From 0fd233d64c6eaac482fabd7dc6b63905715741cd Mon Sep 17 00:00:00 2001 From: chx96642264 Date: Tue, 14 Jul 2026 15:14:35 +0800 Subject: [PATCH 1/5] gsm8k deepseek-r1, 2 --- .../test_npu_deepep_auto_deepseek_R1_0528_w4a8_per_channel.py | 1 + 1 file changed, 1 insertion(+) diff --git a/test/registered/ascend/basic_function/parallel_strategy/expert_parallelism/test_npu_deepep_auto_deepseek_R1_0528_w4a8_per_channel.py b/test/registered/ascend/basic_function/parallel_strategy/expert_parallelism/test_npu_deepep_auto_deepseek_R1_0528_w4a8_per_channel.py index 00a93e1f8c4..0024f65597d 100644 --- a/test/registered/ascend/basic_function/parallel_strategy/expert_parallelism/test_npu_deepep_auto_deepseek_R1_0528_w4a8_per_channel.py +++ b/test/registered/ascend/basic_function/parallel_strategy/expert_parallelism/test_npu_deepep_auto_deepseek_R1_0528_w4a8_per_channel.py @@ -77,6 +77,7 @@ class TestDeepEpAutoDeepseekR1(GSM8KAscendMixin, CustomTestCase): "SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES": "200", "SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "88", "HCCL_BUFFSIZE": "1600", + "DEEPEP_HCCL_BUFFSIZE": "1600", "DEEPEP_NORMAL_LONG_SEQ_ROUND": "10", "DEEPEP_NORMAL_LONG_SEQ_PER_ROUND_TOKENS": "512", "DEEP_NORMAL_MODE_USE_INT8_QUANT": "1", From fb8f8f8f5e7c681fa4c1d47b135b6817eb71031b Mon Sep 17 00:00:00 2001 From: chx96642264 Date: Tue, 14 Jul 2026 15:18:17 +0800 Subject: [PATCH 2/5] gsm8k deepseek-r1, change yml --- .github/workflows/nightly-test-npu.yml | 197 +------------------------ 1 file changed, 3 insertions(+), 194 deletions(-) diff --git a/.github/workflows/nightly-test-npu.yml b/.github/workflows/nightly-test-npu.yml index 90243f7ebdc..602ee973d5a 100644 --- a/.github/workflows/nightly-test-npu.yml +++ b/.github/workflows/nightly-test-npu.yml @@ -5,6 +5,7 @@ on: pull_request: branches: - main + - testcases paths: - ".github/workflows/nightly-test-npu.yml" workflow_dispatch: @@ -103,28 +104,6 @@ jobs: else echo "skip_install_flag=${{ inputs.skip_install_flag }}" >> $GITHUB_OUTPUT fi - nightly-poc-single-node-a2-tests: - name: single-node-poc-a2 - if: ${{ !cancelled() }} - needs: [ set-image-config ] - strategy: - fail-fast: false - max-parallel: 6 - matrix: - test_config: - - name: qwen3_32b_w8a8_2p_in3k5_out1k5_50ms_a2 - runner: linux-aarch64-a2b3-4 - test_case: test/registered/ascend/performance/qwen3_32b/test_npu_qwen3_32b_w8a8_2p_in3k5_out1k5_50ms_a2.py - test_type: 'perf' - uses: ./.github/workflows/nightly-test-npu-e2e-single-node.yml - with: - runner: ${{ matrix.test_config.runner }} - test_type: ${{ matrix.test_config.test_type }} - test_config_name: ${{ matrix.test_config.name }} - test_case: ${{ matrix.test_config.test_case }} - image: ${{ needs.set-image-config.outputs.image_a2 }} - install_sglang_from_source: false - transformers_version: '' nightly-poc-single-node-tests: name: single-node-poc @@ -136,121 +115,10 @@ jobs: matrix: test_config: # qwen3_6_35b_a3b performance tests - - name: qwen3_6_35b_a3b_1p_in3k5_out1k5_50ms - runner: linux-aarch64-a3-2 - test_case: