From 396e97e5dd0766cfb6a593cb71c6c0860838dbf4 Mon Sep 17 00:00:00 2001 From: Zhaoyi Li <36555117+Lzy17@users.noreply.github.com> Date: Sun, 18 Oct 2026 02:28:34 -0500 Subject: [PATCH] [CI] [AMD] ci/slurm: Add a Mooncake KV-transfer test to the MI355X disaggregation nightly (#44409) --- scripts/ci/slurm/launch_mi355x.sh | 12 +++- scripts/ci/slurm/nightly-configs.yaml | 16 ++++++ .../dsv4flash/1k1k/1p1d-mooncake.yaml | 57 +++++++++++++++++++ 3 files changed, 84 insertions(+), 1 deletion(-) create mode 100644 scripts/ci/slurm/recipes/mi355x-fp4/dsv4flash/1k1k/1p1d-mooncake.yaml diff --git a/scripts/ci/slurm/launch_mi355x.sh b/scripts/ci/slurm/launch_mi355x.sh index 33588639dd..c16c36a53d 100755 --- a/scripts/ci/slurm/launch_mi355x.sh +++ b/scripts/ci/slurm/launch_mi355x.sh @@ -194,6 +194,7 @@ emit("DIST_SOCK", rt.get("dist_socket_ifname", "")) # hardcoded value) so EP<=8 recipes stay byte-identical; wide-EP spur recipes # set mooncake, which is the validated cross-node KV path there. emit("XFER", rt.get("kv_transfer_backend", "mori")) +emit("MC_DISABLE_HIP", rt.get("mc_disable_hip", 0)) # SGLANG_USE_ROCM700A toggles the ROCm-7.0.0-alpha codepath. Default 1 (the # pre-wide hardcoded value) keeps EP<=8 recipes byte-identical; the validated # wide-EP run on the rocm720 0715 image needs 0, set via runtime.rocm700a. @@ -292,7 +293,7 @@ if (( (PN_PER > 1 && PW > 1) || (DN_PER > 1 && DW > 1) )); then echo "ERROR: a wide engine (nodes/engine>1) cannot have >1 worker of that role (got PW=$PW PN_PER=$PN_PER DW=$DW DN_PER=$DN_PER)" >&2 exit 1 fi -echo "recipe: image=$IMAGE attn=${ATTN:-$PATTN/$DATTN} ib=$IB ptp=$PTP dtp=$DTP pn_per=$PN_PER dn_per=$DN_PER a2a=${A2A:-} concs=$CONCS isl=$ISL osl=$OSL" +echo "recipe: xfer=$XFER image=$IMAGE attn=${ATTN:-$PATTN/$DATTN} ib=$IB ptp=$PTP dtp=$DTP pn_per=$PN_PER dn_per=$DN_PER a2a=${A2A:-} concs=$CONCS isl=$ISL osl=$OSL" # --------------------------------------------------------------------------- # Shared NFS scratch (visible to login node + compute nodes). Raw bench output @@ -476,6 +477,15 @@ if (( PN_PER > 1 || DN_PER > 1 )); then fi fi +# Mooncake KV transfer: it moves KV over its own RDMA engine, so the mori env above +# is irrelevant to it. MC_GID_INDEX=1 picks the routable RoCEv2 IPv4 GID on the +# routed L3 fabric; MC_DISABLE_HIP (when the recipe sets runtime.mc_disable_hip=1) +# keeps cross-node requests off the intra-node-only hip transport. +if [[ "$XFER" == "mooncake" ]]; then + MORI_ENV="$MORI_ENV -e MC_GID_INDEX=1" + [[ "$MC_DISABLE_HIP" == "1" ]] && MORI_ENV="$MORI_ENV -e MC_DISABLE_HIP=1" +fi + # Model-specific docker `-e` env + sglang server args from the recipe's optional # `model:` block, written as bash arrays to model_flags.sh (sourced by # prefill.sh/decode.sh). DSV4 recipes have no `model:` block -> empty arrays, so diff --git a/scripts/ci/slurm/nightly-configs.yaml