Copies text for another assistant or a person to review. Nothing is deployed.
This is not a fit and not a finished deployment. Source reported; not reproduced by this site.
Hardware matches this pin. 4 nodes. TP 4. EP 4.
Reported revision is pinned.
As-pinned recipe.
serving-repo: GNU Affero General Public License v3.0. Commercial use unknown. The rhys LICENSE file is the AGPL-3.0 license text. Its how-to-apply appendix is not a project header electing later versions. This module does not give legal clearance.checkpoint: MIT on the Hugging Face card snapshot only. Commercial use unknown. Card license MIT was read on snapshot sha dba1be0a, which is not serving revision df42c109. Confirm the pinned revision's own license file.The configured 1M context limit and 4M logical KV pool were not stress-tested at full occupancy.Published tok/s figures are the maintainer's own repeat suite. They are not a forecast.The lock says its base-image pin is not proof of the image used for Mia's reported results.
Do not add these rows together. Main weights overlap subcomponents. CPU and a unified pool are one residence, not extra capacity.
Placements are not applied to the calculator memory verdict.
Logical checkpoint
File ledger. Not a per-rank residence.
File ledger.
File size unknown.
UMA
GPU and CPU share this pool.
Shared UMA pool. Sparse routed matmul.
File size unknown.
Shared UMA pool. Resident matmul.
File size unknown.
Shared UMA pool. Attention cache.
File size unknown.
Shared UMA pool. Resident matmul.
File size unknown.
Shared UMA pool. Runtime scratch.
File size unknown.
Shared UMA pool. Resident matmul.
File size unknown.
NVMe
Disk path. Not a drag-to-fit.
NVMe. Sparse row lookup.
202758032400 bytes (188.833 GiB). cluster total, not one rank. Per-rank average 50689508100 bytes (47.208 GiB). Not measured resident.
Lookup or copy. Not free capacity.
Absent: PLE, Recurrent state. Absence is not zero bytes.
Resident bytes: unknown.
Placement
Local disk path
Not a drag-to-fit control.
Other host paths
Other rejected paths
Other residences
Full config and source evidence
- Main weights
This row is the logical checkpoint ledger. Engram bytes overlap it and stay on their own NVMe row. The ledger is not a claim that the whole checkpoint resides in unified memory, and Engram is not subtracted to invent a resident size.Mixed residence. Non-Engram weights are described in the unified pool. Engram rows stay on NVMe. No exact split was published.File size unknown.Per-rank storage: unknown. File ledger.Subcomponents can overlap this main-weight row. Do not add them into the checkpoint.Access: File ledger.No exact resident byte count was published for the non-Engram weights.The HF snapshot file size is a different revision and includes Engram. - Expert weights
Each rank holds 96 whole MXFP4 experts, not a slice of all 384. They stay in the unified pool.Unified pool. EP4 whole experts. Not on SSD.File size unknown.Per-rank storage: unknown. Overlaps the file ledger.Subcomponents can overlap this main-weight row. Do not add them into the checkpoint.Access: Sparse routed matmul.Expert tensor bytes were not published apart from the checkpoint. - Dense weights
Dense projections stay in the unified pool at the checkpoint's FP8 dense format.Unified pool. FP8 dense. Not on SSD.File size unknown.Per-rank storage: unknown. Overlaps the file ledger.Subcomponents can overlap this main-weight row. Do not add them into the checkpoint.Access: Resident matmul.Dense tensor bytes were not published apart from the checkpoint. - Engram rows
The two n-gram tables stay in packed files on local NVMe. A request copies one 264-byte row. The row cache is off.Local NVMe packed rows. No host row cache. No checkpoint fallback.202758032400 bytes (188.833 GiB). cluster total, not one rank. Per-rank average 50689508100 bytes (47.208 GiB). Not measured resident.Per-rank storage: 50689508100 bytes (47.208 GiB). Overlaps the file ledger.Subcomponents can overlap this main-weight row. Do not add them into the checkpoint.Access: Sparse row lookup.Packed-file bytes including the header were not measured on disk here.Resident cache bytes are zero by configuration, not a measured allocator reading.The lock total matches snapshot weight-plus-scale arithmetic. That match is not a hardware run and not a packed-file measurement. - KV cache
