Qwen3.8-Flash-Next on 1 DGX Spark

SGLang file-backed PLE on one DGX Spark

Qwen3.8-Flash-Next on 1 DGX Spark. One-Spark SGLang report locked to qwen4-main-squashed 4ccff141db, the NVIDIA mixed NVFP4 artifact, and a file-backed PLE table on NVMe. CPU-pinned PLE still uses the unified pool. The dual-Spark TP2 arm is a different recipe.

Community component placement

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. 1 nodes. TP 1. EP unknown.

Weight revision is unresolved.

As-pinned recipe.

  • nvfp4-weights: NVIDIA Open Model License, and the card also references Qwen Community License 1.0. Commercial use unknown. Both terms are named on the current manifest card. The timed file revision is unknown. No commercial recommendation is made.
  • runtime: Unknown. Commercial use unknown. The SGLang runtime license for commit 4ccff141db was not in the cache. It is not the weight license.
  • NVIDIA MTP block 128 does not match TP2's 640/2 = 320 in that report. The dual-Spark recipe borrows a RadixArk BF16 MTP. Do not copy that MTP onto this one-Spark pin.
  • 47.7 GiB for the NVMe table is prose. CPU offload of the same table does not free unified memory.
  • C8 aggregate 71.7 tok/s is one reported cell, not a per-stream promise and not a dual-Spark number.
  • The 6B active label is not the NVFP4 file size.

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.

NVIDIA mixed NVFP4 artifact

File ledger.

132680249378 bytes (123.568 GiB). cluster total, not one rank.

UMA

GPU and CPU share this pool.

Routed experts

Shared UMA pool. Sparse routed matmul.

File size unknown.

Dense layers

Shared UMA pool. Resident matmul.

File size unknown.

Internal MTP

Shared UMA pool. Resident matmul.

File size unknown.

KV cache

Shared UMA pool. Attention cache.

File size unknown.

Recurrent state

Shared UMA pool. Recurrent state.

File size unknown.

Workspace

Shared UMA pool. Runtime scratch.

File size unknown.

NVMe

Disk path. Not a drag-to-fit.

File-backed PLE

NVMe. Sparse row lookup.

File size unknown.

Lookup or copy. Not free capacity.

Unknown

Residence unknown. Not zero.

Vision tower

Unknown residence. Resident matmul.

File size unknown.

Absent: Engram. Absence is not zero bytes.

Resident bytes: unknown.

Placement

Local disk path

Not a drag-to-fit control.

Other rejected paths

Other residences

Full config and source evidence
  • NVIDIA mixed NVFP4 artifact
    The current NVIDIA artifact is the inclusive file ledger: NVFP4 experts, FP8 PLE, block-FP8 MTP, and retained high-precision pieces. The file-backed table places the PLE portion on NVMe. Those file bytes are not measured all-UMA residence, and PLE is not subtracted from the file.
    Mixed residence. The file ledger is the whole artifact. PLE is the NVMe overlap. Other resident pieces are described in the unified pool. Manifest file bytes are not resident bytes.
    132680249378 bytes (123.568 GiB). cluster total, not one rank.
    Per-rank storage: unknown. File ledger.
    Subcomponents can overlap this main-weight row. Do not add them into the checkpoint.
    Access: File ledger.
    The research brief does not tie this current manifest revision to the measured run. These file bytes are not resident memory and are not a fit.
    Do not subtract a 47.7 GiB prose label from this file and call the remainder resident.
  • Routed experts
    The model has 512 experts, with 10 routed and 1 shared per token in the research brief. Those experts are the resident MoE weights. The 6B active label is only the compute activated for a token. It is not the bytes to load.
    Inside the served artifact's main ledger. Not an NVMe expert tier unless that profile says so. Not a 6B file.
    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 bytes were not split out of the main artifact.
    6B active is not a storage figure.
  • Dense layers
    Attention and shared paths stay in the main artifact. This row overlaps that ledger.
    Unified pool, inside the main artifact. 48 layers: 36 Gated DeltaNet plus 12 QSA layers.
    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 bytes were not split out.
  • Engram
    This Qwen profile has no Engram table. A PLE lookup, when this model has one, 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 Engram bytes exist on this profile.
  • File-backed PLE
    One Spark uses a file backend so the PLE table is read from NVMe. That lookup is not Engram. File RSS is not the 8 GiB page-cache budget, and that budget is not a claim that host-resident bytes are zero. A CPU pin of the same table still occupies the unified pool. 47.7 GiB is prose. The same pool still needs headroom.
    NVMe file when the file backend is the selected path. Page cache is a separate budget from file RSS. CPU pin: same unified pool, no freed memory.
    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 PLE file bytes were not printed.
    The 8 GiB page-cache budget is not file RSS and is not stored as bytes.
    Do not compute resident weights as the NVFP4 file minus 47.7 GiB.
  • Vision tower
    The architecture has a vision tower. This one-Spark speed cell does not publish a vision byte size. It stays a separate row with unknown bytes.
    Not given a byte size in this cell. Not the drafter and not the PLE table.
    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 vision byte count was published for this cell.
  • Internal MTP
    The one-Spark cell uses the artifact's own MTP. The dual-Spark recipe's borrowed RadixArk BF16 MTP is a different role and is not added here.
    Inside the mixed NVFP4 artifact. Block size noted as 128 for the NVIDIA MTP. Not a shared sidecar.
    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.
    MTP bytes inside the mixed artifact were not split out.
  • KV cache
    KV length is not capped at the QSA index budget of 2048 tokens. No exact KV byte size was printed for this cell.
    Unified pool. Context and concurrency for the file-backed cell were not printed as a configured maximum.
    File size unknown.
    Per-rank storage: unknown. Runtime, not weights.
    This row can overlap main weights. Do not add it again.
    Access: Attention cache.
    No KV byte count was published.
    This module does not estimate KV from the 2048 budget.
  • Recurrent state
    36 Gated DeltaNet layers keep a fixed recurrent state, including at short context. The state is not the KV cache and not the QSA index budget.
    Unified pool. Fixed state. Bytes unknown.
    File size unknown.
    Per-rank storage: unknown. Runtime, not weights.
    This row can overlap main weights. Do not add it again.
    Access: Recurrent state.
    No integer byte size was printed for the recurrent state.
    Five slots per request is the published cache setting divided by the published request cap. It is not a measured allocation.
  • Workspace
    No workspace byte count was published for this one-Spark cell.
    The same unified pool.
    File size unknown.
    Per-rank storage: unknown. Runtime, not weights.
    This row can overlap main weights. Do not add it again.
    Access: Runtime scratch.
    Workspace bytes are unknown.
  • Pinned source Official cookbook snapshot checked 2026-10-02. This revision field stays unknown: the cache has no git sha for the docs page. Runtime pin 4ccff141db is the profile runtimeCommit and is not a docs git revision. unknown
  • Pinned source Current mixed NVFP4 manifest and license string. Not tied to the timed run. fc694b54fb0174e0913e6adf86691ef85a4ead47
  • Pinned source Official BF16 scope and Qwen Community License card. Not the NVFP4 file. de4b8e4d43b917e7706784d8bb445c9af86a3540

