{"slug":"deepseek-v4-1-flash-on-4x-dgx-spark-tuned-tp4-sglang-dspark","title":"DeepSeek-V4.1-Flash on 4x DGX Spark \u2014 tuned TP4 SGLang/DSpark","summary":"","license":"CC-BY-4.0","tier":"community","beta":false,"revision":{"no":3,"commit_sha":"ddf188add74be5de70f13f6f71735ec18a87c83c","content_hash":"647b167cbdd40dfd95cbaabc81e8624e929d847d","created_at":"2026-10-29 21:32:23.622356-06","notes":"Submitted 2026-10-10 from lar-scout ingest card t_ad3aebd3. Repo @ fa1356384ed246f3040775210668f02a1bba3426 (HEAD verified via git ls-remote 2026-10-10; AGPL-3.0-or-later, upstream README preserved at docs/README-upstream.md with the credit chain kept verbatim). Launch argv transcribed from boot.py serve() + .env.tp4.example EXTRA_SGLANG_ARGS at that sha and flag-verified against boot.py's own env plumbing (spec/DSPARK/tables/nnodes/max-total-tokens/EXTRA_SGLANG_ARGS passthrough; EP_SIZE=1 pairs with DSV41_MOE_B12X_NEXT per .env.tp4.example comment; --min-free-slots-delay 1 matches README; DSPARK_SPS_TABLE=/state/dspark_sps.json only applied when the file exists). Performance rows transcribed from the repo's own bench docs (README 'Current results' + docs/results/validation-20260929-v23.txt sparkDash section) \u2014 measured, not claimed-by-anecdote; all at the 2200 MHz SM clock cap on a switched RoCE fabric (switchless ring slower, see notes). Engine pins from Dockerfile.canary-roce + scripts/fetch-sglang-canary.sh + runtime/b12x_next/b12x_next/SOURCE_COMMIT at the same sha; MiaAI upstream HEAD pinned via git ls-remote same day. HARDWARE NOTE: dgx-spark-x4 is a NEW device_count profile (site had 1x only) \u2014 hardware_profiles row seeded by a 0041-convention INSERT-only migration (family dgx-spark, generation blackwell, 128 GB unified memory each) submitted alongside this ingest, mirroring the 0053 cmp-170hx-x1 and 0047 rtx-3090-x4 precedents so the ingest resolves the seed slug instead of silent-creating a NULL card_family orphan bucket."},"engine_source":{"type":"docker_image","git_url":"https://github.com/sgl-project/sglang","base_engine":"sglang","pip_package":null,"base_version":"dsv4.1 branch f80c91a4b + RoCEnante SG17 + adapters","image_digest":"docker.io/lmsysorg/sglang:dev-dsv41","install_note":"NOT a stock sglang image: Dockerfile.canary-roce FROM lmsysorg/sglang:dev-dsv41, replaces the SGLang python tree with the upstream dsv4.1 branch at f80c91a4b (GitHub tarball via scripts/fetch-sglang-canary.sh), installs sglang-kernel==0.4.7 + sgl-deep-gemm==0.2.0 aarch64 wheels, applies rhys101's SG17 RoCEnante overlay + runtime/roce_tp4_adapt.py, bakes runtime/b12x (RoCEnante, frozen SG17) and runtime/b12x_next (b12x main at a7d7d29b + 64-row-tile patch) under /opt, and copies all adapters. Image tag dsv41-4x-spark:canary-roce, built per node via ./start-tp4.sh build. Requires aarch64 + nvidia-container-toolkit + NVIDIA driver 580.x.","pinned_commit":"f80c91a4b0000000000000000000000000000000"},"credit":[{"name":"knapcio (TP4 profile, v2.1-v2.3 production