GLM-5.3-Flash on 2 DGX Sparks

LimeChain TensorFold C8 comparison

GLM-5.3-Flash on 2 DGX Sparks. A separate two-Spark TensorFold profile on the jayleaton kit at commit 5624dfc, with LimeChain's C8 admission settings. It is not the Mia 53-patch image. The bake-off ran once per arm, and the arms use different quantizations. Those gaps are not a same-precision ranking and not a four-node result.

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. 2 nodes. TP 2. EP unknown.

Reported revision is pinned.

As-pinned recipe.

  • kit: Apache-2.0 for jayleaton/glm53-tensorfold-spark. Commercial use allowed.
  • tensorfold: MIT for ashhart/TensorFold, as credited by LimeChain. Commercial use allowed.
  • exl3-weights: Checkpoint license must be read on the downloaded repo. Commercial use unknown. LimeChain says to check every weight's license. This cache does not include the checkpoint LICENSE text. Do not assume MIT.
  • dflash2: Creative Commons Attribution-NonCommercial-NoDerivatives 4.0. Commercial use restricted. Pinned drafter revision 7d74cdd881ed7e32c31175984a67823127b66cfe. Non-commercial. An empty DRAFTER uses MTP only and is slower. No commercial recommendation.
  • Each bake-off arm ran once. Differences under about 5% are called noise by the source.
  • C1 short was 99 tok/s on this TensorFold profile, 72 on the vLLM Mia EXL3 kit, and 37 on vLLM NVFP4. The quantizations differ. The 72 is not the Mia TensorFold profile's 60.4 tok/s prose cell.
  • C8 short aggregate was 368 versus 251 versus 185 across those same three arms. That is not a same-precision general ranking and not the Mia TensorFold concurrency-4 cell.
  • Eight cold ~23K prompts: this profile's slowest first token was 183 s, against 127 s on the vLLM NVFP4 arm. Cold prefill is not a win for this profile in that cell.
  • C1 prose and code cells were measured on an earlier kit build 80dba15, not on 5624dfc.
  • Not validated at four nodes. Not a double-replica throughput claim.
  • The README requires two cabled Sparks and refuses READY until 8 slots load after a page-cache eviction. That gate is not a memory fit.

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.

EXL3 checkpoint

File ledger.

File size unknown.

UMA

GPU and CPU share this pool.

Routed experts

Shared UMA pool. Sparse routed matmul.

File size unknown.

Dense weights

Shared UMA pool. Resident matmul.

File size unknown.

KV cache

Shared UMA pool. Attention cache.

File size unknown.

DFlash2 drafter

Shared UMA pool. Resident matmul.

File size unknown.

Workspace

Shared UMA pool. Runtime scratch.

File size unknown.

Vision tower

Shared UMA pool. Resident matmul.

File size unknown.

Recurrent state

Shared UMA pool. Recurrent state.

File size unknown.

Absent: Engram, PLE. Absence is not zero bytes.

Resident bytes: unknown.

Placement

Unified pool

Local disk path

Not a drag-to-fit control.

Other rejected paths

Other residences

Full config and source evidence
  • EXL3 checkpoint
    Disk planning in this README is about 165 GB for the checkpoint plus about 84 GB of prepared weights. That is a different sentence from the Mia recipe's 176 GB file.
    Unified pool after the kit prepares weights. Exact bytes unknown.
    File size unknown.
    Per-rank storage: unknown. File ledger.
    Subcomponents can overlap this main-weight row. Do not add them into the checkpoint.
    Access: Resident matmul.
    Exact file bytes were not printed.
    Prepared weights are extra disk, not a second memory pool.
  • Routed experts
    EXL3 experts load with the checkpoint. No NVMe expert path is described.
    Unified pool. Split across two ranks. Per-rank expert count unpublished.
    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.
    Per-rank expert bytes were not published.
  • Dense weights
    Dense weight bytes and the serving dense quant for this kit were not separated in the pinned README.
    Unified pool. Exact dense quant policy for this kit was not copied from the Mia q4 setting.
    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.
    Do not assume Mia's DENSE=q4 French-prompt result applies to this kit.
  • Engram
    No Engram table is part of this GLM profile.
    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 to report.
  • KV cache
    Context per request in the bake-off table is 1,048,576 for this TensorFold arm. A shared-pool byte size was not published.
    Unified pool. Token and byte residency unpublished beyond the context column.
    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 bytes and whether the pool is shared were not published for this kit.
    12.9 GiB is observed free memory, not a KV size.
  • DFlash2 drafter
    The pinned drafter cannot serve as its own model. It is non-commercial. This LimeChain pin is the C8 kit, not Mia's four-request TensorFold profile. MTP-only is the documented way to avoid that download, and it is slower.
    Separate drafter download. License CC-BY-NC-ND-4.0. Exact bytes unknown.
    File size unknown.
    Per-rank storage: unknown. Separate from the file ledger.
    This row can overlap main weights. Do not add it again.
    Access: Resident matmul.
    Drafter file bytes were not printed.
  • Workspace
    The bake-off records lowest available memory. It does not publish a workspace allocation.
    The same unified pool. Free-memory observations are not a fit.
    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 were not published.
  • 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 tower
    The vision tower is a separate row from the language weights and from the drafter. The README's GiB labels stay prose.
    Rank 0, inside the same unified pool. Not a second pool and not an SSD tier.
    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.
    The GiB labels were not published as integer byte counts.
  • Recurrent state
    KDA linear-attention state stays in the unified pool. It is not the KV pool and it is not a per-token cache that this module sizes.
    Unified pool. Fixed recurrent state. Byte size unpublished.
    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 published for the KDA state.
  • Pinned source C8 profile, one-run bake-off, drafter revision and license 441de756601d7ca3b8ec09a900c19e7eabaa4f64
  • Pinned source Official parameter scope only. Not this EXL3 revision. eb9eb208eb0d988989d07a6a12d0fdeb5f52574a

