| 1 | model_name: "RedHatAI/DeepSeek-Coder-V2-Lite-Instruct-FP8" |
| 2 | accuracy_threshold: 0.72 |
| 3 | num_questions: 1319 |
| 4 | num_fewshot: 5 |
| 5 | rocm_request_timeout_seconds: 1800 |
| 6 | # Dense MLA PCP on MI355; DP1/DCP1, eager execution, and Triton MoE. |
| 7 | server_args: >- |
| 8 | --revision efe1ced428db63e7ccbcf367334596f77e9af140 |
| 9 | --code-revision e434a23f91ba5b4923cf6c9d9a238eb4a08e3a11 |
| 10 | --dtype bfloat16 |
| 11 | --enforce-eager |
| 12 | --max-model-len 4096 |
| 13 | --max-num-seqs 8 |
| 14 | --max-num-batched-tokens 4096 |
| 15 | --tensor-parallel-size 1 |
| 16 | --prefill-context-parallel-size 2 |
| 17 | --enable-expert-parallel |
| 18 | --moe-backend triton |
| 19 | --attention-config '{"backend":"ROCM_AITER_MLA","mla_prefill_backend":"ROCM_AITER_FA"}' |
| 20 | --kv-cache-dtype auto |
| 21 | env: |
| 22 | VLLM_USE_V2_MODEL_RUNNER: "1" |
| 23 | VLLM_ROCM_USE_AITER: "1" |
| 24 | VLLM_WORKER_MULTIPROC_METHOD: "spawn" |