialized. Set enable_kv_cache_copy=True in __init__"
+
+        cfg = self._kv_copy_config
+        N_upper = next_power_of_2(N)
+        grid = (self.data_ptrs.numel(), cfg["byte_tiles"])
+
+        copy_all_layer_kv_cache_tiled[grid](
             self.data_ptrs,
             self.data_strides,
             tgt_loc,
             src_loc,
-            len(tgt_loc),
-            next_power_of_2(len(tgt_loc)),
+            N,
+            N_upper,
+            BYTES_PER_TILE=cfg["bytes_per_tile"],
+            num_warps=cfg["num_warps"],
+            num_stages=2,
         )
 
 
@@ -665,12 +808,18 @@ class HybridLinearKVPool(KVCache):
         full_attention_layer_ids: List[int],
         enable_kvcache_transpose: bool,
         device: str,
+        mamba_pool: MambaPool,
     ):
         self.size = size
         self.dtype = dtype
         self.device = device
         self.full_layer_nums = len(full_attention_layer_ids)
         self.page_size = page_size
+        # TODO support pp?
+        self.start_layer = 0
+        self.head_num = head_num
+        self.head_dim = head_dim
+        self.mamba_pool = mamba_pool
         # TODO MHATransposedTokenToKVPool if enable_kvcache_transpose is True
         assert not enable_kvcache_transpose
         if _is_npu:
@@ -699,6 +848,15 @@ class HybridLinearKVPool(KVCache):
     def get_contiguous_buf_infos(self):
         return self.full_kv_pool.get_contiguous_buf_infos()
 
+    def get_state_buf_infos(self):
+        mamba_data_ptrs, mamba_data_lens, mamba_item_lens = (
+            self.mamba_pool.get_contiguous_buf_infos()
+        )
+        return mamba_data_ptrs, mamba_data_lens, mamba_item_lens
+
+    def maybe_get_custom_mem_pool(self):
+        return self.full_kv_pool.maybe_get_custom_mem_pool()
+
     def _transfer_full_attention_id(self, layer_id: int):
         if layer_id not in self.full_attention_layer_id_mapping:
             raise ValueError(
@@ -749,28 +907,57 @@ class SWAKVPool(KVCache):
         self,
         size: int,
         size_swa: int,
+        dtype: torch.dtype,
+        head_num: int,
+        head_dim: int,
         swa_attention_layer_ids: List[int],
         full_attention_layer_ids: List[int],
         enable_kvcache_transpose: bool,
+        device: str,
         token_to_kv_pool_class: KVCache = MHATokenToKVPool,
         **kwargs,
     ):
         self.size = size
         self.size_swa = size_swa
+        self.dtype = dtype
+        self.head_num = head_num
+        self.head_dim = head_dim
+        self.device = device
         self.swa_layer_nums = len(swa_attention_layer_ids)
         self.full_layer_nums = len(full_attention_layer_ids)
+        self.start_layer = 0
+        self.page_size = 1
+
         kwargs["page_size"] = 1
         kwargs["enable_memory_saver"] = False
+        kwargs["head_num"] = head_num
+        kwargs["head_dim"] = head_dim
+        kwargs["device"] = device
         # TODO MHATransposedTokenToKVPool if enable_kvcache_transpose is True
         assert not enable_kvcache_transpose
 
+        # for disagg with nvlink
+        self.enable_custom_mem_pool = get_bool_env_var(
+            "SGLANG_MOONCAKE_CUSTOM_MEM_POOL", "false"
+        )
+        if self.enable_custom_mem_pool:
+            # TODO(shangming): abstract custom allocator class for more backends
+            from mooncake.allocator import NVLinkAllocator
+
+            allocator = NVLinkAllocator.get_allocator(self.device)
+            self.custom_mem_pool = torch.cuda.MemPool(allocator.allocator())
+        else:
+            self.custom_mem_pool = None
+
         self.swa_kv_pool = token_to_kv_pool_class(
             size=size_swa,
+            dtype=dtype,
             layer_num=self.swa_layer_nums,
             **kwargs,
         )
         self.full_kv_pool = token_to_kv_pool_class(
             size=size,
+            dtype=dtype,
             layer_num=self.full_layer_nums,
             **kwargs,
         )
@@ -783,6 +970,9 @@ class SWAKVPool(KVCache):
 
