| --- |
| license: apache-2.0 |
| tags: |
| - deepseek-v4 |
| - vllm |
| - dgx-spark |
| - blackwell |
| - sm120 |
| - sm121 |
| - docker-image |
| - gb10 |
| --- |
| |
| # vllm-w4a16-dsv4:exp — pre-built Docker image for DGX Spark TP=2 |
|
|
| This dataset hosts a pre-built OCI tarball of the Docker image used to serve |
| [`pastapaul/DeepSeek-V4-Flash-W4A16-FP8`](https://huggingface.co/pastapaul/DeepSeek-V4-Flash-W4A16-FP8) |
| on dual DGX Spark GB10 (SM 12.1a) — for users who can't build the image |
| themselves due to network constraints. |
|
|
| The image is also fine for 2× RTX PRO 6000 Blackwell Server (SM 12.0). |
|
|
| ## What's inside |
|
|
| vLLM build pinned to: |
|
|
| - **`jasl/vllm@ds4-sm120-experimental`** (HEAD `c05638d70` after cherry-pick) — SM12x DSV4 support + experimental superset (split-KV decode, GB10 fused-MoE config aliases, tuned MLA graph defaults) |
| - **Cherry-pick** `f910a73a93` from `neuralmagic/vllm@kylesayrs/deepseek-ct` (vLLM PR #41276 — DSV4 compressed-tensors attention) |
| - **Local patch** [`patch_v4_packed_mapping.py`](https://github.com/pasta-paul/dsv4-flash-w4a16-fp8/blob/main/scripts/patch_v4_packed_mapping.py) — adds `packed_modules_mapping` to `DeepseekV4ForCausalLM` |
| - Workspace pre-reservation patch is **not** applied — landed upstream as `jasl/vllm@1d6f5c4` |
|
|
| Compressed tarball: ~9 GB. Uncompressed image: 20.2 GB. ARM64 only. |
|
|
| ## How to use |
|
|
| ```bash |
| # Download |
| huggingface-cli download pastapaul/dsv4-flash-w4a16-spark-image \ |
| vllm-w4a16-dsv4-exp.tar.gz \ |
| --repo-type dataset \ |
| --local-dir . |
| |
| # Load on each Spark |
| gunzip -c vllm-w4a16-dsv4-exp.tar.gz | docker load |
| |
| # Verify |
| docker images vllm-w4a16-dsv4:exp |
| ``` |
|
|
| Once loaded on both Sparks, you can use the bootstrap script with `--skip-build`: |
|
|
| ```bash |
| curl -fsSLO https://raw.githubusercontent.com/pasta-paul/dsv4-flash-w4a16-fp8/main/scripts/bootstrap_dsv4_spark.sh |
| chmod +x bootstrap_dsv4_spark.sh |
| ./bootstrap_dsv4_spark.sh \ |
| --head-host spark-a \ |
| --worker-host spark-b \ |
| --skip-build |
| ``` |
|
|
| The script will then handle network setup, model pre-cache (no HF token needed), container launch on both nodes, and wait for `/health=200`. |
|
|
| ## What you get |
|
|
| Production canonical at **1 M-token context** TP=2 graphs-ON: |
| - decode 12 t/s smoke / 14–15 t/s think-max sustained |
| - NIAH 4/4 retrieval at 200K-token haystack |
| - mini-suite 10/10 PASS · think-max 3/3 PASS |
| - GSM8K 95.00% strict / 94.92% flex (preserved from prior canonical) |
| - tool calling: parallel function calls with structured arguments |
|
|
| ## Reference docs |
|
|
| - Model: https://huggingface.co/pastapaul/DeepSeek-V4-Flash-W4A16-FP8 |
| - Reproduction repo + raw evidence: https://github.com/pasta-paul/dsv4-flash-w4a16-fp8 |
| - Quickstart: https://github.com/pasta-paul/dsv4-flash-w4a16-fp8/blob/main/findings/QUICKSTART_DUAL_SPARK.md |
| - Phase 4e validation report: https://github.com/pasta-paul/dsv4-flash-w4a16-fp8/blob/main/findings/spark_tp2_deployment.md#phase-4e--production-canonical-at-1-m-context-on-ds4-sm120-experimental-2026-05-06 |
|
|
