Add opt-in native two-node MI300X path for Kimi K3 AgentX / 新增 Kimi K3 AgentX 的 MI300X 双节点 native 路径

.github/workflows/test-matrix-logic.yml

@@ -5,6 +5,9 @@ on:
   pull_request:
     paths:
       - 'utils/matrix_logic/**'
+      - 'runners/launch_mi300x-amds*.sh'
+      - 'runners/mi300x_native_node_preflight.sh'
+      - 'benchmarks/multi_node/agentic/kimik3_fp4_mi300x_vllm.sh'
 
 permissions:
   contents: read
@@ -39,3 +42,8 @@ jobs:
         run: |
           cd utils/matrix_logic
           pytest test_validation.py -v
+
+      - name: test_kimik3_mi300x_native tests
+        run: |
+          cd utils/matrix_logic
+          pytest test_kimik3_mi300x_native.py -v

benchmarks/multi_node/agentic/kimik3_fp4_mi300x_vllm.sh

@@ -0,0 +1,114 @@
+#!/usr/bin/env bash
+set -euo pipefail
+set -x
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+source "$SCRIPT_DIR/../../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    MODEL_PATH \
+    PORT \
+    CONC_LIST \
+    PREFILL_NUM_WORKERS \
+    PREFILL_TP \
+    PREFILL_PP_SIZE \
+    PREFILL_EP \
+    PREFILL_DP_ATTN \
+    DECODE_NUM_WORKERS \
+    MULTINODE_NODE_COUNT \
+    MULTINODE_GPUS_PER_NODE \
+    MULTINODE_NODE_RANK \
+    MULTINODE_MASTER_ADDR
+
+require_agentic_kv_offload_none
+
+fail() {
+    echo "ERROR: $*" >&2
+    exit 1
+}
+
+if [[ "$MULTINODE_NODE_COUNT" != "2" ]]; then
+    fail "this entrypoint serves exactly 2 nodes, got MULTINODE_NODE_COUNT=$MULTINODE_NODE_COUNT"
+fi
+if [[ "$MULTINODE_GPUS_PER_NODE" != "8" ]]; then
+    fail "this entrypoint requires 8 GPUs per node, got MULTINODE_GPUS_PER_NODE=$MULTINODE_GPUS_PER_NODE"
+fi
+if [[ "$MULTINODE_NODE_RANK" != "0" && "$MULTINODE_NODE_RANK" != "1" ]]; then
+    fail "MULTINODE_NODE_RANK must be 0 or 1, got '$MULTINODE_NODE_RANK'"
+fi
+if [[ "$PREFILL_TP" != "8" || "$PREFILL_PP_SIZE" != "2" ]]; then
+    fail "this entrypoint serves only TP8 x PP2, got TP$PREFILL_TP x PP$PREFILL_PP_SIZE"
+fi
+if [[ "$PREFILL_EP" != "1" ]]; then
+    fail "this entrypoint serves only EP1, got PREFILL_EP=$PREFILL_EP"
+fi
+if [[ "$PREFILL_DP_ATTN" != "false" ]]; then
+    fail "this entrypoint does not enable DP attention, got PREFILL_DP_ATTN=$PREFILL_DP_ATTN"
+fi
+if [[ "$PREFILL_NUM_WORKERS" != "1" || "$DECODE_NUM_WORKERS" != "0" ]]; then
+    fail "this entrypoint is aggregated: it needs 1 prefill worker and 0 decode workers, got ${PREFILL_NUM_WORKERS}P/${DECODE_NUM_WORKERS}D"
+fi
+
+read -r -a CONCURRENCIES <<< "$CONC_LIST"
+if [[ "${#CONCURRENCIES[@]}" -ne 1 ]]; then
+    fail "one concurrency per allocation is required, got CONC_LIST='$CONC_LIST'"
+fi
+case "${CONCURRENCIES[0]}" in
+    1|2|4|8) ;;
+    *) fail "concurrency must be 1, 2, 4, or 8, got '${CONCURRENCIES[0]}'" ;;
+esac
+
+if [[ -n "${AITER_SITUV2_A8W4+set}" ]]; then
+    if [[ "$AITER_SITUV2_A8W4" != "0" && "$AITER_SITUV2_A8W4" != "1" ]]; then
+        fail "AITER_SITUV2_A8W4 must be 0 or 1 when set, got '$AITER_SITUV2_A8W4'"
+    fi
+fi
+
+export VLLM_ROCM_USE_AITER=1
+export SAFETENSORS_FAST_GPU=1
+export AITER_BF16_FP8_MOE_BOUND=0
+export VLLM_USE_BREAKABLE_CUDAGRAPH=0
+export VLLM_USE_V2_MODEL_RUNNER=0
+export VLLM_ENGINE_READY_TIMEOUT_S="${VLLM_ENGINE_READY_TIMEOUT_S:-7200}"
+export PYTHONNOUSERSITE=1
+
+{ set +x; } 2>/dev/null
+VLLM_CMD=(
+    vllm serve "$MODEL_PATH"
+    --served-model-name "$MODEL"
+    --host 0.0.0.0
+    --port "$PORT"
+    --tensor-parallel-size 8
+    --pipeline-parallel-size 2
+    --nnodes 2
+    --node-rank "$MULTINODE_NODE_RANK"
+    --master-addr "$MULTINODE_MASTER_ADDR"
+    --trust-remote-code
+    --load-format auto
+    --moe-backend auto
+    --gpu-memory-utilization "${VLLM_GPU_MEMORY_UTILIZATION:-0.95}"
+    --max-model-len 1048576
+    --max-num-seqs "$CONC_LIST"
+    --max-num-batched-tokens "${VLLM_MAX_NUM_BATCHED_TOKENS:-4096}"
+    --mm-encoder-tp-mode data
+    --enable-auto-tool-choice
+    --tool-call-parser kimi_k3
+    --reasoning-parser kimi_k3
+    --language-model-only
+)
+if [[ "$MULTINODE_NODE_RANK" == "1" ]]; then
+    VLLM_CMD+=(--headless)
+fi
+
+echo "AITER_SITUV2_A8W4=${AITER_SITUV2_A8W4-unset}"
+printf 'vLLM command:'
+printf ' %q' "${VLLM_CMD[@]}"
+printf '\n'
+
+if [[ "${KIMIK3_VLLM_DRY_RUN:-0}" == "1" ]]; then
+    echo "KIMIK3_VLLM_DRY_RUN=1 set; not starting the server"
+    exit 0
+fi
+
+exec "${VLLM_CMD[@]}"

