Accueil Explorer Aide
Connexion
sm121
/
recipes
2
0
Fork 0
Fichiers Tickets 0 Pull Requests 0 Wiki
Aborescence: 50b1f75ee4
Branches Tags
recipes
/
docker-images
/
mia-tp4
/
docs
/
tp4.md

tp4.md 18 KB
Historique Raw

Four Sparks (TP4): the production line

A measured TP4 serving line for four DGX Spark (GB10) nodes, built on this repository's launcher (start-tp4.sh → start.sh), image layout, NVMe Engram row store, b12x MXFP8 route and memory model. Everything it adds is armed only by ./start-tp4.sh, which sets DSV41_LAUNCHER=tp4 before it execs start.sh. That flag is what turns on the named Dockerfile, EXTRA_CONTAINER_ENV, the anchored worker rsync, EP 1 or 2 on the ring, and the production adapter hooks. The switches still do nothing until the profile sets them. ./start.sh leaves the flag unset, so the 3-Spark profile (.env, start.sh) and the default Dockerfile keep their behaviour.

The per-change history, every earlier measurement and the experiments that were tried and not adopted are in knapcio/DeepSeek-V4.1-Flash-4x-DGX-Spark-TP4 (docs/history.md, CHANGELOG.md). This page is the v2.1 state of that repository.

Results

Production line (the last EXTRA_CONTAINER_ENV line of .env.tp4.example with EP_SIZE=1, Dockerfile.canary-roce image), built from a fresh clone on all four nodes and measured on that image with sparkDash 1.8.8: 256 new tokens, temperature 0, thinking off, idle fleet, 2026-09-25, switched RoCE fabric. Engine start to healthy 160-170 s; KV pool 6.0-6.6M tokens (1M context). Raw output: results/tp4/validation-20260925-v2.txt, results/tp4/validation-20260925-v21.txt.

Decode, aggregate tok/s (per stream in brackets)

prompt type c1 c2 c4 c8 c16
prose 87.7 120.4 (61.9) 163.6 (41.4) 237.6 (30.8) 342.7 (22.2)
code 124.8 175.0 (88.4) 246.8 (63.2) 309.8 (41.4) 438.3 (29.3)
structured 152.4 177.6 (103.3) 240.4 (70.4) 295.5 (44.2) 572.2 (44.4)
json 118.9 174.2 (89.8) 301.7 (76.2) 471.5 (60.2) 659.9 (42.7)

Prose and code c1 are from the v2.1 fresh-clone boot (prose median of five runs 86.7-87.7 after two discarded warm-ups); the other columns are from the v2 sweep, whose decode path is identical. With the deterministic MoE reduction the greedy text is identical run to run, so the sparkDash numbers repeat within about ±1 tok/s. sparkDash's prose c1 is one prompt, and a stack that sums in a different order (another fabric, another all-reduce) follows a different greedy text there, so compare step time or a many-prompt benchmark across stacks. On 45 varied prompts (prose, structured and other catalogs, c1 greedy) the same image runs 58.4 / 94.3 / 71.8 tok/s; decode step 32.4-33.6 ms on prose and 38.9-39.4 ms on code at c1. sparkDash uses a different set of prompts at each concurrency for the non-prose types, so per-stream values are not comparable across columns. Sampled chat at the model card's T=1 / top_p=0.95 with thinking (c1, 18 requests x 800 tokens on two prompt sets) ran 67.2 / 65.9 tok/s on the 2026-09-24 stack (sparkDash benches are greedy, where the draft temperature and block verification do not act).

Prefill, cold, tok/s by prompt length (two passes, v2.1)

4k 16k 32k 64k 128k 262k
4119 / 4840 5891 / 5846 5893 / 5868 5936 / 5903 5793 / 5761 5355 / 5360

sparkDash's prefill filler is one repeated token, so every filler token hits the same Engram row and the row cache inflates these numbers (reported by koldfrontier in #21). On real text (documentation and source code, a unique prefix per prompt so nothing comes from the prefix cache) the same image measured 4202-4685 / 4824-4922 / 4827-4830 / 4942-4970 / 4822-4871 tok/s at ~4k / ~15k / ~32k / ~62k / ~120k tokens. Needle retrieval (a single phrase in varied filler at 37 % depth): PASS at 99k, 198k, 746k and 1,011,084 tokens (the last in 322-325 s, head MemAvailable low-water 7 GiB).

Quality. scripts/qeval.py runs 75 auto-scored tasks (code executed against hidden asserts, JSON schema-checked, numeric answers matched, format constraints enforced, prose checked for degeneration; no LLM judge), one request at a time, temperature 0. The fresh-clone image scores 72 of 75; the three misses (code_interval_intersect, json_escape, math_m9) fail identically on the stock profile. Every speed change is meant to be lossless: same weights, every draft token verified by the target, and each adapter either bit-identical to the stock path (checked at boot or in the in-image tests) or exact in distribution (draft temperature, block verification). RoCEnante and prefill sequence parallel sum in a different order than NCCL's all-reduce, so the numerics are not bit-identical to an NCCL run, which is why the line is also scored. Run qeval from a worker, not from the head (it executes model-generated Python).

