【main】ADXL/HIXL supports FabricMem Mode (#6806)

### What this PR does / why we need it?

### Does this PR introduce _any_ user-facing change?

### How was this patch tested?

- vLLM version: v0.15.0
- vLLM main:
83b47f67b1

---------

Signed-off-by: fems14 <1804143737@qq.com>

docs/source/tutorials/features/pd_colocated_mooncake_multi_instance.md

@@ -121,7 +121,7 @@ Moonshot AI. Installation and compilation guide:
 First, obtain the Mooncake project using the following command:
 
 ```bash
-git clone -b v0.3.8.post1 --depth 1 https://github.com/kvcache-ai/Mooncake.git
+git clone -b v0.3.9 --depth 1 https://github.com/kvcache-ai/Mooncake.git
 cd Mooncake
 git submodule update --init --recursive
 ```

docs/source/tutorials/features/pd_disaggregation_mooncake_multi_node.md

@@ -177,7 +177,7 @@ Mooncake is the serving platform for Kimi, a leading LLM service provided by Moo
 First, we need to obtain the Mooncake project. Refer to the following command:
 
 ```shell
-git clone -b v0.3.8.post1 --depth 1 https://github.com/kvcache-ai/Mooncake.git
+git clone -b v0.3.9 --depth 1 https://github.com/kvcache-ai/Mooncake.git
 ```
 
 (Optional) Replace go install url if the network is poor

docs/source/tutorials/features/pd_disaggregation_mooncake_single_node.md

@@ -98,7 +98,7 @@ Mooncake is the serving platform for Kimi, a leading LLM service provided by Moo
 First, we need to obtain the Mooncake project. Refer to the following command:
 
 ```shell
-git clone -b v0.3.8.post1 --depth 1 https://github.com/kvcache-ai/Mooncake.git
+git clone -b v0.3.9 --depth 1 https://github.com/kvcache-ai/Mooncake.git
 ```
 
 (Optional) Replace go install url if the network is poor.

docs/source/user_guide/feature_guide/kv_pool.md

@@ -42,7 +42,7 @@ export PYTHONHASHSEED=0
         First, we need to obtain the Mooncake project. Refer to the following command:
 
         ```shell
-        git clone -b v0.3.7.post2 --depth 1 https://github.com/kvcache-ai/Mooncake.git
+        git clone -b v0.3.9 --depth 1 https://github.com/kvcache-ai/Mooncake.git
         ```
 
         (Optional) Replace go install url if the network is poor
@@ -85,6 +85,15 @@ export PYTHONHASHSEED=0
         export LD_LIBRARY_PATH=/usr/local/lib64/python3.11/site-packages/mooncake:$LD_LIBRARY_PATH
         ```
 
+### Environment Variables Description
+
+`export ASCEND_ENABLE_USE_FABRIC_MEM=1`: Enable unified memory address direct transmission scheme and only can be used for 800 I/T A3 series. Required supporting hardware versions are as follows:
+
+    HDK >=26.0
+    CANN >= 9.0
+
+`export ASCEND_BUFFER_POOL=4:8`: ASCEND_BUFFER_POOL is the environment variable for configuring the number and size of buffer on NPU Device for aggregation and KV transfer，the value 4:8 means we allocate 4 buffers of size 8MB. It only can be used for 800 I/T A2 series.
+
 ### Run Mooncake Master
 
 #### 1.Configure mooncake.json

vllm_ascend/distributed/kv_transfer/kv_pool/ascend_store/backend/mooncake_backend.py

@@ -7,16 +7,14 @@ from dataclasses import dataclass
 import torch
 
 # Third Party
-from mooncake.store import ReplicateConfig  # type: ignore
 from vllm.config import ParallelConfig
 from vllm.logger import logger
 from vllm.utils.network_utils import get_ip
 
 from vllm_ascend.distributed.kv_transfer.kv_pool.ascend_store.backend.backend import Backend
 from vllm_ascend.distributed.kv_transfer.utils.mooncake_transfer_engine import global_te
-from vllm_ascend.utils import AscendDeviceType, get_ascend_device_type
 
-DEFAULT_GLOBAL_SEGMENT_SIZE = 3355443200  # 3.125 GiB
+DEFAULT_GLOBAL_SEGMENT_SIZE = 1073741824  # 1.0 GiB
 DEFAULT_LOCAL_BUFFER_SIZE = 1073741824  # 1.0 GiB
 
