Skip to content

Sglang render - #312

Open
vdar25 wants to merge 228 commits into
ROCm:developfrom
vdar25:sglang-render
Open

vdar25 wants to merge 228 commits into
ROCm:developfrom
vdar25:sglang-render

Conversation

@vdar25

@vdar25 vdar25 commented Nov 17, 2026

Copy link
Copy Markdown
Contributor

Stacked on #300 (di-refactor): this branch merges its current head (a967711). The changes below are only what this PR adds on top.

Summary

The SGLang disaggregated recipe's server and router commands now come from data: models.yaml holds every model's flags, and render.py turns them into each node's command lines. Adding a model, changing a flag, or moving to a newer SGLang no longer means editing shell. The PR also carries two fixes found bringing the recipe up on Broadcom Thor2 NICs (MI350X), a site setting for Docker daemons that refuse --privileged, and the same fabric portability for the KV cache transfer and large-EP benchmarks.

Changes

Recipes as data (scripts/sglang_disagg/render.py, models.yaml)

  • models.yaml is layered: defaults and per-model entries (base -> modes.<tp|dp_ep> -> roles.<prefill|decode>, with extends), then backends (KV transfer), overlays (agentic radix cache and metrics) and compat.
  • render.py shell --rank N gives a node its role, server argv and router argv. The role script runs those as arrays; its three hand-built command strings, the inline YAML reader and the parallel-size math are gone.
  • render.py validate replaces the hard-coded model lists (VALID_MODELS, MORI_EP_VALID_MODELS, and MORI_DP_MODE1_ALLOWED_MODELS, which the role script duplicated). A model runs DP_MODE=1 if its entry has a modes.dp_ep block.
  • render.py plan previews every node's commands without GPUs; render.py check reports any flag an image's SGLang does not accept.

SGLang version compat (replaces sglang_compat.sh)

  • compat in models.yaml lists renamed and removed flags. render.py applies them only when the installed SGLang's --help no longer accepts the name in the recipe. For SGLang 0.5.20: --cuda-graph-bs -> --cuda-graph-bs-decode, --prefill-round-robin-balance dropped. On 0.5.12 the commands are unchanged.

Broadcom Thor2 under the in-kernel bnxt_re driver (MI350X)

  • The gfx950 image builds rdma-core v61.0 over the base's rdma-core 39 + Broadcom bnxt-rocelib (new rdma_core pin component). The base stack rejects the NICs (kernel ABI 1) or cannot create QPs; v54.0 is the first release that can. Hosts on Broadcom's out-of-tree driver need the base stack instead; that trade-off is open with the image's owners.
  • RDMA preflight: before the container barrier, each node runs a loopback ibv_rc_pingpong on every IB_DEVICES NIC. A NIC that cannot create a working QP fails the job in seconds with the device and the reason, instead of hanging the first KV transfer behind --watchdog-timeout. RDMA_PREFLIGHT=0 skips it.

Site setting

  • CONTAINER_PRIVILEGE_ARGS (cluster.sh, default --privileged) replaces the hard-coded --privileged in node.sh. Where Docker refuses privileged containers: --cap-add IPC_LOCK --ulimit memlock=-1:-1.

Benchmarks on any fabric (the KV cache transfer and large-EP benchmarks hit the same multi-homed Thor2 nodes)

  • kvcache_transfer_bench, on Integrating KV Cache Transfer Benchmark with madengine #269's launcher: NIC, GID and interfaces from cluster.sh (Thor2 and AINIC detected; CX7 unchanged), node addresses on FABRIC_SUBNET_PREFIX, roles matched on short hostnames, MoRI found in the SGLang and large-EP images too, result directories writable on root-squashed storage, and sbatch from the MAD root. The cards no longer pin IBDEVICE=mlx5_0, so the fabric picks the NIC.
  • large-ep-benchmark: the same fabric and addressing; the models share is mounted only where it exists; perf_large_ep_bench.csv is published so its cards report through madengine; cleanup stops only the benchmark's own container.

Smaller

  • common/topology/specs/sglang_disagg_2p2d_tp8.json.
  • Snapshot harness stubs whoami, so the goldens do not depend on who runs them.
  • Docs: docs/sglang-disagg.md (recipes, compat, RDMA_PREFLIGHT), configuration.md, adding-a-model.md, the recipe README.

Compatibility

Snapshot goldens (18 SGLang cards) were re-recorded. Compared with #300's goldens: every card's docker run (flags, mounts, environment) is identical, and every server and router command has the same flags with the same effective values, in a different order. The one textual difference beyond order: DeepSeek prefill passed --mem-fraction-static twice (0.73, then 0.65, the value that took effect) and now passes it once.

Testing

Offline (on the merged head)

  • snapshot.py --catalog sglang_disagg: 18 cards, 0 differ
  • scripts/sglang_disagg/tests: test_render.py 20 tests OK; prr_flag_assert 13, parse_to_csv_assert 19, fault_capture_assert 6, mori_env_assert 6, forward_env_assert 4 passed
  • scripts/common/tests: dispatch 27, launcher_verdict 35, log_names 25, config_layers 27, node_lib 16, launcher_lib 15, cluster_fabric 7 passed, and topology, config and pin-set unit tests OK

GPU: DeepSeek-V3, 1 prefill + 1 decode node, DP_MODE=1 (MoRI EP, DP attention), 1024/1024 sweep
MI300X (OCI, CX7 RoCE), image built from this branch's pin set (SGLang 0.5.12.post1, MoRI 158c7e8).

Path Job Result
Launcher run directly (run_xPyD_models.slurm in the allocation, no madengine) 567099 pending
madengine (madengine run of pyt_sglang_disagg_mori_dp_deepseek-v3, slurm_multi) 567100 pending

Total token throughput (tok/s) by concurrency, against the same card from develop CI on this cluster (image ci-sglang_disagg_..._sglang_disagg_inference, SGLang 0.5.12.post1):

Concurrency develop CI 441863 develop CI 447278 this PR, direct this PR, madengine
8 309.8 311.9
16 585.8 585.7
32 1058.5 1062.4
64 1855.0 1872.8
128 3335.8 3451.0
256 5980.4 6092.3
512 9854.9 10063.2

GPU: Broadcom Thor2, MI350X

  • TODO: result with rdma-core v61.0. Before it: the base gfx950 image's provider rejected every NIC, and rdma-core 39's provider opened them but could not create QPs, so a run hung for 15 minutes.

Cemberk and others added 30 commits August 13, 2026 17:17
… selector

The xPyD harness already models pools that span multiple nodes (xP/yD count
NODES, roles are MASTER/CHILD via --data-parallel-start-rank + --headless), but
the wideEP path hardcoded TP=1: dp_size was xP*GPUS_PER_NODE and the connector
emitted a literal `-tp 1`. That makes any model whose REPLICATED (non-expert)
weights exceed one GPU inexpressible - e.g. Kimi-K3 on MI300X, where TP1/DP16
needs 190.7 GiB/GPU (> 192 GB HBM) and TP2 is required to fit.

- vllm_disagg.sh: topology math is TP-aware. dp_per_node = GPUS_PER_NODE/TP_SIZE
  feeds dp_size, dp_size_local and both start-rank calculations. TP_SIZE defaults
  to 1 and is validated to divide GPUS_PER_NODE.
- moriio.sh: emit -tp ${TP_SIZE}; clamp --api-server-count to dp_size (it must be
  <= data-parallel-size or the frontend DP balancer routes to ranks that do not
  exist - only reachable once TP>1 shrinks dp_size below GPUS_PER_NODE).
- moriio.sh: advertise the peer pool's node IPs as moriio_pod_hosts when a pool
  spans >1 node. Without it the connector falls back to the peer MASTER only and
  KV writes aimed at ranks on a peer CHILD node silently miss, so those ranks
  decode with no context. Emitted only when xP>1 || yD>1.
- run_xPyD_models.slurm: forward TP_SIZE/GPUS_PER_NODE, wire BENCHMARK_SCRIPT=niah
  to the already-present benchmark_niah.sh, forward NIAH_* knobs.

Default-neutral: at xP=1 yD=1 TP_SIZE=1 - the topology every registered entry
uses - DRY_RUN output is byte-identical to develop across all four node roles.
At xP=2 yD=2 TP_SIZE=2 it emits tp=2, dp_size=8, dp_local=4, child start-rank 4
(= EP16 per pool over 16 GPUs).

