[atom-sgl-benchmark] Add manual OOB selection and Minimax-M3 for SGLang MI355 benchmark

.github/benchmark/sglang_benchmark_models.json

@@ -22,7 +22,8 @@
       "deepseek_v4_prefix_cache_common": "SGLANG_DEFAULT_SERVER_ARGS=\nAITER_BF16_FP8_MOE_BOUND=0\nATOM_MOE_GU_ITLV=1\nSGLANG_DEFAULT_THINKING=1\nSGLANG_DSV4_REASONING_EFFORT=max\nSGLANG_USE_AITER=1\nSGLANG_DSV4_FP4_EXPERTS=true\nSGLANG_EXTERNAL_MODEL_PACKAGE=atom.plugin.sglang.models\nSGLANG_ENABLE_TORCH_COMPILE=1\nTORCHINDUCTOR_COMPILE_THREADS=128",
       "glm52_common": "SGLANG_DEFAULT_SERVER_ARGS=\nAITER_QUICK_REDUCE_QUANTIZATION=INT4\nAITER_USE_FLYDSL_MOE_SORTING=1\nSGLANG_USE_AITER=1\nSGLANG_ENABLE_TORCH_COMPILE=1\nSGLANG_EXTERNAL_MODEL_PACKAGE=atom.plugin.sglang.models\nTORCHINDUCTOR_COMPILE_THREADS=128",
       "glm52_mi308_common": "SGLANG_DEFAULT_SERVER_ARGS=\nAITER_QUICK_REDUCE_QUANTIZATION=INT4\nAITER_USE_FLYDSL_MOE_SORTING=1\nSGLANG_AITER_FP8_PREFILL_ATTN=0\nSGLANG_ENABLE_TORCH_COMPILE=1\nSGLANG_USE_AITER=1\nSGLANG_EXTERNAL_MODEL_PACKAGE=atom.plugin.sglang.models\nTORCHINDUCTOR_COMPILE_THREADS=128",
-      "qwen_common": "SGLANG_DEFAULT_SERVER_ARGS=\nSGLANG_EXTERNAL_MODEL_PACKAGE=atom.plugin.sglang.models\nATOM_ENABLE_QK_NORM_ROPE_CACHE_QUANT_FUSION=0"
+      "qwen_common": "SGLANG_DEFAULT_SERVER_ARGS=\nSGLANG_EXTERNAL_MODEL_PACKAGE=atom.plugin.sglang.models\nATOM_ENABLE_QK_NORM_ROPE_CACHE_QUANT_FUSION=0",
+      "minimax_m3_common": "SGLANG_DEFAULT_SERVER_ARGS=\nAITER_QUICK_REDUCE_QUANTIZATION=INT4\nSGLANG_USE_AITER=1\nATOM_FORCE_ATTN_TRITON=1\nSGLANG_AITER_FP8_PREFILL_ATTN=0\nSGLANG_ENABLE_TORCH_COMPILE=1\nTORCHINDUCTOR_COMPILE_THREADS=128\nSGLANG_EXTERNAL_MODEL_PACKAGE=atom.plugin.sglang.models\nSGLANG_EXTERNAL_MM_PROCESSOR_PACKAGE=atom.plugin.sglang.models"
     }
   },
   "models": [
@@ -462,6 +463,31 @@
       "nightly_group": "A",
       "env_vars": "glm52_common"
     },
+    {
+      "display": "MiniMax-M3-MXFP4 TP4",
+      "dashboard_model": "MiniMax-M3-MXFP4-tp4",
+      "workload_label": "SGLang-OOB",
+      "source_path": "amd/MiniMax-M3-MXFP4",
+      "path": "amd/MiniMax-M3-MXFP4",
+      "prefix": "minimax-m3-mxfp4-tp4",
+      "extra_args": [
+        "trust_remote_code",
+        "--tensor-parallel-size 4",
+        "--attention-backend aiter",
+        "--page-size 128",
+        "--mem-fraction-static 0.8",
+        "--context-length 32768",
+        "--max-running-requests 128",
+        "--model-loader-extra-config '{\"online_quant_config\":{\"global_quant_config\":\"ptpc_fp8\",\"exclude_layer\":[\"lm_head\",\"model.embed_tokens\",\"vision_tower\",\"multi_modal_projector\",\"patch_merge_mlp\",\"*block_sparse_moe\"]}}'",
+        "--json-model-override-args '{\"use_index_cache\":true,\"index_topk_freq\":4}'",
+        "--kv-cache-dtype fp8_e4m3",
+        "--disable-radix-cache"
+      ],
+      "bench_args": "",
+      "runner": "atom-mi355-8gpu-aac-runner",
+      "nightly_group": "B",
+      "env_vars": "minimax_m3_common"
+    },
     {
       "display": "GLM-5.2 FP8 TP8 MI308",
       "dashboard_model": "GLM-5.2-FP8-tp8 MI308",