stack, release gates)","url":"https://github.com/knapcio/DeepSeek-V4.1-Flash-4x-DGX-Spark-TP4"},{"name":"MiaAI-Lab (upstream: launcher, Engram store, b12x routing, memory model)","url":"https://github.com/MiaAI-Lab/DeepSeek-v4.1-Flash-DGX-Sparks"},{"name":"0xSero (original 4x RTX PRO 6000 recipe skeleton boot.py derives from)","url":"https://github.com/sybil-solutions/deepseek-v4.1-flash-4x-rtx-pro-6000"},{"name":"kpham-sgl (sglang#39187 bounded dense-indexer prefill transient, backported as adapter/indexer_chunked*.py)","url":"https://github.com/sgl-project/sglang/pull/39187"},{"name":"BBuf + dsv4.1 branch contributors (#39370, #39646, #39648, #39653 decode kernel work in the canary images)","url":"https://github.com/sgl-project/sglang"},{"name":"rhys101 (SG17 RoCEnante overlay + SG18 prefill TP split from DeepSeek-V4.1-Flash-vLLM-DGX-Spark-8)","url":"https://github.com/rhys101/DeepSeek-V4.1-Flash-vLLM-DGX-Spark-8"},{"name":"local-inference-lab \u2014 Luke Alonso & Jason (b12x: RoCEnante one-shot RDMA all-reduce + fused MoE kernels)","url":"https://github.com/local-inference-lab/b12x"},{"name":"FujitsuPolycom + sparkring contributors (switchless-ring opposite-node paths, path-aware RoCEnante)","url":"https://github.com/FujitsuPolycom/sparkring"},{"name":"Saolence (#1/#2/#4/#5/#6: ring as opt-in switch, NCCL/build fixes, first-deployer review)","url":"https://github.com/knapcio/DeepSeek-V4.1-Flash-4x-DGX-Spark-TP4/pull/4"},{"name":"rsync/rchmagos (#8: RoCEnante on the switchless ring, DSV41_ROCE_RING + scripts/ring_mesh/)","url":"https://github.com/knapcio/DeepSeek-V4.1-Flash-4x-DGX-Spark-TP4/pull/8"}],"author":{"name":"localai-recipes","isSite":false},"uploaded_by":"localai-recipes","config":{"launch":["python3","-u","/opt/dsv41/boot.py","run","--model-path","/models/DeepSeek-V4.1-Flash","--served-model-name","deepseek-v4.1-flash","--trust-remote-code","--load-format","safetensors","--tp","4","--ep-size","1","--nnodes","4","--node-rank","0","--dist-init-addr","10.0.0.1:20000","--attention-backend","dsv4","--moe-runner-backend","flashinfer_mxfp4","--mem-fraction-static","0.80","--chunked-prefill-size","4096","--context-length","1048576","--max-running-requests","16","--cuda-graph-max-bs-decode","16","--max-total-tokens","8000000","--random-seed","0","--enable-decoder-swa-bounded-replay","--speculative-algorithm","DSPARK","--speculative-dspark-block-size","5","--tool-call-parser","deepseekv41","--reasoning-parser","deepseek-v41","--host","0.0.0.0","--port","8888","--fp8-gemm-backend","flashinfer_cutlass","--watchdog-timeout","1800","--min-free-slots-delay","1","--enable-deepseek-v4-fp4-indexer","--enable-cache-report","--sleep-on-idle","--model-loader-extra-config","{\"num_threads\":1}"],"runtime":{"env":"EXTRA_CONTAINER_ENV (production v2.3 line, mirrored to every rank): DSV41_INDEXER_CHUNKED=1 SGLANG_DSPARK_FOLDED_SAMPLING=2 SGLANG_ROCE_ALLREDUCE=1 SGLANG_ROCE_MAX_SIZE=2097152 B12X_ROCE_HCA=rocep1s0f0,roceP2p1s0f0 B12X_ROCE_CACHE_DIR=/state/b12x-roce