[Klaud Cold] Migrate all SPEED-Bench AL collectors to native srt-slurm / 将全部 SPEED-Bench AL 收集器迁移到原生 srt-slurm

.github/workflows/speedbench-al.yml

@@ -21,7 +21,7 @@ on: # zizmor: ignore[concurrency-limits]
         type: string
         default: 'deepseek-ai/DeepSeek-V4-Pro'
       model-prefix:
-        description: "Model prefix; drives launcher MODEL_PATH resolution, exp name, collector script, and artifact names"
+        description: "Model prefix; drives srt-slurm recipe lookup, exp name, and artifact names"
         required: false
         type: string
         default: 'dsv4'
@@ -105,8 +105,6 @@ env:
   EP_SIZE: '1'
   DP_ATTENTION: 'false'
   SPEC_DECODING: mtp
-  # Run the AL-matrix collector instead of the auto-selected throughput script.
-  BENCH_SCRIPT_OVERRIDE: benchmarks/single_node/speedbench/${{ inputs.model-prefix }}_fp4_b300_vllm.sh
   SALLOC_TIME_LIMIT: ${{ inputs.salloc-time }}
   # Matrix-collector tunables (propagated into the container via srun --export=ALL).
   MTP_LIST: ${{ inputs.mtp-list }}
@@ -239,7 +237,70 @@ jobs:
           if [[ -f runners/runtime_settings.sh ]]; then
             source runners/runtime_settings.sh
           fi
-          bash ./runners/launch_${RUNNER_NAME%%_*}.sh
+
+          # All collectors use the srt-slurm single-node path with per-model
+          # recipes under srt-slurm-recipes/<prefix>/vllm/b300-fp4-speedbench/.
+          RECIPE_PATH="benchmarks/single_node/srt-slurm-recipes/${MODEL_PREFIX}/vllm/b300-fp4-speedbench/speedbench.yaml"
+          if [[ ! -f "$RECIPE_PATH" ]]; then
+            echo "ERROR: SPEED-Bench recipe not found: $RECIPE_PATH" >&2
+            exit 1
+          fi
+          # Plain TP8 is incompatible with the Qwen3.8-Flash-Next FP8 checkpoint's
+          # 128-wide quantization blocks; its recipe runs TP4 (as the legacy
+          # collector did), so bind the matrix TP and GPU count to match.
+          if [[ "$MODEL_PREFIX" == qwen3.8next ]]; then
+            export TP=4 GPU_COUNT=4
+          fi
+          mkdir -p speedbench_results
+          ALL_FAILED=true
+          # Harmless values for launch_srt_single_node env checks; the collector
+          # recipe and client do not read them.
+          export ISL=256
+          export OSL=256
+          export RANDOM_RANGE_RATIO=0.0
+          export CONC=1
+          export PP_SIZE=1
+          export DCP_SIZE=1
+          export PCP_SIZE=1
+          for mode in $THINKING_MODES; do
+            for mtp in $MTP_LIST; do
+              IDX=$((mtp - 1))
+              export THINKING="$mode"
+              export MTP="$mtp"
+              export SRT_RECIPE="${RECIPE_PATH}:zip_override_mtp[${IDX}]"
+              export RESULT_FILENAME="speedbench_${mode}_mtp${mtp}"
+              echo ""
+              echo "=========================================="
+              echo "  srt-slurm cell: thinking=${mode} MTP=${mtp}"
+              echo "  SRT_RECIPE=${SRT_RECIPE}"
+              echo "=========================================="
+              CELL_RC=0
+              bash ./runners/launch_"${RUNNER_NAME%%_*}".sh || CELL_RC=$?
+              # Collect per-cell artifacts into the results directory.
+              if [[ -f "${RESULT_FILENAME}.json" ]]; then
+                mv "${RESULT_FILENAME}.json" "speedbench_results/${RESULT_FILENAME}.json"
+                ALL_FAILED=false
+              fi
+              if [[ -f srt-single-node-logs.tar.gz ]]; then
+                mv srt-single-node-logs.tar.gz "speedbench_results/srt-logs_${mode}_mtp${mtp}.tar.gz"
+              fi
+              if [[ "$CELL_RC" -ne 0 ]]; then
+                echo "  -> cell failed (rc=$CELL_RC), recording N/A"
+              fi
+            done
+          done
+          if [[ "$ALL_FAILED" == true ]]; then
+            echo "ERROR: every SPEED-Bench cell failed" >&2
+            exit 1
+          fi
+          # Aggregate per-cell result JSONs into the reference YAML.
+          PYTHONPATH="${GITHUB_WORKSPACE}${PYTHONPATH:+:$PYTHONPATH}" \
+            python3 -m infx.workflows.speedbench_matrix \
+            --result-dir speedbench_results \
+            --model-key "$(basename "$(echo "$MODEL" | tr '[:upper:]' '[:lower:]')")" \
+            --thinking-modes "$THINKING_MODES" \
+            --mtp-list "$MTP_LIST" \
+            > speedbench-reference-al.yaml
 
           if [ ! -f "speedbench-reference-al.yaml" ]; then
             echo "AL collection failed: speedbench-reference-al.yaml not produced." >&2
@@ -292,7 +353,9 @@ jobs:
         uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
         with:
           name: speedbench_server_logs-${{ inputs.model-prefix }}
-          path: speedbench_results/server_*.log
+          path: |
+            speedbench_results/server_*.log
+            speedbench_results/srt-logs_*.tar.gz
           if-no-files-found: ignore
 
       # Per-request benchmark detail (vllm bench serve --save-detailed): includes