@tduncan / Nvidia-Optimization-Stack-VLLM
Code Commits Issues Pulls Cloud Agent CI/CD
main
main
yaml 153 lines 5.48 KB
Raw
1 group: Quantization
2 depends_on:
3 - image-build
4 steps:
5 - label: ":nvidia: (H200 MIG 35GB) Quantization Shard %N"
6 device: h200_35gb
7 key: quantization
8 timeout_in_minutes: 40
9 parallelism: 4
10 env:
11 VLLM_USE_V2_MODEL_RUNNER: "0"
12 source_file_dependencies:
13 - csrc/
14 - vllm/model_executor/layers/quantization
15 - tests/quantization
16 commands:
17 # Pin torchao to the version compatible with the CI PyTorch and CUDA stack.
18 - uv pip install --system torchao==0.17.0 --index-url https://download.pytorch.org/whl/cu130
19 - uv pip install --system conch-triton-kernels
20 # The SM90-only checkpoint currently contains a removed weight_chan_scale
21 # parameter. It was not exercised by the previous L4 job.
22 - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ --ignore quantization/test_blackwell_moe.py --ignore quantization/test_rocm_moe.py -k 'not test_compressed_tensors_w4a8_fp8' --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
23 mirror:
24 amd:
25 label: ":amd: (MI300) Quantization Shard %N"
26 dind: false
27 device: mi300_1
28 timeout_in_minutes: 115
29 working_dir: "/vllm-workspace/tests"
30 depends_on:
31 - image-build-amd
32 source_file_dependencies:
33 - csrc/
34 - vllm/model_executor/layers/quantization
35 - tests/quantization
36 - vllm/_aiter_ops.py
37 - vllm/platforms/rocm.py
38 - vllm/model_executor/layers/fused_moe/oracle/unquantized.py
39 - vllm/model_executor/layers/fused_moe/unquantized_fused_moe_method.py
40 - tests/rocm/test_moe_weight_replay.py
41 commands:
42 # Use the ROCm default instead of the NVIDIA runner override.
43 - unset VLLM_USE_V2_MODEL_RUNNER
44 - uv pip install --system torchao==0.17.0
45 - uv pip install --system conch-triton-kernels
46 - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ --ignore quantization/test_blackwell_moe.py --ignore quantization/test_rocm_moe.py --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
47 - if [ "$$BUILDKITE_PARALLEL_JOB" = "0" ]; then pytest -v -s rocm/test_moe_weight_replay.py; fi
48
49 - label: ":nvidia: (H200 MIG 35GB) Quantized Fusions"
50 device: h200_35gb
51 key: quantized-fusions
52 timeout_in_minutes: 20
53 source_file_dependencies:
54 - tests/fusion
55 - vllm/model_executor/layers/fusion
56 - vllm/model_executor/kernels/linear
57 - vllm/model_executor/layers/quantization/compressed_tensors
58 - vllm/model_executor/layers/quantization/modelopt.py
59 commands:
60 - pytest -v -s fusion/
61 mirror:
62 amd:
63 label: ":amd: (MI300) Quantized Fusions"
64 dind: false
65 device: mi300_1
66 timeout_in_minutes: 45
67 working_dir: "/vllm-workspace/tests"
68 depends_on:
69 - image-build-amd
70 source_file_dependencies:
71 - tests/fusion
72 - vllm/model_executor/layers/fusion
73 - vllm/model_executor/kernels/linear
74 - vllm/model_executor/layers/quantization/compressed_tensors
75 - vllm/model_executor/layers/quantization/modelopt.py
76 - vllm/platforms/rocm.py
77
78 - label: ":nvidia: (B200) Quantized MoE"
79 key: quantized-moe-test-b200
80 timeout_in_minutes: 120
81 working_dir: "/vllm-workspace/"
82 device: b200-k8s
83 source_file_dependencies:
84 - tests/quantization/test_blackwell_moe.py
85 - vllm/model_executor/models/deepseek_v2.py
86 - vllm/model_executor/models/gpt_oss.py
87 - vllm/model_executor/models/llama4.py
88 - vllm/model_executor/layers/fused_moe
89 - vllm/model_executor/layers/quantization/compressed_tensors
90 - vllm/model_executor/layers/quantization/modelopt.py
91 - vllm/model_executor/layers/quantization/mxfp4.py
92 - vllm/v1/attention/backends/flashinfer.py
93 commands:
94 - pytest -s -v tests/quantization/test_blackwell_moe.py
95 mirror:
96 amd:
97 label: ":amd: (MI355) Quantized MoE"
98 dind: false
99 device: mi355_1
100 timeout_in_minutes: 45
101 working_dir: "/vllm-workspace/"
102 depends_on:
103 - image-build-amd
104 source_file_dependencies:
105 - tests/quantization/test_gfx950_moe.py
106 - tests/quantization/test_rocm_moe.py
107 - tests/utils.py
108 - vllm/model_executor/models/deepseek_v2.py
109 - vllm/model_executor/models/qwen3_moe.py
110 - vllm/model_executor/models/gpt_oss.py
111 - vllm/model_executor/model_loader/
112 - vllm/model_executor/layers/fused_moe/
113 - vllm/model_executor/layers/quantization/
114 - vllm/model_executor/kernels/linear/
115 - vllm/v1/attention/
116 - vllm/v1/worker/gpu/
117 - vllm/_aiter_ops.py
118 - vllm/platforms/rocm.py
119 - vllm/envs.py
120 commands:
121 - pytest -v -s tests/quantization/test_rocm_moe.py
122 - pytest -v -s tests/quantization/test_gfx950_moe.py
123
124 - label: ":nvidia: (H200 MIG 35GB) Quantized Models"
125 device: h200_35gb
126 key: quantized-models-test
127 timeout_in_minutes: 65
128 env:
129 VLLM_USE_V2_MODEL_RUNNER: "0"
130 source_file_dependencies:
131 - vllm/model_executor/layers/quantization
132 - tests/models/quantization
133 commands:
134 - pytest -v -s models/quantization
135 mirror:
136 amd:
137 label: ":amd: (MI355 DPX) Quantized Models"
138 dind: false
139 device: mi355_dpx
140 timeout_in_minutes: 80
141 working_dir: "/vllm-workspace/tests"
142 depends_on:
143 - image-build-amd
144 source_file_dependencies:
145 - vllm/model_executor/layers/quantization
146 - tests/models/quantization
147 - vllm/_aiter_ops.py
148 - vllm/platforms/rocm.py
149 - vllm/model_executor/model_loader/
150 commands:
151 # Use the ROCm default instead of the NVIDIA runner override.
152 - unset VLLM_USE_V2_MODEL_RUNNER
153 - pytest -v -s models/quantization