B12X_COMPILE_CACHE_DIR=/state/b12x-compile SPARK_PREFILL_TP_SPLIT=1 SPARK_PREFILL_TP_MIN_CONTEXT=32768 SPARK_PREFILL_TP_MIN_ROWS=1024 SGLANG_RUST_BUILD_MODE=never DSV41_FAST_LOAD=1 DSV41_FAST_LOAD_EP_SIZE=1 DSV41_FAST_LOAD_TP_SLICE=auto DSV41_FAST_LOAD_N_EXPERTS=384 DSV41_FAST_LOAD_INFLIGHT_GB=6 DSV41_ENGRAM_PREFETCH=1 DSV41_SHARED_PAD_K=1 DSV41_WO_A_W8=1 DSV41_DRAFT_TAU=0.7 DSV41_DRAFT_HEAD_FP8=1 DSV41_BLOCK_VERIFY=1 DSV41_FOLDED_FENCE=1 DSV41_VERIFY_CAP=conf:0.1 DSV41_AUTOTUNE_KEEP=1 DSV41_WO_A_W8_MID=1 DSV41_WO_A_W8_DROP=1 DSV41_REPLICATED_SPLIT=wqkv_a,engram.wkv DSV41_DRAFT_MAIN_PROJ_SPLIT=1 DSV41_ROCE_GATHER=2097152 DSV41_ROUTER_LIVE=1 DSV41_MOE_B12X_NEXT=1 DSV41_MOE_B12X_NEXT_DETERMINISTIC=1 DSV41_HC_FUSED=1 DSV41_PREFILL_SP=1 DSV41_PREFILL_SP_FP8=1 DSV41_L2_PREFETCH=1 DSV41_L2_PREFETCH_WOA=1 DSV41_SPEC_SYNC_FREE=all DSV41_EAGER_GLUE=all DSV41_SPLIT_COMPACT_GATHER=1 DSV41_CERT_HEAD=1 DSV41_CERT_HEAD_M=6,12,18,24,30,36,42,48 | Also per .env.tp4.example: EP_SIZE=1 DSV41_MXFP8_BACKEND=b12x DSV41_SHARED_PAD_K=1 DSV41_CACHE_GIB=4 DSV41_CACHE_WAYS=16 DSV41_IO_THREADS=96 DSV41_ENGRAM_PREFETCH=1 DSV41_TP_PAD=0 DSV41_MAX_NEW_TOKENS=32768 DSV41_LOOP_ABORT=1 SGLANG_DSV41_REASONING_EFFORT=75 PYTORCH_CUDA_ALLOC_CONF=expandable_segments:False DSV41_PREFILL_EMPTY_CACHE_TOKENS=8192 SGLANG_FLASHINFER_MOE_FUSED_FINALIZE=0 NCCL_BUFFSIZE=1048576 NCCL_PROTO=^LL128 NCCL_MAX_NCHANNELS=8 B12X_ROCE_HCA=rocep1s0f0,roceP2p1s0f0 | Launch: IMAGE=dsv41-4x-spark:canary-roce, BUILD_DOCKERFILE=Dockerfile.canary-roce, EP_SIZE=1, port 8888 + dist port 20000 + NFS 2049/111 node-to-node.","notes":"Run ./start-tp4.sh (doctor|download|build|share|pack|serve): boot.py profile bootstrap, .env.tp4 from .env.tp4.example (HEAD_IP/WORKER_IPS/MODEL_DIR/fabric), dockerized 4-rank fleet --gpus all --privileged --ipc host, ranks 1-3 over ssh, head rank 0 on :8888; boot log must show the DSV41 arming lines (shared-expert padding, indexer chunked, DSpark gamma=5, max_running_requests=16, RoCEnante ready, moe_b12x_next armed EP1, hc_fused bit-identical). Weights deepseek-ai/DeepSeek-V4.1-Flash @ fb2764a5 (476 GiB, 48 shards) at the pinned revision, packed Engram rows ~48 GiB NVMe/node (./start-tp4.sh pack). Fabric: switched RoCE (both CX-7 rails rocep1s0f0,roceP2p1s0f0) for the production line; switchless ring is opt-in (docs/switchless-ring.md, DSV41_ROCE_RING + patched NCCL). Boot ~556 s cold / ~120 s with DSV41_FAST_LOAD; KV pool 6.8-7.3M tokens at 1M context; MEM_FRACTION_STATIC=0.80 keeps ~24 GiB/rank outside the static pool for the bounded prefill transient (indexer chunked 2 GiB rows, sglang#39187 backport). MEASURED FABRIC CAVEAT: headline numbers are on a switched RoCE fabric; a switchless ring without opposite-node hardware paths falls back to NCCL and decodes slower, and ring prefill is lower either way (one-link bisection). Known limits: prompt-token logprobs unavailable (--enable-decoder-swa-bounded-replay, requests 400, DSV41_REPLAY_GUARD), single-stream prose bounded by DSpark acceptance ~3 tok/step, fast loader leaves KV pool 3-13% smaller. Rollback: every layer is an env change (see README Rollback table); upstream MiaAI profile = MAX_RUNNING_REQUESTS=8, CHUNKED_PREFILL_SIZE=1024, EP_SIZE=4, EXTRA_CONTAINER_ENV unset."