[PD] Add custom memory pool option to support Mooncake PD with NVLink (#7264)

Signed-off-by: Shangming Cai <caishangming@linux.alibaba.com>

python/sglang/srt/disaggregation/mooncake/memory_pool.py

+import os
+import threading
+from importlib import resources
+from typing import Dict, Final, Optional
+
+import torch
+from torch.cuda.memory import CUDAPluggableAllocator
+
+
+# TODO(shangming): move this class into mooncake's package for more general use cases
+class MooncakeNVLinkAllocator:
+    _instances: Dict[torch.device, CUDAPluggableAllocator] = {}
+    _lock: Final = threading.Lock()
+
+    @classmethod
+    def _get_so_path(cls) -> str:
+        """Dynamically locate hook.so in the mooncake package installation"""
+        try:
+            # Attempt to locate package resource
+            with resources.path("mooncake", "hook.so") as so_path:
+                if so_path.exists():
+                    return str(so_path)
+        except (ImportError, FileNotFoundError, TypeError):
+            pass
+
+        # Fallback strategy: check in package location via import metadata
+        try:
+            import mooncake
+
+            base_path = os.path.dirname(os.path.abspath(mooncake.__file__))
+            so_path = os.path.join(base_path, "hook.so")
+            if os.path.exists(so_path):
+                return so_path
+        except (ImportError, FileNotFoundError, TypeError):
+            raise ImportError(
+                "SGLANG_MOONCAKE_CUSTOM_MEM_POOL require mooncake-transfer-engine >= 0.3.3.post2."
+            )
+
+    @classmethod
+    def get_allocator(cls, device: torch.device) -> CUDAPluggableAllocator:
+        with cls._lock:
+            if device not in cls._instances:
+                so_path = cls._get_so_path()
+                cls._instances[device] = CUDAPluggableAllocator(
+                    so_path, "mc_nvlink_malloc", "mc_nvlink_free"
+                )
+            return cls._instances[device]

python/sglang/srt/disaggregation/utils.py

@@ -6,6 +6,7 @@ import random
 import threading
 import warnings
 from collections import deque
+from contextlib import nullcontext
 from enum import Enum
 from typing import TYPE_CHECKING, List, Optional
 
@@ -84,24 +85,37 @@ class ReqToMetadataIdxAllocator:
 
 
 class MetadataBuffers:
-    def __init__(self, size: int, max_top_logprobs_num: int = 128):
-        # TODO: abort top_logprobs_num > 128 in PD
-
-        # We transfer the metadata of first output token to decode
-        # The minimal size for RDMA is 64Bytes, so we pad it to > 64Bytes
-        self.output_ids = torch.zeros((size, 16), dtype=torch.int32, device="cpu")
-        self.output_token_logprobs_val = torch.zeros(
-            (size, 16), dtype=torch.float32, device="cpu"
-        )
-        self.output_token_logprobs_idx = torch.zeros(
-            (size, 16), dtype=torch.int32, device="cpu"
-        )
-        self.output_top_logprobs_val = torch.zeros(
-            (size, max_top_logprobs_num), dtype=torch.float32, device="cpu"
-        )
-        self.output_top_logprobs_idx = torch.zeros(
-            (size, max_top_logprobs_num), dtype=torch.int32, device="cpu"
-        )
+    def __init__(
+        self,
+        size: int,
+        max_top_logprobs_num: int = 128,
+        custom_mem_pool: torch.cuda.MemPool = None,
+    ):
+        self.custom_mem_pool = custom_mem_pool
+        device = "cuda" if self.custom_mem_pool else "cpu"
+
+        with (
+            torch.cuda.use_mem_pool(self.custom_mem_pool)
+            if self.custom_mem_pool
+            else nullcontext()
+        ):
+            # TODO: abort top_logprobs_num > 128 in PD
+
+            # We transfer the metadata of first output token to decode
+            # The minimal size for RDMA is 64Bytes, so we pad it to > 64Bytes
+            self.output_ids = torch.zeros((size, 16), dtype=torch.int32, device=device)
+            self.output_token_logprobs_val = torch.zeros(
+                (size, 16), dtype=torch.float32, device=device
+            )
+            self.output_token_logprobs_idx = torch.zeros(
+                (size, 16), dtype=torch.int32, device=device
+            )
+            self.output_top_logprobs_val = torch.zeros(
+                (size, max_top_logprobs_num), dtype=torch.float32, device=device
+            )
+            self.output_top_logprobs_idx = torch.zeros(
+                (size, max_top_logprobs_num), dtype=torch.int32, device=device
+            )
 
     def get_buf_infos(self):
         ptrs = [

python/sglang/srt/managers/scheduler.py

@@ -622,7 +622,10 @@ class Scheduler(
             self.req_to_metadata_buffer_idx_allocator = ReqToMetadataIdxAllocator(
                 buffer_size
             )
-            self.disagg_metadata_buffers = MetadataBuffers(buffer_size)
+            self.disagg_metadata_buffers = MetadataBuffers(
+                buffer_size,
+                custom_mem_pool=self.token_to_kv_pool_allocator.get_kvcache().maybe_get_custom_mem_pool(),
+            )
 
