EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Fix potential memory fault issue and ncclSystemError in CI test (#8681)

Browse Source 
Co-authored-by: wunhuang <wunhuang@amd.com>
This commit is contained in:
kk
2025-08-06 03:19:37 +08:00
committed by GitHub
parent 4f4e0e4162
commit 32d9e39a29
3 changed files with 7 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
.github/workflows/pr-test-amd.yml vendored
View File

@@ -291,7 +291,7 @@ jobs:
bash scripts/amd_ci_exec.sh python3 run_suite.py --suite per-commit-8-gpu-amd --timeout-per-file 3600 bash scripts/amd_ci_exec.sh python3 run_suite.py --suite per-commit-8-gpu-amd --timeout-per-file 3600
- name: Run CustomAllReduce test - name: Run CustomAllReduce test
timeout-minutes: 10 timeout-minutes: 20
run: | run: |
bash scripts/amd_ci_exec.sh -e CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 -m unittest test_custom_allreduce.TestCustomAllReduce bash scripts/amd_ci_exec.sh -e CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 -m unittest test_custom_allreduce.TestCustomAllReduce

13
python/sglang/srt/layers/attention/aiter_backend.py
View File

@@ -720,11 +720,6 @@ class AiterIndicesUpdaterPrefill:
self.req_to_token = model_runner.req_to_token_pool.req_to_token self.req_to_token = model_runner.req_to_token_pool.req_to_token
self.update = self.update_single_wrapper self.update = self.update_single_wrapper
# get the last index of the pool
self.pool_size = (
model_runner.token_to_kv_pool.size + model_runner.token_to_kv_pool.page_size
) - 1
self.kv_indices = None self.kv_indices = None
self.max_q_len = 0 self.max_q_len = 0
self.max_kv_len = 0 self.max_kv_len = 0
@@ -769,9 +764,8 @@ class AiterIndicesUpdaterPrefill:
# but the 0 location will be made nan (noqa) in cuda graph capture mode # but the 0 location will be made nan (noqa) in cuda graph capture mode
# this will cause the output tensor value becomes nan # this will cause the output tensor value becomes nan
# WA is to assure that last index of pool not changed # WA is to assure that last index of pool not changed
kv_indices = torch.full( kv_indices = torch.empty(
(paged_kernel_lens_sum + 128,), paged_kernel_lens_sum + 256,
self.pool_size,
dtype=torch.int32, dtype=torch.int32,
device=req_pool_indices.device, device=req_pool_indices.device,
) )
@@ -785,6 +779,9 @@ class AiterIndicesUpdaterPrefill:
self.req_to_token.shape[1], self.req_to_token.shape[1],
) )
token_num = kv_indptr[-1]
kv_indices[token_num:] = kv_indices[0]
self.max_kv_len = torch.max(paged_kernel_lens).item() self.max_kv_len = torch.max(paged_kernel_lens).item()
extend_lens = seq_lens - prefix_lens extend_lens = seq_lens - prefix_lens

1
scripts/amd_ci_start_container.sh
View File

@@ -124,6 +124,7 @@ echo "Starting container: ci_sglang"
docker run -dt --user root --device=/dev/kfd $DEVICE_FLAG \ docker run -dt --user root --device=/dev/kfd $DEVICE_FLAG \
-v "${GITHUB_WORKSPACE:-$PWD}:/sglang-checkout" \ -v "${GITHUB_WORKSPACE:-$PWD}:/sglang-checkout" \
--ipc=host --group-add video \ --ipc=host --group-add video \
--shm-size 32g \
--cap-add=SYS_PTRACE \ --cap-add=SYS_PTRACE \
-e HF_TOKEN="${HF_TOKEN:-}" \ -e HF_TOKEN="${HF_TOKEN:-}" \
--security-opt seccomp=unconfined \ --security-opt seccomp=unconfined \