},"kv_pool_tokens":3670000,"max_context_len":1048576},"models":[{"slug":"deepseek-v4-1-flash","name":"DeepSeek-V4.1-Flash","hf_repo":"deepseek-ai/DeepSeek-V4.1-Flash","quant":"native-weights-mxfp4-experts-fp8-dense-deepseek-ai-release"}],"hardware":[{"slug":"dgx-spark-x4","vendor":"nvidia","card":"DGX Spark (GB10, unified memory)","card_family":"dgx-spark","vram_gb":128,"device_count":4}],"generation":"blackwell","tags":["dspark","long-context","multi-node","rdma","roce","spark-cluster","speculative","switched-fabric","tp4"],"tasks":["chat","code"],"aa":{"aa_index":39,"aa_provenance":"measured","aa_source_url":"https://artificialanalysis.ai/models/deepseek-v4-1-flash","aa_index_version":"Index v4.3.2","aa_methodology_version":"Index v4.3.2 (2026-10-18)","aa_published_fetched_at":"2026-10-20 21:49:11.309778-06","aa_family_slug":"deepseek-v4-1-flash","aa_resolved_slug":"deepseek-v4-1-flash"},"performance":[{"claimed_by":"submitter","metrics":"{\"notes\": \"sparkDash 1.8.8 at the fleet's 2200 MHz SM clock cap (nvidia-smi -lgc 0,2200, timer re-applied), idle fleet, thinking off, temperature 0, 256 new tokens; production stack v2.3 (canary-roce image, EP_SIZE=1, all v2.3 switches on), fresh clone @ 5f11a97, 2026-09-29, docs/results/validation-20260929-v23.txt. c1 = median of three runs; c4 median of two; c2/c8/c16 single runs. Non-prose columns use different prompts per concurrency (per-stream not comparable across columns). sparkDash prefill filler is one repeated token so the Engram row cache inflates prefill (MiaAI-Lab#21): on real text v2.2 measured 3915/4743/4822/4946/4800 tok/s at ~4k/15k/30k/56k/123k, unchanged in v2.3. Decode gains almost nothing above the 2200 MHz cap (uncapped: -1.5% prose, -0.7% code).\", \"tok_s\": 89.7000000000000028421709430404007434844970703125, \"points\": [{\"notes\": \"sparkDash prose c1 (median 89.72/89.0/90.23), thinking off\", \"concurrency\": 1, \"context_len\": 1048576, \"decode_tok_s\": 89.7000000000000028421709430404007434844970703125}, {\"notes\": \"sparkDash code c1 (131.52/131.92/132.05)\", \"concurrency\": 1, \"context_len\": 1048576, \"decode_tok_s\": 131.900000000000005684341886080801486968994140625}, {\"notes\": \"sparkDash structured c1 (157.77/160.11/155.66)\", \"concurrency\": 1, \"context_len\": 1048576, \"decode_tok_s\": 157.80000000000001136868377216160297393798828125}, {\"notes\": \"sparkDash json c1 (130.65/130.4/130.31)\", \"concurrency\": 1, \"context_len\": 1048576, \"decode_tok_s\": 130.400000000000005684341886080801486968994140625}, {\"notes\": \"sparkDash prose c2 aggregate (per-stream 64.2)\", \"concurrency\": 2, \"context_len\": 1048576, \"decode_tok_s\": 125.2000000000000028421709430404007434844970703125}, {\"notes\": \"sparkDash code c2 aggregate (per-stream 