             # The decode requests polling kv cache
             self.disagg_decode_transfer_queue = DecodeTransferQueue(
@@ -669,7 +672,10 @@ class Scheduler(
             self.req_to_metadata_buffer_idx_allocator = ReqToMetadataIdxAllocator(
                 buffer_size
             )
-            self.disagg_metadata_buffers = MetadataBuffers(buffer_size)
+            self.disagg_metadata_buffers = MetadataBuffers(
+                buffer_size,
+                custom_mem_pool=self.token_to_kv_pool_allocator.get_kvcache().maybe_get_custom_mem_pool(),
+            )
 
             self.disagg_prefill_bootstrap_queue = PrefillBootstrapQueue(
                 token_to_kv_pool=self.token_to_kv_pool_allocator.get_kvcache(),

python/sglang/srt/mem_cache/memory_pool.py

@@ -26,6 +26,8 @@ KVCache actually holds the physical kv cache.
 
 import abc
 import logging
+import os
+from contextlib import nullcontext
 from typing import List, Optional, Tuple, Union
 
 import numpy as np
@@ -34,7 +36,7 @@ import triton
 import triton.language as tl
 
 from sglang.srt.layers.radix_attention import RadixAttention
-from sglang.srt.utils import debug_timing, is_cuda, next_power_of_2
+from sglang.srt.utils import debug_timing, get_bool_env_var, is_cuda, next_power_of_2
 
 logger = logging.getLogger(__name__)
 
@@ -260,6 +262,22 @@ 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:
+            from sglang.srt.disaggregation.mooncake.memory_pool import (
+                MooncakeNVLinkAllocator,
+            )
+
+            # TODO(shangming): abstract custom allocator class for more backends
+            allocator = MooncakeNVLinkAllocator.get_allocator(self.device)
+            self.custom_mem_pool = torch.cuda.MemPool(allocator.allocator())
+        else:
+            self.custom_mem_pool = None
+
         self._create_buffers()
 
         # used for chunked cpu-offloading
@@ -275,24 +293,29 @@ class MHATokenToKVPool(KVCache):
 
     def _create_buffers(self):
         with self.memory_saver_adapter.region():
-            # [size, head_num, head_dim] for each layer
-            # The padded slot 0 is used for writing dummy outputs from padded tokens.
-            self.k_buffer = [
-                torch.zeros(
-                    (self.size + self.page_size, self.head_num, self.head_dim),
-                    dtype=self.store_dtype,
-                    device=self.device,
-                )
-                for _ in range(self.layer_num)
-            ]
-            self.v_buffer = [
-                torch.zeros(
-                    (self.size + self.page_size, self.head_num, self.head_dim),
-                    dtype=self.store_dtype,
-                    device=self.device,
-                )
-                for _ in range(self.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
+                # The padded slot 0 is used for writing dummy outputs from padded tokens.
+                self.k_buffer = [
+                    torch.zeros(
+                        (self.size + self.page_size, self.head_num, self.head_dim),
+                        dtype=self.store_dtype,
+                        device=self.device,
+                    )
+                    for _ in range(self.layer_num)
+                ]
+                self.v_buffer = [
+                    torch.zeros(
+                        (self.size + self.page_size, self.head_num, self.head_dim),
+                        dtype=self.store_dtype,
+                        device=self.device,
+                    )
+                    for _ in range(self.layer_num)
+                ]
 
         self.data_ptrs = torch.tensor(
             [x.data_ptr() for x in self.k_buffer + self.v_buffer],
@@ -349,6 +372,9 @@ 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 = []
@@ -569,16 +595,36 @@ class MLATokenToKVPool(KVCache):
         self.kv_lora_rank = kv_lora_rank
         self.qk_rope_head_dim = 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:
+            from sglang.srt.disaggregation.mooncake.memory_pool import (
+                MooncakeNVLinkAllocator,
+            )
+
+            # TODO(shangming): abstract custom allocator class for more backends
+            allocator = MooncakeNVLinkAllocator.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():
-            # The padded slot 0 is used for writing dummy outputs from padded tokens.
-            self.kv_buffer = [
-                torch.zeros(
-                    (size + page_size, 1, kv_lora_rank + qk_rope_head_dim),
-                    dtype=self.store_dtype,
-                    device=device,
-                )
-                for _ in range(layer_num)
-            ]
+            with (
+                torch.cuda.use_mem_pool(self.custom_mem_pool)
+                if self.custom_mem_pool
+                else nullcontext()
+            ):
+                # The padded slot 0 is used for writing dummy outputs from padded tokens.
+                self.kv_buffer = [
+                    torch.zeros(
+                        (size + page_size, 1, kv_lora_rank + qk_rope_head_dim),
+                        dtype=self.store_dtype,
+                        device=device,
+                    )
+                    for _ in range(layer_num)
+                ]
 
         self.layer_transfer_counter = None
 
@@ -604,6 +650,9 @@ 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)