[ROCm][CI] Add the MI355 TurboQuant t3nc mirror (#58432)
Signed-off-by: Andreas Karatzas <andreas.karatzas@protonmail.com> Co-authored-by: OpenAI Codex <noreply@openai.com>
Andreas Karatzas committed
Sep 24, 2026 at 02:16 UTC
c3f5270270aedf598b5f633088623c194c8f0b6c
4 files changed
+62
.buildkite/test-amd.yaml
+27
@@ -4126,6 +4126,33 @@ steps:
4126
commands:
4127
- pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/models-turboquant-k3v4nc-rocm.txt
4128
4129
+- label: ":amd: (MI355) LM Eval TurboQuant t3nc"
4130
+ timeout_in_minutes: 30
4131
+ mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
4132
+ dind: false
4133
+ agent_pool: mi355_1
4134
+ num_gpus: 1
4135
+ optional: true
4136
+ working_dir: "/vllm-workspace/tests"
4137
+ source_file_dependencies:
4138
+ - .buildkite/test-amd.yaml
4139
+ - tests/evals/gsm8k/
4140
+ - tests/utils.py
4141
+ - tests/conftest.py
4142
+ - vllm/model_executor/layers/quantization/turboquant/
4143
+ - vllm/v1/attention/backends/turboquant_attn.py
4144
+ - vllm/v1/attention/backends/triton_attn.py
4145
+ - vllm/v1/attention/ops/triton_unified_attention.py
4146
+ - vllm/v1/attention/selector.py
4147
+ - vllm/v1/attention/ops/triton_turboquant_decode.py
4148
+ - vllm/v1/attention/ops/triton_turboquant_store.py
4149
+ - vllm/v1/attention/ops/flydsl_turboquant_decode.py
4150
+ - vllm/v1/attention/ops/flydsl_kernels/
4151
+ - vllm/v1/attention/ops/turboquant_soa/
4152
+ - vllm/platforms/rocm.py
4153
+ commands:
4154
+ - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/models-turboquant-t3nc-rocm.txt
4155
+
4156
- label: ":amd: (MI355) LM Eval Small Models Distributed"
4157
timeout_in_minutes: 55
4158
mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
.buildkite/test_areas/lm_eval.yaml
+28
@@ -766,6 +766,34 @@ steps:
766
- vllm/v1/attention/ops/triton_turboquant_store.py
767
commands:
768
- pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/models-turboquant-t3nc.txt
769
+ mirror:
770
+ amd:
771
+ label: ":amd: (MI355) LM Eval TurboQuant t3nc"
772
+ dind: false
773
+ device: mi355_1
774
+ num_devices: 1
775
+ timeout_in_minutes: 30
776
+ working_dir: /vllm-workspace/tests
777
+ depends_on:
778
+ - image-build-amd
779
+ source_file_dependencies:
780
+ - .buildkite/test_areas/lm_eval.yaml
781
+ - tests/evals/gsm8k/
782
+ - tests/utils.py
783
+ - tests/conftest.py
784
+ - vllm/model_executor/layers/quantization/turboquant/
785
+ - vllm/v1/attention/backends/turboquant_attn.py
786
+ - vllm/v1/attention/backends/triton_attn.py
787
+ - vllm/v1/attention/ops/triton_unified_attention.py
788
+ - vllm/v1/attention/selector.py
789
+ - vllm/v1/attention/ops/triton_turboquant_decode.py
790
+ - vllm/v1/attention/ops/triton_turboquant_store.py
791
+ - vllm/v1/attention/ops/flydsl_turboquant_decode.py
792
+ - vllm/v1/attention/ops/flydsl_kernels/
793
+ - vllm/v1/attention/ops/turboquant_soa/
794
+ - vllm/platforms/rocm.py
795
+ commands:
796
+ - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/models-turboquant-t3nc-rocm.txt
797
798
- label: ":nvidia: (H100) GPQA Eval (GPT-OSS)"
799
key: gpqa-eval-gpt-oss-2xh100
tests/evals/gsm8k/configs/Qwen3-4B-TQ-t3nc-ROCm.yaml
new
+6
@@ -0,0 +1,6 @@
1
+model_name: "Qwen/Qwen3-4B"
2
+accuracy_threshold: 0.75
3
+num_questions: 1319
4
+num_fewshot: 5
5
+# Triton serves the unquantized boundary layers with TurboQuant-compatible layout.
6
+server_args: "--kv-cache-dtype turboquant_3bit_nc --max-model-len 4096 --attention-backend TRITON_ATTN"
tests/evals/gsm8k/configs/models-turboquant-t3nc-rocm.txt
new
+1
@@ -0,0 +1 @@
1
+Qwen3-4B-TQ-t3nc-ROCm.yaml