{"slug":"deepseek-v4-1-flash-1017-sage-exl3-1-62bpw-1x-dgx-spark-dspark","title":"DeepSeek-V4.1-Flash-1017 SAGE EXL3 1.62bpw 1x DGX Spark DSpark","summary":"DeepSeek-V4.1-Flash-1017 at 1.62 bpw (SAGE MixedK EXL3 routed experts, EXL3 MTP draft experts, EXL3 K8 output head, block-32 FP8 non-routed linears) served on one NVIDIA DGX Spark (GB10, 121.7 GiB unified). The full pack does not fit, so the 1,418 least-routed experts plus the embedding are gathered into an 8.59 GiB read-only cold-expert alias side file while every hot expert stays device-resident; the ~111 GB Engram n-gram tables are streamed from NVMe. vLLM TP1 with TileLang attention/linear/MoE kernels, DSpark spec decode k=5, 524,288-token window, 1 GiB KV pool (959,768 tokens), armed memory guard. Author-measured on the serving build: 35.1 tok/s mean decode at 1 request (137.4 ms/step, 4.73 accepted tok/step), ~720 tok/s real-text prefill, 3 min 11 s to API-ready, 2.9-3.2 GiB MemAvailable at 524K. recipe by vcruz305.","license":"MIT","tier":"community","beta":false,"revision":{"no":1,"commit_sha":"06d13ad43d104509b61008999405f0b106753fe2","content_hash":"3bbab54f4d76f4f5c0ced95fb11bbe09f0043f20","created_at":"2026-11-04 21:55:48.655707-07","notes":"Community ingest by localai-recipes (lar-scout 2026-11-05 sweep) of vcruz305's DS41 SAGE EXL3 1.62bpw recipe @ 8036aa74 - HEAD at sweep == HEAD at ingest, drift-free; all perf numbers are the author's own measurements at this pin (T0 claimed). NOT a dupe: the only other DS41-on-Spark row is the 4x-DGX-Spark TP4 SGLang tuned recipe (4 nodes); every other DS41 card is 3x/4x CMP 170HX, 3x RTX PRO 6000, 2x DGX Station or GB300. This is the first DS41 1.62bpw SAGE EXL3 single-Spark vLLM recipe on the site (cold-expert ATS alias side file + Engram streamed from NVMe + TileLang kernels + memory guard). CORRECTIONS to the scout card: (1) the card's flags string dropped measured flags present in packs/1.62bpw/scripts/serve_inner.sh at the pin (--speculative-config dspark k=5 with adaptive verification, --override-generation-config, --max-cudagraph-capture-size, --reasoning-parser/--tool-call-parser/--enable-auto-tool-choice, --async-scheduling, --no-scheduler-reserve-full-isl, --no-enable-flashinfer-autotune, --language-model-only, --mm-processor-cache-gb 0, --hf-overrides) - launch argv above is transcribed from serve_inner.sh + serve.sh at 8036aa74; (2) the card quoted 31.3-34.0 tok/s (pre-W2 tile-K 16 row) as the best measured - at this pin the serving build is the W2 row, mean 35.1 tok/s at 524288 (39.6 with prefix caching on), both rows kept as perf points; (3) the cold-alias selection has 1,418 experts over 43 keys (40 layers + 3 MTP) in configs/cold-alias-selection.json. scripts/env.sh + packs/1.62bpw/scripts/env.sh are the pin authority (BASE_HF_REV, FORK_REF, PLUGIN_REF, MOE_TILESIZE_K)."},"engine_source":{"type":"docker_image","git_url":"https://github.com/vcruz305/DSV4.1-Flash-1017-EXL3-Spark-recipe","base_engine":"vllm","pip_package":null,"base_version":"dsv41-1017-exl3:r6w (derived image, built locally)","image_digest":"sha256:8d743923e09ad9d435255f9436b5aa1a42d2e88cced04d121b238f8a726990cc","install_note":"Derived image dsv41-1017-exl3:r6w = Christopher Owen's spark-ds41f r6w base (upstreams pinned at 2dc95329) + vcruz305/exllamav3 feat/moe-warp-private-stages @ 909ad8564fcb + vcruz305/vllm-exl3 feat/dsv41-1017-tp1 @ 30e8b941376c + quant-config shim + TP1 decode tiles; MoE tile-K compiled at 16; built by scripts/build_image.sh (base fetched from vcruz305/DSV4.1-Flash-1017-Spark-base-image @ 17ed433666f4, archive sha256 97d31344b0e5f8e17bf69e05d679a25b3256a452a7275bc294eb6bd149f58e9a, layer sha256:8d743923e09a...). Needs passwordless sudo -n docker, zstd, huggingface_hub[hf_xet]; sudoers must grant docker, sysctl for the three vm.