91.7)\", \"concurrency\": 2, \"context_len\": 1048576, \"decode_tok_s\": 181.400000000000005684341886080801486968994140625}, {\"notes\": \"sparkDash structured c2 aggregate (per-stream 106.9)\", \"concurrency\": 2, \"context_len\": 1048576, \"decode_tok_s\": 184.599999999999994315658113919198513031005859375}, {\"notes\": \"sparkDash json c2 aggregate (per-stream 92.8)\", \"concurrency\": 2, \"context_len\": 1048576, \"decode_tok_s\": 179.900000000000005684341886080801486968994140625}, {\"notes\": \"sparkDash prose c4 aggregate median (41.9/stream)\", \"concurrency\": 4, \"context_len\": 1048576, \"decode_tok_s\": 165.900000000000005684341886080801486968994140625}, {\"notes\": \"sparkDash code c4 aggregate median (63.5/stream)\", \"concurrency\": 4, \"context_len\": 1048576, \"decode_tok_s\": 248}, {\"notes\": \"sparkDash structured c4 aggregate median (70.6/stream)\", \"concurrency\": 4, \"context_len\": 1048576, \"decode_tok_s\": 242.099999999999994315658113919198513031005859375}, {\"notes\": \"sparkDash json c4 aggregate (80.9/stream)\", \"concurrency\": 4, \"context_len\": 1048576, \"decode_tok_s\": 319.80000000000001136868377216160297393798828125}, {\"notes\": \"sparkDash prose c8 aggregate (31.7/stream)\", \"concurrency\": 8, \"context_len\": 1048576, \"decode_tok_s\": 244.599999999999994315658113919198513031005859375}, {\"notes\": \"sparkDash code c8 aggregate (44.0/stream)\", \"concurrency\": 8, \"context_len\": 1048576, \"decode_tok_s\": 328.6000000000000227373675443232059478759765625}, {\"notes\": \"sparkDash structured c8 aggregate (44.6/stream)\", \"concurrency\": 8, \"context_len\": 1048576, \"decode_tok_s\": 300.6000000000000227373675443232059478759765625}, {\"notes\": \"sparkDash json c8 aggregate (61.0/stream)\", \"concurrency\": 8, \"context_len\": 1048576, \"decode_tok_s\": 477.30000000000001136868377216160297393798828125}, {\"notes\": \"sparkDash prose c16 aggregate (23.2/stream); MAX_RUNNING_REQUESTS=16\", \"concurrency\": 16, \"context_len\": 1048576, \"decode_tok_s\": 357.19999999999998863131622783839702606201171875}, {\"notes\": \"sparkDash code c16 aggregate (30.5/stream)\", \"concurrency\": 16, \"context_len\": 1048576, \"decode_tok_s\": 456.30000000000001136868377216160297393798828125}, {\"notes\": \"sparkDash structured c16 aggregate (47.1/stream)\", \"concurrency\": 16, \"context_len\": 1048576, \"decode_tok_s\": 603.3999999999999772626324556767940521240234375}, {\"notes\": \"sparkDash json c16 aggregate (44.8/stream)\", \"concurrency\": 16, \"context_len\": 1048576, \"decode_tok_s\": 693.200000000000045474735088646411895751953125}, {\"notes\": \"sparkDash prefill bench cold (salted prompts), median of runs 2-3 of three, 2200 MHz cap; filler caveat above\", \"context_len\": 4096, \"prefill_tok_s\": 4688}, {\"notes\": \"sparkDash prefill bench cold, 2200 MHz cap\", \"context_len\": 16384, \"prefill_tok_s\": 5583}, {\"notes\": \"sparkDash prefill bench cold, 2200 MHz cap; SPARK_PREFILL_TP_SPLIT from 32k\", \"context_len\": 