vllm-ascend vnpu v1

vllm_ascend/device_allocator/camem.py

@@ -20,10 +20,12 @@ import dataclasses
 import os
 from contextlib import contextmanager
 from typing import Any, Callable, Dict, Optional, Tuple, Union
+import time
 
 import torch
-from acl.rt import memcpy  # type: ignore # noqa: F401
+from acl.rt import memcpy, memset  # type: ignore # noqa: F401
 from vllm.logger import logger
+import vllm_ascend.envs as envs_ascend
 
 from vllm_ascend.platform import NPUPlatform
 
@@ -56,8 +58,20 @@ def find_loaded_library(lib_name) -> Optional[str]:
 
 camem_available = False
 try:
-    from vllm_ascend.vllm_ascend_C import (  # type: ignore # noqa: F401
-        init_module, python_create_and_map, python_unmap_and_release)
+    if envs_ascend.VLLM_ASCEND_ENABLE_IDLE_OFFLOAD:
+        from vllm_ascend.vllm_ascend_C import (  # type: ignore # noqa: F401
+            init_module_offload as init_module,
+            python_create_and_map_offload as python_create_and_map,python_unmap_and_release_offload as python_unmap_and_release,
+            python_get_mem_info_offload as python_get_mem_info,
+            python_lock_gpu_offload as python_lock_gpu,
+            python_unlock_gpu_offload as python_unlock_gpu
+            )
+    else:
+        from vllm_ascend.vllm_ascend_C import (  # type: ignore # noqa: F401
+            init_module, python_create_and_map, python_unmap_and_release)
+        python_get_mem_info = None
+        python_lock_gpu = None
+        python_unlock_gpu = None
     lib_name = find_loaded_library("vllm_ascend_C")
     camem_available = True
 except ImportError as e:
@@ -66,6 +80,9 @@ except ImportError as e:
     init_module = None
     python_create_and_map = None
     python_unmap_and_release = None
+    python_get_mem_info = None
+    python_lock_gpu = None
+    python_unlock_gpu = None
     lib_name = None
     libcudart = None
 
@@ -93,8 +110,14 @@ def get_pluggable_allocator(
     python_free_func: Callable[[int], tuple[int, int, int, int]]
 ) -> torch.npu.memory.NPUPluggableAllocator:
     init_module(python_malloc_fn, python_free_func)
-    new_alloc = torch.npu.memory.NPUPluggableAllocator(lib_name, 'my_malloc',
-                                                       'my_free')
+    if envs_ascend.VLLM_ASCEND_ENABLE_IDLE_OFFLOAD:
+        new_alloc = torch.npu.memory.NPUPluggableAllocator(
+            lib_name, 'my_malloc_offload', 'my_free_offload'
+        )
+    else:
+        new_alloc = torch.npu.memory.NPUPluggableAllocator(
+            lib_name, 'my_malloc', 'my_free'
+        )
     return new_alloc
 
 
@@ -153,6 +176,7 @@ class CaMemAllocator:
         self.pointer_to_data: Dict[int, AllocationData] = {}
         self.current_tag: str = CaMemAllocator.default_tag
         self.allocator_and_pools: Dict[str, Any] = {}
+        # self.requested_vram_size = 0
 
     def python_malloc_callback(self, allocation_handle: HandleType) -> None:
         """
@@ -254,6 +278,9 @@ class CaMemAllocator:
             # to avoid the issue, we keep a reference of the data.
             # see https://github.com/pytorch/pytorch/issues/146431 .
             self.allocator_and_pools[tag] = data
+            # lock gpu
+            if envs_ascend.VLLM_ASCEND_ENABLE_IDLE_OFFLOAD:
+                self.vnpu_lock_gpu()
             yield
             # PyTorch's bug, calling torch.cuda.empty_cache() will error
             # when using pluggable allocator, see
@@ -265,6 +292,8 @@ class CaMemAllocator:
             # allocate memory.
             # TODO: we need to find a way to release the memory,
             # i.e. calling torch.cuda.empty_cache()
+            if envs_ascend.VLLM_ASCEND_ENABLE_IDLE_OFFLOAD:
+                self.vnpu_unlock_gpu()
             self.current_tag = old_tag
 
