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 1.
Reported revision is pinned.
As-pinned recipe.
serving-repo: GNU Affero General Public License v3.0. Commercial use unknown. The Mia repository LICENSE file is the AGPL-3.0 license text. Its how-to-apply appendix is not a project header electing later versions. Weights are not in the repository. 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 fb2764a5. Confirm that revision's license file.The reported ~77 GiB resident weights per rank are maintainer prose. 476 GiB minus 189 GiB is 287 GiB, not the README's ~305 GiB, and neither quotient is an exact byte ledger.A 1,011,084-token needle passed. Sixteen request slots and an 8M-token pin are not a measured 16 by 1,048,576 occupancy.The cited validation boot reported a KV pool of 6.0-6.6M tokens, which is not the same integer as the 8,000,000 pin.b12x routed MoE is described as numerically close to FlashInfer, not bit-identical.The fast loader leaves a smaller and more variable KV pool than the stock loader.Boot memory moves with page cache. A KV pin the budget cannot cover can fail the boot or clamp. That warning is not a fit calculation.
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.
File size unknown.
Lookup or copy. Not free capacity.
Absent: PLE, Recurrent state. Absence is not zero bytes.
Resident bytes: unknown.
Placement
Unified pool
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 overlaps it on NVMe. The prose about what remains in unified memory is not a subtraction, and it is not a claim that the whole checkpoint resides in that pool.Mixed residence. The ledger is the checkpoint file. Engram rows are the NVMe overlap. Reported ~77 GiB per rank is not an exact byte count.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.Runtime pieces inside the ~77 GiB prose are not itemized.Do not compute resident bytes as (476 - 189) / 4. - Expert weights
Each rank holds a quarter of every expert, 576 wide, in the unified pool. NVMe is not an expert tier in this profile.Unified pool. EP1 quarter-experts. Not 96 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.Exact expert bytes per rank were not published. - Dense weights
FP8 dense projections stay in the unified pool and use the b12x MXFP8 route on this line.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.Exact dense bytes were not published. - Engram rows
Packed NVMe rows, a 4 GiB host cache knob, and a fallback to the checkpoint if the packed file is absent. Use the packed file for the measured line.Local NVMe when packed. 4 GiB host cache knob, 16 ways, inside the unified pool. Checkpoint/NFS only if the packed file is missing.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 row lookup.Exact packed-file bytes for revision fb2764a5 were not published.Per-layer cache bytes after the code's divide-by-2 are not a measured allocation. - KV cache
KV is small and lives in the unified pool. The pin and the boot that was measured are different numbers, and neither is 16 full 1M contexts.Unified pool. Pin 8,000,000 tokens. One needle at 1,011,084 tokens. Boot note 6.0-6.6M 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.Resident KV bytes move with MemAvailable and were not published as an exact allocation.16 x 1,048,576 concurrent occupancy was not the needle test. - DSpark draft
DSpark block size 5 on this TP4 line. The draft stays with the resident weights.Unified pool. Block size 5. Not a separate non-commercial drafter.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 were not given a separate file size. - Workspace and reserve
NCCL buffers, the indexer transient, and free memory all share the unified pool. Free-memory prose is not a fit.The same unified pool. Page cache and the indexer transient compete with weights.File size unknown.Per-rank storage: unknown. Runtime, not weights.This row can overlap main weights. Do not add it again.Access: Runtime scratch.Free-memory prose is not measured workspace bytes.Head-node processes share the pool with rank 0. - 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
TP4 settings, prose memory figures, and checkpoint revision cad252b7d3000cf21d919764e70a6912b6d3b0b0 - Pinned source
EP1 quarter-experts, cache knob, needle, rollback cad252b7d3000cf21d919764e70a6912b6d3b0b0 - Pinned source
SGLang f80c91a4 and b12x commits cad252b7d3000cf21d919764e70a6912b6d3b0b0 - Pinned source
O_DIRECT row size, header, optional DRM cache cad252b7d3000cf21d919764e70a6912b6d3b0b0 - Pinned source
Packed path, checkpoint fallback, cache budget divisor cad252b7d3000cf21d919764e70a6912b6d3b0b0 - Pinned source
AGPL-3.0 license text cad252b7d3000cf21d919764e70a6912b6d3b0b0 - 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. The Mia README also describes each box as 121.7 GiB unified. That prose is not a second pool and is not added to a 128 GB label.. Weight pin unset.