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…Kimi-K3 recipe

Two parts.

1) moriio.sh wideEP dropped ${model_args[@]} on the floor. models.yaml documents
   per-model dp: tuning as supported ("both connectors now append it") and rixl's
   deepep path does append it, but the moriio wideEP branch parsed the flags,
   logged them, and then emitted an argv without them. Verified with a sentinel
   flag: present in the log line, absent from the command. Now appended, matching
   rixl. Every existing wideEP model has an empty dp: block, so emitted argv is
   unchanged for all of them.

2) Kimi-K3 recipe, entirely as data: env: block for the gfx942 knobs
   (VLLM_ROCM_USE_AITER_MLA=0 - the AITER MLA kernel is gfx950-only -
   AITER_SITUV2_A8W4=1 for the packed-int4 SiTUv2 MoE path, KDA conv state layout,
   40e9 KV cache for >600K contexts, per-role cudagraph + MoRI backends), and a
   dp: block for the serve flags the connector does not emit (reasoning parser,
   1M max-model-len, batched-tokens pinned at 2048, MoE quantization-config).
   Registered in MORI_EP_VALID_MODELS and WIDE_EP_ONLY_MODELS.

   The JSON quantization-config carries its own shell quotes: the launcher
   word-splits dp: via eval, which would otherwise strip the JSON's double quotes
   and hand vLLM invalid JSON. Verified the emitted value parses as JSON.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
docker/pyt_vllm_kimi_k3_mi300x.ubuntu.amd.Dockerfile
  The shared vllm_disagg_inference stack with the pins K3 needs on gfx942
  (MoRI 1.2.2, AITER 0.1.19 + flydsl 0.2.4, the K3+MoRIIO vLLM commit) plus the
  K3-aware AITER graft from the public vendor image - without that graft the K3
  MoE profiling shape finds no tuned FlyDSL config, falls back to a heuristic
  kernel and aborts LLVM inside determine_available_memory.

  Differences from the recipe this is ported from:
  - every source pinned to an immutable SHA, not a fork branch name (branches on
    personal forks can be force-pushed; MAD needs the image rebuildable later)
  - GH_TOKEN build-arg dropped: all three repos are public, and a token passed
    this way is recorded in image metadata
  - the flydsl 0.2.4 re-pin happens at build time, so the launcher no longer runs
    pip install inside every container at serve time
  - WITH_NIXL defaults to 0 (K3 uses the moriio connector only)
  - no runtime patchers: all connector/KDA fixes are committed in the pinned vLLM

scripts/vllm_multinode/
  Colocated (single-instance) multi-node serving - the counterpart to
  vllm_dissag, for models that do not fit one node but want lowest single-request
  latency rather than disagg throughput. TP within a node, PP across nodes.

  It owns only node discovery, the container launch and the head/worker split;
  models.yaml, socket_barrier.py, benchmark_xPyD.sh, benchmark_niah.sh and
  parse_to_csv.py are reused from vllm_dissag (scripts/ is mounted whole), so the
  model recipe and the CSV pipeline have one home.

  The three K3 colocated variants collapse to this one script: dry-run confirms
  pp2xtp8 / allgather / moriep argv differ only by --enable-expert-parallel and
  --all2all-backend, so they become three models.json entries rather than three
  near-identical run.sh copies.

Not built here - this environment has no docker and one 8-GPU node; the harness
is verified by DRY_RUN argv inspection only.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
models.json: four entries, all skip_gpu_arch gfx950 (the existing four Kimi-K3
entries are gfx950/TP8 single-node and all carry skip_gpu_arch gfx942, so K3 is
currently skipped entirely on MI300X - this is that gap):

  pyt_vllm_disagg_mori_kimi-k3               -N 4  xP2 yD2 TP2 -> EP16/pool
  pyt_vllm_kimi-k3_mi300x_pp2xtp8            -N 2  TP8 x PP2, no EP
  pyt_vllm_kimi-k3_mi300x_wideep_allgather   -N 2  + EP, allgather_reducescatter
  pyt_vllm_kimi-k3_mi300x_wideep_moriep      -N 2  + EP, mori_low_latency

The three colocated entries differ only in ENABLE_EP / ALL2ALL_BACKEND /
AITER_SITUV2_A8W4, which is why they share one launcher.

Result reporting: BENCHMARK_SCRIPT=niah previously produced logs and nothing
else, so a model declaring multiple_results would have reported no metric.
- parse_to_csv.py gains --niah: parses benchmark_niah.py output and writes the
  same 29-column madengine perf.csv as the throughput path (verified identical
  column sets). One row per context size, performance = needles found /10. A
  size that errored is written as a FAILURE row with performance 0 rather than
  dropped, so a pass->crash regression is visible instead of silent.
- benchmark_niah.sh calls it.
- Both slurm launchers copy /run_logs/$SLURM_JOB_ID/perf.csv to
  ./perf_$MODEL_NAME.csv on completion: madengine resolves multiple_results
  against the job directory it launched from, not the container log mount.
  No-op for entries that do not declare multiple_results.

Verified: madengine discover goes 133 -> 137 with exactly these four added and
none removed; every referenced dockerfile and script exists; -N matches each
topology; the int4 quantization-config survives shell word-splitting as valid
JSON; DeepSeek moriio and rixl/deepep DRY_RUN output is byte-identical to
develop.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Ports the writeups from PR ROCm#193 next to the existing gfx950 K3 doc, keyed to the
madengine tags rather than to shell scripts: topology rationale (why MI300X needs
multi-node and why the disagg pool needs TP2), the gfx942 knobs and where each
one lives now, the NIAH results, and the three root-cause fixes.

Corrections carried in:
- the NIAH ceiling is 900K throughout. The upstream READMEs still said 300K in
  seven places after two later commits raised it to 500K and then 900K.
- states that the NIAH harness sizes context in WORDS (~1.3x tokens) while the
  results table is in tokens, so the two columns are not comparable.

benchmark/kimi_k3/README.md gains a pointer: all four entries there carry
skip_gpu_arch gfx942, which is exactly the gap the MI300X page fills.

Every tag, node count and knob value in the new page is checked against
models.json and models.yaml.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…metadata

The four MI300X Kimi-K3 entries could not run correctly on a cluster. Three
separate defects, all from the colocated launcher inheriting conventions that
only hold for the disaggregated one.

Node sizing. The entries carried `"args": "-N 4 -n 4"` on the assumption that
madengine forwards args to sbatch. It does not — args are appended to
`bash <model>.slurm`, and neither launcher parses $@, so the flags were dropped
and the job was submitted with the `slurm.nodes` default of 1. Sizing now uses
`distributed.nnodes`, which madengine reconciles into `#SBATCH --nodes`, and the
inert args are removed. `slurm.nodes` is deliberately left unset: setting it
selects the multi-node preset, whose `NCCL_SOCKET_IFNAME=eth0` is forwarded into
the container by run_xPyD_models.slurm and would override the fabric interface
on an RDMA cluster. `slurm.time` is set explicitly instead, since the
single-node preset's 12 h default is short for a full NIAH sweep.

NIAH model tag. benchmark_niah.sh requested MODEL_PATH as the model name,
correct for the disagg path where vLLM defaults served_model_name to the serve
argument. serve_colocated.sh passes `--served-model-name "$MODEL_NAME"`, so
every request 404'd and all three colocated entries — which default to
BENCHMARK_SCRIPT=niah — would have recorded a full sweep of FAILURE rows. The
harness now honors NIAH_MODEL when a launcher sets it, and the colocated
launcher sets it to the tag it actually serves.

Colocated run metadata. parse_to_csv derived nnodes, n_gpus, deployment_type and
tags from xP/yD. serve_colocated.sh exports xP=1 yD=0 only so the shared log
filenames stay unique, so a 2-node 16-GPU colocated run reported itself as a
1-node 8-GPU `disagg_1P0D` with a nixl backend it never used. Node count now
comes from NNODES (exported by both launchers), and a launcher whose shape is
not "xP prefill + yD decode" states its own identity via PERF_DEPLOYMENT_TYPE /
PERF_TAGS. The disagg path is byte-identical to before.