32768, \"prefill_tok_s\": 5650}, {\"notes\": \"sparkDash prefill bench cold, 2200 MHz cap\", \"context_len\": 65536, \"prefill_tok_s\": 5664}, {\"notes\": \"sparkDash prefill bench cold, 2200 MHz cap\", \"context_len\": 131072, \"prefill_tok_s\": 5398}], \"context_len\": 1048576}"}],"perf_points":[{"id":2035,"context_len":4096,"prefill_tok_s":4688,"decode_tok_s":null,"ttft_ms":null,"concurrency":null,"claimed_by":"submitter","notes":"sparkDash prefill bench cold (salted prompts), median of runs 2-3 of three, 2200 MHz cap; filler caveat above"},{"id":2036,"context_len":16384,"prefill_tok_s":5583,"decode_tok_s":null,"ttft_ms":null,"concurrency":null,"claimed_by":"submitter","notes":"sparkDash prefill bench cold, 2200 MHz cap"},{"id":2037,"context_len":32768,"prefill_tok_s":5650,"decode_tok_s":null,"ttft_ms":null,"concurrency":null,"claimed_by":"submitter","notes":"sparkDash prefill bench cold, 2200 MHz cap; SPARK_PREFILL_TP_SPLIT from 32k"},{"id":2038,"context_len":65536,"prefill_tok_s":5664,"decode_tok_s":null,"ttft_ms":null,"concurrency":null,"claimed_by":"submitter","notes":"sparkDash prefill bench cold, 2200 MHz cap"},{"id":2039,"context_len":131072,"prefill_tok_s":5398,"decode_tok_s":null,"ttft_ms":null,"concurrency":null,"claimed_by":"submitter","notes":"sparkDash prefill bench cold, 2200 MHz cap"},{"id":2030,"context_len":1048576,"prefill_tok_s":null,"decode_tok_s":89.7000000000000028421709430404007434844970703125,"ttft_ms":null,"concurrency":1,"claimed_by":"submitter","notes":"sparkDash prose c1 (median 89.72/89.0/90.23), thinking off"},{"id":2031,"context_len":1048576,"prefill_tok_s":null,"decode_tok_s":125.2000000000000028421709430404007434844970703125,"ttft_ms":null,"concurrency":2,"claimed_by":"submitter","notes":"sparkDash prose c2 aggregate (per-stream 64.2)"},{"id":2032,"context_len":1048576,"prefill_tok_s":null,"decode_tok_s":165.900000000000005684341886080801486968994140625,"ttft_ms":null,"concurrency":4,"claimed_by":"submitter","notes":"sparkDash prose c4 aggregate median (41.9/stream)"},{"id":2033,"context_len":1048576,"prefill_tok_s":null,"decode_tok_s":244.599999999999994315658113919198513031005859375,"ttft_ms":null,"concurrency":8,"claimed_by":"submitter","notes":"sparkDash prose c8 aggregate (31.7/stream)"},{"id":2034,"context_len":1048576,"prefill_tok_s":null,"decode_tok_s":357.19999999999998863131622783839702606201171875,"ttft_ms":null,"concurrency":16,"claimed_by":"submitter","notes":"sparkDash prose c16 aggregate (23.2/stream); MAX_RUNNING_REQUESTS=16"}],"agent":{"prompt":"Fetch https://localairecipes.com/api/v1/recipes/deepseek-v4-1-flash-on-4x-dgx-spark-tuned-tp4-sglang-dspark.json and apply the inference recipe it describes (engine flags, model, quant, context) to my setup. Ask before overwriting any existing config.","curl":"curl -fsSL https://localairecipes.com/api/v1/recipes/deepseek-v4-1-flash-on-4x-dgx-spark-tuned-tp4-sglang-dspark.json"},"requirements":{"requires":["Docker (or a compatible OCI container runtime)","NVIDIA Container