     def get_current_usage(self) -> int:
@@ -276,3 +305,100 @@ class CaMemAllocator:
             handle = data.handle
             sum_bytes += handle[1]
         return sum_bytes
+
+    
+    def vnpu_lock_gpu(self) -> bool:
+        if python_lock_gpu:
+            return python_lock_gpu()
+        else:
+            return False
+
+
+    def vnpu_unlock_gpu(self):
+        if python_unlock_gpu:
+            python_unlock_gpu()
+
+    def get_pool_mem_info(self) -> int:
+        """
+        get available memory in reserved pool."""
+        return python_get_mem_info()
+
+    def offload_vram(
+            self,
+            offload_tags: Optional[Union[Tuple[str, ...],
+                                         str]] = None) -> None:
+        """
+        Put the allocator in sleep mode.
+        All data in the memory allocation with the specified tag will be 
+        offloaded to CPU memory, and others will be discarded.
+        :param offload_tags: The tags of the memory allocation that will be
+            offloaded. The rest of the memory allocation will be discarded.
+        """
+        if offload_tags is None:
+            # by default, allocated tensors are offloaded
+            # when the allocator sleeps
+            offload_tags = (CaMemAllocator.default_tag, )
+        elif isinstance(offload_tags, str):
+            offload_tags = (offload_tags, )
+
+        assert isinstance(offload_tags, tuple)
+
+        sz_weights = 0
+        sz_kvcache = 0
+
+        for ptr, data in self.pointer_to_data.items():
+            handle = data.handle
+            if data.tag in offload_tags:
+                size_in_bytes = handle[1]
+                if data.cpu_backup_tensor is None:
+                    cpu_backup_tensor = torch.empty(
+                        size_in_bytes,
+                        dtype=torch.uint8,
+                        device='cpu',
+                        pin_memory=NPUPlatform.is_pin_memory_available())
+                    cpu_ptr = cpu_backup_tensor.data_ptr()
+                    ACL_MEMCPY_DEVICE_TO_HOST = 2
+                    dest_max = cpu_ptr + size_in_bytes * 2
+                    memcpy(cpu_ptr, dest_max, ptr, size_in_bytes,
+                        ACL_MEMCPY_DEVICE_TO_HOST)
+                    data.cpu_backup_tensor = cpu_backup_tensor
+                unmap_and_release(handle)
+                sz_weights += size_in_bytes
+            else:
+                size_in_bytes = handle[1]
+                unmap_and_release(handle)
+                sz_kvcache += size_in_bytes
+        # self.requested_vram_size = sz_weights + sz_kvcache
+
+        self.vnpu_unlock_gpu()
+        # logger.info(f"offload: tags {offload_tags}: {sz_weights/(1024**3):.2f} GB, discard kv cache: {sz_kvcache/(1024**3):.2f} GB")
+
+    def reload_vram(self, tags: Optional[list[str]] = None) -> bool:
+        """
+        Wake up the allocator from sleep mode.
+        All data that is previously offloaded will be loaded back to GPU 
+        memory, and the rest of the data will have empty memory."""
+        prev_is_self = self.vnpu_lock_gpu()
+        if prev_is_self:
+            # nothing to do
+            return True
+
+        for ptr, data in self.pointer_to_data.items():
+            handle = data.handle
+            if tags is None or data.tag in tags:
+                create_and_map(handle)
+                if data.cpu_backup_tensor is not None:
+                    cpu_backup_tensor = data.cpu_backup_tensor
+                    size_in_bytes = cpu_backup_tensor.numel(
+                    ) * cpu_backup_tensor.element_size()
+                    cpu_ptr = cpu_backup_tensor.data_ptr()
+                    ACL_MEMCPY_HOST_TO_DEVICE = 1
+                    dest_max = ptr + size_in_bytes * 2
+                    memcpy(ptr, dest_max, cpu_ptr, size_in_bytes,
+                        ACL_MEMCPY_HOST_TO_DEVICE)
+                    # data.cpu_backup_tensor = None
+                # TO check: no need to re-memset if we reset_prefix_cache
+                # else:
+                #     size_in_bytes = handle[1]
+                #     memset(ptr, size_in_bytes, 0, size_in_bytes)
+        return False