b/scripts/ci/slurm/nightly-configs.yaml index 5149c42542..d8c1e4e59e 100644 --- a/scripts/ci/slurm/nightly-configs.yaml +++ b/scripts/ci/slurm/nightly-configs.yaml @@ -636,3 +636,19 @@ glm52-fp4-mi355x-ep16-mtp-sglang: search-space: - conc-list: [1, 8, 16, 32, 64, 128, 256] config_file: scripts/ci/slurm/recipes/mi355x-fp4/glm52/1k1k/2p1d-ep16-mtp.yaml + +dsv4flash-fp4-mi355x-mooncake-sglang: + model: deepseek-ai/DeepSeek-V4-Flash + model-prefix: dsv4flash + model_path: /it-share/model_coverage/models--deepseek-ai--DeepSeek-V4-Flash + runner: mi355x + precision: fp4 + framework: sglang + multinode: true + disagg: true + seq-len-configs: + - isl: 1024 + osl: 1024 + search-space: + - conc-list: [1, 8, 16, 32, 64, 128, 256] + config_file: scripts/ci/slurm/recipes/mi355x-fp4/dsv4flash/1k1k/1p1d-mooncake.yaml diff --git a/scripts/ci/slurm/recipes/mi355x-fp4/dsv4flash/1k1k/1p1d-mooncake.yaml b/scripts/ci/slurm/recipes/mi355x-fp4/dsv4flash/1k1k/1p1d-mooncake.yaml new file mode 100644 index 0000000000..cb1fa8942d --- /dev/null +++ b/scripts/ci/slurm/recipes/mi355x-fp4/dsv4flash/1k1k/1p1d-mooncake.yaml @@ -0,0 +1,57 @@ +# MI355X DeepSeek-V4-Flash (FP4) 2-node 1P1D disaggregation recipe, Mooncake KV transfer. +# Same as 1p1d.yaml except kv_transfer_backend. Needs an image with Mooncake >= v0.3.14-rc1. +# +# FP4 enables SGLANG_DSV4_FP4_EXPERTS in launch_mi355x.sh (driven by PRECISION). +# +# Consumed by: +# * scripts/ci/slurm/process_result.py reads `resources` and +# `backend.sglang_config` (TP/EP/DP + worker counts) for the summary table. +# * scripts/ci/slurm/launch_mi355x.sh reads `runtime` and `bench`. + +resources: + prefill_workers: 1 + decode_workers: 1 + +backend: + sglang_config: + prefill: + tensor-parallel-size: 8 + expert-parallel-size: 1 + data-parallel-size: 1 + decode: + tensor-parallel-size: 8 + expert-parallel-size: 1 + data-parallel-size: 1 + +runtime: + image: lmsysorg/sglang-rocm:v0.5.13.post1-rocm720-mi35x-20260623 + attention_backend: dsv4 + kv_transfer_backend: mooncake + # RoCE HCAs used for cross-node KV transfer. + ib_devices: rdma0,rdma1,rdma2,rdma3 + prefill_port: 30025 + decode_port: 30026 + prefill_bootstrap_port: 8998 + decode_bootstrap_port: 9001 + lb_port: 8000 + mem_fraction_static: 0.90 + page_size: 256 + max_running_requests: 256 + max_total_tokens: 8551168 + chunked_prefill_size: 8192 + swa_full_tokens_ratio: 0.1 + +bench: + # bench_serving --max-concurrency sweep; one result JSON per concurrency. + concurrencies: [1, 8, 16, 32, 64, 128, 256] + num_prompts_factor: 4 # num-prompts = concurrency * factor + random_range_ratio: 1.0 + + # Correctness gate run through the PD path before the perf sweep + # (full GSM8K, 8-shot, accuracy > 0.91). A regression here fails the nightly + # even when throughput looks fine ("fast but wrong"). + accuracy: + enabled: true + num_shots: 8 + num_questions: 1319 # full GSM8K test set + threshold: 0.91