Checkpoint deepseek-ai/DeepSeek-V4.1-Flash revision fb2764a5cf321eaa5070ca8f9e892818f477c16d. Manifest unknown. Runtime SGLang dsv4.1 branch f80c91a4b97e52e17fbcff088aa6b47b6270f6b9 in Dockerfile.canary-roce.. Commit f80c91a4b97e52e17fbcff088aa6b47b6270f6b9.
DSpark block 5, 16 slots, chunk 4096 with the chunked indexer. Context pin 1048576 is printed as an integer. The KV pin is printed as 8M, not as the integer 8000000, so that token count stays null. The validation note reports a 6.0-6.6M pool on that boot. Needle labels are 99k, 198k, 746k, and the exact integer 1,011,084. The rounded labels are not stored as token counts. The pass is one prompt, not 16 concurrent 1M prompts.
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 Mia repository LICENSE file is the AGPL-3.0 license text. Its how-to-apply appendix is not a project header electing later versions. Weights are not in the repository. 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 fb2764a5. Confirm that revision's license file.
EP1 on the production image: every rank holds a quarter of all 384 experts and streams a 576-wide slice. That is not 96 whole experts. The README's 768-wide EP1 sentence is the 3-Spark profile, not this TP4 pin. 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 fb2764a5cf321eaa5070ca8f9e892818f477c16d and SGLang f80c91a4b97e52e17fbcff088aa6b47b6270f6b9.Use the production image path: EP_SIZE=1, Dockerfile.canary-roce, and the b12x_next MoE at a7d7d29b.Keep DSpark block 5, 16 request slots, prefill chunk 4096, and the chunked indexer patch. Chunk 4096 without that patch is outside the documented safe point.Pack local Engram shards before serving. A missing shard falls back to the checkpoint or NFS, which is not the measured line.Read full_token from the boot log. The KV pin is clamped or can fail the boot when the budget moves with page cache.Treat a switchless ring as a different qualification.
Validation
Maintainer needle retrieval passed at the printed labels 99k, 198k, 746k, and the exact integer 1,011,084 tokens on the production line.Maintainer qeval score on the fresh-clone image was 72 of 75, with the three misses also failing on the stock profile.This repository did not rerun those checks.
Rollback
The TP4 page lists env gates back to an earlier example: EP_SIZE=2 without the b12x-next MoE, fast load off, RoCEnante off, prefill chunk 1024, DSpark block 3, cache knob 0, and EP 4. Those gates are source notes, not a launch script from this module.
Sources
- TP4 settings, prose memory figures, and checkpoint revision
mia-ds-readme cad252b7d3000cf21d919764e70a6912b6d3b0b0 checked 2026-10-02 - EP1 quarter-experts, cache knob, needle, rollback
mia-ds-tp4 cad252b7d3000cf21d919764e70a6912b6d3b0b0 checked 2026-10-02 - SGLang f80c91a4 and b12x commits
mia-ds-manifest cad252b7d3000cf21d919764e70a6912b6d3b0b0 checked 2026-10-02 - O_DIRECT row size, header, optional DRM cache
mia-ds-row-store cad252b7d3000cf21d919764e70a6912b6d3b0b0 checked 2026-10-02 - Packed path, checkpoint fallback, cache budget divisor
mia-ds-engram cad252b7d3000cf21d919764e70a6912b6d3b0b0 checked 2026-10-02 - AGPL-3.0 license text
mia-ds-license cad252b7d3000cf21d919764e70a6912b6d3b0b0 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.
Prose, 1 stream: 87.7 tok/s. sparkDash 1.8.8, greedy, thinking off, switched fabric, fresh-clone production line, 2026-09-25. Not a forecast.Needle: PASS at 1,011,084 tokens. Single phrase at 37% depth. Not a 16-way 1M occupancy test.
- Do not blend with deepseek-v4.1-flash-rhys-tp4-ep4-nvme
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.