benchmarks/multi_node/agentic_srt.sh

@@ -2,8 +2,8 @@
 set -euo pipefail
 set -x
 
-# Client-only agentic trace replay for srt-slurm multinode jobs.
-# srt-slurm owns server startup; this script runs as benchmark.type=custom
+# Client-only agentic trace replay for externally managed multi-node jobs.
+# The caller owns server startup; this script runs as benchmark.type=custom
 # against the already-ready frontend on the head node.
 
 INFMAX_CONTAINER_WORKSPACE="${INFMAX_CONTAINER_WORKSPACE:-/infmax-workspace}"

configs/amd-master.yaml

@@ -2244,6 +2244,36 @@ minimaxm3-fp4-mi355x-vllm-agentic:
       search-space:
       - { tp: 4, kv-offloading: dram, kv-offload-backend: { name: vllm-simple }, conc-list: [1, 4, 8, 16] }
 
+kimik3-fp4-mi300x-vllm-agentic:
+  image: vllm/vllm-openai-rocm:kimi-k3
+  model: moonshotai/Kimi-K3
+  model-prefix: kimik3
+  runner: cluster:mi300x-amds
+  precision: fp4
+  framework: vllm
+  multinode: true
+  disagg: false
+  scenarios:
+    agentic-coding:
+    - search-space:
+      - spec-decoding: none
+        kv-offloading: none
+        conc-list: [1, 2, 4, 8]
+        prefill:
+          num-worker: 1
+          tp: 8
+          pp: 2
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "NATIVE_MULTINODE=1"
+        decode:
+          num-worker: 0
+          tp: 8
+          pp: 2
+          ep: 1
+          dp-attn: false
+
 dsv4-fp4-mi355x-sglang-disagg-agentic-hicache:
   image: lmsysorg/sglang-rocm:v0.5.14-rocm720-mi35x-20260710
   model: deepseek-ai/DeepSeek-V4-Pro

perf-changelog.yaml

@@ -5339,3 +5339,13 @@
     - "Pass --use-chat-template to the benchmark, as required for EAGLE-style speculative decoding"
     - "Validated on the node through the real launcher, every request successful: TEP4 concurrency 64 640/640 at 6895 tok/s (11% above the non-MTP arm's 6198 tok/s, mean TTFT 27.2s to 9.6s), and TEP4 concurrency 1 10/10 at 1233 tok/s (83% above the non-MTP arm's 672 tok/s, mean TPOT 12.6ms to 6.5ms)"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2312
+
+- config-keys:
+    - kimik3-fp4-mi300x-vllm-agentic
+  description:
+    - "Add an opt-in native Slurm path for aggregated Kimi K3 AgentX on two 8xMI300X nodes with TP8 x PP2, EP1, and concurrency 1/2/4/8"
+    - "Image: vllm/vllm-openai-rocm:kimi-k3; model moonshotai/Kimi-K3"
+    - "Fail closed on the exact aggregate topology, complete and revision-matched node-local model snapshots, eight gfx942 GPUs per node, and independently validated node-local Enroot squash images"
+    - "Keep AITER_SITUV2_A8W4 caller-configurable pending exact-shape gfx942 validation, and never download the approximately 1.5 TB target checkpoint inside a benchmark job"
+    - "Preserve the existing single-node MI300X launcher unless NATIVE_MULTINODE=1, with bounded host-owned artifact handoff and signal/failure cleanup"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2401