| 1 | # DeepSeek-V4-Flash MXFP4 (AMD Quark) accuracy test. |
| 2 | # Runs on ROCm (gfx950) with the AITER MoE + fused shared-experts path. |
| 3 | # Reference GSM8K (8-shot, lm-eval): strict 0.9522 / flexible 0.9515 measured |
| 4 | # on gfx950 TP=4 (AITER_MXFP4_BF16 backend); default rtol=0.08. |
| 5 | model_name: "amd/DeepSeek-V4-Flash-MXFP4" |
| 6 | tasks: |
| 7 | - name: "gsm8k" |
| 8 | metrics: |
| 9 | - name: "exact_match,strict-match" |
| 10 | value: 0.95 |
| 11 | - name: "exact_match,flexible-extract" |
| 12 | value: 0.95 |
| 13 | limit: 1319 |
| 14 | num_fewshot: 8 |
| 15 | max_model_len: 8192 |
| 16 | kv_cache_dtype: fp8 |
| 17 | moe_backend: aiter |
| 18 | enforce_eager: false |
| 19 | trust_remote_code: true |
| 20 | tokenizer_mode: deepseek_v4 |
| 21 | required_gpu_arch: |
| 22 | - gfx950 |
| 23 | env_vars: |
| 24 | VLLM_ROCM_USE_AITER: "1" |
| 25 | VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS: "1" |