Also corrects the README, which documented the args-to-sbatch behavior that does
not exist and a quick start that cannot work now that the "<supply-your-image>"
placeholder is rejected rather than pulled.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
NUM_NODES is derived from the requested topology (xP+yD), not from the
allocation, and the node list was then truncated with `head -n $NUM_NODES`. A
short allocation therefore produced a silently wrong topology: pools were built
from nodes that were never allocated, and the run failed much later as a
connector handshake timeout with no indication of the real cause.

This mattered little while these models were always launched from a matching
`salloc`, but the sbatch path now sizes the allocation from the model card, so a
mismatch between card and cluster is a reachable state. Check up front and name
the three ways to fix it.

The colocated launcher already fails on the equivalent mismatch via its
TP*PP == NNODES*GPUS check, so it needs no counterpart.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…ntract

parse_to_csv.py hand-wrote a full 29-column perf.csv, assembling node counts, GPU
counts, launcher, image and tags from the environment. Most of that is not the
workload's to know: madengine already owns it, and the guesses were wrong for the
colocated launcher, which sets xP=1 yD=0 only to keep log filenames unique.

The NIAH writer now emits a narrow CSV — model, performance, metric, status, plus
descriptive columns — and madengine merges in the run metadata via the entries'
`multiple_results` declaration. The columns that genuinely describe the workload's
own configuration rather than its placement (tp, pp, ep_backend, prefill_decode)
move to descriptive columns, matching how scripts/vllm/run_vllm.py already reports
tp/dtype/bs on the templated path. Nothing is lost; it is relocated to whichever
side actually knows it.

`status` stays explicit: an errored context size scores 0, and deriving status from
performance would file that real failure as a SUCCESS.

Scoped to leave every other workload alone. The NIAH writer is reachable only by the
four Kimi entries, which all declare `multiple_results`. The throughput writer keeps
its full-schema output by default, because the twelve disagg cards that use it
declare no `multiple_results` and madengine reads their CSV directly with no
metadata to merge — a narrow CSV there would drop every descriptive column. Their
output is byte-identical; `--narrow` is the opt-in for migrating one.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…uted.nnodes

madengine reads only slurm.nodes when emitting `#SBATCH --nodes`
(deployment/slurm.py: `self.nodes = slurm_config.get("nodes", 1)`), and
nothing maps distributed.nnodes across — build_orchestrator copies nnodes
into deployment_config.distributed for launcher detection only. The four
MI300X model cards carried nnodes but no slurm.nodes, so every one of them
would have been submitted as a 1-node job. The launcher's
TP*PP == NNODES*GPUS_PER_NODE assertion catches it, but only after the
allocation is granted.

Add nodes/gpus_per_node to the four slurm blocks, and ship the site configs
that carry the values a model card cannot:

  - results_dir, absent from the key list madengine copies out of models.json,
    and required because the slurm_multi collector globs results_dir for
    perf*.csv rather than resolving multiple_results
  - MODEL_DIR / LOG_PATH, which otherwise default to /shared_inference

*.json is gitignored repo-wide, so the templates need explicit negations or
they never reach a clone.

Correct the two README claims that did not match madengine's behavior
(nnodes-sizes-the-allocation, multiple_results-is-resolved-first), document
that --account/--qos are unwired for slurm_multi and need SBATCH_ACCOUNT,
and drop --keep-model-dir from the SLURM examples — it is a local-Docker
flag that madengine ignores with a warning on SLURM.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Four defects, each found by running the recipes end to end on a real
MI300X SLURM cluster for the first time. None of them could have survived
a single successful run, which is consistent with these cards never having
been executed through madengine.

1. Build was not reproducible. Every source is pinned to a commit SHA, but
   pip resolves build dependencies in an isolated environment from PyPI at
   build time, so nothing pinned the toolchain. setuptools >= 80 added
       assert isinstance(self.compiler, CCompiler)
   to distutils' build_ext.build_extension, which MoRI's legacy
   Cython.Distutils.build_ext path violates, so the amd_mori wheel failed
   with "AssertionError: run() must precede build_extension()" while every
   pinned SHA was still correct (job 223849). PIP_CONSTRAINT is the only
   mechanism that reaches inside pip's build isolation -- installing
   setuptools in the image does not. Set globally so the AITER, vLLM and
   router stages cannot regress the same way. Build then completed in
   19 minutes, all 18 stages (job 223872).

2. RDMA library mounts used -e, which is true for directories. Docker
   creates a DIRECTORY at a bind-mount source that does not exist, so the
   first run on a node lacking e.g. libionic.so.1 leaves an empty directory
   behind, and every later run on that node mounts a directory over a file
   inside the image:
       OCI runtime create failed: ... not a directory
   Self-propagating and per-node, so it presents as intermittent. The
   surviving node then waits at socket_barrier.py forever for a peer that
   already died -- 4h10m of "Waiting for nodes. . ." before the wall clock
   (job 223909). Use -f so only regular files are mounted.

3. pyt_vllm_kimi-k3_mi300x_pp2xtp8 omitted the gfx942 MoE requirement that
   the two wideep variants carry. gfx942 has no scaled-MXFP4 MFMA and the
   a16w4 SiTUv2 heuristic FlyDSL kernel cannot codegen there, so the MoE
   must be requantized to packed int4. Without it the worker aborts inside
   determine_available_memory with
       LLVM ERROR: Do not know how to expand this operator's operand!
   and quantization_config=None in the engine config (job 224132). This is
   a hardware requirement, not an EP-specific tuning, so all three
   colocated cards now carry AITER_SITUV2_A8W4=1 and the int4 config.

4. A run that produced no results exited 0. The launchers warned about a
   missing perf CSV and returned success anyway, so madengine recorded
   exit_code=0 in its completion marker and the collector reported
   "0 successful, 0 failed" -- a hard engine crash was indistinguishable
   from a clean run. For a benchmark repo that is the worst possible
   reporting outcome. Both launchers now exit 1.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Each of these cost hours to diagnose on a real cluster, and in every case
the symptom is a long way from the cause -- an intermittent multi-node
hang that is actually a poisoned bind-mount path, a wheel build that fails
while every pinned SHA is correct, an LLVM abort that is a missing
quantization flag. Record them so the next person reads instead of
rediscovers.

Also record the cluster facts a model card structurally cannot carry: the
SLURM association wall-time cap (which silently parks a job as PENDING
rather than failing), account requirements, docker on compute nodes, and
the fact that a registry-less cluster has no supported path to distribute
an image -- --build-on-compute requires --registry.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Adds Kimi-K3-MXFP4 (MI300X gfx942 MXFP4 MoE) to the vllm_dissag framework:
- models.yaml: K3 recipe env anchor + model entry (TP2×DP8, MoRI-EP)
- run_xPyD_models.slurm: allowlists + 2P/2D validation + JIT cache split
- moriio.sh: K3-gated topology (pod_hosts, api-server-count, headless KV)
- vllm_disagg.sh: pod hosts export + _model_config_to_array helper
- argv_assert.sh: K3 offline tests

Config-only integration; image build documented in README.

Co-Authored-By: Claude <noreply@anthropic.com>
…lsifiable

Both found on the first end-to-end MI300X run (job 224200), which served
Kimi-K3 and produced results -- but results that could not be trusted, from
a job that then would not exit.

Deadlock on teardown. serve_colocated.sh has the head kill only its own
server and exit, while every worker sits in `wait` on a headless vLLM that
nothing stops. srun waits on those tasks, so the job holds all nodes until
the wall clock: 4.5 hours of two idle exclusive MI300X nodes AFTER the
benchmark had already written perf.csv. The head now raises a sentinel on
the shared log volume and workers watch for it. The sentinel is raised from
an EXIT trap so a head that fails or times out cannot strand workers either.

NIAH could not distinguish truncation from retrieval failure. Scoring reads
content + reasoning_content, and K3 is a reasoning model served with
--reasoning-parser kimi_k3, so a trace that exhausts max_tokens is cut off
and only the earliest needles survive. At the old 2048 default, two of four
context sizes scored 1/10 and both listed exactly ANIMALS[0] -- the needle
placed first in the haystack. That signature is a truncated trace, not a
model failure, but finish_reason was never recorded so the two were
indistinguishable, and every row was written SUCCESS regardless.