         k_size, v_size = self.get_kv_size_bytes()
         self.mem_usage = (k_size + v_size) / GB
+        logger.info(
+            f"SWAKVPool mem usage: {self.mem_usage} GB, swa size: {self.size_swa}, full size: {self.size}"
+        )
 
     def get_kv_size_bytes(self):
         k_size, v_size = self.full_kv_pool.get_kv_size_bytes()
@@ -793,15 +983,19 @@ class SWAKVPool(KVCache):
         full_kv_data_ptrs, full_kv_data_lens, full_kv_item_lens = (
             self.full_kv_pool.get_contiguous_buf_infos()
         )
+
+        kv_data_ptrs = full_kv_data_ptrs
+        kv_data_lens = full_kv_data_lens
+        kv_item_lens = full_kv_item_lens
+
+        return kv_data_ptrs, kv_data_lens, kv_item_lens
+
+    def get_state_buf_infos(self):
         swa_kv_data_ptrs, swa_kv_data_lens, swa_kv_item_lens = (
             self.swa_kv_pool.get_contiguous_buf_infos()
         )
 
-        kv_data_ptrs = full_kv_data_ptrs + swa_kv_data_ptrs
-        kv_data_lens = full_kv_data_lens + swa_kv_data_lens
-        kv_item_lens = full_kv_item_lens + swa_kv_item_lens
-
-        return kv_data_ptrs, kv_data_lens, kv_item_lens
+        return swa_kv_data_ptrs, swa_kv_data_lens, swa_kv_item_lens
 
     def get_key_buffer(self, layer_id: int):
         layer_id_pool, is_swa = self.layers_mapping[layer_id]
@@ -1057,19 +1251,6 @@ class MLATokenToKVPool(KVCache):
             else (kv_lora_rank + qk_rope_head_dim)
         )
 
-        # for disagg with nvlink
-        self.enable_custom_mem_pool = get_bool_env_var(
-            "SGLANG_MOONCAKE_CUSTOM_MEM_POOL", "false"
-        )
-        if self.enable_custom_mem_pool:
-            # TODO(shangming): abstract custom allocator class for more backends
-            from mooncake.allocator import NVLinkAllocator
-
-            allocator = NVLinkAllocator.get_allocator(self.device)
-            self.custom_mem_pool = torch.cuda.MemPool(allocator.allocator())
-        else:
-            self.custom_mem_pool = None
-
         with self.memory_saver_adapter.region(GPU_MEMORY_TYPE_KV_CACHE):
             with (
                 torch.cuda.use_mem_pool(self.custom_mem_pool)
@@ -1091,7 +1272,9 @@ class MLATokenToKVPool(KVCache):
             dtype=torch.uint64,
             device=self.device,
         )
-        self._finalize_allocation_log(size)
+        if not use_nsa:
+            # NSA will allocate indexer KV cache later and then log the total size
+            self._finalize_allocation_log(size)
 
     def get_kv_size_bytes(self):
         assert hasattr(self, "kv_buffer")
@@ -1110,9 +1293,6 @@ class MLATokenToKVPool(KVCache):
         ]
         return kv_data_ptrs, kv_data_lens, kv_item_lens
 
-    def maybe_get_custom_mem_pool(self):
-        return self.custom_mem_pool
-
     def get_key_buffer(self, layer_id: int):
         if self.layer_transfer_counter is not None:
             self.layer_transfer_counter.wait_until(layer_id - self.start_layer)
@@ -1212,6 +1392,9 @@ class MLATokenToKVPool(KVCache):
 
 
 class NSATokenToKVPool(MLATokenToKVPool):
+    quant_block_size = 128
+    index_k_with_scale_buffer_dtype = torch.uint8
+
     def __init__(
         self,
         size: int,
@@ -1245,27 +1428,33 @@ class NSATokenToKVPool(MLATokenToKVPool):
         # num head == 1 and head dim == 128 for index_k in NSA
         assert index_head_dim == 128
 