test/registered/ascend/performance/qwen3_6_35b_a3b/test_npu_qwen3_6_35b_a3b_1p_in3k5_out1k5_50ms.py - test_type: 'perf' - - name: qwen3_6_35b_a3b_1p_aime26 - runner: linux-aarch64-a3-2 - test_case: test/registered/ascend/accuracy/qwen3_6_35b_a3b/test_npu_qwen3_6_35b_a3b_1p_aime26.py - - name: qwen3_6_35b_a3b_1p_in64k_out1k_50ms - runner: linux-aarch64-a3-2 - test_case: test/registered/ascend/performance/qwen3_6_35b_a3b/test_npu_qwen3_6_35b_a3b_1p_in64k_out1k_50ms.py - test_type: 'perf' - - name: qwen3_6_35b_a3b_1p_in128k_out1k_50ms - runner: linux-aarch64-a3-2 - test_case: test/registered/ascend/performance/qwen3_6_35b_a3b/test_npu_qwen3_6_35b_a3b_1p_in128k_out1k_50ms.py - test_type: 'perf' - - name: qwen3_6_35b_a3b_1p_in64k_out1k_prefix90_50ms_aime26 - runner: linux-aarch64-a3-2 - test_case: test/registered/ascend/performance/qwen3_6_35b_a3b/test_npu_qwen3_6_35b_a3b_1p_in64k_out1k_prefix90_50ms_aime26.py - test_type: 'perf' - - name: qwen3_6_35b_a3b_1p_in128k_out1k_prefix90_50ms - runner: linux-aarch64-a3-2 - test_case: test/registered/ascend/performance/qwen3_6_35b_a3b/test_npu_qwen3_6_35b_a3b_1p_in128k_out1k_prefix90_50ms.py - test_type: 'perf' - # qwen3_6_27b performance tests - - name: qwen3_6_27b_w8a8_1p_in3k5_out1k5_50ms_gpqa - runner: linux-aarch64-a3-2 - test_case: test/registered/ascend/performance/qwen3_6_27b/test_npu_qwen3_6_27b_w8a8_1p_in3k5_out1k5_50ms_gpqa.py - test_type: 'perf' - - name: qwen3_6_27b_w8a8_2p_in16k_out1k_50ms - runner: linux-aarch64-a3-4 - test_case: test/registered/ascend/performance/qwen3_6_27b/test_npu_qwen3_6_27b_w8a8_2p_in16k_out1k_50ms.py - test_type: 'perf' - - name: qwen3_6_27b_w8a8_1p_in64k_out1k_50ms - runner: linux-aarch64-a3-4 - test_case: test/registered/ascend/performance/qwen3_6_27b/test_npu_qwen3_6_27b_w8a8_1p_in64k_out1k_50ms.py - test_type: 'perf' - - name: qwen3_6_27b_w8a8_2p_in128k_out1k_50ms - runner: linux-aarch64-a3-4 - test_case: test/registered/ascend/performance/qwen3_6_27b/test_npu_qwen3_6_27b_w8a8_2p_in128k_out1k_50ms.py - test_type: 'perf' - - name: qwen3_6_27b_2p_in64k_out1k_prefix90_50ms - runner: linux-aarch64-a3-4 - test_case: test/registered/ascend/performance/qwen3_6_27b/test_npu_qwen3_6_27b_2p_in64k_out1k_prefix90_50ms.py - test_type: 'perf' - - name: qwen3_6_27b_1p_in1024x1024_30_out1024_50ms - runner: linux-aarch64-a3-2 - test_case: test/registered/ascend/performance/qwen3_6_27b/test_npu_qwen3_6_27b_1p_in1024x1024_30_out1024_50ms.py - test_type: 'perf' - - name: qwen3_6_27b_1p_in1080p_30_out256_50ms - runner: linux-aarch64-a3-2 - test_case: test/registered/ascend/performance/qwen3_6_27b/test_npu_qwen3_6_27b_1p_in1080p_30_out256_50ms.py - test_type: 'perf' - - name: qwen3_6_27b_1p_gpqa - runner: linux-aarch64-a3-2 - test_case: test/registered/ascend/accuracy/qwen3_6_27b/test_npu_qwen3_6_27b_1p_gpqa.py - # qwen3_32b performance tests - - name: qwen3_32b_w8a8_2p_in3k5_out1k5_50ms_gpqa - runner: linux-aarch64-a3-4 - test_case: test/registered/ascend/performance/qwen3_32b/test_npu_qwen3_32b_w8a8_2p_in3k5_out1k5_50ms_gpqa.py - test_type: 'perf' - - name: qwen3_32b_bf16_8p_in18k_out4k_6ms - runner: