[atom-sgl-vllm-ci] Modify Qwen3.5-35B-A3B-FP8 runner

.github/actionlint.yaml

@@ -5,7 +5,9 @@ self-hosted-runner:
     - atom-mi355-8gpu.predownload
     - atom-mi355-8gpu-aac-runner
     - atom-mi355-8gpu-conductor-sgl-runner
+    - atom-mi355-8gpu-vllm-sgl-ci
     - atom-mi308-8gpu-plugins-benchmark
+    - atom-mi308-8gpu-vllm-sgl-ci
     - atom-plugin-acc-validation-runner
     - build-only-atom
     - linux-atom-do-mi350x-8

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

@@ -252,7 +252,7 @@ jobs:
               SGLANG_ENABLE_TORCH_COMPILE=1
               TORCHINDUCTOR_COMPILE_THREADS=128
             accuracy_test_threshold: 0.91
-            runner: linux-atom-mi35x-4
+            runner: atom-mi355-8gpu-vllm-sgl-ci
           - model_name: "DeepSeek-R1-FP4 TP4"
             model_path: "amd/DeepSeek-R1-0528-MXFP4-v2"
             extra_args: "--trust-remote-code --tensor-parallel-size 4 --attention-backend aiter --kv-cache-dtype fp8_e4m3 --mem-fraction-static 0.85 --page-size 1 --disable-radix-cache"
@@ -265,7 +265,7 @@ jobs:
               SGLANG_ENABLE_TORCH_COMPILE=1
               TORCHINDUCTOR_COMPILE_THREADS=128
             accuracy_test_threshold: 0.91
-            runner: linux-atom-mi35x-4
+            runner: atom-mi355-8gpu-vllm-sgl-ci
           - model_name: "Qwen3.5-35B-A3B-FP8 TP2"
             model_path: "Qwen/Qwen3.5-35B-A3B-FP8"
             extra_args: "--tensor-parallel-size 2 --mem-fraction-static 0.9 --reasoning-parser qwen3 --disable-radix-cache"
@@ -274,7 +274,7 @@ jobs:
               SGLANG_EXTERNAL_MODEL_PACKAGE=atom.plugin.sglang.models
               ATOM_ENABLE_QK_NORM_ROPE_CACHE_QUANT_FUSION=0
             accuracy_test_threshold: 0.76
-            runner: linux-atom-mi35x-4
+            runner: atom-mi355-8gpu-vllm-sgl-ci
           - model_name: "DeepSeek-V4-Pro TP8"
             model_path: "deepseek-ai/DeepSeek-V4-Pro"
             extra_args: "--trust-remote-code --tensor-parallel-size 8 --kv-cache-dtype fp8_e4m3 --mem-fraction-static 0.9 --swa-full-tokens-ratio 0.1 --max-running-requests 256 --page-size 256 --disable-radix-cache --disable-shared-experts-fusion --tool-call-parser deepseekv4 --reasoning-parser deepseek-v4"
@@ -291,7 +291,7 @@ jobs:
             lm_eval_num_fewshot: 5
             lm_eval_num_concurrent: 8
             accuracy_test_threshold: 0.94
-            runner: atom-mi355-8gpu.predownload
+            runner: atom-mi355-8gpu-vllm-sgl-ci
     runs-on: ${{ matrix.runner }}
     timeout-minutes: 180
     env:
@@ -464,8 +464,17 @@ jobs:
           if [ -d "/models" ]; then
             MODEL_CACHE_MOUNT="-v /models:/models"
             MODEL_CACHE_DESC="/models (host mount)"
+          elif [ -d "/it-share/models" ]; then
+            MODEL_CACHE_MOUNT="-v /it-share/models:/models"
+            MODEL_CACHE_DESC="/it-share/models (host path)"
+          elif [ -d "/mnt/dcgpuval/models" ]; then
+            MODEL_CACHE_MOUNT="-v /mnt/dcgpuval/models:/models"
+            MODEL_CACHE_DESC="/mnt/dcgpuval/models (host path)"
+          elif [ -d "/shareddata/models" ]; then
+            MODEL_CACHE_MOUNT="-v /shareddata/models:/models"
+            MODEL_CACHE_DESC="/shareddata/models (host path)"
           else
-            echo "Warning: /models directory not found on runner; using container-local /models."
+            echo "Warning: /models and /it-share/models and /mnt/dcgpuval/models and /shareddata/models directory not found on runner; using container-local /models."
           fi
 
           echo "Using model cache backend: ${MODEL_CACHE_DESC}"
@@ -487,7 +496,7 @@ jobs:
             -v "${GITHUB_WORKSPACE:-$PWD}":/workspace \
             $MODEL_MOUNT \
             -w /workspace \
-            --ipc=host --group-add video \
+            --ipc=host --network=host --group-add video \
             --shm-size=16G \
             --privileged \
             --cap-add=SYS_PTRACE \
@@ -585,7 +594,15 @@ jobs:
           fi
 
           echo "RESULT_FILE: $result_file"
-          flexible_extract_value=$(jq '.results.gsm8k["exact_match,flexible-extract"]' "$result_file")
+          flexible_extract_value=$(python3 - "$result_file" <<'PY'
+          import json
+          import sys
+
+          with open(sys.argv[1], encoding="utf-8") as f:
+              data = json.load(f)
+          print(data["results"]["gsm8k"]["exact_match,flexible-extract"])
+          PY
+          )
           echo "Flexible extract value: $flexible_extract_value"
           echo "Accuracy test threshold: ${{ matrix.accuracy_test_threshold }}"
 