vllm_ascend/envs.py

@@ -166,7 +166,8 @@ env_variables: Dict[str, Callable[[], Any]] = {
     lambda: bool(int(os.getenv("VLLM_ASCEND_ENABLE_MLAPO", '0'))),
     # Whether to enable transpose weight and cast format to FRACTAL_NZ.
     "VLLM_ASCEND_ENABLE_NZ":
-    lambda: int(os.getenv("VLLM_ASCEND_ENABLE_NZ", 1)),
+    lambda: int(os.getenv("VLLM_ASCEND_ENABLE_NZ", 0)),
+    "VLLM_ASCEND_ENABLE_IDLE_OFFLOAD": lambda: int(os.getenv("VLLM_ASCEND_ENABLE_IDLE_OFFLOAD", 1)),
 }
 
 # end-env-vars-definition

vllm_ascend/patch/platform/__init__.py

@@ -28,3 +28,4 @@ if os.getenv("DYNAMIC_EPLB", "false") == "true" or os.getenv(
 if os.getenv("SHM_BARRIER", "true") == "true":
     import vllm_ascend.patch.platform.patch_core  # noqa
     import vllm_ascend.patch.platform.patch_message_queue  # noqa
+    import vllm_ascend.patch.platform.patch_executor  # noqa

vllm_ascend/patch/platform/patch_core.py

@@ -1,12 +1,19 @@
 import signal
 from typing import Optional
+from logging import DEBUG
+import time
 
-from vllm.config import ParallelConfig
+from vllm.config import ParallelConfig, VllmConfig
 from vllm.logger import logger
 from vllm.transformers_utils.config import \
     maybe_register_config_serialize_by_value
 from vllm.utils import decorate_logs, set_process_title
-from vllm.v1.engine.core import DPEngineCoreProc, EngineCoreProc
+from vllm.v1.kv_cache_interface import KVCacheConfig
+from vllm.v1.core.kv_cache_utils import (generate_scheduler_kv_cache_config,
+                                         get_kv_cache_configs)
+from vllm.v1.engine.core import DPEngineCoreProc, EngineCoreProc, EngineCore
+
+import vllm_ascend.envs as envs_ascend
 
 
 def run_engine_core(*args, dp_rank: int = 0, local_dp_rank: int = 0, **kwargs):
@@ -66,3 +73,101 @@ def run_engine_core(*args, dp_rank: int = 0, local_dp_rank: int = 0, **kwargs):
 