What runs

Layer Setting What it does
Image Dockerfile.canary-roce upstream SGLang dsv4.1 branch at f80c91a4b + rhys101's RoCEnante overlay adapted to TP4 + b12x main as b12x_next + all adapters; third-party sources fetched at build (runtime/sources.manifest)
Slots MAX_RUNNING_REQUESTS=16 adds the c16 tier
Experts EP_SIZE=1, DSV41_MOE_B12X_NEXT=1, DSV41_MOE_B12X_NEXT_DETERMINISTIC=1 every rank holds a quarter of all 384 experts and the routed MoE runs on b12x main: no expert-group straggler at the MoE all-reduce (all-reduce 5.9 → ~2.6 ms per step), MoE time unchanged. The deterministic reduction (per-slot buffer and a fixed-order sum instead of atomics) makes greedy output bit-identical run to run
Engram DSV41_CACHE_GIB=4, DSV41_CACHE_WAYS=16, DSV41_ENGRAM_PREFETCH=1 row cache (67–76 % hits) and row lookups on a side stream, rows bit-identical
Draft DSPARK_BLOCK_SIZE=5, SGLANG_DSPARK_FOLDED_SAMPLING=2 k=5 wins on code, ties on prose at TP4; sampled requests stay in the CUDA graph
Draft sampling DSV41_DRAFT_TAU=0.7, DSV41_BLOCK_VERIFY=1 sharper draft proposals and block verification for sampled rows, both exact in distribution
Verify length DSV41_VERIFY_CAP=conf:0.1, DSV41_ROUTER_LIVE=1 verifies only the drafts the confidence head expects to survive; the other verify rows reuse the anchor row's experts inside the router kernel, so they read no extra weights; greedy outputs identical
Kernels DSV41_SHARED_PAD_K=1, DSV41_WO_A_W8=1, DSV41_WO_A_W8_MID=1, DSV41_WO_A_W8_DROP=1, DSV41_DRAFT_HEAD_FP8=1, DSV41_DRAFT_HEAD_FP8_IMPL=tp4 shared expert kept on the b12x kernel; wo_a and the draft LM head read exact fp8 twins of their weights
Replicated linears DSV41_REPLICATED_SPLIT=wqkv_a,engram.wkv, DSV41_DRAFT_MAIN_PROJ_SPLIT=1 layers every rank computed in full are split by columns across ranks and all-gathered; enabled only where bit-identical on all ranks
Transport SGLANG_ROCE_ALLREDUCE=1, SGLANG_ROCE_MAX_SIZE=2097152, DSV41_ROCE_GATHER=2097152, B12X_ROCE_HCA=rocep1s0f0,roceP2p1s0f0 TP all-reduces and all-gathers up to 2 MiB over RoCEnante's one-shot RDMA kernel on both rails
Prefill CHUNKED_PREFILL_SIZE=4096, DSV41_INDEXER_CHUNKED=1, SPARK_PREFILL_TP_SPLIT=1 bounded indexer transient (sglang#39187) plus the SG18 row split across ranks from 32k context
Loading DSV41_FAST_LOAD=1, DSV41_FAST_LOAD_TP_SLICE=auto, DSV41_AUTOTUNE_KEEP=1 engine start 343 s → ~120 s; at EP1 each rank reads only its slice of every expert, into 256 MiB pinned slabs; MoE autotune cache kept across boots (fast-load.md)
Prefill hc DSV41_HC_FUSED=1 the hyper-connection mix statistics of prefill chunks in one pass over K instead of 80 partial slices: ~1.46 → ~0.83 ms per call at 4096 rows, bit-identical (checked on the first live call)
Prefill sequence parallel DSV41_PREFILL_SP=1, DSV41_PREFILL_SP_FP8=1 at prefill chunks of >= 2048 rows the per-layer all-reduces become reduce-scatter + all-gather and the per-row work between them runs on each rank's quarter of the rows: prefill ~+20 % from 16k to 262k, decode untouched. DSV41_PREFILL_SP_EXACT=1 keeps the all-reduce (bit-identical to the unsharded path) for ~+10 %; _FP8 gathers the attention input as the MXFP8 bytes wqkv_a would compute itself (bit-identical per chunk size), +1.5-3 %
L2 prefetch DSV41_L2_PREFETCH=1 during each RoCE collective of a decode step a side-stream kernel prefetches the first 6 MB of the weights that follow into L2; data untouched; decode step -1 ms at c1
Correctness DSV41_FOLDED_FENCE=1 closes the sglang#40919 D2H race on the folded verify path
Serving --enable-cache-report, --sleep-on-idle, SGLANG_RUST_BUILD_MODE=never cached-token usage for clients, idle CPU 47 % → 14 %, avoids a cargo hang at start on the branch images
Optional DSV41_ENGRAM_DRM_NODE=/dev/dri/card0 one Engram layer's cache in the GB10 display reservation, ~1.8 GiB more headroom; needs a host change (display-reserve.md)

Each adapter is described in tp4-adapters.md; the image layers on their own and the line on a switchless ring are in tp4-optional-setups.md.