* keys, swapoff/swapon, tee /proc/sys/vm/drop_caches and pkill. The 326.6 GB pack plus 8.6 GiB alias side file plus ~37 GB images need ~400 GB free NVMe. HF token required (model repo is gated - accept terms once). guard.sh MUST be armed.","pinned_commit":"8036aa741e2c2328035ed546f1f844b6797775f6"},"credit":[{"name":"vcruz305 (recipe author, SAGE quant, packs, measurements, MIT)","url":"https://github.com/vcruz305/DSV4.1-Flash-1017-EXL3-Spark-recipe"},{"name":"Christopher Owen (spark-ds41f r6w base image, upstreams pinned at 2dc95329)","url":"https://github.com/christopherowen/spark-ds41f"},{"name":"vcruz305/exllamav3 feat/moe-warp-private-stages @ 909ad8564c - mixed-K unified MoE + warp-private pipeline stages","url":"https://github.com/vcruz305/exllamav3"},{"name":"vcruz305/vllm-exl3 feat/dsv41-1017-tp1 @ 30e8b94137 - Exl3MoEMethod / ATS loader","url":"https://github.com/vcruz305/vllm-exl3"},{"name":"AgentSparko (REXL3 overlay path used by the sibling 1.55bpw pack)","url":"https://github.com/vcruz305/DSV4.1-Flash-1017-EXL3-Spark-recipe"},{"name":"deepseek-ai/DeepSeek-V4.1-Flash-1017 - upstream weights (MIT)","url":"https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017"}],"author":{"name":"localai-recipes","isSite":false},"uploaded_by":"localai-recipes","config":{"launch":["bash packs/1.62bpw/scripts/download_and_verify.sh (hf_xet download of vcruz305/DSV4.1-Flash-1017-SAGE-EXL3-1.62bpw @ c852150e06f5... + sha256 verify vs packs/1.62bpw/scripts/SHA256SUMS - the SUMS file is the authority, the pinned revision may stop resolving); bash scripts/build_image.sh; bash packs/1.62bpw/scripts/build_cold_alias.sh (8.59 GiB read-only side file, ~10-20 s)","bash packs/1.62bpw/scripts/serve.sh: host wrapper - sets vm.compaction_proactiveness=0 vm.watermark_boost_factor=0 vm.swappiness=0, swapoff/swapon, drop_caches, refuses a stale alias manifest (serve.sh compares ALIAS_DIR/manifest.json against configs/cold-alias-selection.json) or an image whose build labels do not match the pins (PLUGIN_REF 30e8b941376c, FORK_REF 909ad8564fcb, MOE_TILESIZE_K=16), then runs the docker command below and arms scripts/guard.sh. DRY_RUN=1 prints the plan.","Env knobs of the measured run (scripts/env.sh defaults): CONTAINER=dsv41-tp1 API_PORT=8000 SERVED_MODEL_NAME=dsv41 MAXLEN=524288 KVB=1073741824 MNBT=2112 LPT=2047 MTP_K=5 CG_MODE=FULL_DECODE_ONLY CAPTURE_SIZES=1,2,3,4,5,6,7 PREFIX_CACHING=1 DEFAULT_MAX_TOKENS=32768 DEFAULT_TEMPERATURE=0.6 DEFAULT_TOP_P=0.95 ALIAS_DIR=$HOME/.cache/dsv41-ds41f/alias-cold MOE_TILESIZE_K=16","sudo -n docker run -d --name dsv41-tp1 --gpus all --memory=118g --memory-swap=118g --oom-score-adj 1000 --ulimit core=0 --shm-size=2g --security-opt=seccomp=unconfined --security-opt=label=disable --ulimit=memlock=-1:-1 -v $HOME/models:/models:ro -v $HOME/.cache/dsv41-ds41f:/cache:rw -v $HOME/.cache/dsv41-ds41f/alias-cold:/alias:ro -v <RUN>/ctr:/mem:rw -v packs/1.62bpw/scripts/serve_inner.sh:/root/serve.sh:ro -p 8000:8000 --entrypoint /bin/bash dsv41-1017-exl3:r6w /root/serve.sh","container env: HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 B12X_AUTOTUNE=0 B12X_BIND_COMPILE_WORKERS=2 B12X_STATE_COMPILE_WORKERS=2 B12X_WEIGHTS_COMPILE_WORKERS=4 B12X_DENSE_SPLITK_TURBO=1 B12X_W4A8_TINY_DECODE=0 CUDA_DEVICE_MAX_CONNECTIONS=32 VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 VLLM_USE_FLASHINFER_SAMPLER=1 VLLM_ENGINE_READY_TIMEOUT_S=3600 VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS=900","container env (EXL3/ATS): VLLM_EXL3_MOE_KERNEL=exllamav3 VLLM_EXL3_V41_NATIVE_MOE=0 VLLM_EXL3_MIXEDK_LEGACY=1 VLLM_EXL3_MIXEDK_MODE=unified VLLM_EXL3_TRIM_AFTER_LAYER=1 VLLM_EXL3_MODEL_DIR=/models/dsv41-1.62bpw-v3 VLLM_EXL3_ATS_ALIAS=1 VLLM_EXL3_ATS_ALIAS_DIR=/alias VLLM_EXL3_ATS_ALIAS_LAYERS=38,39 VLLM_EXL3_ATS_ALIAS_MTP=1 VLLM_EXL3_ATS_ALIAS_EMBED=1 VLLM_EXL3_ATS_PREFETCH=1","container env (kernels/stops): VLLM_DS41_ROPE_TRIM=1 VLLM_DS41_ROPE_TRIM_MAX_POS=524288 VLLM_DS41_MARKOV_NVFP4=1 VLLM_DS41_L2_PREFETCH=0 VLLM_DS41_ATTENTION_COMPUTE=auto VLLM_DS41_KERNEL_BACKEND=tilelang VLLM_DS41_HEAD_DTYPE=float32 DSV41_DSML_STOP=1 DSV41_DSML_PARAM_REPAIR=1 DSV41_REP_STOP=64:512:8; JIT caches under $HOME/.cache/dsv41-ds41f/kkref/jit (mounted /cache)","scripts/guard.sh watch dsv41-tp1 <stage> <guard.log> 524288 <base_swap>: MUST be armed - 0.5 GiB MemAvailable floor, 64 MiB swap-growth allowance, 0.05 s sampling; kills the container on breach. guard.sh stop restores the sysctls","vllm serve /models/dsv41-1.62bpw-v3","--served-model-name","dsv41","--host","0.0.0.0","--port","8000","--tokenizer-mode","deepseek_v41","--dtype","bfloat16","--tensor-parallel-size","1","--distributed-executor-backend","uni","--quantization","exl3_dsv41","--block-size","256","--enable-prefix-caching","--enable-chunked-prefill","--max-model-len","524288","--max-num-seqs","1","--max-parallel-prefills","1","--max-num-batched-tokens","2112","--gpu-memory-utilization","0.90","--kv-cache-memory-bytes","1073741824","--generation-config","vllm","--override-generation-config","{\"max_new_tokens\": 32768,\"temperature\": 0.6,\"top_p\": 0.95}","--engram-config","{\"table_memory\":\"disk\",\"disk_resident_scales\":false,\"projection_tp\":true}","--attention-backend","tilelang","--linear-backend","tilelang","--moe-backend","tilelang","--jit-monitor-mode","warn","--compilation-config","{\"cudagraph_mode\":\"FULL_DECODE_ONLY\",\"custom_ops\":[\"all\"],\"cudagraph_capture_sizes\":[1,2,3,4,5,6,7]}","--max-cudagraph-capture-size","7","--speculative-config","{\"method\":\"dspark\",\"num_speculative_tokens\":5,\"draft_tensor_parallel_size\":1,\"attention_backend\":\"tilelang\",\"draft_sample_method\":\"greedy\",\"rejection_sample_method\":\"block\",\"enable_adaptive_verification\":true,\"adaptive_verification_row_cost\":0.01}","--reasoning-parser","deepseek_v41","--tool-call-parser","deepseek_v41","--enable-auto-tool-choice","--long-prefill-token-threshold","2047","--async-scheduling","--no-scheduler-reserve-full-isl","--no-enable-flashinfer-autotune","--language-model-only","--mm-processor-cache-gb","0","--hf-overrides","{\"quantization_config\":{\"quant_method\":\"exl3\",\"bits\":2,\"codebook\":\"mcg\",\"scope\":\"deepseek_v41_routed_experts\",\"non_routed_quantization\":{\"quant_method\":\"deepseek_v41_fp8\",\"weight_block_size\":[32,32],\"is_checkpoint_fp8_serialized\":true,\"activation_scheme\":\"dynamic\"},\"mtp_experts\":\"exl3\"}}","Do