Add long-context attention profile closure

runs/frontier-fidelity-envelope-v1/jobs_attention_long_context.toml

@@ -0,0 +1,43 @@
+version = 1
+
+[[jobs]]
+name = "qwen30-attention-long-context-tp1-20260717-v1"
+gpus = 1
+gpu_model = "H20"
+hosts = ["dash0"]
+command = "cd /home/admin/cpfs/wjh/aituner/aituner-qwen30-vllm020-profile-v1/runs/frontier-fidelity-envelope-v1 && timeout --signal=TERM --kill-after=30s 900 bash run_flashattn_long_context.sh"
+artifacts = ["artifacts/attention-long-context-tp1-v1"]
+
+[jobs.env]
+HOME = "/tmp/wjh"
+XDG_CACHE_HOME = "/tmp/wjh/.cache"
+TP = "1"
+OUTPUT_ROOT = "/home/admin/cpfs/wjh/aituner/gpu-fleet-fidelity-envelope-v1/artifacts/attention-long-context-tp1-v1"
+
+[[jobs]]
+name = "qwen30-attention-long-context-tp2-20260717-v1"
+gpus = 1
+gpu_model = "H20"
+hosts = ["dash0"]
+command = "cd /home/admin/cpfs/wjh/aituner/aituner-qwen30-vllm020-profile-v1/runs/frontier-fidelity-envelope-v1 && timeout --signal=TERM --kill-after=30s 900 bash run_flashattn_long_context.sh"
+artifacts = ["artifacts/attention-long-context-tp2-v1"]
+
+[jobs.env]
+HOME = "/tmp/wjh"
+XDG_CACHE_HOME = "/tmp/wjh/.cache"
+TP = "2"
+OUTPUT_ROOT = "/home/admin/cpfs/wjh/aituner/gpu-fleet-fidelity-envelope-v1/artifacts/attention-long-context-tp2-v1"
+
+[[jobs]]
+name = "qwen30-attention-long-context-tp4-20260717-v1"
+gpus = 1
+gpu_model = "H20"
+hosts = ["dash0"]
+command = "cd /home/admin/cpfs/wjh/aituner/aituner-qwen30-vllm020-profile-v1/runs/frontier-fidelity-envelope-v1 && timeout --signal=TERM --kill-after=30s 900 bash run_flashattn_long_context.sh"
+artifacts = ["artifacts/attention-long-context-tp4-v1"]
+
+[jobs.env]
+HOME = "/tmp/wjh"
+XDG_CACHE_HOME = "/tmp/wjh/.cache"
+TP = "4"
+OUTPUT_ROOT = "/home/admin/cpfs/wjh/aituner/gpu-fleet-fidelity-envelope-v1/artifacts/attention-long-context-tp4-v1"