What the profile changes

Against the earlier TP4 example, all of it in .env.tp4.example plus gated adapters:

Setting Before Now Why (measured)
EP_SIZE 4 1 (production), 2 (base and canary images) On the production image every rank streams a 576-wide slice of every touched expert, so no expert group waits for another at the MoE all-reduce. The base and canary images stay at EP_SIZE=2 (FlashInfer's MXFP4 MoE needs the per-rank width to be a multiple of 128), which already halves the EP4 straggler wait: NCCL time per step 16.5 → 10.2 ms
DSV41_CACHE_GIB / WAYS 0 / 4 4 / 16 Engram rows repeat (bigram/trigram heads): 67–76 % hit rate, 4x fewer NVMe reads
DSPARK_BLOCK_SIZE 3 5 on TP4 k=5 wins on code by ~10 % and ties on prose
MAX_RUNNING_REQUESTS 8 16 CUDA graphs to bs 16 cost ~5.6 GB and add a c16 tier (+64 % aggregate over c8)
CHUNKED_PREFILL_SIZE 1024 4096 safe only together with the indexer backport below
DSV41_SHARED_PAD_K – 1 pads the shared expert's K 576 → 640 so it stays on the b12x MXFP8 kernel: −0.9 ms/step, bit-identical
DSV41_INDEXER_CHUNKED – 1 sglang#39187: indexer logits scored in ≤ 2 GiB row chunks, tail-only candidate masks; a 262k cold prefill keeps ≥ 7 GiB free on the head
SPARK_PREFILL_TP_SPLIT (canary images) – 1 SG18: the dense prefill indexer's query rows are partitioned across the four ranks from 32k context
DSV41_SERIAL_WEIGHT_LOAD 1 commented, in the ring block the switched line was measured with the default loader and the fast loader's own pacing
--enable-deepseek-v4-fp4-indexer off on no effect on V4.1 (no ratio-4 layers; its only indexer is fp4 by design); kept so the argument line matches the measured runs

Unchanged: memory fraction 0.80, the 8M-token KV pin, 1M context, the NFS/Engram layout, the OpenAI serving fixes, --min-free-slots-delay 1.

How to run

cp .env.tp4.example .env.tp4          # HEAD_IP / WORKER_* / MODEL_DIR / fabric, as for any TP4 profile
# production: in .env.tp4 uncomment IMAGE=dsv41-4x-spark:canary-roce, BUILD_DOCKERFILE=Dockerfile.canary-roce
# and the last EXTRA_CONTAINER_ENV line, and set EP_SIZE=1 (the file's EP_SIZE=2 is for the base/canary images)
./start-tp4.sh doctor
./start-tp4.sh build                  # every node: fetches the pinned sources, bakes the adapters, runs the in-image tests
./start-tp4.sh share && ./start-tp4.sh pack   # first time only (NFS share, Engram shards)
./start-tp4.sh serve                  # ./start-tp4.sh stop | status | logs | smoke

build needs github.com during the image's fetch stage (and pypi.org, or BUILD_ARGS="--build-arg PIP_INDEX=<mirror>/simple"). The fetched trees can be inspected on a host with scripts/fetch_runtime.sh runtime/fetched.

The boot log must show these lines, otherwise the line is not active:

DSV41 shared-expert padding K: ... (5120, 576) -> (5120, 640)
DSV41 indexer chunked (sglang#39187 backport) ARMED: ...
Initialized DSpark draft runner. ... gamma=5, verify_num_draft_tokens=6
max_total_num_tokens=..., chunked_prefill_size=4096, ... max_running_requests=16
RoCEnante ready: world=4 hcas=...
[moe_b12x_next] armed: routed MoE on b12x_next a7d7d29b ...
[moe_b12x_next] INFO: routed MoE at EP_SIZE=1 ...
DSV41 hc_fused: first call (4096 rows) bit-identical to stock

Measuring

  • Bench with sparkDash's decode and prefill benches; discard the first two runs after a boot (cold Engram row cache, first-run warm-up).
  • Any #running-req in the engine log above the concurrency being benched means foreign traffic landed in the window.
  • A dashboard polling nvidia-smi every 2 s on every node costs ~0.6 ms per decode step; poll every 10 s.
  • sparkDash's fixed prompts are greedy: a change that alters rounding can flip a near-tie early in the 256 tokens and move c1 by ±10 % without changing speed. Kernel A/Bs use step time on several prompts as well.
  • Greedy text equality is not a correctness gate for prefill changes: identical cold prompts of ~70k tokens produce different greedy continuations run to run. Correctness of the backports rests on the bitwise CPU tests, the needle tests and qeval.