Speed figures in the copied brief are maintainer observations, not a measurement by this site.

Related

Hashes, preflight, and reported observations

Hardware nvidia-dgx-spark-128gb-unified. Pool One DGX Spark, one unified pool. CPU offload does not add a second pool and does not free the first one.. Weight pin manifest-not-tied-to-report.

Checkpoint nvidia/Qwen3.8-Flash-Next-NVFP4 revision unknown. Manifest fc694b54fb0174e0913e6adf86691ef85a4ead47. Runtime qwen4-main-squashed 4ccff141db on dev-qwen38-next-local. Not a floating latest tag. SGLang base support is named at 0.5.20 or newer in the research brief.. Commit 4ccff141db.

Official nvfp4-nvda single-Spark low-latency flags: TP 1, page size 64, chunk 4096, context-length 262144, max-running-requests 8, max-mamba-cache-size 40 (5 slots per request at that concurrency), NEXTN steps 3, topk 1, draft tokens 4, flashinfer_cutlass for the MoE runner and the FP4 GEMM, mem-fraction-static 0.85. Request context is not the shared KV max. Dual-Spark TP2 at 47.4 tok/s is not this pin.

Commercial recommendation: false. Memory fit claimed: false. Deployment done: false.

  • nvfp4-weights: NVIDIA Open Model License, and the card also references Qwen Community License 1.0. SPDX unknown. Commercial use unknown. Both terms are named on the current manifest card. The timed file revision is unknown. No commercial recommendation is made.
  • runtime: Unknown. SPDX unknown. Commercial use unknown. The SGLang runtime license for commit 4ccff141db was not in the cache. It is not the weight license.

512 experts, 10 routed plus 1 shared, on one rank. Per-rank byte split unknown. 6B active is not the resident size. Official BF16 safetensors sum to 360,000,192,888 bytes at manifest revision de4b8e4d. No exact integer was printed for 125B, 51B, 4B, 6B, or 180B, so those stay labels. The served MLX or NVFP4 file is not this official BF16 file. Checkpoint file 360000192888 bytes (335.276 GiB).

Preflight

  • Pin runtime 4ccff141db. Do not substitute latest. The docs page has no git sha in the cache; that runtime pin is not a docs revision.
  • Use the published file PLE backend for the one-Spark NVMe table. A CPU pin stays in the unified pool. The file RSS and the 8 GiB page-cache budget are different, and neither is zero host use.
  • The cookbook warns that a boot rewrites the PLE table and a populated file rewrites slowly. This profile does not delete files.
  • The measured NVFP4 revision is unknown. Manifest fc694b54 is the current card only.
  • Stay on one Spark. TP2 and TP4, including two TP2 pairs, are outside this pin.
  • QSA index budget 2048 does not mean the KV cache holds only 2048 tokens.
  • Read both the NVIDIA Open Model License and the Qwen Community terms. No commercial recommendation is made.

Validation

  • The research brief reports single-Spark MTP at 27.5 tok/s and a C8 aggregate of 71.7 tok/s.
  • This site did not run those cells.

Rollback

  • Return to the previous SGLang checkout and the previous weight files. This module does not check out code or restart a server.

Sources

Reported observations

Maintainer observations, not a measurement by this site, and not a forecast.

  • Reported one-Spark MTP: 27.5 tok/s, C8 aggregate 71.7 tok/s. Research brief, runtime 4ccff141db, file-backed PLE. Dual-Spark 47.4 tok/s is not this pin. Source report only.
  • Do not blend with qwen3.8-flash-next-spark-tensorfold-mlx
  • NVIDIA mixed NVFP4 artifact: resident bytes are unknown.
  • Routed experts: resident bytes are unknown.
  • Dense layers: resident bytes are unknown.
  • Engram: resident bytes are unknown.
  • File-backed PLE: resident bytes are unknown.
  • Vision tower: resident bytes are unknown.
  • Internal MTP: resident bytes are unknown.
  • KV cache: resident bytes are unknown.
  • Recurrent state: resident bytes are unknown.
  • Workspace: resident bytes are unknown.
  • Runtime claimed by this site: false. Summing components creates a fit: false. The generic calculator verdict is unchanged.