runs/frontier-fidelity-envelope-v1/run_flashattn_long_context.sh

@@ -0,0 +1,65 @@
+#!/usr/bin/env bash
+
+set -euo pipefail
+
+TP="${TP:?TP must be set to 1, 2, or 4}"
+case "${TP}" in
+  1|2|4) ;;
+  *) echo "ERROR: invalid TP=${TP}" >&2; exit 1 ;;
+esac
+
+OUTPUT_ROOT="${OUTPUT_ROOT:?OUTPUT_ROOT must be set}"
+VENV_ROOT="${VENV_ROOT:-/tmp/wjh/venvs/vllm-0.20.0-cu129-profiler-v1}"
+VLLM_SOURCE="${VLLM_SOURCE:-/home/admin/cpfs/wjh/agentic-kv/third_party/vllm_v20_build}"
+MODEL="${MODEL:-/home/admin/cpfs/wjh/models/Qwen/Qwen3-30B-A3B}"
+CAMPAIGN_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+PROFILE_SCRIPT="${CAMPAIGN_ROOT}/../frontier-qwen30-vllm020-profile-v1/profile_vllm020_flashattn.py"
+BATCH_SPECS=(
+  q1s16k 8q1s16k 16q1s16k 32q1s16k 64q1s16k
+  q1s32k 8q1s32k 16q1s32k 32q1s32k 64q1s32k
+  q1s40k 8q1s40k 16q1s40k 32q1s40k 64q1s40k
+  q512_8q1s16k q1k_16q1s16k q2k_32q1s16k
+  q4k_16q1s32k q2k_32q1s32k q1k_64q1s32k
+)
+
+mkdir -p "${OUTPUT_ROOT}/logs" "${OUTPUT_ROOT}/provenance" "${OUTPUT_ROOT}/raw"
+exec > >(tee -a "${OUTPUT_ROOT}/logs/long-context.log") 2>&1
+
+IFS=',' read -r -a GPU_IDS <<< "${CUDA_VISIBLE_DEVICES:?fleet GPU is required}"
+if [[ "${#GPU_IDS[@]}" -ne 1 ]]; then
+  echo "ERROR: expected exactly one GPU, got ${CUDA_VISIBLE_DEVICES}" >&2
+  exit 1
+fi
+
+echo "PROFILE_LAUNCH_ECHO host=$(hostname) gpu=${CUDA_VISIBLE_DEVICES} model=${MODEL} runtime=vLLM-0.20.0+cu129 operator=FlashAttention3 tp=${TP} batch_specs=${BATCH_SPECS[*]} profile_script=${PROFILE_SCRIPT} output=${OUTPUT_ROOT} expected_wall=1-5m hard_wall=900s hard_gpu_cap=0.25_H20h"
+date -u +"START_UTC=%Y-%m-%dT%H:%M:%SZ"
+nvidia-smi --query-gpu=index,name,driver_version,memory.used,utilization.gpu --format=csv,noheader
+
+test -x "${VENV_ROOT}/bin/python"
+test -f "${VLLM_SOURCE}/benchmarks/attention_benchmarks/runner.py"
+test -f "${MODEL}/config.json"
+test -f "${PROFILE_SCRIPT}"
+
+git -C "${CAMPAIGN_ROOT}/../.." rev-parse HEAD > "${OUTPUT_ROOT}/provenance/aituner.commit"
+git -C "${VLLM_SOURCE}" rev-parse HEAD > "${OUTPUT_ROOT}/provenance/vllm-source.commit"
+sha256sum "${PROFILE_SCRIPT}" "${BASH_SOURCE[0]}" > "${OUTPUT_ROOT}/provenance/source.sha256"
+uv pip freeze --python "${VENV_ROOT}/bin/python" > "${OUTPUT_ROOT}/provenance/pip-freeze.txt"
+nvidia-smi --query-gpu=index,uuid,name,driver_version,memory.total --format=csv,noheader > "${OUTPUT_ROOT}/provenance/gpus.csv"
+printf '%s\n' "${BATCH_SPECS[@]}" > "${OUTPUT_ROOT}/provenance/batch-specs.txt"
+
+timeout --signal=TERM --kill-after=30s 780 \
+  "${VENV_ROOT}/bin/python" "${PROFILE_SCRIPT}" \
+  --vllm-source "${VLLM_SOURCE}" \
+  --model "${MODEL}" \
+  --output "${OUTPUT_ROOT}/raw/flashattn-long-context-tp${TP}.json" \
+  --tp "${TP}" \
+  --batch-specs "${BATCH_SPECS[@]}" \
+  --warmup-iters 5 \
+  --repeats 10 \
+  --profile-kv-update
+
+test -s "${OUTPUT_ROOT}/raw/flashattn-long-context-tp${TP}.json"
+sha256sum "${OUTPUT_ROOT}/raw/flashattn-long-context-tp${TP}.json" "${OUTPUT_ROOT}/provenance"/* > "${OUTPUT_ROOT}/artifacts.sha256"
+nvidia-smi --query-gpu=index,name,memory.used,utilization.gpu --format=csv,noheader
+date -u +"END_UTC=%Y-%m-%dT%H:%M:%SZ"
+echo "FLASHATTN_LONG_CONTEXT_COMPLETE tp=${TP} cases=${#BATCH_SPECS[@]}"

runs/frontier-fidelity-envelope-v1/run_frontier_qwen30_exact_trace_surface.py

@@ -276,6 +276,7 @@ def main() -> None:
         loads = []
         config_knobs = BASE.knobs(config, paths, args.output_root / "cache")
         config_knobs["enable_prefix_caching"] = args.prefix_caching
+        config_knobs["prediction_max_tokens_per_request"] = 40960
         for trace in traces:
             run_dir = args.output_root / "runs" / config.name / trace["label"]
             result_path = run_dir / "result.json"