NOT add --enforce-eager (the 74.4 ms/token no-draft path); do NOT raise --max-num-batched-tokens past 2112 (8192 OOMs); --gpu-memory-utilization is INERT because serve_inner.sh always passes --kv-cache-memory-bytes; MAXLEN=32768 is 1-5% faster on short prompts.","DEFAULT_MAX_TOKENS/DEFAULT_TEMPERATURE/DEFAULT_TOP_P (32768/0.6/0.95) exist because agent clients send none - without them replies ran 906 vs 277 completion tokens and 83% vs 89% valid tool calls. The repo also carries a sibling SAGE REXL3 1.55bpw pack (packs/1.55bpw-rexl3/, AgentSparko overlay, MAX_NUM_SEQS>1, vision) sharing the same engine - not part of this measured recipe. scripts/env.sh + packs/1.62bpw/scripts/env.sh are the pin authority."],"layouts":["tp"],"kv_pool_tokens":959768,"max_context_len":524288,"storage_requirements":{"note":"326.6 GB pack + 8.59 GiB alias side file + ~37 GB images need ~400 GB free NVMe; the ~111 GB Engram tables and the cold experts are read from NVMe on the serving path.","nvme_gb":400}},"models":[{"slug":"deepseek-v4-1-flash-1017-sage-exl3-1-62bpw","name":"DeepSeek-V4.1-Flash-1017-SAGE-EXL3-1.62bpw","hf_repo":"vcruz305/DSV4.1-Flash-1017-SAGE-EXL3-1.62bpw","quant":"exl3-sage-mixedk-1-62bpw"}],"hardware":[{"slug":"dgx-spark","vendor":"nvidia","card":"DGX Spark (GB10, unified memory)","card_family":"dgx-spark","vram_gb":128,"device_count":1}],"generation":"blackwell","tags":["cold-expert-alias","dspark","engram-nvme","exl3","speculative","tilelang","tp1"],"tasks":["chat","code"],"aa":{"aa_index":45,"aa_provenance":"measured","aa_source_url":"https://artificialanalysis.ai/models/deepseek-v4-1-flash-1017","aa_index_version":null,"aa_methodology_version":"v4.3.2","aa_published_fetched_at":"2026-10-19 12:00:00-06","aa_family_slug":"deepseek-v4-1-flash-1017","aa_resolved_slug":"deepseek-v4-1-flash-1017"},"performance":[{"claimed_by":"submitter","metrics":"{\"notes\": \"Author-measured on one DGX Spark 2026-10 (README speed tables at recipe pin 8036aa74). Decode = client-observed tok/s, engine's own spec-decode counters; bench7.py single stream 512 tokens. Serving build (W2 warp-private MoE kernel, shipped cold set 1,418 experts/43 keys, MAXLEN=524288): four boots mean 137.4 ms/step, 4.73 tok/step, 35.1 tok/s. Cold set at MAXLEN=32768 (another GB10 box): 129.7/131.3 ms/step, 4.69, 35.6/36.3 tok/s. Pre-W2 tile-K 16 k=5: 140.0-145.6 ms/step, 4.35-4.74 tok/step, 31.3-34.0 tok/s; 39.6 tok/s with prefix caching on at 524288 (127.6 ms/step, 5.07). Prefill ~720 tok/s varied real text; 1,179 tok/s at 131,034-token needle (229,376, TTFT 111.1 s); 519,984-token prefill at 524,288 TTFT 516.4 s, decode 29.3, min MemAvailable 2.23 GiB. Bring-up 3 min 11 s to /v1/models 200. Memory: engine reservation 110.5 GiB at ready; steady MemAvailable ~4.0 GiB at 32768, 2.9-3.2 at 524288, 0 swap growth. Honest limit: 50-60 tok/s is NOT reached; ~31-39 is the best measured.\", \"tok_s\": 35.10000000000000142108547152020037174224853515625, \"points\": [{\"notes\": \"HEADLINE decode c1, serving build (W2 warp-private MoE kernel, shipped cold set), mean of four boots at MAXLEN=524288: 137.4 ms/step, 4.73 accepted tok/step (bench7, 512 tokens).