refactor: migrate fixed-sequence recipes to SRT-Slurm / 将定长配方迁移至 SRT-Slurm

.github/workflows/benchmark-tmpl.yml

@@ -110,6 +110,7 @@ env:
   HF_HUB_CACHE: '/mnt/hf_hub_cache/'
   EXP_NAME: ${{ fromJSON(inputs.config).exp-name }}
   RECIPE_FINGERPRINT: ${{ fromJSON(inputs.config).recipe-fingerprint || '' }}
+  SRT_RECIPE: ${{ fromJSON(inputs.config).srt-recipe || '' }}
   MODEL: ${{ fromJSON(inputs.config).model }}
   THINKING_MODE: thinking_on
   MODEL_PREFIX: ${{ fromJSON(inputs.config).model-prefix }}
@@ -399,6 +400,9 @@ jobs:
             server.log
             results/*.log
             results/*_config.json
+            srt-single-node-logs.tar.gz
+            srt-single-node-submission.json
+            srt-slurm-sha.txt
           if-no-files-found: ignore
 
       - name: Upload GPU metrics

MODELS.md

@@ -58,6 +58,8 @@ Rationale: `dsv4` carries the largest single-turn footprint in the repository. 4
 
 **Deprecation parity audit (2026-09-21):** Active master configs and benchmark-script locations match the enacted retirements above and in the support matrix below. GLM-5.1 B200 TileRT remains the documented exception to the earlier GLM-5/5.1 and 1k1k retirements. Conditional A/B baseline retirement remains pending; non-speculative Pareto contributors remain supported. The broader routing audit also removed stale retired-model branches from launchers/runtime settings and a GLM-5-only environment override, and corrected workflow/agent guidance that still recommended retired coverage. SPEED-Bench collectors, historical result readers, and the explicitly retained recipe YAMLs remain available. Deprecated configs are consolidated in [`configs/deprecated/amd-master.yaml`](configs/deprecated/amd-master.yaml) and [`configs/deprecated/nvidia-master.yaml`](configs/deprecated/nvidia-master.yaml).
 
+**Single-node SRT-only cutover (2026-09-22):** Active single-node fixed-sequence recipes now use SRT-Slurm. The two Docker-only Qwen3.5 RTX PRO 6000 FP4 configs (with and without MTP) are retired, with their original settings preserved in `configs/deprecated/nvidia-master.yaml` and their scripts in `benchmarks/single_node/fixed_seq_len/deprecated/`. The unused `rtx6000pro-lat` runner mappings, launcher, and runtime settings are removed. Qwen3.5 remains active on the other supported Slurm pools; AgentX and multi-node coverage are unchanged.
+
 ## Scenarios
 
 | Scenario | ISL/OSL | Status |

MODELS_zh.md

@@ -58,6 +58,8 @@ InferenceX-e2e 运行在数量固定且有限的 GPU 资源池上，并由一支
 
 **弃用状态一致性核查（2026-09-21）：** 启用的主配置及基准测试脚本位置与上述已执行的退役事项和下方支持矩阵一致。GLM-5.1 B200 TileRT 仍是文档明确保留的例外，不受此前 GLM-5/5.1 和 1k1k 退役范围限制。有条件的 A/B 基线退役仍待执行；对 Pareto 前沿有贡献的非投机解码配置继续受支持。进一步的路由核查还移除了启动器和运行时设置中遗留的退役模型分支及 GLM-5 专用环境覆盖，并修正了仍推荐退役配置的工作流和智能体指南。SPEED-Bench 采集器、历史结果读取逻辑及明确保留的配方 YAML 继续保留。弃用配置现统一归档至 [`configs/deprecated/amd-master.yaml`](configs/deprecated/amd-master.yaml) 和 [`configs/deprecated/nvidia-master.yaml`](configs/deprecated/nvidia-master.yaml)。
 