KV lives in the same unified pool as the weights. The README's 1M context and 4M logical pool were not filled in the reported retrieval.Unified pool. Configured 1M context and 4M logical pool. Verified retrieval input 299102 tokens.File size unknown.Per-rank storage: unknown. Runtime, not weights.This row can overlap main weights. Do not add it again.Access: Attention cache.KV resident bytes were not published.Full 1M / 4M occupancy was not stress-tested. - DSpark draft
The checkpoint's own DSpark drafter runs at block size 5. It is not a separate non-commercial drafter.Unified pool, with the resident weights. 32 local draft experts per rank.File size unknown.Per-rank storage: unknown. Overlaps the file ledger.Subcomponents can overlap this main-weight row. Do not add them into the checkpoint.Access: Resident matmul.Draft bytes are inside the HF parameter total's scope and were not given a separate file size. - Workspace and reserve
Activations, kernels, and the OS reserve share the unified pool. The only published reserve figure is an observed minimum, not a budget formula.The same unified pool. No separate host RAM reserve is added on top.File size unknown.Per-rank storage: unknown. Runtime, not weights.This row can overlap main weights. Do not add it again.Access: Runtime scratch.19.52 GiB is an observed minimum reserve, not measured workspace bytes. - PLE
This checkpoint has no Qwen-style n-gram PLE table. Engram, when this model has it, is a different component.Not present.File size unknown.Per-rank storage: unknown. Not in this recipe.This row can overlap main weights. Do not add it again.Access: Not in this recipe.No PLE bytes exist on this profile. - Vision
Image checks are part of the reported serving profile. No separate vision-file size was published, so this row overlaps the inclusive checkpoint ledger and is not added again.Unified pool, inside the checkpoint ledger. Bytes were not split out.File size unknown.Per-rank storage: unknown. Overlaps the file ledger.Subcomponents can overlap this main-weight row. Do not add them into the checkpoint.Access: Resident matmul.No separate vision byte count was published. - Recurrent state
The pinned DeepSeek profile does not describe a Gated DeltaNet state. The Qwen fixed-state rule does not apply here.Not present.File size unknown.Per-rank storage: unknown. Not in this recipe.This row can overlap main weights. Do not add it again.Access: Not in this recipe.No recurrent-state bytes were published.
- Pinned source
Serving profile, retrieval tokens, and tok/s observations 1fbfa3c641d4556f1e255fd967cb48c40d74e9a2 - Pinned source
Checkpoint, SGLang, EP4, and Engram byte pins 1fbfa3c641d4556f1e255fd967cb48c40d74e9a2 - Pinned source
O_DIRECT row size, header, and cache code 1fbfa3c641d4556f1e255fd967cb48c40d74e9a2 - Pinned source
Packed-file requirement and zero-cache guard 1fbfa3c641d4556f1e255fd967cb48c40d74e9a2 - Pinned source
AGPL-3.0 license text 1fbfa3c641d4556f1e255fd967cb48c40d74e9a2 - Pinned source
Parameter-total scope only. Not the serving revision. dba1be0a40aa45a94ad051997016db3960a90277
Speed figures in the copied brief are maintainer observations, not a measurement by this site.
Hashes, preflight, and reported observations
Hardware nvidia-dgx-spark-128gb-unified. Pool Each DGX Spark has one unified memory pool. Host RAM is that pool. Two or four nodes are separate pools, not one contiguous capacity.. Weight pin unset.
Checkpoint deepseek-ai/DeepSeek-V4.1-Flash revision df42c109f1defefcbfcedbe7d905718a12266e40. Manifest unknown. Runtime SGLang commit e087e662ba1ac4ef7747537e2a9141085efd4561. Measured runtime commit 25e7cb82f6524e3bbc46db6a9a54125e75e36d0e is a separate pin.. Commit e087e662ba1ac4ef7747537e2a9141085efd4561.
DSpark block 5, eight request slots, 2048-token prefill chunks. The README prints a configured 1M context and a 4M logical pool without an exact token integer. Full occupancy of that pool was not stress-tested. The longest reported retrieval input is 299102 tokens.