-        self.quant_block_size = 128
-
         assert self.page_size == 64
-        self.index_k_with_scale_buffer = [
-            torch.zeros(
-                # Layout:
-                #     ref: test_attention.py :: kv_cache_cast_to_fp8
-                #     shape: (num_pages, page_size 64 * head_dim 128 + page_size 64 * fp32_nbytes 4)
-                #     data: for page i,
-                #         * buf[i, :page_size * head_dim] for fp8 data
-                #         * buf[i, page_size * head_dim:].view(float32) for scale
-                (
-                    (size + page_size + 1) // self.page_size,
-                    self.page_size
-                    * (index_head_dim + index_head_dim // self.quant_block_size * 4),
-                ),
-                dtype=torch.uint8,
-                device=device,
-            )
-            for _ in range(layer_num)
-        ]
+        with (
+            torch.cuda.use_mem_pool(self.custom_mem_pool)
+            if self.custom_mem_pool
+            else nullcontext()
+        ):
+            self.index_k_with_scale_buffer = [
+                torch.zeros(
+                    # Layout:
+                    #     ref: test_attention.py :: kv_cache_cast_to_fp8
+                    #     shape: (num_pages, page_size 64 * head_dim 128 + page_size 64 * fp32_nbytes 4)
+                    #     data: for page i,
+                    #         * buf[i, :page_size * head_dim] for fp8 data
+                    #         * buf[i, page_size * head_dim:].view(float32) for scale
+                    (
+                        (size + page_size + 1) // self.page_size,
+                        self.page_size
+                        * (
+                            index_head_dim + index_head_dim // self.quant_block_size * 4
+                        ),
+                    ),
+                    dtype=self.index_k_with_scale_buffer_dtype,
+                    device=device,
+                )
+                for _ in range(layer_num)
+            ]
+        self._finalize_allocation_log(size)
 
     def get_index_k_with_scale_buffer(self, layer_id: int) -> torch.Tensor:
         if self.layer_transfer_counter is not None:
@@ -1307,6 +1496,24 @@ class NSATokenToKVPool(MLATokenToKVPool):
             pool=self, buf=buf, loc=loc, index_k=index_k, index_k_scale=index_k_scale
         )
 
+    def get_state_buf_infos(self):
+        data_ptrs = [
+            self.index_k_with_scale_buffer[i].data_ptr() for i in range(self.layer_num)
+        ]
+        data_lens = [
+            self.index_k_with_scale_buffer[i].nbytes for i in range(self.layer_num)
+        ]
+        item_lens = [
+            self.index_k_with_scale_buffer[i][0].nbytes for i in range(self.layer_num)
+        ]
+        return data_ptrs, data_lens, item_lens
+
+    def get_kv_size_bytes(self):
+        kv_size_bytes = super().get_kv_size_bytes()
+        for index_k_cache in self.index_k_with_scale_buffer:
+            kv_size_bytes += get_tensor_size_bytes(index_k_cache)
+        return kv_size_bytes
+
 
 class AscendMLAPagedTokenToKVPool(MLATokenToKVPool):
     def __init__(
@@ -1531,27 +1738,38 @@ class DoubleSparseTokenToKVPool(KVCache):
         )
 
         with self.memory_saver_adapter.region(GPU_MEMORY_TYPE_KV_CACHE):
-            # [size, head_num, head_dim] for each layer
-            self.k_buffer = [
-                torch.zeros(
-                    (size + page_size, head_num, head_dim), dtype=dtype, device=device
-                )
-                for _ in range(layer_num)
-            ]
-            self.v_buffer = [
-                torch.zeros(
-                    (size + page_size, head_num, head_dim), dtype=dtype, device=device
-                )
-                for _ in range(layer_num)
-            ]
+            with (
+                torch.cuda.use_mem_pool(self.custom_mem_pool)
+                if self.enable_custom_mem_pool
+                else nullcontext()
+            ):
+                # [size, head_num, head_dim] for each layer
+                self.k_buffer = [
+                    torch.zeros(
+                        (size + page_size, head_num, head_dim),
+                        dtype=dtype,
+                        device=device,
+                    )
+                    for _ in range(layer_num)
+                ]
+                self.v_buffer = [
+                    torch.zeros(
+                        (size + page_size, head_num, head_dim),
+                        dtype=dtype,
+                        device=device,
+                    )
+                    for _ in range(layer_num)
+                ]
 