linux-aarch64-a3-16 - test_case: test/registered/ascend/performance/qwen3_32b/test_npu_qwen3_32b_bf16_8p_in18k_out4k_6ms.py - test_type: 'perf' - - name: qwen3_32b_bf16_8p_gpqa - runner: linux-aarch64-a3-16 - test_case: test/registered/ascend/accuracy/qwen3_32b/test_npu_qwen3_32b_bf16_8p_gpqa.py - # qwen3_30b_a3b performance tests - - name: qwen3_30b_w8a8_1p_in3k5_out1k5_50ms_aime25 + - name: deepep_auto_deepseek_R1_0528_w4a8_per_channel runner: linux-aarch64-a3-2 - test_case: test/registered/ascend/performance/qwen3_30b_a3b/test_npu_qwen3_30b_w8a8_1p_in3k5_out1k5_50ms_aime25.py + test_case: test/registered/ascend/basic_function/parallel_strategy/expert_parallelism/test_npu_deepep_auto_deepseek_R1_0528_w4a8_per_channel.py test_type: 'perf' - # qwen3-8b performance tests - - name: qwen3_8b_w8a8_1p_in3k5_out1k5_50ms_gpqa - runner: linux-aarch64-a3-2 - test_case: test/registered/ascend/performance/qwen3-8b/test_npu_qwen3_8b_w8a8_1p_in3k5_out1k5_50ms_gpqa.py - test_type: 'perf' - - name: qwen3_8b_w8a8_1p_in6k_out1k5_bs16_gpqa - runner: linux-aarch64-a3-2 - test_case: test/registered/ascend/performance/qwen3-8b/test_npu_qwen3_8b_w8a8_1p_in6k_out1k5_bs16_gpqa.py - test_type: 'perf' - # qwen3_next_80b_a3b_instruct performance tests - - name: qwen3_next_80b_w8a8_2p_in6k_out1k5_bs16_aime25 - runner: linux-aarch64-a3-4 - test_case: test/registered/ascend/performance/qwen3_next_80b_a3b_instruct/test_npu_qwen3_next_80b_w8a8_2p_in6k_out1k5_bs16_aime25.py - test_type: 'perf' - # minimax_m2_5 performance tests - - name: minimax_m2_5_w8a8_8p_in3k5_out1k5_50ms_gpqa - runner: linux-aarch64-a3-16 - test_case: test/registered/ascend/performance/minimax_m2_5/test_npu_minimax_m2_5_w8a8_8p_in3k5_out1k5_50ms_gpqa.py - test_type: 'perf' - - name: minimax_m2_5_w8a8_4p_in64k_out1k_prefix90_50ms_gpqa - runner: linux-aarch64-a3-16 - test_case: test/registered/ascend/performance/minimax_m2_5/test_npu_minimax_m2_5_w8a8_4p_in64k_out1k_prefix90_50ms_gpqa.py - test_type: 'perf' - # deepseek_v3_2 accuracy tests - - name: deepseek_v3_2_8p_aime25 - runner: linux-aarch64-a3-16 - test_case: test/registered/ascend/accuracy/deepseek_v3_2/test_npu_deepseek_v3_2_8p_aime25.py - # glm4_7_flash accuracy tests - - name: glm4_7_flash_1p_aime25 - runner: linux-aarch64-a3-2 - test_case: test/registered/ascend/accuracy/glm4_7_flash/test_npu_glm4_7_flash_1p_aime25.py - # glm4_6v_flash accuracy tests - - name: glm4_6v_flash_1p_mmmu - runner: linux-aarch64-a3-2 - test_case: test/registered/ascend/accuracy/glm4_6v_flash/test_npu_glm4_6v_flash_1p_mmmu.py - # qwen3_vl_8b_thinking accuracy tests - - name: qwen3_vl_8b_thinking_1p_mmmu - runner: linux-aarch64-a3-2 - test_case: test/registered/ascend/accuracy/qwen3_vl_8b_thinking/test_npu_qwen3_vl_8b_thinking_1p_mmmu.py - # qwen3_vl_30b_a3b_thinking accuracy tests - - name: qwen3_vl_30b_a3b_thinking_1p_mmmu - runner: linux-aarch64-a3-2 - test_case: test/registered/ascend/accuracy/qwen3_vl_30b_a3b_thinking/test_npu_qwen3_vl_30b_a3b_thinking_1p_mmmu.py uses: ./.github/workflows/nightly-test-npu-e2e-single-node.yml with: runner: ${{ matrix.test_config.runner }} @@ -261,69 +129,10 @@ jobs: install_sglang_from_source: false transformers_version: '' - nightly-poc-multi-node-tests: - name: multi-node-poc - if: ${{ !cancelled() }} - needs: [set-image-config, nightly-poc-single-node-tests] - strategy: - fail-fast: false - max-parallel: 1 - matrix: - test_config: - # glm5_1 performance tests - - name: glm5_1_w4a8_1p1d_32p_in64k_out1k_50ms_aime26 - prefill_size: 2 - decode_size: 2 - router_size: 1 - test_case: test/registered/ascend/performance/glm5_1/test_npu_glm5_1_w4a8_1p1d_32p_in64k_out1k_50ms_aime26.py - test_type: 'perf' - prefill_decode_deployment: 'separation' - uses: ./.github/workflows/nightly-test-npu-e2e-multi-node.yml - with: - runner: linux-amd64-cpu-8 - test_type: ${{ matrix.test_config.test_type }} - test_config_name: ${{ matrix.test_config.name }} - prefill_size: ${{ matrix.test_config.prefill_size }} - decode_size: ${{ matrix.test_config.decode_size }} - router_size: ${{ matrix.test_config.router_size }} - test_case: ${{ matrix.test_config.test_case }} - image: ${{ needs.set-image-config.outputs.image_a3 }} - install_sglang_from_source: false - prefill_decode_deployment: ${{ matrix.test_config.prefill_decode_deployment }} - transformers_version: '' - - nightly-poc-multi-node-mix-tests: - name: multi-node-mix-poc - if: ${{ !cancelled() }} - needs: [set-image-config, nightly-poc-single-node-tests, nightly-poc-multi-node-tests] - strategy: - fail-fast: false - max-parallel: 1 - matrix: - test_config: - # kimi_k2_6 performance tests - - name: kimi_k2_6_w4a8_16p_in64k_out1k_100ms_aime25 - node_size: 2 - test_case: test/registered/ascend/performance/kimi_k2_6/test_npu_kimi_k2_6_w4a8_16p_in64k_out1k_100ms_aime25.py - test_type: 'perf' - uses: ./.github/workflows/nightly-test-npu-e2e-multi-node.yml - with: - runner: linux-amd64-cpu-8 - test_type: ${{ matrix.test_config.test_type }} - test_config_name: ${{ matrix.test_config.name }} - node_size: ${{ matrix.test_config.node_size }} - test_case: ${{ matrix.test_config.test_case }} - image: ${{ needs.set-image-config.outputs.image_a3 }} - install_sglang_from_source: false - prefill_decode_deployment: 'mix' - transformers_version: '' check-all-jobs: if: ${{ !cancelled() && github.event_name != 'workflow_call' }} needs: - - nightly-poc-single-node-a2-tests - nightly-poc-single-node-tests - - nightly-poc-multi-node-tests - - nightly-poc-multi-node-mix-tests runs-on: ubuntu-latest steps: - name: Download all metrics From 8bda8213daa9b4c0a69672b72571758e5241b13c Mon Sep 17 00:00:00 2001 From: chx96642264 Date: Tue, 14 Jul 2026 17:26:12 +0800 Subject: [PATCH 3/5] gsm8k deepseek-r1, change yml --- .github/workflows/nightly-test-npu.yml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/.github/workflows/nightly-test-npu.yml b/.github/workflows/nightly-test-npu.yml index 602ee973d5a..d502910a5b4 100644 --- a/.github/workflows/nightly-test-npu.yml +++ b/.github/workflows/nightly-test-npu.yml @@ -116,7 +116,7 @@ jobs: test_config: # qwen3_6_35b_a3b performance tests - name: deepep_auto_deepseek_R1_0528_w4a8_per_channel - runner: linux-aarch64-a3-2 + runner: linux-aarch64-a3-16 test_case: test/registered/ascend/basic_function/parallel_strategy/expert_parallelism/test_npu_deepep_auto_deepseek_R1_0528_w4a8_per_channel.py test_type: 'perf' uses: ./.github/workflows/nightly-test-npu-e2e-single-node.yml From 035b2a5a3fb51bc1cafdb728f35c778a701d002e Mon Sep 17 00:00:00 2001 From: chx96642264 Date: Wed, 15 Jul 2026 11:33:11 +0800 Subject: [PATCH 4/5] change --- .../test_npu_deepep_auto_deepseek_R1_0528_w4a8_per_channel.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/test/registered/ascend/basic_function/parallel_strategy/expert_parallelism/test_npu_deepep_auto_deepseek_R1_0528_w4a8_per_channel.py b/test/registered/ascend/basic_function/parallel_strategy/expert_parallelism/test_npu_deepep_auto_deepseek_R1_0528_w4a8_per_channel.py index 0024f65597d..a1ae76fccef 100644 --- a/test/registered/ascend/basic_function/parallel_strategy/expert_parallelism/test_npu_deepep_auto_deepseek_R1_0528_w4a8_per_channel.py +++ b/test/registered/ascend/basic_function/parallel_strategy/expert_parallelism/test_npu_deepep_auto_deepseek_R1_0528_w4a8_per_channel.py @@ -76,7 +76,7 @@ class TestDeepEpAutoDeepseekR1(GSM8KAscendMixin, CustomTestCase): "SGLANG_SCHEDULER_DECREASE_PREFILL_IDLE": "1", "SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES": "200", "SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "88", - "HCCL_BUFFSIZE": "1600", + # "HCCL_BUFFSIZE": "1600", "DEEPEP_HCCL_BUFFSIZE": "1600", "DEEPEP_NORMAL_LONG_SEQ_ROUND": "10", "DEEPEP_NORMAL_LONG_SEQ_PER_ROUND_TOKENS": "512", From 7d9ab79132512d9ad4eb248f007dfcf0fe8f8f4f Mon Sep 17 00:00:00 2001 From: chx96642264 Date: Wed, 15 Jul 2026 17:19:46 +0800 Subject: [PATCH 5/5] gsm8k, 2 --- ...est_npu_deepep_auto_deepseek_R1_0528_w4a8_per_channel.py | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/test/registered/ascend/basic_function/parallel_strategy/expert_parallelism/test_npu_deepep_auto_deepseek_R1_0528_w4a8_per_channel.py b/test/registered/ascend/basic_function/parallel_strategy/expert_parallelism/test_npu_deepep_auto_deepseek_R1_0528_w4a8_per_channel.py index a1ae76fccef..bf0170b7733 100644 --- a/test/registered/ascend/basic_function/parallel_strategy/expert_parallelism/test_npu_deepep_auto_deepseek_R1_0528_w4a8_per_channel.py +++ b/test/registered/ascend/basic_function/parallel_strategy/expert_parallelism/test_npu_deepep_auto_deepseek_R1_0528_w4a8_per_channel.py @@ -41,7 +41,9 @@ class TestDeepEpAutoDeepseekR1(GSM8KAscendMixin, CustomTestCase): "21", "22", "--mem-fraction-static", - "0.725", + "0.85", + "--max-total-tokens", + "20480", "--max-running-requests", "352", "--disable-radix-cache", @@ -76,7 +78,7 @@ class TestDeepEpAutoDeepseekR1(GSM8KAscendMixin, CustomTestCase): "SGLANG_SCHEDULER_DECREASE_PREFILL_IDLE": "1", "SGLANG_PREFILL_DELAYER_MAX_DELAY_PASSES": "200", "SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "88", - # "HCCL_BUFFSIZE": "1600", + "HCCL_BUFFSIZE": "1600", "DEEPEP_HCCL_BUFFSIZE": "1600", "DEEPEP_NORMAL_LONG_SEQ_ROUND": "10", "DEEPEP_NORMAL_LONG_SEQ_PER_ROUND_TOKENS": "512",