.github/workflows/atom-sglang-benchmark.yaml

@@ -26,6 +26,10 @@ on:
         options:
           - none
           - all
+          - OOB_P0
+          - OOB_P1
+          - OOB_P2
+          - OOB_P0+OOB_P1
           - DeepSeek-V4-Pro TP8
           - DeepSeek-V4-Pro MTP1 TP8
           - DeepSeek-V4-Pro MTP3 TP8
@@ -50,6 +54,7 @@ on:
           - GLM-5.1 FP8 TP8
           - GLM-5.2 FP8 TP4
           - GLM-5.2 FP4 TP4
+          - MiniMax-M3-MXFP4 TP4
       model_mi308:
         description: "Manual MI308 benchmark model selection"
         required: false
@@ -664,6 +669,7 @@ jobs:
               "GLM-5.1 FP8 TP8": "glm-5-1-fp8-tp8",
               "GLM-5.2 FP8 TP4": "glm-5-2-fp8-tp4",
               "GLM-5.2 FP4 TP4": "glm-5-2-fp4-tp4",
+              "MiniMax-M3-MXFP4 TP4": "minimax-m3-mxfp4-tp4",
               "MI308 Qwen3.5-397B-A17B-FP8 TP4": "qwen3-5-397b-a17b-fp8-tp4-mi308",
               "MI308 Qwen3.5-397B-A17B-FP8 TP8": "qwen3-5-397b-a17b-fp8-tp8-mi308",
               "MI308 Qwen3.5-35B-A3B-FP8 TP1": "qwen3-5-35b-a3b-fp8-tp1-mi308",
@@ -691,9 +697,9 @@ jobs:
               "deepseek-v4-pro-tp8",
               "deepseek-v4-pro-mtp1-tp8",
               "deepseek-v4-pro-mtp3-tp8",
-              "glm-5-1-fp8-tp8",
               "glm-5-2-fp8-tp4",
               "glm-5-2-fp4-tp4",
+              "minimax-m3-mxfp4-tp4",
           ]
           OOB_P1_PREFIX_ORDER = [
               "deepseek-v3-2-fp8-tp4",
@@ -715,6 +721,7 @@ jobs:
               "deepseek-r1-fp4-mtp3-tp8-dp8-ep8",
               "qwen3-5-397b-a17b-fp8-tp4",
               "qwen3-5-397b-a17b-fp8-tp8",
+              "glm-5-1-fp8-tp8",
           ]
 
           def is_oob_mi308(model):
@@ -740,40 +747,70 @@ jobs:
               if schedule_cron not in schedule_to_tier:
                   return "SKIP-UNKNOWN-DAY", []
               selected_group, tier_order = schedule_to_tier[schedule_cron]
+              return selected_group, select_oob_models([tier_order], include_mi308=True)
+
+          def select_oob_models(tier_orders, include_mi308):
               by_prefix = first_model_by_prefix(models)
-              tier_prefixes = frozenset(tier_order)
+              tier_prefixes = frozenset(
+                  prefix
+                  for tier_order in tier_orders
+                  for prefix in tier_order
+              )
               selected = []
-              for prefix in tier_order:
-                  model = by_prefix.get(prefix)
-                  if model is not None:
-                      selected.append(model)
+              seen_prefixes = set()
+              for tier_order in tier_orders:
+                  for prefix in tier_order:
+                      if prefix in seen_prefixes:
+                          continue
+                      seen_prefixes.add(prefix)
+                      model = by_prefix.get(prefix)
+                      if model is not None:
+                          selected.append(model)
+              if not include_mi308:
+                  return selected
               for prefix in sorted(by_prefix):
                   if prefix in tier_prefixes:
                       continue
                   model = by_prefix[prefix]
                   if is_oob_mi308(model):
                       selected.append(model)
-              return selected_group, selected
+              return selected
+
+          MANUAL_OOB_SELECTIONS = {
+              "OOB_P0": ("OOB_P0", [OOB_P0_PREFIX_ORDER]),
+              "OOB_P1": ("OOB_P1", [OOB_P1_PREFIX_ORDER]),
+              "OOB_P2": ("OOB_P2", [OOB_P2_PREFIX_ORDER]),
+              "OOB_P0+OOB_P1": ("OOB_P0+OOB_P1", [OOB_P0_PREFIX_ORDER, OOB_P1_PREFIX_ORDER]),
+          }
 
           if event == "schedule":
               schedule_cron = os.environ.get("SCHEDULE_CRON", "")
               selected_group, selected = scheduled_oob_selection(schedule_cron)
           else:
               model_mi355 = os.environ.get("MODEL_MI355") or "none"
               model_mi308 = os.environ.get("MODEL_MI308") or "none"
-              selected = [
+              selected = []
+              manual_oob_selection = MANUAL_OOB_SELECTIONS.get(model_mi355)
+              if manual_oob_selection:
+                  selected_group, tier_orders = manual_oob_selection
+                  selected.extend(select_oob_models(tier_orders, include_mi308=False))
+              else:
+                  selected.extend(
+                      model
+                      for model in models
+                      if matches_workload(model)
+                      and manually_selected(model, model_mi355, expected_mi308=False)
+                  )
+              selected.extend(
                   model
                   for model in models
                   if matches_workload(model)
-                  and (
-                      manually_selected(model, model_mi355, expected_mi308=False)
-                      or manually_selected(model, model_mi308, expected_mi308=True)
-                  )
-              ]
+                  and manually_selected(model, model_mi308, expected_mi308=True)
+              )
 
           print(f"SGLang workload label: {workload_label}", file=sys.stderr)
           if selected_group:
-              print(f"Scheduled SGLang benchmark group: {selected_group}", file=sys.stderr)
+              print(f"SGLang benchmark group: {selected_group}", file=sys.stderr)
           if not selected:
               print("No models selected for SGLang benchmark.", file=sys.stderr)
           else:
@@ -792,7 +829,7 @@ jobs:
           MODELS_JSON: ${{ steps.load.outputs.models_json }}
         run: |
           if [[ -n "${SELECTED_GROUP}" ]]; then
-            echo "Scheduled SGLang benchmark group: ${SELECTED_GROUP}"
+            echo "SGLang benchmark group: ${SELECTED_GROUP}"
           fi
           printf 'Selected models: %s\n' "${MODELS_JSON}"