+**单节点切换为仅使用 SRT（2026-09-22）：** 活跃的单节点定长配方现统一使用 SRT-Slurm。两个仅支持 Docker 的 Qwen3.5 RTX PRO 6000 FP4 配置（启用和关闭 MTP）已退役，原始设置保留在 `configs/deprecated/nvidia-master.yaml`，脚本保留在 `benchmarks/single_node/fixed_seq_len/deprecated/`。已移除不再使用的 `rtx6000pro-lat` runner 映射、启动器及运行时设置。Qwen3.5 在其他受支持的 Slurm 池上继续启用；AgentX 和多节点覆盖保持不变。
+
 ## 场景
 
 | 场景 | ISL/OSL | 状态 |

benchmarks/benchmark_lib.sh

@@ -425,6 +425,24 @@ GPU_METRICS_CSV="${GPU_METRICS_CSV:-gpu_metrics.csv}"
 NVIDIA_GPU_MONITOR_QUERY="timestamp,index,power.draw,temperature.gpu,clocks.current.sm,clocks.current.memory,utilization.gpu,utilization.memory"
 export GPU_METRICS_CSV
 
+# Keep one AMD CSV header and forward each complete row immediately. Some awk
+# implementations buffer pipe input even with fflush(), losing the final ticks
+# when the monitor stops.
+_filter_amd_smi_metrics() {
+    local line header_seen=false
+    while IFS= read -r line; do
+        if [[ "$line" == timestamp,* ]]; then
+            if [[ "$header_seen" == true ]]; then
+                continue
+            fi
+            header_seen=true
+        fi
+        if [[ "$header_seen" == true ]]; then
+            printf '%s\n' "$line"
+        fi
+    done
+}
+
 # Background nvidia-smi/amd-smi sampler writing CSV.
 # Usage: start_gpu_monitor [--output /path/to/output.csv] [--interval 1]
 start_gpu_monitor() {
@@ -457,10 +475,9 @@ start_gpu_monitor() {
     elif command -v amd-smi &>/dev/null; then
         GPU_MONITOR_VENDOR="amd"
         # amd-smi is Python and block-buffers stdout; without PYTHONUNBUFFERED the
-        # trailing ticks were lost at kill (measured on MI355X). awk keeps the first
-        # CSV header, drops repeated ones, and flushes every row for the same reason.
+        # trailing ticks were lost at kill (measured on MI355X).
         PYTHONUNBUFFERED=1 amd-smi metric -p -c -t -u -w "$interval" --csv 2>/dev/null \
-            | awk '/^timestamp,/{if(!h){print;h=1};next} h{print;fflush()}' > "$output" &
+            | _filter_amd_smi_metrics > "$output" &
         GPU_MONITOR_PID=$!
         # Hardware energy-accumulator + identity snapshots; the end-side twin in
         # stop_gpu_monitor lets auditors cross-check the integrated energy
@@ -800,6 +817,7 @@ run_benchmark_serving() {
     local model=""
     local port=""
     local backend=""
+    local base_url=""
     local endpoint=""
     local input_len=""
     local output_len=""
@@ -834,6 +852,10 @@ run_benchmark_serving() {
                 endpoint="$2"
                 shift 2
                 ;;
+            --base-url)
+                base_url="$2"
+                shift 2
+                ;;
             --input-len)
                 input_len="$2"
                 shift 2
@@ -954,12 +976,16 @@ run_benchmark_serving() {
         num_prompts="$max_concurrency"
     fi
 
+    if [[ -z "$base_url" ]]; then
+        base_url="http://0.0.0.0:$port"
+    fi
+
     local benchmark_cmd=(
         env PYTHONPATH="$workspace_dir${PYTHONPATH:+:$PYTHONPATH}"
         python3 -m infx.bench_serving.benchmark_serving
         --model "$model"
         --backend "$backend"
-        --base-url "http://0.0.0.0:$port"
+        --base-url "$base_url"
         --dataset-name random
         --random-input-len "$input_len"
         --random-output-len "$output_len"