 
@@ -35,18 +33,34 @@ class MooncakeBackend(Backend):
         self.rank = parallel_config.rank
         if self.config.protocol == "ascend":
             local_hostname = get_ip()
-            transfer_engine = global_te.get_transfer_engine(local_hostname, device_name=None)
+            # ASCEND_ENABLE_USE_FABRIC_MEM: Enable unified memory address direct transmission scheme
-            self.local_seg = local_hostname + ":" + str(transfer_engine.get_rpc_port())
+            # and only can be used for 800 I/T A3 series.
-            ret = self.store.setup(
+            # Required supporting hardware versions are as follows:
-                self.local_seg,
+            if os.getenv("ASCEND_ENABLE_USE_FABRIC_MEM", "0") != "1":
-                self.config.metadata_server,
+                transfer_engine = global_te.get_transfer_engine(local_hostname, device_name=None)
-                self.config.global_segment_size,
+                self.local_seg = local_hostname + ":" + str(transfer_engine.get_rpc_port())
-                self.config.local_buffer_size,
+                ret = self.store.setup(
-                self.config.protocol,
+                    self.local_seg,
-                self.config.device_name,
+                    self.config.metadata_server,
-                self.config.master_server_address,
+                    self.config.global_segment_size,
-                transfer_engine.get_engine(),
+                    self.config.local_buffer_size,
-            )
+                    self.config.protocol,
+                    self.config.device_name,
+                    self.config.master_server_address,
+                    transfer_engine.get_engine(),
+                )
+            else:
+                self.local_seg = local_hostname
+                ret = self.store.setup(
+                    self.local_seg,
+                    self.config.metadata_server,
+                    self.config.global_segment_size,
+                    0,
+                    self.config.protocol,
+                    self.config.device_name,
+                    self.config.master_server_address,
+                )
+
         if ret != 0:
             msg = "Initialize mooncake failed."
             logger.error(msg)
@@ -57,21 +71,15 @@ class MooncakeBackend(Backend):
         torch.npu.set_device(device)
 
     def register_buffer(self, ptrs: list[int], lengths: list[int]):
-        global_te.register_buffer(ptrs, lengths)
+        if os.getenv("ASCEND_ENABLE_USE_FABRIC_MEM", "0") != "1":
+            global_te.register_buffer(ptrs, lengths)
 
     def exists(self, keys: list[str]) -> list[int]:
         return self.store.batch_is_exist(keys)
 
     def put(self, keys: list[str], addrs: list[list[int]], sizes: list[list[int]]):
         try:
-            soc_version = get_ascend_device_type()
+            res = self.store.batch_put_from_multi_buffers(keys, addrs, sizes)
-            if soc_version in {AscendDeviceType.A2}:
-                res = self.store.batch_put_from_multi_buffers(keys, addrs, sizes)
-            else:
-                config = ReplicateConfig()
-                config.preferred_segment = self.local_seg
-                config.prefer_alloc_in_same_node = True
-                res = self.store.batch_put_from_multi_buffers(keys, addrs, sizes, config)
             for value in res:
                 if value < 0:
                     logger.error(f"Failed to put key {keys},res:{res}")
@@ -80,11 +88,7 @@ class MooncakeBackend(Backend):
 
     def get(self, keys: list[str], addrs: list[list[int]], sizes: list[list[int]]):
         try:
-            soc_version = get_ascend_device_type()
+            res = self.store.batch_get_into_multi_buffers(keys, addrs, sizes)
-            if soc_version in {AscendDeviceType.A2}:
-                res = self.store.batch_get_into_multi_buffers(keys, addrs, sizes)
-            else:
-                res = self.store.batch_get_into_multi_buffers(keys, addrs, sizes, True)
             for value in res:
                 if value < 0:
                     logger.error(f"Failed to get key {keys}, res:{res}")

vllm_ascend/distributed/kv_transfer/kv_pool/ascend_store/pool_worker.py

@@ -30,7 +30,6 @@ from vllm_ascend.distributed.kv_transfer.kv_pool.ascend_store.kv_transfer import
     KVCacheStoreSendingThread,
     KVTransferThread,
 )
-from vllm_ascend.utils import AscendDeviceType, get_ascend_device_type
 
 backend_map = {
     "mooncake": {
@@ -98,12 +97,6 @@ class KVPoolWorker:
             self.head_or_tp_rank = self.tp_rank
             self.put_step = 1
 
-        soc_version = get_ascend_device_type()
-        # be removed later
-        if self.backend == "mooncake" and soc_version in {AscendDeviceType.A3}:
-            self.head_or_tp_rank = self.tp_rank
-            self.put_step = 1
-
         self.metadata = KeyMetadata(
             model_config.model.rstrip("/").split("/")[-1],
             self.head_or_tp_rank,