-            # [size, head_num, heavy_channel_num] for each layer
-            self.label_buffer = [
-                torch.zeros(
-                    (size + 1, head_num, heavy_channel_num), dtype=dtype, device=device
-                )
-                for _ in range(layer_num)
-            ]
+                # [size, head_num, heavy_channel_num] for each layer
+                self.label_buffer = [
+                    torch.zeros(
+                        (size + 1, head_num, heavy_channel_num),
+                        dtype=dtype,
+                        device=device,
+                    )
+                    for _ in range(layer_num)
+                ]
 
     def get_key_buffer(self, layer_id: int):
         return self.k_buffer[layer_id - self.start_layer]
@@ -1584,38 +1802,36 @@ class DoubleSparseTokenToKVPool(KVCache):
 
 
 @triton.jit
-def copy_all_layer_kv_cache(
+def copy_all_layer_kv_cache_tiled(
     data_ptrs,
     strides,
     tgt_loc_ptr,
     src_loc_ptr,
     num_locs,
     num_locs_upper: tl.constexpr,
+    BYTES_PER_TILE: tl.constexpr,
 ):
-    BLOCK_SIZE: tl.constexpr = 128
-
+    """2D tiled kernel. Safe for in-place copy."""
     bid = tl.program_id(0)
+    tid = tl.program_id(1)
+
     stride = tl.load(strides + bid)
+    base_ptr = tl.load(data_ptrs + bid)
+    base_ptr = tl.cast(base_ptr, tl.pointer_type(tl.uint8))
 
-    data_ptr = tl.load(data_ptrs + bid)
-    data_ptr = tl.cast(data_ptr, tl.pointer_type(tl.uint8))
+    byte_off = tid * BYTES_PER_TILE + tl.arange(0, BYTES_PER_TILE)
+    mask_byte = byte_off < stride
+    tl.multiple_of(byte_off, 16)
 
-    num_locs_offset = tl.arange(0, num_locs_upper)
-    tgt_locs = tl.load(tgt_loc_ptr + num_locs_offset, mask=num_locs_offset < num_locs)
-    src_locs = tl.load(src_loc_ptr + num_locs_offset, mask=num_locs_offset < num_locs)
+    loc_idx = tl.arange(0, num_locs_upper)
+    mask_loc = loc_idx < num_locs
 
-    # NOTE: we cannot parallelize over the tgt_loc_ptr dim with cuda blocks
-    # because this copy is an inplace operation.
+    src = tl.load(src_loc_ptr + loc_idx, mask=mask_loc, other=0)
+    tgt = tl.load(tgt_loc_ptr + loc_idx, mask=mask_loc, other=0)
 
-    num_loop = tl.cdiv(stride, BLOCK_SIZE)
-    for i in range(num_loop):
-        copy_offset = tl.arange(0, BLOCK_SIZE) + i * BLOCK_SIZE
-        mask = (num_locs_offset < num_locs)[:, None] & (copy_offset < stride)[None, :]
-        value = tl.load(
-            data_ptr + src_locs[:, None] * stride + copy_offset[None, :], mask=mask
-        )
-        tl.store(
-            data_ptr + tgt_locs[:, None] * stride + copy_offset[None, :],
-            value,
-            mask=mask,
-        )
+    src_ptr = base_ptr + src[:, None] * stride + byte_off[None, :]
+    tgt_ptr = base_ptr + tgt[:, None] * stride + byte_off[None, :]
+
+    mask = mask_loc[:, None] & mask_byte[None, :]
+    vals = tl.load(src_ptr, mask=mask)
+    tl.store(tgt_ptr, vals, mask=mask)