From 4bf9bdf28fc61f9bc588be9d6329b8e04e6d02bf Mon Sep 17 00:00:00 2001
From: Gahow Wang <gahow.wang@gmail.com>
Date: Thu, 16 Jul 2026 21:52:43 +0800
Subject: [PATCH] Add full Qwen30 MoE profiling matrix

---
 .../jobs_allreduce_smoke_tp2.toml             |  6 +--
 .../jobs_moe_full.toml                        | 18 +++++++
 .../run_moe_full.sh                           | 51 +++++++++++++++++++
 3 files changed, 72 insertions(+), 3 deletions(-)
 create mode 100644 runs/frontier-qwen30-vllm020-profile-v1/jobs_moe_full.toml
 create mode 100644 runs/frontier-qwen30-vllm020-profile-v1/run_moe_full.sh

diff --git a/runs/frontier-qwen30-vllm020-profile-v1/jobs_allreduce_smoke_tp2.toml b/runs/frontier-qwen30-vllm020-profile-v1/jobs_allreduce_smoke_tp2.toml
index 802066e..7a9d10f 100644
--- a/runs/frontier-qwen30-vllm020-profile-v1/jobs_allreduce_smoke_tp2.toml
+++ b/runs/frontier-qwen30-vllm020-profile-v1/jobs_allreduce_smoke_tp2.toml
@@ -1,12 +1,12 @@
 version = 1
 
 [[jobs]]
-name = "qwen30-vllm020-flashinfer-allreduce-smoke-tp2-20260716-v1"
+name = "qwen30-vllm020-flashinfer-allreduce-smoke-tp2-20260716-v2-config-context"
 gpus = 2
 gpu_model = "H20"
 hosts = ["dash0"]
 command = "cd /home/admin/cpfs/wjh/aituner/aituner-qwen30-vllm020-profile-v1/runs/frontier-qwen30-vllm020-profile-v1 && timeout --signal=TERM --kill-after=30s 840 bash run_allreduce_profile.sh"
-artifacts = ["artifacts/allreduce-smoke-tp2"]
+artifacts = ["artifacts/allreduce-smoke-tp2-v2"]
 
 [jobs.env]
 HOME = "/tmp/wjh"
@@ -14,6 +14,6 @@ XDG_CACHE_HOME = "/tmp/wjh/.cache"
 VLLM_CACHE_ROOT = "/tmp/wjh/.cache/vllm"
 TP = "2"
 NUM_TOKENS = "8"
-OUTPUT_ROOT = "/home/admin/cpfs/wjh/aituner/aituner-qwen30-vllm020-profile-fleet/artifacts/allreduce-smoke-tp2"
+OUTPUT_ROOT = "/home/admin/cpfs/wjh/aituner/aituner-qwen30-vllm020-profile-fleet/artifacts/allreduce-smoke-tp2-v2"
 VENV_ROOT = "/tmp/wjh/venvs/vllm-0.20.0-cu129-profiler-v1"
 VLLM_SOURCE = "/home/admin/cpfs/wjh/agentic-kv/third_party/vllm_v20_build"
diff --git a/runs/frontier-qwen30-vllm020-profile-v1/jobs_moe_full.toml b/runs/frontier-qwen30-vllm020-profile-v1/jobs_moe_full.toml
new file mode 100644
index 0000000..aaeae66
--- /dev/null
+++ b/runs/frontier-qwen30-vllm020-profile-v1/jobs_moe_full.toml
@@ -0,0 +1,18 @@
+version = 1
+
+[[jobs]]
+name = "qwen30-vllm020-moe-full-20260716-v1-local-shard"
+gpus = 1
+gpu_model = "H20"
+hosts = ["dash0"]
+command = "cd /home/admin/cpfs/wjh/aituner/aituner-qwen30-vllm020-profile-v1/runs/frontier-qwen30-vllm020-profile-v1 && timeout --signal=TERM --kill-after=30s 1620 bash run_moe_full.sh"
+artifacts = ["artifacts/moe-full-v1"]
+
+[jobs.env]
+HOME = "/tmp/wjh"
+XDG_CACHE_HOME = "/tmp/wjh/.cache"
+VLLM_CACHE_ROOT = "/tmp/wjh/.cache/vllm"
+OUTPUT_ROOT = "/home/admin/cpfs/wjh/aituner/aituner-qwen30-vllm020-profile-fleet/artifacts/moe-full-v1"
+VENV_ROOT = "/tmp/wjh/venvs/vllm-0.20.0-cu129-profiler-v1"
+VLLM_SOURCE = "/home/admin/cpfs/wjh/agentic-kv/third_party/vllm_v20_build"
+MODEL = "/home/admin/cpfs/wjh/models/Qwen/Qwen3-30B-A3B"
diff --git a/runs/frontier-qwen30-vllm020-profile-v1/run_moe_full.sh b/runs/frontier-qwen30-vllm020-profile-v1/run_moe_full.sh
new file mode 100644
index 0000000..374a78b
--- /dev/null
+++ b/runs/frontier-qwen30-vllm020-profile-v1/run_moe_full.sh
@@ -0,0 +1,51 @@
+#!/usr/bin/env bash
+
+set -euo pipefail
+
+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}"
+TOKENS=(1 8 16 32 64 128 256 512 1024 2048 4096 8192)
+mkdir -p "${OUTPUT_ROOT}/logs" "${OUTPUT_ROOT}/provenance" "${OUTPUT_ROOT}/raw"
+exec > >(tee -a "${OUTPUT_ROOT}/logs/full.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=FlashInfer-CUTLASS-unquantized-MoE tp_local_shapes=1,2,4 tokens=${TOKENS[*]} routing=uniform_random_logits,hotset8 reference_check=on dtype=BF16 output=${OUTPUT_ROOT} expected_wall=4-12m hard_wall=1500s hard_gpu_cap=0.42_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 "${VENV_ROOT}/bin/python"
+test -f "${MODEL}/config.json"
+git rev-parse HEAD > "${OUTPUT_ROOT}/provenance/aituner.commit"
+git -C "${VLLM_SOURCE}" rev-parse HEAD > "${OUTPUT_ROOT}/provenance/vllm-source.commit"
+sha256sum profile_vllm020_moe.py run_moe_full.sh \
+  > "${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 > "${OUTPUT_ROOT}/provenance/gpus.csv"
+printf '%s\n' "${TOKENS[@]}" > "${OUTPUT_ROOT}/provenance/tokens.txt"
+
+timeout --signal=TERM --kill-after=30s 1380 \
+  "${VENV_ROOT}/bin/python" profile_vllm020_moe.py \
+  --vllm-source "${VLLM_SOURCE}" \
+  --model "${MODEL}" \
+  --output "${OUTPUT_ROOT}/raw/moe-full.json" \
+  --tp 1 2 4 \
+  --num-tokens "${TOKENS[@]}" \
+  --routing-modes uniform_random_logits hotset8 \
+  --warmup-iters 5 \
+  --repeats 10 \
+  --check-reference
+
+test -s "${OUTPUT_ROOT}/raw/moe-full.json"
+sha256sum "${OUTPUT_ROOT}/raw/moe-full.json" "${OUTPUT_ROOT}/provenance"/* \
+  > "${OUTPUT_ROOT}/artifacts.sha256"
+date -u +"END_UTC=%Y-%m-%dT%H:%M:%SZ"
+echo "MOE_FULL_COMPLETE rows=72"