EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

add env vars & misc

Browse Source
This commit is contained in:
starkwj
2026-02-11 06:27:58 +00:00
parent 739d074b0c
commit 389030a8f8
128 changed files with 89 additions and 59 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
vllm_ascend/patch/platform/patch_core.py
View File

@@ -84,12 +84,12 @@ def run_busy_loop(self):
# 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:
if envs_ascend.VLLM_ASCEND_ENABLE_VNPU 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:
if envs_ascend.VLLM_ASCEND_ENABLE_VNPU and not self.scheduler.has_requests() and not self.model_executor.is_offloaded:
self.model_executor.offload_vram()
def _process_input_queue(self):
@@ -101,7 +101,7 @@ def _process_input_queue(self):
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:
if envs_ascend.VLLM_ASCEND_ENABLE_VNPU and not self.model_executor.is_offloaded:
self.model_executor.offload_vram()
req = self.input_queue.get()
self._handle_client_request(*req)
@@ -128,10 +128,10 @@ def _initialize_kv_caches(
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:
if envs_ascend.VLLM_ASCEND_ENABLE_VNPU:
# get available memory in idle offload mode
available_gpu_memory = (
self.model_executor.determine_available_memory_idle_offload_mode())
self.model_executor.determine_available_memory_vnpu_offload_mode())
self.available_gpu_memory_for_kv_cache = \
available_gpu_memory[0]
elif os.environ.get("VLLM_ELASTIC_EP_SCALE_UP_LAUNCH") == "1":

6
vllm_ascend/patch/platform/patch_executor.py
View File

@@ -43,11 +43,11 @@ def reload_vram(self) -> bool:
time.sleep(0.001)
def determine_available_memory_idle_offload_mode(self) -> int:
return self.collective_rpc("determine_available_memory_idle_offload_mode")
def determine_available_memory_vnpu_offload_mode(self) -> int:
return self.collective_rpc("determine_available_memory_vnpu_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
ExecutorBase.determine_available_memory_vnpu_offload_mode = determine_available_memory_vnpu_offload_mode