  - record finish_reason and print it as finish=<reason>
  - raise the default max_tokens to 8192
  - keep the needle count as the metric, always reported, never dropped
  - mark a truncated row FAILURE and annotate the metric, so a measurement
    artifact cannot enter the results as if it were a model result

9/10 stays SUCCESS: the README documents ~9/10 at >=20K as a known RDMA
residual, so a strict 10/10 gate would flag known-acceptable behaviour.
Logs predating the finish= field still parse.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Job 239755 (Kimi-K3, 2x8 on oci-64) died in vLLM's startup barrier. Node 0
finished loading weights at 23:21:41 and entered in_the_same_node_as ->
torch.distributed.barrier; node 1 was still constructing MoE layers ~50 min
later. gloo gave up at 23:51:41, exactly 1800s in, and the master's TCPStore
died with it, so node 1 then failed with "Broken pipe" to :29500. The barrier
is all-or-nothing across all 16 ranks, so what has to stay small is the GAP
between nodes, not the absolute load time.

Five problems, addressed here:

1. --distributed-timeout-seconds does not reach the barrier that failed. It
   feeds ParallelConfig.distributed_timeout_seconds, which only configures the
   device/NCCL groups. The startup barrier runs on the gloo CPU group, built by
   GroupCoordinator from the separate cpu_distributed_timeout_seconds field.
   Unset, it falls back to PyTorch's stock 1800s -- which is exactly the
   deadline in the traceback, despite 7200 being passed. Pass
   --cpu-distributed-timeout-seconds too.

2. Cold AITER JIT compile on every rank, every run. The image points
   AITER_JIT_DIR/TRITON_CACHE_DIR/VLLM_CACHE_ROOT/COMGR_CACHE_DIR at
   /opt/vllm_cache, but this launcher mounted only /tmp/vllm_cache, a path
   nothing reads, inside --rm containers. Mount a host-persistent cache at
   /opt/vllm_cache keyed by image ID, as vllm_dissag/run_xPyD_models.slurm
   already does. Per-node cold compile is precisely the skew the barrier
   cannot absorb.

3. WORKER_PID was tee's PID. After `a | b &` the shell reports b, so every
   kill in this script hit tee and left vLLM running, and no liveness check
   was possible. Launch through process substitution instead.

4. A dead engine was indistinguishable from a slow one. The head polled the
   log for the full 4000s after its writer had already died, then reported a
   timeout rather than the traceback. Both wait loops now check liveness --
   via a helper that also rejects zombies, since an exited-but-unreaped child
   still answers kill -0. This also fixes a latent hang where a worker would
   spin forever on a dead engine.

5. Container env was a closed allowlist. Tuning any of the above from CI meant
   a new -e line and another MAD PR. COLOCATED_FORWARD_ENV names the variables
   to forward, and they are passed via docker --env-file so values containing
   spaces survive. The file is node-local, mode 600, and unlike the wrapper
   scripts is never archived as a build artifact.

Also adds an opt-out checkpoint page-cache pre-warm before the container
barrier (PREWARM_CHECKPOINT=0 to skip). Its per-node duration is the real
diagnostic: a large spread means the storage path is the problem and no
timeout will fix it.

Verified: bash -n clean on both files, and on the single-quoted srun body
extracted in isolation; env-file forwarding preserves spaces and skips unset
vars at mode 600; the liveness helper correctly reports an unreaped child as
dead; DRY_RUN=1 emits the new flag and honours an override, with
COLOCATED_EXTRA_ARGS still appended last.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Job 240624 spent its entire 7200s allocation in the pre-warm added by the
previous commit. Both nodes printed "[prewarm] reading ..." and neither ever
printed "[prewarm] done", so vllm serve was never launched -- the run timed
out having tested nothing.

The flaw: it reads the WHOLE checkpoint on EVERY node, while PP2xTP8 means a
node only ever loads its own shard. On a 1453 GiB checkpoint that is roughly
4x the necessary I/O against a single NFS export, with both nodes competing
for it. It manufactured exactly the contention it was meant to relieve.

Default it to off. It stays available, because the per-node duration is a
useful storage probe -- a large spread between nodes means the storage path is
the problem and no timeout will fix it -- but it is now opt-in via
PREWARM_CHECKPOINT=1 and bounded by PREWARM_TIMEOUT_SECONDS (default 900). On
expiry it warns that the filesystem could not deliver the checkpoint in the
window and proceeds, rather than burning the whole allocation.

Note this run did not exercise --cpu-distributed-timeout-seconds: vLLM never
reached the startup barrier, so the absence of the 1800000ms gloo timeout in
the log says nothing either way. That fix is still untested.

Verified: default unset is silent; PREWARM_CHECKPOINT=1 runs and reports its
duration; a stalled read hits the budget, warns, and exits 0 so startup
continues.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
mori_ep_env.sh:72-73 sets

    export NCCL_DEBUG="${NCCL_DEBUG:-INFO}"
    export NCCL_DEBUG_SUBSYS="${NCCL_DEBUG_SUBSYS:-INIT,NET,GRAPH}"

The ${VAR:-default} idiom advertises that a caller can override these. No
caller can. run_xPyD_models.slurm hands the container a fixed 22-name -e
list, and NCCL_DEBUG is not on it, so whatever the submitter exports is
dropped at the container boundary and the default always wins. The override
contract in mori_ep_env.sh is dead code as far as anything outside the
container is concerned.

Two things follow from that, and neither is hypothetical -- both come from
job 246618, a Qwen3-32B 1P1D run that completed successfully:

1. Log volume. That run archived 19 MB of per-rank logs (decode_NODE1.log
   9.6M, prefill_NODE0.log 4.4M, proxy_NODE0.log 2.7M), overwhelmingly NCCL
   channel-tuning triples -- pre-adjustment / minNChannels / post-adjustment
   -- emitted by all 8 ranks for every collective size.

2. Corrupted lines. Those 8 ranks share one stdout with no line atomicity,
   so at INFO volume the writes interleave mid-token:

       useocp[2026-09-07 16:09:52 TP0] Prefill batch, ...
       justment based on threadThreshold
       80 nc:112

   Any grep-based marker extraction over these logs is unreliable, which
   matters because CI reads them to decide pass/fail.

Narrowing NCCL_DEBUG_SUBSYS is not the fix. This run already had the
narrowed value INIT,NET,GRAPH and still produced all of the above; the
channel-tuning lines are GRAPH. Only NCCL_DEBUG=WARN suppresses them.

Forward the three NCCL debug knobs conditionally, using the same
${VAR:+-e VAR=$VAR} idiom vllm_dissag/run_xPyD_models.slurm already uses for
AITER_JIT_DIR and VLLM_ROCM_USE_AITER_MLA. When the variable is unset the
expansion is empty and the docker run is byte-identical to before, so the
default INFO behaviour is unchanged for anyone who does not opt out. Add the
same treatment for any further knob by adding one line.

The submitter's environment reaches the compute node because sbatch defaults
to --export=ALL and srun propagates it, so an `export NCCL_DEBUG=WARN` before
sbatch is now enough.

Deliberately not changing the INFO default here. INFO is a poor default for a
benchmark harness that archives its logs, but that is a call for the people
who use this interactively; this commit only makes the documented override
work.
models.json forced USE_CX7_NICS=0 on every sglang_disagg entry, which sends
prefill->decode KV transfer over mlx5_1, the 100G management NIC, instead of
the 8x 400G CX7 rail NICs that mori_ep_env.sh defaults to. That is the
critical path of a disaggregated run, so the measured throughput does not
represent the hardware. Two runs on different node pairs (AMD-internal
247073 and 247580) both show --disaggregation-ib-device mlx5_1.

Flipped to 1 on the four models this evidence covers: llama-3.1-8b,
qwen3-32b, llama-3.3-70b-fp8 and llama-3.1-405b-fp8. The remaining twelve
entries keep 0 until someone benchmarks them.

The rail NICs require same-rail nodes, and the file header has been citing
select_same_rail_nodes() as the thing that enforces that -- a function which
was never written. Written now: it maps each allocated node to its Level=0
switch via scontrol show topology and aborts before any container starts if
they do not share one. A cross-rail pair hangs in NCCL ibv_modify_qp rather
than failing, and base_flags sets --watchdog-timeout 1000000, so without
this the job burns its whole allocation instead of failing in seconds.