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 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 brandonmusic/GLM-5.3-Flash-tr3-4bpw revision 5ab363a8dcf6405955fd5f99671e01a1c9fb124b. Manifest unknown. Runtime jayleaton/glm53-tensorfold-spark @ 5624dfc6fce32d747727eaf4dbda98999f868066 plus the LimeChain C8 profile. Not Mia TensorFold v0.6.0 with 53 patches.. Commit 5624dfc6fce32d747727eaf4dbda98999f868066.

C8 means eight slots on this LimeChain profile. Mia's default is four and only with DFlash2. The two slot counts are different pins. KV bytes for this kit were not published.

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

  • kit: Apache-2.0 for jayleaton/glm53-tensorfold-spark. SPDX Apache-2.0. Commercial use allowed. The kit license is not the weight license.
  • tensorfold: MIT for ashhart/TensorFold, as credited by LimeChain. SPDX MIT. Commercial use allowed. Engine credit only. It does not license the EXL3 checkpoint.
  • exl3-weights: Checkpoint license must be read on the downloaded repo. SPDX unknown. Commercial use unknown. LimeChain says to check every weight's license. This cache does not include the checkpoint LICENSE text. Do not assume MIT.
  • dflash2: Creative Commons Attribution-NonCommercial-NoDerivatives 4.0. SPDX CC-BY-NC-ND-4.0. Commercial use restricted. Pinned drafter revision 7d74cdd881ed7e32c31175984a67823127b66cfe. Non-commercial. An empty DRAFTER uses MTP only and is slower. No commercial recommendation.

Official config still lists 288 routed experts. This kit does not publish a per-rank expert count. Do not copy Mia's serving settings onto this kit. The official HF total remains 321,323,031,390 parameters for snapshot eb9eb208. It is not the EXL3 file, not a 169B FP4 count, and not an 18B active count. Those latter two figures were not in the pinned README. Checkpoint file 328326771576 bytes (305.778 GiB).

Preflight

  • Use kit commit 5624dfc6fce32d747727eaf4dbda98999f868066 and weight revision 5ab363a8dcf6405955fd5f99671e01a1c9fb124b.
  • Confirm DFlash2 revision 7d74cdd881ed7e32c31175984a67823127b66cfe and its non-commercial terms, or leave the drafter unset for MTP only.
  • Plan disk as about 165 GB for the checkpoint plus about 84 GB for prepared weights. That pair is not the Mia recipe's 176 GB figure.
  • Evict page cache before expecting all 8 slots. The source says a node that just copied weights can boot with 1 slot.
  • Read the checkpoint license before use. It was not in this cache.

Validation

  • The bake-off is one run per arm on one pair of Sparks, temperature 0, on 2026-09-30.
  • Correctness gates in that README are part of the same one-run comparison. They are not a license or a memory fit.
  • This repository did not rerun the bake-off.

Rollback

  • This profile does not supply a rollback for the Mia TensorFold image. Stopping the LimeChain service does not change the Mia pins.

Sources

Reported observations

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

  • C1 short decode: 99 vs 72 vs 37 tok/s. One run. This TensorFold profile, then the vLLM Mia EXL3 kit, then vLLM NVFP4. Different quantizations. The middle arm is not the Mia TensorFold profile. Not a forecast.
  • C8 short aggregate: 368 vs 251 vs 185 tok/s. One run. Per stream 46 vs 34 vs 24. Same three arms. Not a same-precision ranking.
  • C8 cold ~23K slowest first token: 183 s vs 187 s vs 127 s. This TensorFold profile, the vLLM Mia EXL3 kit, then vLLM NVFP4. Cold prefill is a different result from the short-prompt cells.
  • Do not blend with glm-5.3-flash-mia-tensorfold-tp2
  • EXL3 checkpoint: resident bytes are unknown.
  • Routed experts: resident bytes are unknown.
  • Dense weights: resident bytes are unknown.
  • Engram: resident bytes are unknown.
  • KV cache: resident bytes are unknown.
  • DFlash2 drafter: resident bytes are unknown.
  • Workspace: resident bytes are unknown.
  • PLE: resident bytes are unknown.
  • Vision tower: 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.