Commercial recommendation: false. Memory fit claimed: false. Deployment done: false.
serving-repo: GNU Affero General Public License v3.0. SPDX AGPL-3.0-only. Commercial use unknown. The rhys LICENSE file is the AGPL-3.0 license text. Its how-to-apply appendix is not a project header electing later versions. This module does not give legal clearance.checkpoint: MIT on the Hugging Face card snapshot only. SPDX MIT. Commercial use unknown. Card license MIT was read on snapshot sha dba1be0a, which is not serving revision df42c109. Confirm the pinned revision's own license file.
EP4 keeps 96 whole target experts on each rank, local intermediate 2304, and 32 local draft experts. This is not a quarter-width slice of every expert. Hugging Face snapshot dba1be0a reports 763,205,315,794 parameters and 510,286,023,000 checkpoint file bytes. That snapshot is a different revision from either serving pin, and the parameter total mixes stored F8, I8, BF16, and F32 tensors, including Engram and drafter tensors. A backbone-only count was not printed in the pinned caches, so it stays unknown. Snapshot Engram tables hold 196,613,849,600 weight elements; scales are extra. Checkpoint file 510286023000 bytes (475.241 GiB).
Preflight
Confirm checkpoint revision df42c109f1defefcbfcedbe7d905718a12266e40 and SGLang commit e087e662ba1ac4ef7747537e2a9141085efd4561.Require OFFLOAD_MODE=nvme, DSV41_CACHE_GIB=0, and resident scales off.Require the local packed Engram directory and a successful O_DIRECT attach for every rank. A missing shard is a failed preflight.Confirm four Spark nodes, TP4, EP4, one rank per node, and the RoCEnante link this profile used.Read the model-card license on revision df42c109 before any use decision.
Validation
Maintainer retrieval passed at 32,870, 131,174, and 299,102 input tokens.Maintainer reports both Engram tables bit-exact against checkpoint bytes in the GPU lookup and changing-graph tests.Maintainer reports all four ranks finished the serving validation without an OOM. This repository did not rerun that validation.
Rollback
The pinned README and versions.lock do not print a rollback procedure. Treat rollback as unknown until that repository's own notes are read.
Sources
- Serving profile, retrieval tokens, and tok/s observations
rhys-readme 1fbfa3c641d4556f1e255fd967cb48c40d74e9a2 checked 2026-10-02 - Checkpoint, SGLang, EP4, and Engram byte pins
rhys-lock 1fbfa3c641d4556f1e255fd967cb48c40d74e9a2 checked 2026-10-02 - O_DIRECT row size, header, and cache code
rhys-row-store 1fbfa3c641d4556f1e255fd967cb48c40d74e9a2 checked 2026-10-02 - Packed-file requirement and zero-cache guard
rhys-engram 1fbfa3c641d4556f1e255fd967cb48c40d74e9a2 checked 2026-10-02 - AGPL-3.0 license text
rhys-license 1fbfa3c641d4556f1e255fd967cb48c40d74e9a2 checked 2026-10-02 - Parameter-total scope only. Not the serving revision.
hf-deepseek-snapshot dba1be0a40aa45a94ad051997016db3960a90277 checked 2026-10-02
Reported observations
Maintainer observations, not a measurement by this site, and not a forecast.
Coding C1 decode after the long retrieval: 82.73 tok/s. Repeat-suite mean on 12 September 2026 after a 299,102-token retrieval. Not a forecast.Coding C8 aggregate after the long retrieval: 340.41 tok/s. Same repeat suite. Aggregate timing includes the whole batch. Not a forecast.
- Do not blend with deepseek-v4.1-flash-mia-tp4-ep1
Main weights: resident bytes are unknown.Expert weights: resident bytes are unknown.Dense weights: resident bytes are unknown.Engram rows: resident bytes are unknown.KV cache: resident bytes are unknown.DSpark draft: resident bytes are unknown.Workspace and reserve: resident bytes are unknown.PLE: resident bytes are unknown.Vision: resident bytes are unknown.Recurrent state: resident bytes are unknown.- Runtime claimed by this site: false. Summing components creates a fit: false. The generic calculator verdict is unchanged.