Use this page as the stable entry point for DeepSeek-V4-Flash and DSpark on RTX PRO 6000 Blackwell. Release pages are immutable measurement and deployment specifications.
| Need | Page |
|---|---|
| Serve the qualified fixed-K5 source composition | DeepSeek-V4-Flash-0731 Infernal Invocation r21 |
| Test the target-only 2.11M-KV capacity profile | DeepSeek-V4-Flash-0731 Infernal Invocation r19 preview |
| Serve the 0731 DSpark checkpoint | DeepSeek-V4-Flash-0731 Infernal Invocation r18 |
| Test topology-calibrated B12X transport | B12X PCIe transport calibration |
| Inspect the Gilded Gnosis baseline | DeepSeek-V4-Flash-0731 Gilded Gnosis r33 |
| Inspect the Fathomless TP2/TP4 sweep | DeepSeek-V4-Flash v10 Fathomless Validation |
| Inspect the full DSpark and standard-MTP sweep | DeepSeek-V4-Flash and DSpark v9 |
| Diagnose empty reasoning before tool calls | DS4 empty-think troubleshooting |
| Area | Specification |
|---|---|
| Qualified DSpark K5 image line | Infernal Invocation r21 for deepseek-ai/DeepSeek-V4-Flash-0731 |
| Target-only capacity study | Infernal Invocation r19 for deepseek-ai/DeepSeek-V4-Flash-0731 |
| DSpark checkpoint | deepseek-ai/DeepSeek-V4-Flash-0731 |
| Standard-MTP checkpoint | deepseek-ai/DeepSeek-V4-Flash |
| Archived DSpark checkpoint | deepseek-ai/DeepSeek-V4-Flash-DSpark |
| General-purpose DSpark profile | fixed probabilistic K5, InstantTensor BUFFERED, B12X W4A8, FP8 compressed MLA KV |
| KV offload | Native vLLM CPU/filesystem tiers or LMCache L1/filesystem L2; use one ownership model per server |
| Backend family | B12X; archived pages also describe SparkInfer, Lucifer, and CUTLASS implementations |
| Speculative decoding | Standard MTP and DSpark use different checkpoint, graph, and verifier contracts |
| Source line | Revision namespace | Serving specification |
|---|---|---|
dev/infernal-invocation |
Infernal Invocation r* |
r21 qualified source composition, r19 capacity study |
dev/gilded-gnosis |
Gilded Gnosis v20-r* |
Gilded Gnosis r33 |
| Fathomless Firmament | v9 and v10 |
v10, v9 |
| Eldritch Enlightenment | DS4 Flash v1-v6 |
v6, v5, v4, v3, v2, v1 |
Gilded Gnosis pages remain historical specifications. Infernal Invocation
revision numbers do not continue the Gilded Gnosis v20-r* sequence because
the source branches have different identities and merge contracts.
- Standard MTP and DSpark are not interchangeable.
- Confirm model revision, backend markers, graph coverage, and source-tree labels before comparing performance.
- Keep
NCCL_GRAPH_FILEunset unless it names an existing NCCL XML topology file. - Reuse release-scoped JIT caches. CuTe and FlashInfer compilation can dominate the first startup for an uncovered shape.