 
 EngineCoreProc.run_engine_core = run_engine_core
+
+
+def run_busy_loop(self):
+    """Core busy loop of the EngineCore."""
+
+    # Loop until process is sent a SIGINT or SIGTERM
+    while True:
+        # 1) Poll the input queue until there is work to do.
+        self._process_input_queue()
+        # 2) Step the engine core and return the outputs.
+        if envs_ascend.VLLM_ASCEND_ENABLE_IDLE_OFFLOAD and self.scheduler.has_requests() and self.model_executor.is_offloaded:
+            prev_is_self = self.model_executor.reload_vram()
+            if not prev_is_self:
+                self.reset_prefix_cache()
+        self._process_engine_step()
+        if envs_ascend.VLLM_ASCEND_ENABLE_IDLE_OFFLOAD and not self.scheduler.has_requests() and not self.model_executor.is_offloaded:
+            self.model_executor.offload_vram()
+
+def _process_input_queue(self):
+    """Exits when an engine step needs to be performed."""
+
+    waited = False
+    while not self.engines_running and not self.scheduler.has_requests() \
+            and not self.batch_queue:
+        if logger.isEnabledFor(DEBUG) and self.input_queue.empty():
+            logger.debug("EngineCore waiting for work.")
+            waited = True
+            if envs_ascend.VLLM_ASCEND_ENABLE_IDLE_OFFLOAD and not self.model_executor.is_offloaded:
+                self.model_executor.offload_vram()
+        req = self.input_queue.get()
+        self._handle_client_request(*req)
+
+    if waited:
+        logger.debug("EngineCore loop active.")
+
+    # Handle any more client requests.
+    while not self.input_queue.empty():
+        req = self.input_queue.get_nowait()
+        self._handle_client_request(*req)
+
+
+EngineCoreProc.run_busy_loop = run_busy_loop
+EngineCoreProc._process_input_queue = _process_input_queue
+
+
+def _initialize_kv_caches(
+        self, vllm_config: VllmConfig) -> tuple[int, int, KVCacheConfig]:
+    start = time.time()
+
+    # Get all kv cache needed by the model
+    kv_cache_specs = self.model_executor.get_kv_cache_specs()
+
+    has_kv_cache = any(kv_cache_spec for kv_cache_spec in kv_cache_specs)
+    if has_kv_cache:
+        if envs_ascend.VLLM_ASCEND_ENABLE_IDLE_OFFLOAD:
+            # get available memory in idle offload mode
+            available_gpu_memory = (
+                self.model_executor.determine_available_memory_idle_offload_mode())
+            self.available_gpu_memory_for_kv_cache = \
+                available_gpu_memory[0]
+        elif os.environ.get("VLLM_ELASTIC_EP_SCALE_UP_LAUNCH") == "1":
+            dp_group = getattr(self, "dp_group", None)
+            assert dp_group is not None
+            self.available_gpu_memory_for_kv_cache = \
+                ParallelConfig.sync_kv_cache_memory_size(dp_group, -1)
+            available_gpu_memory = [
+                self.available_gpu_memory_for_kv_cache
+            ] * len(kv_cache_specs)
+        else:
+            # Profiles the peak memory usage of the model to determine how
+            # much memory can be allocated for kv cache.
+            available_gpu_memory = (
+                self.model_executor.determine_available_memory())
+            self.available_gpu_memory_for_kv_cache = \
+                available_gpu_memory[0]
+    else:
+        # Attention free models don't need memory for kv cache
+        available_gpu_memory = [0] * len(kv_cache_specs)
+
+    assert len(kv_cache_specs) == len(available_gpu_memory)
+
+    kv_cache_configs = get_kv_cache_configs(vllm_config, kv_cache_specs,
+                                            available_gpu_memory)
+    scheduler_kv_cache_config = generate_scheduler_kv_cache_config(
+        kv_cache_configs)
+    num_gpu_blocks = scheduler_kv_cache_config.num_blocks
+    num_cpu_blocks = 0
+
+    # Initialize kv cache and warmup the execution
+    self.model_executor.initialize_from_config(kv_cache_configs)
+
+    elapsed = time.time() - start
+    logger.info(("init engine (profile, create kv cache, "
+                    "warmup model) took %.2f seconds"), elapsed)
+    return num_gpu_blocks, num_cpu_blocks, scheduler_kv_cache_config
+
+
+EngineCore._initialize_kv_caches = _initialize_kv_caches

vllm_ascend/patch/platform/patch_executor.py

@@ -0,0 +1,44 @@
+import time
+
+from vllm.executor.executor_base import logger, ExecutorBase
+
+
+original_init = ExecutorBase.__init__
+def init(self, *args, **kwargs):
+    original_init(self, *args, **kwargs)
+    self.is_offloaded = False
+
+
+def offload_vram(self):
+    if self.is_offloaded:
+        logger.warning("Executor is already offloaded.")
+        return
+    time_before_offload = time.perf_counter()
+    self.collective_rpc("offload_vram")
+    time_after_offload = time.perf_counter()
+
+    self.is_offloaded = True
+    logger.info(f"Offloading VRAM costs {time_after_offload - time_before_offload:.6f} seconds.")
+
+
+def reload_vram(self) -> bool:
+    if not self.is_offloaded:
+        logger.warning("Executor is not offloaded.")
+        return True
+
+    time_before_reload = time.perf_counter()
+    prev_is_self = self.collective_rpc("reload_vram")
+    time_after_reload = time.perf_counter()
+    self.is_offloaded = False
+    logger.info(f"Reloading VRAM costs {time_after_reload - time_before_reload:.6f} seconds.")
+    return prev_is_self
+
+
+def determine_available_memory_idle_offload_mode(self) -> int:
+    return self.collective_rpc("determine_available_memory_idle_offload_mode")
+
+
+ExecutorBase.__init__ = init
+ExecutorBase.offload_vram = offload_vram
+ExecutorBase.reload_vram = reload_vram
+ExecutorBase.determine_available_memory_idle_offload_mode = determine_available_memory_idle_offload_mode