\", \"concurrency\": 1, \"context_len\": 524288, \"decode_tok_s\": 35.10000000000000142108547152020037174224853515625}, {\"notes\": \"Best measured decode c1 with prefix caching on (default), bench7 at 524288: 127.6 ms/step, 5.07 tok/step. context_len is the tuple marker 524288+10240 (prefix-cache row).\", \"concurrency\": 1, \"context_len\": 534528, \"decode_tok_s\": 39.60000000000000142108547152020037174224853515625}, {\"notes\": \"Pre-W2 tile-K 16 build k=5 at MAXLEN=32768: 140.0-145.6 ms/step, 4.35-4.74 tok/step, range 31.3-34.0 tok/s (34.0 = best of the range). This is the configuration the scout card quoted.\", \"concurrency\": 1, \"context_len\": 32768, \"decode_tok_s\": 34}, {\"notes\": \"Shipped cold set (1,418 experts) at MAXLEN=32768 on another GB10 box: 129.7/131.3 ms/step, 4.69 tok/step, 35.6/36.3 tok/s. context_len is the tuple marker 32768+10240 (32768 row).\", \"concurrency\": 1, \"context_len\": 43008, \"decode_tok_s\": 36.2999999999999971578290569595992565155029296875}, {\"notes\": \"Long-context decode at depth: 519,984-token prompt (needle depth 0.5) at MAXLEN=524288, retrieved, min MemAvailable 2.23 GiB.\", \"concurrency\": 1, \"context_len\": 519984, \"decode_tok_s\": 29.300000000000000710542735760100185871124267578125}, {\"notes\": \"Real-text varied prefill c1 (prefill_varied_probe.py, python stdlib source): ~720 tok/s at ~14,000 prompt tokens (533 at 1.6K, 718 at 7K, 963 at 24K). concurrency omitted (NULL).\", \"context_len\": 14000, \"prefill_tok_s\": 720}, {\"notes\": \"Needle probe: 131,034-token prompt at depth 0.5, MAXLEN=229,376, TTFT 111.1 s, retrieved, min MemAvailable 3.13 GiB. concurrency omitted (NULL).\", \"ttft_ms\": 111100, \"context_len\": 131034, \"prefill_tok_s\": 1179}, {\"notes\": \"Full prefill of the 519,984-token prompt at MAXLEN=524288: TTFT 516.4 s, 1,007 tok/s, retrieved. concurrency omitted (NULL).\", \"ttft_ms\": 516400, \"context_len\": 519984, \"prefill_tok_s\": 1007}, {\"notes\": \"Shipped cold set: 518,103-token full prefill at MAXLEN=524288, TTFT 470.2 s, 1,102 tok/s, retrieved, min MemAvailable 2.19 GiB. concurrency omitted (NULL).\", \"ttft_ms\": 470200, \"context_len\": 518103, \"prefill_tok_s\": 1102}], \"context_len\": 524288}"}],"perf_points":[{"id":2438,"context_len":14000,"prefill_tok_s":720,"decode_tok_s":null,"ttft_ms":null,"concurrency":null,"claimed_by":"submitter","notes":"Real-text varied prefill c1 (prefill_varied_probe.py, python stdlib source): ~720 tok/s at ~14,000 prompt tokens (533 at 1.6K, 718 at 7K, 963 at 24K). concurrency omitted (NULL)."},{"id":2435,"context_len":32768,"prefill_tok_s":null,"decode_tok_s":34,"ttft_ms":null,"concurrency":1,"claimed_by":"submitter","notes":"Pre-W2 tile-K 16 build k=5 at MAXLEN=32768: 140.0-145.6 ms/step, 4.35-4.74 tok/step, range 31.3-34.0 tok/s (34.0 = best of the range). This is the configuration the scout card quoted."},{"id":2436,"context_len":43008,"prefill_tok_s":null,"decode_tok_s":36.2999999999999971578290569595992565155029296875,"ttft_ms":null,"concurrency":1,"claimed_by":"submitter","notes":"Shipped cold set (1,418 experts) at MAXLEN=32768 on another GB10 box: 129.7/131.3 ms/step, 4.69 tok/step, 35.6/36.3 tok/s. context_len is the tuple marker 32768+10240 (32768 row)."