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

[Bugfix] Fix multi-instance serving OOM on single card (#7427)

Browse Source 
### What this PR does / why we need it?
Fix https://github.com/vllm-project/vllm-ascend/issues/7308.

Subtracting `init_non_torch_memory` (maybe used by the first instance)
from the total `non_torch_memory` when calculating
`available_kv_cache_memory`.

Directly use `non_torch_memory_increase` (contained in
`non_kv_cache_memory`) to calculate `available_kv_cache_memory`.

### Does this PR introduce _any_ user-facing change?
No.

### How was this patch tested?
Launch tow vllm-ascend instances sequentially on single card.

```bash
# Launch first instance
vllm serve /root/.cache/modelscope/hub/models/Qwen/Qwen3-0.6B \
--port 8100 \
--host 0.0.0.0 \
--additional-config='{"enable_cpu_binding":true}'  \
--gpu-memory-utilization 0.3 \
--max-num-seqs 1 \
--max-model-len 2048 \
--max-num-batched-tokens 2048 \
--no-enable-prefix-caching \
--enforce-eager

# Launch second instance
vllm serve /root/.cache/modelscope/hub/models/Qwen/Qwen3-0.6B \
--port 8101 \
--host 0.0.0.0 \
--additional-config='{"enable_cpu_binding":true}'  \
--gpu-memory-utilization 0.3 \
--max-num-seqs 1 \
--max-model-len 2048 \
--max-num-batched-tokens 2048 \
--no-enable-prefix-caching \
--enforce-eager
```

**Before this PR:**

```bash
# First instance:
------------------------------------------------------------------
requested_memory: 18.287109375 GiB
non_kv_cache_memory: 1.2340388298034668 GiB
init_non_torch_memory: 0.3616676330566406 GiB
non_torch_memory_before_empty_cache: 0.3896217346191406 GiB
non_torch_memory_increase: 0.0279541015625 GiB
non_torch_memory_cleared_by_empty_cache: 0.3616676330566406 GiB
------------------------------------------------------------------

# Second instance:
------------------------------------------------------------------
requested_memory: 18.287109375 GiB
non_kv_cache_memory: 1.2336344718933105 GiB
init_non_torch_memory: 18.37220001220703 GiB
non_torch_memory_before_empty_cache: 18.399906158447266 GiB
non_torch_memory_increase: 0.02754974365234375 GiB
non_torch_memory_cleared_by_empty_cache: 18.372356414794922 GiB
------------------------------------------------------------------
# available_kv_cache_memory = requested_memory - non_kv_cache_memory - non_torch_memory_cleared_by_empty_cache
Available KV cache memory: -1.32 GiB
```

**After this PR:**

```bash
# First instance:
------------------------------------------------------------------
requested_memory: 18.287109375 GiB
non_kv_cache_memory: 1.2340540885925293 GiB
init_non_torch_memory: 0.36182403564453125 GiB
non_torch_memory_before_empty_cache: 0.38979339599609375 GiB
non_torch_memory_increase: 0.0279693603515625 GiB
non_torch_memory_cleared_by_empty_cache: 0.0 GiB
------------------------------------------------------------------

# Second instance:
------------------------------------------------------------------
requested_memory: 18.287109375 GiB
non_kv_cache_memory: 1.233344554901123 GiB
init_non_torch_memory: 18.74309539794922 GiB
non_torch_memory_before_empty_cache: 18.770355224609375 GiB
non_torch_memory_increase: 0.02725982666015625 GiB
non_torch_memory_cleared_by_empty_cache: 0.0 GiB
------------------------------------------------------------------
# available_kv_cache_memory = requested_memory - non_kv_cache_memory - non_torch_memory_cleared_by_empty_cache
Available KV cache memory: 17.05 GiB
```

- vLLM version: v0.17.0
- vLLM main:
4497431df6

---------

Signed-off-by: shen-shanshan <467638484@qq.com>
Signed-off-by: Shanshan Shen <87969357+shen-shanshan@users.noreply.github.com>
This commit is contained in:
Shanshan Shen
2026-03-23 14:22:59 +08:00
committed by GitHub
parent 44ef9a36ac
commit 5c0d02f689
5 changed files with 348 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
vllm_ascend/worker/worker.py
View File

@@ -341,13 +341,6 @@ class NPUWorker(WorkerBase):
weights_memory=int(self.model_runner.model_memory_usage),
) as profile_result:
self.model_runner.profile_run()
free_memory, total_memory = torch.npu.mem_get_info()
torch_memory = torch.npu.memory_reserved()
non_torch_memory_before_empty_cache = total_memory - free_memory - torch_memory
self.non_torch_memory = profile_result.non_torch_increase
self.peak_activation_memory = profile_result.torch_peak_increase
non_torch_memory_cleared_by_empty_cache = non_torch_memory_before_empty_cache - self.non_torch_memory
free_gpu_memory = profile_result.after_profile.free_memory
assert self.init_snapshot.free_memory > free_gpu_memory, (
@@ -359,16 +352,12 @@ class NPUWorker(WorkerBase):
"To fix this, ensure consistent GPU memory allocation or "
"isolate vLLM in its own container."
)
self.available_kv_cache_memory_bytes = (
self.requested_memory - profile_result.non_kv_cache_memory - non_torch_memory_cleared_by_empty_cache
)
self.available_kv_cache_memory_bytes = self.requested_memory - profile_result.non_kv_cache_memory
logger.debug(profile_result)
logger.info_once(
"Available KV cache memory: %.2f GiB",
GiB(self.available_kv_cache_memory_bytes),
scope="local",
"Available KV cache memory: %.2f GiB", GiB(self.available_kv_cache_memory_bytes), scope="local"
)
return int(self.available_kv_cache_memory_bytes)
def execute_model(