vllm_ascend/worker/worker_v1.py

@@ -258,6 +258,17 @@ class NPUWorker(WorkerBase):
         )
         return available_kv_cache_memory
 
+    def determine_available_memory_idle_offload_mode(self) -> int:
+        allocator = CaMemAllocator.get_instance()
+        free, total = allocator.get_pool_mem_info()
+        available_kv_cache_memory = int(
+            total * self.cache_config.gpu_memory_utilization - (total - free))
+        available_kv_cache_memory = int(max(available_kv_cache_memory, 0))
+        logger.info(
+            f"Available memory (idle offload mode): {available_kv_cache_memory}, total memory: {total}"
+        )
+        return available_kv_cache_memory
+
     def execute_model(
         self,
         scheduler_output: "SchedulerOutput",
@@ -306,11 +317,46 @@ class NPUWorker(WorkerBase):
                 "Sleep mode can only be "
                 "used for one instance per process.")
             context = allocator.use_memory_pool(tag="weights")
+        elif envs_ascend.VLLM_ASCEND_ENABLE_IDLE_OFFLOAD:
+            if not sleep_mode_enabled():
+                raise ValueError(
+                    "Sleep mode is not enabled. Please compile vllm-ascend with COMPILE_CUSTOM_KERNELS=1."
+                )
+            if is_enable_nz():
+                raise ValueError(
+                    "FRACTAL_NZ mode is enabled. This may cause model parameter precision issues "
+                    "in the RL scenarios. Please set VLLM_ASCEND_ENABLE_NZ=0.")
+            allocator = CaMemAllocator.get_instance()
+            assert allocator.get_current_usage() == 0, (
+                "Idle offload mode can only be "
+                "used for one instance per process.")
+            context = allocator.use_memory_pool(tag="weights")
         else:
             from contextlib import nullcontext
             context = nullcontext()  # type: ignore
         with context:
             self.model_runner.load_model()
+            if envs_ascend.VLLM_ASCEND_ENABLE_IDLE_OFFLOAD:
+                # save memory to host with lock
+                self.offload_vram()
+                self.reload_vram()
+
+    def offload_vram(self) -> None:
+        # free_bytes_before_offload = NPUPlatform.mem_get_info()[0]
+        allocator = CaMemAllocator.get_instance()
+        allocator.offload_vram(offload_tags=("weights", ))
+        # free_bytes_after_offload, total = NPUPlatform.mem_get_info()
+        # freed_bytes = free_bytes_after_offload - free_bytes_before_offload
+        # used_bytes = total - free_bytes_after_offload
+        # assert freed_bytes >= 0, "Memory usage increased after sleeping."
+        # logger.info(
+        #     "Offloading freed %.2f GiB memory, "
+        #     "%.2f GiB memory is still in use.", freed_bytes / GiB_bytes,
+        #     used_bytes / GiB_bytes)
+
+    def reload_vram(self) -> bool:
+        allocator = CaMemAllocator.get_instance()
+        return allocator.reload_vram(tags=None)
 
     def compile_or_warm_up_model(self) -> None:
         # Note: need to adapt for graph mode.
@@ -351,6 +397,9 @@ class NPUWorker(WorkerBase):
         if self.vllm_config.model_config.enable_sleep_mode:
             allocator = CaMemAllocator.get_instance()
             context = allocator.use_memory_pool(tag="kv_cache")
+        elif envs_ascend.VLLM_ASCEND_ENABLE_IDLE_OFFLOAD:
+            allocator = CaMemAllocator.get_instance()
+            context = allocator.use_memory_pool(tag="kv_cache")
         else:
             from contextlib import nullcontext
             context = nullcontext()  # type: ignore