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

[CI]Fix oom of deepseek-eplb nigtly test. (#3884)

Browse Source 
### What this PR does / why we need it?
Fix oom of deepseek-eplb nigtly test

- vLLM version: v0.11.0rc3
- vLLM main:
83f478bb19

---------

Signed-off-by: offline0806 <3337230449@qq.com>
Co-authored-by: offline0806 <3337230449@qq.com>
This commit is contained in:
offline893
2025-10-30 10:18:07 +08:00
committed by GitHub
parent dc960e798e
commit 14ca1e5cb2
4 changed files with 4 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
tests/e2e/nightly/models/test_deepseek_r1_w8a8_eplb.py
View File

@@ -85,7 +85,7 @@ async def test_models(model: str, tp_size: int, dp_size: int) -> None:
"--quantization", "ascend", "--gpu-memory-utilization", "0.9", "--quantization", "ascend", "--gpu-memory-utilization", "0.9",
"--additional-config", '{"enable_weight_nz_layout":true, ' "--additional-config", '{"enable_weight_nz_layout":true, '
'"torch_air_graph_config":{"enabled": true, "enable_multistream_mla": true, "graph_batch_size": [16], "use_cached_graph": true},' '"torch_air_graph_config":{"enabled": true, "enable_multistream_mla": true, "graph_batch_size": [16], "use_cached_graph": true},'
'"dynamic_eplb": true, "num_iterations_eplb_update": 1000, "num_wait_worker_iterations": 200, "init_redundancy_expert": 16}' '"dynamic_eplb": true, "num_iterations_eplb_update": 1000, "num_wait_worker_iterations": 200'
] ]
request_keyword_args: dict[str, Any] = { request_keyword_args: dict[str, Any] = {
**api_keyword_args, **api_keyword_args,

3
tests/e2e/nightly/models/test_qwen3_235b_a22b_w8a8_eplb.py
View File

@@ -82,8 +82,7 @@ async def test_models(model: str, tp_size: int) -> None:
"--quantization", "ascend", "--gpu-memory-utilization", "0.9", "--quantization", "ascend", "--gpu-memory-utilization", "0.9",
"--additional-config", "--additional-config",
'{"enable_weight_nz_layout":true, "dynamic_eplb": true, ' '{"enable_weight_nz_layout":true, "dynamic_eplb": true, '
'"num_iterations_eplb_update": 1000, "num_wait_worker_iterations": 200, ' '"num_iterations_eplb_update": 1000, "num_wait_worker_iterations": 200}'
'"init_redundancy_expert": 16}'
] ]
request_keyword_args: dict[str, Any] = { request_keyword_args: dict[str, Any] = {
**api_keyword_args, **api_keyword_args,

2
vllm_ascend/eplb/core/eplb_device_transfer_loader.py
View File

@@ -126,7 +126,7 @@ class D2DExpertWeightLoader:
local_expert_to_replace, local_expert_to_replace,
buffer_tensor_id) buffer_tensor_id)
logger.info( logger.debug(
f"[EPLB] finished update expert weight for layer: {self.layer_id}") f"[EPLB] finished update expert weight for layer: {self.layer_id}")
self.recv_expert_list = [] self.recv_expert_list = []

1
vllm_ascend/eplb/eplb_updator.py
View File

@@ -77,6 +77,7 @@ class EplbUpdator:
self.cur_iterations += 1 self.cur_iterations += 1
if self.cur_iterations == (self.num_iterations_eplb_update + \ if self.cur_iterations == (self.num_iterations_eplb_update + \
self.num_wait_worker_iterations + self.num_moe_layers): self.num_wait_worker_iterations + self.num_moe_layers):
logger.info("Finish expert parallel load balancing.")
if self.expert_map_record_path is not None: if self.expert_map_record_path is not None:
self.adaptor._export_tensor_to_file( self.adaptor._export_tensor_to_file(
self.shared_dict["expert_maps"], self.shared_dict["expert_maps"],