The check is skipped entirely when USE_CX7_NICS != 1, degrades to a warning
when no topology is configured, and can be bypassed with REQUIRE_SAME_RAIL=0
-- topology.conf may describe the management network rather than the CX7
fabric, in which case the check is conservative rather than correct.

--switches=2 is deliberately unchanged: both observed allocations were
already single-leaf, so constraining placement is the one change here that
could move the numbers.
decode.tp asked for --cuda-graph-bs $(seq 1 512), so 512 CUDA graphs were
captured at startup while benchmark_xPyD.sh only ever sweeps
CON="8 16 32 64 128 256 512". Startup was 51-52% of wall clock across two
runs (426s of 810s, then 401s of 784s).

Set the list to those seven values on the three models that used the full
range. Llama-3.1-405B keeps $(seq 1 8): it is already small and raising it
has memory implications this evidence does not cover.
py-spy, flask and pyyaml were installed from PyPI on every node at the start
of every run, so the pinned image was mutated before each benchmark: two
runs of the same tag are not the same environment, resolution can pick a
different version than the one requested, and a PyPI outage fails the run.
Observed in AMD-internal 247073/247580, where pip resolved click 8.5.0 and
installed 8.4.1, and the force-reinstall surfaced an unrelated xgrammar /
apache-tvm-ffi conflict in the log on both nodes.

Baked into docker/sglang_disagg_inference.ubuntu.amd.Dockerfile. The
launcher scripts now install only what is genuinely missing, so a MAD-built
image does no work there while an image supplied through DOCKER_IMAGE_NAME
still self-heals. --ignore-installed --force-reinstall is dropped.
Brings the branch onto current develop (652bad5) so the sglang_disagg fixes
and the kimi-k3 work can ship in one PR. Eight files conflicted:

models.json
  develop's 61b5301 migrated the registry from the root file to per-directory
  models.json. Took develop's empty root and moved this branch's four entries to
  where they now belong: pyt_vllm_disagg_mori_kimi-k3 into
  scripts/vllm_dissag/models.json, and the three pyt_vllm_kimi-k3_mi300x_*
  entries into a new scripts/vllm_multinode/models.json. The 26 jax_maxtext and
  primus_train entries in the old root file were NOT migrated: develop generates
  those from scripts/{jax-maxtext,primus_train}/get_models_json.py.

scripts/sglang_disagg/run_xPyD_models.slurm
  Union of develop's 25 conditional -e forwards and this branch's three NCCL_DEBUG
  forwards (28 total). The CX7 flip that was applied to the old root models.json is
  re-applied to scripts/sglang_disagg/models.json here.

scripts/vllm_dissag/models.yaml
  Append/append. Kept both recipe blocks (develop's GLM-5.1-FP8, this branch's Kimi-K3).

scripts/vllm_dissag/connectors/moriio.sh
  Took develop's argument order (model_args last) from a33b622/b33923b, so recipe
  flags win over connector defaults. Compatible with the Kimi-K3 recipe.

scripts/vllm_dissag/benchmark_niah.py and .sh
  The two sides added orthogonal features: develop added multi-seed sweeps with
  mean/min/max plus a warmup pass, this branch added finish_reason so a truncated
  answer is distinguishable from a retrieval miss. Combined both: _request now
  returns (message, finish, error) and the cross-seed summary flags TRUNCATED.
  Kept develop's NIAH_MAXTOK=2048 default -- the truncation detector makes the
  smaller default safe, and NIAH_MAXTOK is -e forwarded for models that need more.

scripts/vllm_dissag/parse_to_csv.py
  The auto-merge left two parse_niah_log and two save_niah_perf_csv definitions;
  the later pair (this branch's found= parser) silently shadowed develop's. Since
  the merged summary emits develop's mean=/min=/max= format, dropped this branch's
  pair and kept develop's parser. Kept the --narrow flag.

scripts/vllm_dissag/run_xPyD_models.slurm
  Six hunks, all unions: MORI_EP_VALID_MODELS and WIDE_EP_ONLY_MODELS gain both
  GLM-5.1-FP8 and Kimi-K3; the BENCHMARK_SCRIPT case gains both agentic and niah;
  the -e list gains develop's NIAH_SEEDS/NIAH_WARMUP/SHAPE_WARMUP/
  USE_INDUCTOR_GRAPH_PARTITION alongside this branch's NIAH_* forwards.

Verified: no conflict markers; bash -n on all eight touched shell/slurm files;
py_compile on both Python files; json.load on all 21 registries and yaml.safe_load
on both models.yaml; benchmark_niah.py run against a stub server on both the
scoring and truncation paths and its log fed through parse_to_csv.py --niah to a
valid perf.csv; and the sglang rail guard replayed against the real job-247580
topology across all six cases.
…d step

Loading Qwen3-32B off /shared_inference measured 683s in job 249926 -- 67% of a
1015s bring-up, ~92 MB/s per node with both nodes pulling the same 63 GB through
the same NFS mount. scripts/common/preflight_weights.sh:2-5 records what the other
path is: /mnt/m2m_nobackup/models_blog is per-node local NVMe, non-uniform, while
/shared_inference is the uniform NFS share.

An earlier commit on this branch reordered the sglang launcher to probe NFS first,
to avoid a wasted srun round trip and the "<jobid>.N bash FAILED 1:0" row a missed
probe leaves in sacct. That trade is backwards: it spends 683s of NFS reads to save
a few seconds of probing, and it means a node that does have the weights locally can
never use them. Restored local-NVMe-first.

The sacct noise that motivated the reorder is fixed properly instead. check_model_path
now runs a probe that always exits 0 and decides from its output, so a miss in a
fallback chain is not recorded as a step failure. Two latent bugs fall out of the
same rewrite -- the old version read srun's exit code alone, so an srun that exited 0
while reporting from fewer nodes than were allocated, or one that exited 0 with no
output at all, both returned "available on ALL nodes" and sent the job on to load
weights from a path some rank cannot see. The new version requires one Found line per
allocated node.

Applied to both scripts/sglang_disagg and scripts/vllm_dissag; only sglang needed the
ordering change, vllm_dissag was already local-NVMe-first.

Verified by extracting the function and replaying it against a stubbed srun across
five cases (all-found, one-node-miss, srun-fails, short output, silent exit-0). New:
0,1,1,1,1. Old, same stub: 0,-,1,0,0 -- it accepts short and silent.

Note this does not by itself make Qwen3-32B load fast: jobs 247073, 247580 and 249926
all fell through to NFS, so that model is not currently staged on the local NVMe of
the nodes CI draws from. This makes the fast path reachable and cheap to probe; the
staging itself is a cluster-side action.
The merge moved four entries out of the root models.json into the per-directory
registries develop's 61b5301 introduced, but kept their root-relative paths. All
twenty pre-existing per-directory registries resolve dockerfile as ../../docker/X
and scripts as a bare filename, so the migrated entries pointed at
scripts/vllm_multinode/scripts/vllm_multinode/run_multinode.slurm and would not
have resolved at all.

All 117 entries across the 21 registries now resolve from their own directory.
…t Kimi-K3 offline

The merge brought develop's multi-seed NIAH support into benchmark_niah.sh, which
now reads NIAH_SEEDS and NIAH_WARMUP alongside NIAH_WORDS/MAXTOK/TIMEOUT.
scripts/vllm_dissag forwards all of them; scripts/vllm_multinode forwarded only
three, so a colocated niah run silently fell back to the defaults for seeds and
warmup no matter what was set at submit time. Also forwards SHAPE_WARMUP, which
benchmark_xPyD.sh reads and which is reachable from the colocated path via
BENCHMARK_SCRIPT=sweep. Both launchers now forward every NIAH_* the benchmark reads.

Kimi-K3 is the model this branch adds and the only cell argv_assert.sh did not
cover -- it tested moriio+TP on Llama and moriio+wideEP on DeepSeek-V3. Added a
Kimi-K3 section over both roles: the int4_per_group_32 requant (without which
gfx942 cannot codegen the a16w4 SiTUv2 kernel and vLLM dies with "LLVM ERROR: Do
not know how to expand this operator's operand!" in determine_available_memory,
per benchmark/kimi_k3/mi300x/README.md:160), exactly one --compilation-config per
role, the kimi_k3 reasoning parser, the prefill/decode split across cudagraph_mode
(NONE vs PIECEWISE), all2all backend (mori_high_throughput vs mori_low_latency) and
kv_role, and the two model-gate arrays.