Rollback

Every layer is an env change: EP_SIZE=2 without DSV41_MOE_B12X_NEXT returns the routed MoE to FlashInfer, DSV41_FAST_LOAD=0 restores the stock loader, SGLANG_ROCE_ALLREDUCE=0 drops the RDMA transport, SPARK_PREFILL_TP_SPLIT=0 the row split, IMAGE=dsv41-4x-spark:canary the RoCEnante overlay, IMAGE=dsv41-4x-spark:local the branch. The earlier TP4 example: MAX_RUNNING_REQUESTS=8, CHUNKED_PREFILL_SIZE=1024, DSPARK_BLOCK_SIZE=3, EXTRA_CONTAINER_ENV="", EP_SIZE=4, DSV41_CACHE_GIB=0. The adapters stay in the image but do nothing when their gate is unset.

Known limits

  • The numbers above are on a switched RoCE fabric. On a switchless ring RoCEnante needs a path to the opposite node, built in the neighbours' ConnectX-7 hardware (switchless-ring.md); without it the collectives go through NCCL and decode is slower. Prefill is lower on a ring either way (one-link bisection).
  • The EP1 line needs the Dockerfile.canary-roce image (it carries b12x_next). The routed MoE on b12x is numerically equivalent to FlashInfer's, not bit-identical (relative error against an fp32 reference 4.5–4.8 % for both); qeval and the needle tests are unchanged.
  • Single-stream prose speed is bounded by DSpark acceptance (~3 accepted tokens per step on prose against ~6 on code).
  • The fast loader leaves the KV pool smaller than the stock loader and more variable between boots; DSV41_FAST_LOAD=0 restores it at ~220 s per boot (fast-load.md).
  • RoCEnante is a new transport in the decode path; b12x has one open report of a rank wedging under long mixed-context traffic on an earlier revision (b12x#313). The overlay's result-boundary health check fails the step instead of hanging, and NCCL is one env change away.

Credits

  • Mia (MiaAI-Lab), this repository: the launcher, boot.py, the Engram NVMe row store and packed shards, the MXFP8 b12x routing, the chunked-prefill memory model, the thinking alias, the output cap, the loop abort, the TP4 profile this line started from, and sparkDash, the benchmark behind every number here.
  • local-inference-lab / Luke Alonso and Jason (original-el8), b12x: RoCEnante, the one-shot RDMA all-reduce (the SG17 revision), and the fused MoE kernels that run the routed experts (b12x main at a7d7d29b, installed as b12x_next beside the SG17 copy, with a two-line patch that admits 64-row tiles for 576-wide experts at prefill sizes).
  • rhys101, DeepSeek-V4.1-Flash-vLLM-DGX-Spark-8: the SG17 SGLang overlay that routes small tensor-parallel all-reduces to RoCEnante (with a TP4 adaptation in runtime/roce_tp4_adapt.py), and the SG18 native prefill TP split (adapter/spark_prefill_dense.py, combined with the indexer backport in adapter/indexer_chunked_v3.py).
  • luxingcom (LuZ), LuZ DGX Spark TP4 ring: the first integration of b12x's fused MoE into SGLang on a four-Spark fleet, which showed the route.
  • sumsliu, dgx-spark-deepseek-v41: the eight-Spark measurement that moving to expert tensor parallelism removes most of the all-reduce wait.
  • FujitsuPolycom and the sparkring contributors, sparkring, and rsync (@rchmagos): the hardware-forwarded opposite-node paths and the path-aware RoCEnante (b12x.comm.roce_ring) that run the line on a switchless ring; the NCCL transport patch for the ring.
  • @Saolence: the switchless-ring launcher and dual-plane findings, the BUILD_DOCKERFILE / BUILD_ARGS build step, the anchored worker rsync and the pip-index build argument.
  • kpham-sgl, sgl-project/sglang#39187: the bounded dense-indexer prefill transient, backported as adapter/indexer_chunked*.py.
  • BBuf and the SGLang dsv4.1 branch contributors (#39370, #39646, #39648, #39653): the decode kernel work in the canary images.
  • hushengkai, for independently reproducing the EP2 / Engram cache / shared-expert padding changes on a second 4x GB10 fleet; koldfrontier for the sparkDash prefill-filler finding.