.github/workflows/atom-vllm-test.yaml

@@ -250,7 +250,7 @@ jobs:
               ATOM_MOE_GU_ITLV=1
             lm_eval_num_fewshot: 20
             accuracy_test_threshold: 0.94
-            runner: atom-mi35x-8gpu-oot-acc
+            runner: atom-mi355-8gpu-vllm-sgl-ci
           - display_name: "gpt-oss-120b TP1"
             model_name: "gpt-oss-120b"
             model_path: "openai/gpt-oss-120b"
@@ -259,15 +259,15 @@ jobs:
             env_vars: "ATOM_ENABLE_QK_NORM_ROPE_CACHE_QUANT_FUSION=1\nVLLM_USE_V2_MODEL_RUNNER=1"
             lm_eval_num_fewshot: 3
             accuracy_test_threshold: 0.88
-            runner: atom-mi35x-8gpu-oot-acc
+            runner: atom-mi355-8gpu-vllm-sgl-ci
           - display_name: "Kimi-K2.5-MXFP4 TP4"
             model_name: "Kimi-K2.5-MXFP4"
             model_path: "amd/Kimi-K2.5-MXFP4-AttnFP8"
-            extra_args: "--tensor-parallel-size 4"
-            env_vars: ""
+            extra_args: "--trust-remote-code --tensor-parallel-size 4 --max-num-batched-tokens 16384 --max-model-len 16384"
+            env_vars: "AITER_QUICK_REDUCE_QUANTIZATION=INT4"
             lm_eval_num_fewshot: 3
             accuracy_test_threshold: 0.92
-            runner: atom-mi35x-8gpu-oot-acc
+            runner: atom-mi355-8gpu-vllm-sgl-ci
           - display_name: "Qwen3.5-35B-A3B-FP8 TP2"
             model_name: "Qwen3.5-35B-A3B-FP8"
             model_path: "Qwen/Qwen3.5-35B-A3B-FP8"
@@ -276,8 +276,8 @@ jobs:
               ATOM_ENABLE_QK_NORM_ROPE_CACHE_QUANT_FUSION=1
               ATOM_USE_CUSTOM_ALL_GATHER=0
             lm_eval_num_fewshot: 3
-            accuracy_test_threshold: 0.76
-            runner: atom-mi35x-8gpu-oot-acc
+            accuracy_test_threshold: 0.70
+            runner: atom-mi355-8gpu-vllm-sgl-ci
     runs-on: ${{ matrix.runner }}
     timeout-minutes: 180
     env:
@@ -467,8 +467,17 @@ jobs:
           if [ -d "/models" ]; then
             MODEL_CACHE_MOUNT="-v /models:/models"
             MODEL_CACHE_DESC="/models (host mount)"
+          elif [ -d "/it-share/models" ]; then
+            MODEL_CACHE_MOUNT="-v /it-share/models:/models"
+            MODEL_CACHE_DESC="/it-share/models (host path)"
+          elif [ -d "/mnt/dcgpuval/models" ]; then
+            MODEL_CACHE_MOUNT="-v /mnt/dcgpuval/models:/models"
+            MODEL_CACHE_DESC="/mnt/dcgpuval/models (host path)"
+          elif [ -d "/shareddata/models" ]; then
+            MODEL_CACHE_MOUNT="-v /shareddata/models:/models"
+            MODEL_CACHE_DESC="/shareddata/models (host path)"
           else
-            echo "Warning: /models directory not found on runner; using container-local /models."
+            echo "Warning: /models and /it-share/models and /mnt/dcgpuval/models and /shareddata/models directory not found on runner; using container-local /models."
           fi
 
           echo "Using model cache backend: ${MODEL_CACHE_DESC}"
@@ -490,7 +499,7 @@ jobs:
             -v "${GITHUB_WORKSPACE:-$PWD}":/workspace \
             $MODEL_MOUNT \
             -w /workspace \
-            --ipc=host --group-add video \
+            --ipc=host --network=host --group-add video \
             --shm-size=16G \
             --privileged \
             --cap-add=SYS_PTRACE \
@@ -593,7 +602,15 @@ jobs:
           fi
 
           echo "RESULT_FILE: $result_file"
-          flexible_extract_value=$(jq '.results.gsm8k["exact_match,flexible-extract"]' "$result_file")
+          flexible_extract_value=$(python3 - "$result_file" <<'PY'
+          import json
+          import sys
+
+          with open(sys.argv[1], encoding="utf-8") as f:
+              data = json.load(f)
+          print(data["results"]["gsm8k"]["exact_match,flexible-extract"])
+          PY
+          )
           echo "Flexible extract value: $flexible_extract_value"
           echo "Accuracy test threshold: ${{ matrix.accuracy_test_threshold }}"