The gate assertions are scoped to the arrays themselves rather than grepped over
the whole file: "Kimi-K3" appears in both MORI_EP_VALID_MODELS and
WIDE_EP_ONLY_MODELS, so a file-wide grep passed even with the model removed from
one of them. MORI_EP_VALID_MODELS is line-continued, so it is extracted as a block.

49 passed, 0 failed. Each new assertion group was mutation-checked against the
merged tree: removing --quantization-config from the Kimi-K3 recipe fails 3,
swapping the decode all2all backend fails 2, dropping Kimi-K3 from either gate
array fails 1 each. The first version of the gate assertion did not fail under its
mutation and was rewritten.
benchmark_xPyD.sh looped over "{1..$BENCHMARK_ITR}". Brace expansion runs before
parameter expansion, so that is the literal 6-character string, and the loop body
executed exactly once no matter what the variable held. Job 252775 ran with
BENCHMARK_ITR=2, printed "iter: {1..2}" into its own benchmark log, and swept each
concurrency once. scripts/vllm_dissag/benchmark_xPyD.sh already used seq.

The fix uncovers a second bug that the first one had been hiding. parse_to_csv.py
starts at "RUNNING: the benchserving script for iter: 1" and pairs each result block
with the RUNNING line in the section before it -- but an `if i > 1` guard skipped the
very first pairing. That was invisible while the marker never appeared: the search
failed, parsing began at byte 0, and the skipped section was the pre-warmup preamble,
which carries no RUNNING line anyway. Restore the marker and the skipped section
becomes the first sweep cell. Replaying job 252775's log with the loop fixed and the
parser untouched gives 6 rows instead of 7, silently missing con=8.

So: seq for the loop, drop the guard, and accept an optional prompt count in the
RUNNING regex so archived logs in the older format still parse. Replayed against
job 252775's benchmark log, the fixed parser reproduces the CSV that build 44
published, byte for byte, from a log in either format.

Fixing the loop alone would also have doubled every sglang_disagg sweep, since
run_xPyD_models.slurm defaulted the knob to 2. That default was only ever decorative,
so it moves to 1: observed behaviour is unchanged and the knob now works.
These comments cite the job IDs and measurements that motivated each fix. On a
long-lived branch those numbers are unresolvable -- a reader six months from now
cannot look up job 252775 -- while the mechanism they illustrate is permanent.
Keep the mechanism, drop the identifiers.

Comment-only: no changed line is code.
run_multinode.slurm runs under `set -euo pipefail` and expands
COLOCATED_FORWARD_ENV at line 153 and again inside the srun body. Nothing
assigns it - not the pipeline, not any model card - and `export NAME` on an
unset variable marks the name for export without giving it a value, so it stays
unset. Under `set -u` that expansion is fatal.

It is an optional knob, so unset is the normal case, and the launcher aborted on
the normal case: SLURM job 354061 printed its banner and died with
"COLOCATED_FORWARD_ENV: unbound variable" before touching a model. Every
vllm_multinode run has been failing this way, which is every Kimi card.

Introduced in db276a6, which added the forwarding mechanism and both use sites
but no default.

Reproduced under `set -euo pipefail`: unset -> exit 1 with that exact message;
with the default -> exit 0 and an empty forward list, and a set value still
forwards, single and comma-separated.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
run_multinode.slurm hardcoded MODEL_DIR=/shared_inference/models_blog, so a
colocated run always pulled its checkpoint over NFS -- even on nodes that already
held a local NVMe copy. It is the only one of the three launchers that did this;
sglang_disagg/run_xPyD_models.slurm and vllm_dissag/run_xPyD_models.slurm have
both probed /mnt/m2m_nobackup first for a while.

It also never bind-mounted /mnt/m2m_nobackup into the container, so pointing
MODEL_PATH there by hand could not have worked either.

This matters most for exactly the models people complain about. A cluster survey
found Kimi-K3 staged on local NVMe on all ten nodes that answered, at full size,
while every colocated run read it over NFS. Small models hide the difference
because they fit in page cache; a 1.5 TB checkpoint cannot.

The probe requires the path on EVERY allocated node -- ranks share one MODEL_PATH,
so a partial hit is useless -- and falls back to NFS otherwise. An explicit
MODEL_DIR or MODEL_PATH still wins. The srun body always exits 0 so a miss does
not leave a FAILED step in sacct, and the mount is guarded on -d because docker
auto-creates a directory at a missing bind source, which would shadow the model
with an empty one.

Verified against the real function with per-node fake roots: NVMe on both nodes
-> NVMe; on one node only -> NFS; on neither -> NFS; explicit MODEL_DIR and
MODEL_PATH both bypass the probe; srun failing -> NFS. bash -n clean.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
We are about to run the same cards two ways -- STANDALONE (the pipeline sbatches
the card's script) and MADENGINE (madengine generates the sbatch). Today the two
agree only by coincidence, because each launcher hardcodes the site's facts
separately. That drift has already cost real runs: run_multinode.slurm hardcoded
the NFS model dir and never probed local NVMe, so every colocated run read its
checkpoint over NFS while a full local copy sat on the node. Both siblings had
probed NVMe for a while; counting m2m_nobackup across the three gave 5/6/1, and
the odd one out was the bug.

scripts/common/cluster.sh follows the pattern vLLM uses at
.buildkite/amd-disagg/cluster.sh: sourced, every value defaulted with :-, so the
layering is always "environment beats site default". That single property is what
makes the two paths interchangeable, because both speak only environment
variables -- STANDALONE through standalone_env.sh and sbatch --export=ALL,
MADENGINE through deployment/slurm.py merging the card's env_vars into the sbatch
it writes. cluster.sh only fills what neither set, so neither path changes.

It also absorbs the two byte-identical copies of check_model_path, and drops a
default of SLURM_SUBMIT_DIR that named one person's home directory.

Card schema: the node count lived in one of two mutually exclusive places -- 39
cards had it only in args (the -N flag), which madengine never parses, and 4 had
distributed.nnodes + slurm.nodes with args empty. madengine reads
distributed.nnodes (deployment/slurm.py:677), so on that path the 39 fell back to
a default and pyt_large_ep_bench_1n/_2n/_4n would all have run on the same node
count. nnodes is DERIVED from the -N already in args, never invented; args is
untouched, and the pipeline only consults nnodes when args is blank, so
STANDALONE is unaffected.

Verified locally, no cluster needed:
- both-paths-agree: for all 43 multinode cards, the env STANDALONE produces and
  the env MADENGINE produces resolve identically through cluster.sh -- with no
  overrides, with image/model overrides, and with site overrides.
- resolver replayed against per-node fake roots under both shell regimes
  (set -euo pipefail, and no flags at all): NVMe on all nodes picks NVMe; on a
  subset falls back to NFS; on neither is fatal; an empty staged directory loses
  to a working NFS copy; an explicit MODEL_PATH bypasses the probe; a distinct
  MODEL_DIR is honoured last; srun failing moves to the next candidate.
- orphan-variable scan across all three launchers is clean, and re-introducing
  the COLOCATED_FORWARD_ENV bug makes it fail, so it would have caught d9a22b7.
- every name a launcher stopped defining is provided by cluster.sh; the six that
  are not are function-local scratch from the deleted duplicates, with zero
  remaining references.
- the -N in args equals distributed.nnodes equals slurm.nodes for all 43 cards;
  every models.json round-trips byte-exactly, so the diff is additive only.
- bash -n on all four shell/slurm files.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
benchmark_xPyD.sh sweeps CON="8 16 32 64 128 256 512" -- seven concurrencies.
Mixtral-8x7B, DeepSeek-V3 and DeepSeek-R1 still declared
decode.tp: --cuda-graph-bs $(seq 1 512), so the server captured 512 CUDA graphs
at startup and used seven of them.

