minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[NPU] Set DEEPEP_HYBRID_DEPLOYMENT for new DeepEP tests; switch glm5_2 to w8a8; tune nightly timeouts (#38775)

Browse Source
This commit is contained in:
pllimax
2026-09-10 23:32:54 +08:00
committed by GitHub
parent 12771786f2
commit 92dffebe16
7 changed files with 37 additions and 31 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.github/workflows/nightly-test-npu.yml
+19 -18
View File
@@ -131,7 +131,7 @@ jobs:
self_name: nightly-1-npu-a3
runner_config: linux-aarch64-a3-2-
image: ${{ needs.set-image-config.outputs.image_a3 }}
run_timeout_minutes: '60'
run_timeout_minutes: '90'
timeout_per_file: '3600'
is_nightly_pipeline_job: true
skip_pr_test_health_check: 'true'
@@ -146,7 +146,7 @@ jobs:
self_name: nightly-2-npu-a3
runner_config: linux-aarch64-a3-2-
image: ${{ needs.set-image-config.outputs.image_a3 }}
run_timeout_minutes: '60'
run_timeout_minutes: '90'
timeout_per_file: '3600'
is_nightly_pipeline_job: true
skip_pr_test_health_check: 'true'
@@ -292,20 +292,21 @@ jobs:
matrix:
test_config:
# glm5_1 performance tests
- name: glm5_1_w4a8_1p1d_32p_in64k_out1k_50ms
prefill_size: 2
decode_size: 2
router_size: 1
test_case: test/registered/npu/performance/glm5_1/test_npu_glm5_1_w4a8_1p1d_32p_in64k_out1k_50ms.py
test_type: 'perf'
prefill_decode_deployment: 'separation'
- name: glm5_1_w4a8_1p1d_32p_in64k_out1k_50ms_aime26
prefill_size: 2
decode_size: 2
router_size: 1
test_case: test/registered/npu/accuracy/glm5_1/test_npu_glm5_1_w4a8_1p1d_32p_in64k_out1k_50ms_aime26.py
test_type: 'accuracy'
prefill_decode_deployment: 'separation'
# Disable test case temporarily due to feature refactor in PR #38293.
# - name: glm5_1_w4a8_1p1d_32p_in64k_out1k_50ms
# prefill_size: 2
# decode_size: 2
# router_size: 1
# test_case: test/registered/npu/performance/glm5_1/test_npu_glm5_1_w4a8_1p1d_32p_in64k_out1k_50ms.py
# test_type: 'perf'
# prefill_decode_deployment: 'separation'
# - name: glm5_1_w4a8_1p1d_32p_in64k_out1k_50ms_aime26
# prefill_size: 2
# decode_size: 2
# router_size: 1
# test_case: test/registered/npu/accuracy/glm5_1/test_npu_glm5_1_w4a8_1p1d_32p_in64k_out1k_50ms_aime26.py
# test_type: 'accuracy'
# prefill_decode_deployment: 'separation'
# deepseek_v4_flash performance tests
- name: deepseek_v4_flash_w8a8_1p1d_16p_in8k_out1k_50ms
prefill_size: 1
@@ -339,9 +340,9 @@ jobs:
matrix:
test_config:
# glm_5_2 accuracy tests
- name: glm_5_2_w4a8_16p_gpqa
- name: glm_5_2_w8a8_16p_gpqa
node_size: 2
test_case: test/registered/npu/accuracy/glm5_2/test_npu_glm_5_2_w4a8_16p_gpqa.py
test_case: test/registered/npu/accuracy/glm5_2/test_npu_glm_5_2_w8a8_16p_gpqa.py
test_type: 'accuracy'
# kimi_k3 accuracy tests
- name: kimi_k3_w4a8_32p_gpqa