Toolkit (nvidia-container-toolkit) - exposes host GPUs to the container"],"driver_floor":{"cuda":null,"linux_min":null,"text":"unknown - check the image tag","source":"NVIDIA CUDA Toolkit Release Notes - driver compatibility tables (https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/, fetched 2026-10-03)"},"placeholders":["Hardware-bound values in this recipe (device counts, bind-mount paths, utilization, ports) describe the author's machine, not yours - map them to your own rig before launching. Keep the engine flags, model, quant, and parallelism strategy; adjust paths and ports to the hardware you actually have.","GPU identity: list your GPUs with `nvidia-smi -L` and substitute your own GPU UUIDs/indices wherever the recipe pins the author's.","Model directory: bind-mount paths are author-side - point them at your own mount where the model files live.","Host port: pick a free port on your machine and publish the container there; keep the in-container port unchanged."]},"artifacts":[{"kind":"model_weights","pointer_type":"hf_model","pointer":"https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash","pin":"fb2764a5cf321eaa5070ca8f9e892818f477c16d \u2014 476 GiB, 48 shards, native MXFP4/FP8 release"},{"kind":"engine_image","pointer_type":"git_commit","pointer":"https://github.com/knapcio/DeepSeek-V4.1-Flash-4x-DGX-Spark-TP4","pin":"fa1356384ed246f3040775210668f02a1bba3426 @ HEAD 'docs: bring-up pitfalls, ring recovery notes, results from other fleets' \u2014 Dockerfile.canary-roce + .env.tp4.example full production EXTRA_CONTAINER_ENV line"},{"kind":"reference","pointer_type":"git_commit","pointer":"https://github.com/MiaAI-Lab/DeepSeek-v4.1-Flash-DGX-Sparks","pin":"cad252b7d3000cf21d919764e70a6912b6d3b0b0 @ main (verified via git ls-remote 2026-10-10) \u2014 upstream recipe: launcher, Engram store, b12x routing, docs/README-upstream.md preserved verbatim in the downstream repo"},{"kind":"reference","pointer_type":"git_commit","pointer":"https://github.com/sgl-project/sglang","pin":"f80c91a4b \u2014 dsv4.1 branch (2026-09-16), staged by scripts/fetch-sglang-canary.sh (SGLANG_CANARY_REF default); carries sgl-project#39370/#39646/#39648/#39653 branch kernels; #39187 bounded dense-indexer prefill backported as adapter/indexer_chunked*.py"},{"kind":"reference","pointer_type":"git_commit","pointer":"https://github.com/local-inference-lab/b12x","pin":"a7d7d29b2ef8869086e0ceaa787321f17544e3c9 \u2014 b12x main (runtime/b12x_next, two-line 64-row-tile patch); runtime/b12x = RoCEnante frozen at the rhys101 SG17 revision"}],"links":{"self":"/api/v1/recipes/deepseek-v4-1-flash-on-4x-dgx-spark-tuned-tp4-sglang-dspark.json","alternate":"/r/deepseek-v4-1-flash-on-4x-dgx-spark-tuned-tp4-sglang-dspark","describedby":"/openapi.json"},"updated_at":"2026-10-29 21:32:23.622356-06","created_at":"2026-10-09 19:53:14.926777-06","speeds":{"decode_tok_s":89.7000000000000028421709430404007434844970703125,"prefill_tok_s":5650,"leaderboard":"/api/v1/leaderboard.json"}}