Measured on a 2-node DeepSeek-V3 run in tp mode (DP_MODE=0, so decode.tp is the
line in play), from the decode node's own log:

    Init torch distributed ends     30 s     5%
    Load weight end                 83 s    13%
    Memory pool end                 22 s     3%
    Capture cuda graph end         512 s    79%   <-- 511.58 s, 512 sizes
    ----------------------------------------
    total to ready                 651 s

Capture was four fifths of bring-up. The three sibling models at the top of this
file (Llama-3.1-8B, Qwen3-32B, Llama-3.3-70B-FP8) already ship exactly this seven
-value list in the same slot, and the decode.dp variants already use a short
explicit list too, so this makes the file internally consistent rather than
introducing a new convention.

Llama-3.1-405B-Instruct-FP8-KV keeps $(seq 1 8) deliberately: it is already small
and raising it has memory implications on a 405B model that this evidence does
not cover.

Safe for batch sizes that are not captured: sglang pads a running batch up to the
nearest captured size, which is why the seven-value list already works for the
three models using it.

Verified: every decode.tp capture list now equals CON exactly except the 405B
entry; models.yaml still parses; no $(seq 1 512) remains.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
vdar25 and others added 22 commits November 9, 2026 20:07
models.yaml is now layered --flag: value maps with backends, overlays and compat,
and the model lists and sglang_compat.sh are gone. The vLLM recipe keeps its flag
strings, so docs/configuration.md describes the two formats separately. The SGLang
unit tests and the compat assert join the offline checks.

Co-authored-by: Cursor <cursoragent@cursor.com>
…iver hosts

The gfx950 base image (rocm/sgl-dev v0.5.20 mi35x) loads Broadcom's
bnxt-rocelib 235.2.86.0 from /usr/local/lib, which accepts only kernel
ABI 7-8 (Broadcom's out-of-tree bnxt_re driver), and keeps rdma-core's
own provider (ABI 1) only as libbnxt_re-rdmav34.so-inbox. On a host with
the in-kernel bnxt_re driver (ABI 1), libibverbs skips all eight NICs and
MoRI fails at start-up with "no active RDMA device on this host".

Seen on an MI350X cluster (Thor2
bnxt_re0..7, in-kernel driver): a run died in prefill during MoRI create_backend. No
setting can change which provider the container loads: node.sh has no
hook for extra mounts or env, IBV_DRIVERS cannot name a "-inbox" file,
and the host's provider (/usr/lib64, rdma-core rdmav59) needs the host's
libibverbs.

mori_ep_env.sh is sourced in every container before the servers start.
It now restores the inbox provider only when ibv_devices reports the
bnxt_re ABI mismatch, so hosts where Broadcom's library works are
unchanged.
The _way4 card carries only MAD_CONFIG; its MODEL_NAME and the rest are in
mad-config.kimi-k3.yaml, which madengine resolves and plain sbatch does not.
Submitted with sbatch, the launcher stopped on "set MODEL_NAME" with no hint
why. It now stops before any work and says the file is a madengine layered
config. Tested in launcher_verdict_assert.sh.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Brings ROCm#242's review fixes into ROCm#293: the agentic error rate counted as
validate_agentic_result.sh counts it, sweep (conc<N>) aggregates in perf.csv,
the colocated head's benchmark status and barrier failures, a failed worker
failing the job, EP_TP_SIZE forwarded by run_interactive.sh, header-only
perf CSVs failing the verdict, the mktemp forwarded-env file, and the
MAD_CONFIG refusal in the colocated launcher.

Conflicts were the same fixes applied on both sides: kept this branch's
cluster_expand_nodelist, CONTAINER_ENV/GPU_MEM_TRACE and long-context tests,
and took ROCm#242's launcher directory resolution, header-only verdict and
test-sleep cleanup.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The agent on every pod runs any command sent to its port, and with host
networking anyone who could reach that port could run commands in the pod. The
submitter now puts a random per-run token (MAD_K8S_AGENT_TOKEN) in every pod's
environment; the agent refuses to start without one and refuses any request
that does not carry it. Same files as madengine's k8s_shim.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 5f4c642)
…the container

It mounted a caller-set JIT_CACHE_HOST but did not pass the variable in, so
_clear_stale_jit_locks in vllm_disagg.sh took the cache for the job's own and
removed every AITER lock, including fresh ones of another job sharing the
cache. Both values now go in as the batch launchers pass them. Checked in
jit_lock_assert.sh.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The _way4 card keeps MODEL_NAME and its other values in MAD_CONFIG, which
madengine resolves. submit.py reads only env_vars, so its pods started
without them and the launcher stopped on "set MODEL_NAME". It now refuses
such a card (MAD_CONFIG, or a mad-config.yaml next to the script) and says
madengine is needed. Tested in test_submit.py.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…alue

The per-architecture stack left `ARG BASE_DOCKER=` empty and chose the base
through stage names. Tools that read the base image from that ARG got
nothing: the mad-validate workflow's Dockerfile renderer stopped with "no
base image ... could not read ARG BASE_DOCKER from
docker/sglang_disagg_inference.ubuntu.amd.Dockerfile", and tools/utils.py
get_base_docker returned "". BASE_DOCKER now holds the gfx942 base (the
default architecture, the same image as before the split) and
BASE_DOCKER_GFX950 the gfx950 one; MAD_SYSTEM_GPU_ARCHITECTURE still picks
between them. Checked with `docker buildx build --check` for both
architectures and by rendering every template on a merge with develop.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The previous commit put the JIT_CACHE_HOST and JIT_LOCK_STALE_MIN options on
the EP_TP_SIZE line. Same options, one per continuation line like the rest.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Brings ROCm#293's head, which now carries ROCm#242: the MAD_CONFIG refusal in the
colocated launcher, run_interactive.sh passing JIT_CACHE_HOST and
JIT_LOCK_STALE_MIN, and ROCm#242's merge of develop (kvcache_transfer_bench and
the other develop changes it carried).

Resolved in favour of this branch's structure: the Kubernetes agent and its
token stay removed (pods dispatch per pod), submit.py keeps its own layered
config handling (resolved through madengine when installed), the colocated
srun body stays in node.sh, the per-model SGLang disagg Dockerfile stays
retired (docker/di_sglang + docker/pins/sglang-disagg.json), and
docs/sglang-disagg.md keeps the pin-set build instructions.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The colocated launcher (from ROCm#242 through the spurupdate merge) refuses a
MAD_CONFIG card that has no MODEL_NAME and says madengine resolves the file;
the golden recorded the old "set MODEL_NAME" failure.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
RCCL CI parses these files by name. Checked against ROCm/MAD develop: every
launcher writes the same names in the same place after the refactor. The
node programs (node.sh) tee to pd_vllm_bench_NODE<rank>.log /
pd_sglang_bench_serving.sh_NODE<rank>.log as the launchers' srun bodies did,
the role scripts name prefill_/decode_/proxy_/vllm_router_/
agentic_models_shim_NODE<rank>.log with the same expressions, LOG_PATH is
still mounted at /run_logs, and NODE_RANK is still SLURM_PROCID; the job
array and Kubernetes dispatchers export SLURM_JOB_ID / SLURM_PROCID for the
run, and their snapshot equivalence runs give the same docker commands. The
only additions are new files (gpu_fault_NODE<rank>.log, the opt-in
gpu_mem_trace); the removed sglang_disagg_server.sh was not run by any
launcher. The colocated launcher is new in this stack; its names are
pinned as ROCm#242 added them.

log_names_assert.sh checks every rank of every snapshot golden (mount,
NODE_RANK, the node log's name) and the exact expressions in the role
scripts, so a rename fails a test. docs/multinode-running.md says the
names are a contract and lists the pattern.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
No textual conflicts once spurupdate (which carries ROCm#242's merge of develop)
was in. One semantic one: develop's mad-validate workflow renders every
Dockerfile template with a default base image read from a reference
Dockerfile, and this branch retires two of them
(docker/sglang_disagg_inference, docker/vllm_disagg_inference) for
docker/di_sglang / docker/di_vllm built from pin sets. render_dockerfile.py
now reads those two frameworks' default base image from their pin set's
gfx942 base (docker/pins/sglang-disagg.json, vllm-mori121.json, the sets
that replace them), and the skill's table says so. Checked by running the
workflow's steps (model validator, skill linter, every template rendered
and validated, madengine discover --tags all) on the merge.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ernel-driver hosts"