},{"id":2439,"context_len":131034,"prefill_tok_s":1179,"decode_tok_s":null,"ttft_ms":111100,"concurrency":null,"claimed_by":"submitter","notes":"Needle probe: 131,034-token prompt at depth 0.5, MAXLEN=229,376, TTFT 111.1 s, retrieved, min MemAvailable 3.13 GiB. concurrency omitted (NULL)."},{"id":2441,"context_len":518103,"prefill_tok_s":1102,"decode_tok_s":null,"ttft_ms":470200,"concurrency":null,"claimed_by":"submitter","notes":"Shipped cold set: 518,103-token full prefill at MAXLEN=524288, TTFT 470.2 s, 1,102 tok/s, retrieved, min MemAvailable 2.19 GiB. concurrency omitted (NULL)."},{"id":2440,"context_len":519984,"prefill_tok_s":1007,"decode_tok_s":null,"ttft_ms":516400,"concurrency":null,"claimed_by":"submitter","notes":"Full prefill of the 519,984-token prompt at MAXLEN=524288: TTFT 516.4 s, 1,007 tok/s, retrieved. concurrency omitted (NULL)."},{"id":2437,"context_len":519984,"prefill_tok_s":null,"decode_tok_s":29.300000000000000710542735760100185871124267578125,"ttft_ms":null,"concurrency":1,"claimed_by":"submitter","notes":"Long-context decode at depth: 519,984-token prompt (needle depth 0.5) at MAXLEN=524288, retrieved, min MemAvailable 2.23 GiB."},{"id":2433,"context_len":524288,"prefill_tok_s":null,"decode_tok_s":35.10000000000000142108547152020037174224853515625,"ttft_ms":null,"concurrency":1,"claimed_by":"submitter","notes":"HEADLINE decode c1, serving build (W2 warp-private MoE kernel, shipped cold set), mean of four boots at MAXLEN=524288: 137.4 ms/step, 4.73 accepted tok/step (bench7, 512 tokens)."},{"id":2434,"context_len":534528,"prefill_tok_s":null,"decode_tok_s":39.60000000000000142108547152020037174224853515625,"ttft_ms":null,"concurrency":1,"claimed_by":"submitter","notes":"Best measured decode c1 with prefix caching on (default), bench7 at 524288: 127.6 ms/step, 5.07 tok/step. context_len is the tuple marker 524288+10240 (prefix-cache row)."}],"agent":{"prompt":"Fetch https://localairecipes.com/api/v1/recipes/deepseek-v4-1-flash-1017-sage-exl3-1-62bpw-1x-dgx-spark-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-1017-sage-exl3-1-62bpw-1x-dgx-spark-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":"reference","pointer_type":"git_commit","pointer":"https://github.com/vcruz305/DSV4.1-Flash-1017-EXL3-Spark-recipe","pin":"8036aa741e2c2328035ed546f1f844b6797775f6 (main, MIT; pack scripts + guard + bench suite; SHA256SUMS is the pack authority)"},{"kind":"model_weights","pointer_type":"hf_model","pointer":"https://huggingface.co/vcruz305/DSV4.1-Flash-1017-SAGE-EXL3-1.62bpw","pin":"c852150e06f5e5ee8c4c14f9cd1a9afc6bb21960 (gated; history replaced multiple times - SHA256SUMS decides)"},{"kind":"other","pointer_type":"hf_model","pointer":"https://huggingface.co/vcruz305/DSV4.1-Flash-1017-Spark-base-image","pin":"17ed433666f4cc6dfa86bcc1d3f4448255fd58fb (prebuilt spark-ds41f r6w base, archive sha256 97d31344b0e5f8e17bf69e05d679a25b3256a452a7275bc294eb6bd149f58e9a)"}],"links":{"self":"/api/v1/recipes/deepseek-v4-1-flash-1017-sage-exl3-1-62bpw-1x-dgx-spark-dspark.json","alternate":"/r/deepseek-v4-1-flash-1017-sage-exl3-1-62bpw-1x-dgx-spark-dspark","describedby":"/openapi.json"},"updated_at":"2026-11-04 21:55:48.655707-07","created_at":"2026-11-04 21:55:48.655707-07","speeds":{"decode_tok_s":35.10000000000000142108547152020037174224853515625,"prefill_tok_s":1102,"leaderboard":"/api/v1/leaderboard.json"}}