diff --git a/.github/benchmark/sglang_benchmark_models.json b/.github/benchmark/sglang_benchmark_models.json index 9dcd6766a..ffc089a44 100644 --- a/.github/benchmark/sglang_benchmark_models.json +++ b/.github/benchmark/sglang_benchmark_models.json @@ -479,6 +479,22 @@ "nightly_group": "B", "env_vars": "qwen_common" }, + { + "display": "Qwen3.5-35B-A3B-PTPC-FP8 TP1 MI308", + "dashboard_model": "Qwen3.5-35B-A3B-PTPC-FP8 MI308", + "workload_label": "SGLang-OOB", + "source_path": "amd/Qwen3.5-35B-A3B-PTPC-FP8", + "path": "amd/Qwen3.5-35B-A3B-PTPC-FP8", + "prefix": "qwen3-5-35b-a3b-ptpc-fp8-tp1-mi308", + "extra_args": [ + "--tensor-parallel-size 1", + "qwen_reasoning" + ], + "bench_args": "", + "runner": "atom-mi308-8gpu-plugins-benchmark", + "nightly_group": "B", + "env_vars": "qwen_common" + }, { "display": "Qwen3-32B-FP8 TP1 MI308", "dashboard_model": "Qwen3-32B-FP8 MI308", diff --git a/.github/benchmark/sglang_models_accuracy.json b/.github/benchmark/sglang_models_accuracy.json index 28a3c5b30..3b80f2116 100644 --- a/.github/benchmark/sglang_models_accuracy.json +++ b/.github/benchmark/sglang_models_accuracy.json @@ -344,6 +344,18 @@ "accuracy_baseline_model": "Qwen/Qwen3.5-35B-A3B-FP8", "_baseline_note": "Threshold aligned with the SGLANG accuracy validation workflow target for gsm8k." }, + { + "model_name": "MI308 Qwen3.5-35B-A3B-PTPC-FP8 TP1", + "model_path": "amd/Qwen3.5-35B-A3B-PTPC-FP8", + "extraArgs": "--tensor-parallel-size 1 --mem-fraction-static 0.9 --reasoning-parser qwen3 --disable-radix-cache", + "env_vars": "SGLANG_DEFAULT_SERVER_ARGS=\nSGLANG_EXTERNAL_MODEL_PACKAGE=atom.plugin.sglang.models\nATOM_ENABLE_QK_NORM_ROPE_CACHE_QUANT_FUSION=0", + "runner": "atom-mi308-8gpu-plugins-accuracy", + "test_level": "nightly", + "accuracy_threshold": 0.76, + "accuracy_baseline": null, + "accuracy_baseline_model": "amd/Qwen3.5-35B-A3B-PTPC-FP8", + "_baseline_note": "Threshold aligned with the SGLANG accuracy validation workflow target for gsm8k." + }, { "model_name": "MI308 Qwen3-32B-FP8 TP1", "model_path": "Qwen/Qwen3-32B-FP8", diff --git a/.github/workflows/atom-sglang-accuracy-validation.yaml b/.github/workflows/atom-sglang-accuracy-validation.yaml index a34556556..fa22f79da 100644 --- a/.github/workflows/atom-sglang-accuracy-validation.yaml +++ b/.github/workflows/atom-sglang-accuracy-validation.yaml @@ -50,6 +50,7 @@ on: - MI308 Qwen3.5-397B-A17B-FP8 TP4 - MI308 Qwen3.5-397B-A17B-FP8 TP8 - MI308 Qwen3.5-35B-A3B-FP8 TP1 + - MI308 Qwen3.5-35B-A3B-PTPC-FP8 TP1 - MI308 Qwen3-32B-FP8 TP1 - MI308 Qwen3-32B-FP8 TP8 - MI308 GLM-5.1-FP8 TP8 @@ -384,6 +385,14 @@ jobs: "env_vars": "SGLANG_DEFAULT_SERVER_ARGS=\nSGLANG_EXTERNAL_MODEL_PACKAGE=atom.plugin.sglang.models\nATOM_ENABLE_QK_NORM_ROPE_CACHE_QUANT_FUSION=0", "runner": "atom-mi308-8gpu-plugins-accuracy", }, + { + "model_name": "MI308 Qwen3.5-35B-A3B-PTPC-FP8 TP1", + "model_path": "amd/Qwen3.5-35B-A3B-PTPC-FP8", + "extra_args": "--tensor-parallel-size 1 --mem-fraction-static 0.9 --reasoning-parser qwen3 --disable-radix-cache", + "accuracy_test_threshold": 0.76, + "env_vars": "SGLANG_DEFAULT_SERVER_ARGS=\nSGLANG_EXTERNAL_MODEL_PACKAGE=atom.plugin.sglang.models\nATOM_ENABLE_QK_NORM_ROPE_CACHE_QUANT_FUSION=0", + "runner": "atom-mi308-8gpu-plugins-accuracy", + }, { "model_name": "MI308 Qwen3-32B-FP8 TP1", "model_path": "Qwen/Qwen3-32B-FP8", diff --git a/.github/workflows/atom-sglang-benchmark.yaml b/.github/workflows/atom-sglang-benchmark.yaml index b1b2cd50b..1752e2a4d 100644 --- a/.github/workflows/atom-sglang-benchmark.yaml +++ b/.github/workflows/atom-sglang-benchmark.yaml @@ -46,6 +46,7 @@ on: - "qwen3-5-397b-a17b-fp8-tp4-mi308 (1024x1024/8192x1024: [4,8,16,32,64])" - "qwen3-5-397b-a17b-fp8-tp8-mi308 (1024x1024/8192x1024: [4,8,16,32,64])" - "qwen3-5-35b-a3b-fp8-tp1-mi308 (1024x1024/8192x1024: [4,8,16,32,64])" + - "qwen3-5-35b-a3b-ptpc-fp8-tp1-mi308 (1024x1024/8192x1024: [4,8,16,32,64])" - "qwen3-32b-fp8-tp1-mi308 (1024x1024/8192x1024: [4,8,16,32,64])" - "qwen3-32b-fp8-tp8-mi308 (1024x1024/8192x1024: [4,8,16,32,64])" - "deepseek-v4-flash-fp8-mi308 (1024x1024/8192x1024: [4,8,16,32,64])"