This reverts commit 924a1e8. Restoring the base image's rdma-core 39
provider lets the bnxt_re devices open, but that provider cannot create
QPs on Thor2 under the in-kernel driver (ibv_rc_pingpong: "Couldn't
create QP"), so MoRI KV transfers stall instead of failing at start-up:
one run hung for 15 minutes. rdma-core v54.0 is the first release
that works there; the image now takes it as a build arg instead.
The gfx950 base (rocm/sgl-dev v0.5.20 mi35x) ships rdma-core 39 plus
Broadcom's bnxt-rocelib 235.2.86.0, which only works with Broadcom's
out-of-tree bnxt_re driver. On an MI350X cluster (Thor2 BCM57608,
in-kernel bnxt_re driver) no RDMA works in
the container: bnxt-rocelib rejects every NIC ("does not support the
kernel ABI of 1 (supports 7 to 8)"), and the base's own rdma-core 39
provider opens them but cannot create QPs. Cross-node ibv_rc_pingpong
in the image fails with rdma-core 39, 50 and 53 and passes with 54.0
and 61.0 (likely 5f008dbf, variable-size send WQEs, first in v54.0).
With v61.0, MoRI IO GPU-memory batch writes validate at 40.9 GB/s over
one NIC.

rdma_core is a pin component like MoRI: pins.py emits RDMA_CORE_BUILD,
_REPO and _COMMIT, and the Dockerfile builds it over the base's stack
where the pin set enables it (gfx950 here; gfx942's base is unchanged).
linux-rdma is rdma-core's upstream, so the no-fork test accepts it.

Hosts on Broadcom's out-of-tree driver need the base's stack, which
this replaces; that trade-off is open with the image's owners.
MoRI's start-up only opens the RDMA devices, and the recipes carry
--watchdog-timeout 1000000, so an image whose RDMA user space cannot
create QPs on the host's NICs started both servers and then hung on the
first KV transfer: one run (rdma-core 39 on Thor2 under the in-kernel
bnxt_re driver) aborted 72 transfers and was cancelled after 15 minutes.

Before the container barrier, each node now runs a loopback
ibv_rc_pingpong on every IB_DEVICES NIC at MORI_IB_GID_INDEX. A failure
goes through _job_fail, so the peers' barrier sees the abort file and
the whole job stops in seconds with the device and the reason. With the
base gfx950 image it reports "IB device bnxt_reN not found"; with
rdma-core 39's provider, "Couldn't create QP"; rdma-core v61.0 passes.

Skipped under DRY_RUN and when the image has no ibv_rc_pingpong.
RDMA_PREFLIGHT=0 skips it; node.sh forwards it only when set, so the
rendered docker flags are unchanged otherwise (snapshot: 18 cards, 0
differ).
node.sh hard-coded --privileged, and a Docker daemon whose policy refuses
privileged containers then refuses every node's container, with no setting to
change it. CONTAINER_PRIVILEGE_ARGS (cluster.sh, default --privileged) is what the
container gets instead; on such a site, the capability RDMA needs:

    CONTAINER_PRIVILEGE_ARGS="--cap-add IPC_LOCK --ulimit memlock=-1:-1"

The default leaves every card's docker run unchanged (goldens identical).

Co-authored-by: Cursor <cursoragent@cursor.com>
docs/configuration.md still had SGLang composing PREFILL_MODEL_CONFIG /
DECODE_MODEL_CONFIG and the launcher owning --tp-size and the transfer
backend; render.py and models.yaml do both now. One role-script comment
still named DECODE_NODE_RANK.

node.sh does not source cluster.sh, so it now defaults
CONTAINER_PRIVILEGE_ARGS to --privileged itself: a dispatcher that does not
forward the variable would otherwise start the containers unprivileged.

Comments the port added are trimmed to what the code does not say.
Snapshot: 18 cards, 0 differ.

Co-authored-by: Cursor <cursoragent@cursor.com>
run_benchmark.sbatch ran `docker stop $(docker ps -aq) && docker rm
$(docker ps -aq)` on every node before starting, which stops and deletes
every container on the node, including other users' on a shared
cluster. The lines before it already stop and remove the ep-bench
container this script starts, and the cleanup step at the end does the
same, so only the node-wide one is removed.
The launcher assumed one CX7 cluster and could not be told otherwise:
it runs a fixed docker command with no environment passthrough. On an
MI350X cluster (Thor2 bnxt_re, fenic0, multi-homed nodes, FQDN
hostnames, root-squashed /home) it could not run:
- run_node.sh forced GLOO_SOCKET_IFNAME=eth0; there is no eth0;
- each node mapped its name to its first `hostname -I` address, a /31
  RDMA rail the peer cannot reach over TCP;
- role detection compared the short Slurm name with an FQDN
  `hostname`, so neither node took the target role;
- the MoRI backend only looked for MoRI in /workspace/mori;
- `set +x` above the #SBATCH lines made sbatch ignore every directive
  (-N 2, --output), and there was no madengine card.

Inside MAD the launcher now sources common/cluster.sh and
launcher_lib.sh like the other multinode launchers: --ibdevice defaults
to mlx5_0 on cx7/unknown and the first detected rail on thor2/ainic;
caller-set fabric values always reach the container, and on thor2/ainic
the detected ones too, so cx7 keeps its container defaults; both nodes
resolve both names to the FABRIC_SUBNET_PREFIX address. Run standalone
(no MAD around it) it behaves as before. run_node.sh honours a forwarded
GLOO_SOCKET_IFNAME and compares short names; the MoRI backend takes
MORI_ROOT or the first of /workspace/mori, /sgl-workspace/mori,
/app/mori with MoRI's IO benchmark, and installs prettytable if missing.

For madengine: the launcher falls back to DOCKER_IMAGE_NAME for the
image, pre-creates the results directory writable for a root-squashed
container, merge_results.py --perf-csv writes the perf schema, and
cluster_job_verdict publishes perf_kvcache_transfer_bench.csv (the new
card pyt_kvcache_transfer_bench_mori's multiple_results) and fails the
job on a node failure or a size with no throughput.
run_benchmark.sbatch assumed one CX7 cluster: IBDEVICES defaulted to
mlx5_0, each node was addressed by its first `hostname -I` address
(a /31 RDMA rail on multi-homed nodes, which the peer cannot reach over
TCP), /shared_inference/models was bind-mounted unconditionally (docker
creates a missing source as an empty root-owned directory), and the
perf.csv run_ep_bench.sh writes was never handed to madengine, so the
cards reported nothing.

Like the other multinode launchers it now resolves its directory under
sbatch's spool copy and, inside MAD, sources common/cluster.sh and
launcher_lib.sh: IBDEVICES defaults to mlx5_0 on cx7/unknown and the
first detected rail on thor2/ainic; caller-set fabric values always
reach the container, and on thor2/ainic the detected ones too, so cx7
keeps its container defaults; nodes are addressed on
FABRIC_SUBNET_PREFIX. The models share is mounted only where it exists,
the job log directory is writable for a root-squashed container, and
cluster_job_verdict publishes perf_large_ep_bench.csv, which every card
now declares as multiple_results, and fails the job on a node failure
or an empty result.
ROCm#300 brings ROCm#269, which moved the KV cache transfer benchmark to
run.slurm, NIXL, its own madengine cards and to_perf_csv.py. The fabric
commit on this branch was written against the old launcher; its changes
are kept on ROCm#269's: the cluster.sh fabric (ibdevice, forwarded NCCL / GLOO
/ MoRI settings), node addresses on FABRIC_SUBNET_PREFIX, the project
directory under sbatch from the MAD root, result directories writable for
a root-squashed container, short-name roles and the MoRI root. Its own
card and perf.csv writer give way to ROCm#269's. The cards drop
IBDEVICE=mlx5_0 so the fabric picks the NIC; CX7 still gets mlx5_0.

Co-authored-by: Cursor <cursoragent@cursor.com>
render.py passes the same server flags as before in another order, and
once a flag the old command gave twice (with the value that took effect),
so re-recording rewrote all 18 SGLang goldens for no change in what runs.
The snapshot now compares SGLang server commands as argparse reads them:
option order is free, and a repeated option counts once with its last
value. The goldens stay ROCm#300's. Docker runs, router commands and every
vLLM command are still compared exactly, and a changed or dropped flag
still fails.

Co-authored-by: Cursor <cursoragent@cursor.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants