GLM-5.2 O14 profiles on 4× DGX Spark

This repository is the public profile index for the O14 GLM-5.2 runtime. Machine-readable selection is in profiles/o14-profiles.json; automation rules are in AGENTS.md.

Current profiles

Profile	Status	Topology	Total KV	Deploy?	Runtime artifact
O14 Fast	READY	TP4 / DCP1 / PP1	250K	Yes	Build from the pinned canonical O14 Harness source
O14 Balanced	TESTING / DO NOT DEPLOY	TP4 / DCP2 / PP1	500K target	No	None; separate immutable runtime required after acceptance

Matched-speed results

Balanced speed remains TBD until the exact retained DCP1 fixture set is run against the accepted DCP2 runtime. No DCP2 speed value is inferred.

Profile	Prefill	Peak decode	Deep-context decode	Fixture status
O14 Fast	819 tok/s	42.3 tok/s	29–33 tok/s at ~53K prompt tokens	Retained legacy exp1 DCP1 fixtures
O14 Balanced	TBD	TBD	TBD	Run the exact DCP1 fixtures after DCP2 acceptance

O14 Fast

O14 Fast is READY and deployable at 250K total KV.

Machine field	Exact value
Allocator total KV	250,023 tokens
KV_CACHE_MEMORY_BYTES	7,995,534,848 bytes per rank
MAX_MODEL_LEN	249,000
MAX_NUM_SEQS	4
MAX_NUM_BATCHED_TOKENS	2,048
Topology	TP4 / DCP1 / PP1

The selector pins canonical Harness commit 9ba54db8f7c852eabb90372f7dcf47f87eafbe20, which contains the named Fast selector and exact recipe values.

There is no published O14 Fast OCI image in phase 1. Do not invent or pull a Fast tag. Build from the pinned canonical source recipe and keep model weights outside the runtime image.

O14 Balanced

O14 Balanced is TESTING / DO NOT DEPLOY. Its target is 500K total KV with an expected allocator total of 500,237 tokens, 8,000,000,000 KV bytes per rank, MAX_MODEL_LEN=490000, and TP4/DCP2/PP1.

Balanced speed is TBD. Its source recipe, image reference, and digest are intentionally null. Acceptance requires a separate immutable DCP2 runtime image and reproducible recipe. Agents must never synthesize a Balanced tag, reuse Fast's image, or overwrite Fast.

Safe selection

1. Read the requested profile from profiles/o14-profiles.json.
2. Continue only when status == "READY" and deployable == true.
3. If an image reference or digest is null, do not pull or invent one.
4. Pin the exact source commit, verify checksums, and run the canonical validators before building.
5. Supply model paths, endpoints, ports, and network interfaces explicitly for the target environment.

See AGENTS.md for exact pull/build behavior.

Runtime components

The current source stack combines:

• NVFP4 DS-MLA KV cache (nvfp4_ds_mla, 368-byte record)
• B12X sparse-MLA attention and sparse indexer for GB10 / sm_121a
• MTP-5 speculative decode
• TP4 plus decode-context parallelism
• full and piecewise CUDA graph execution

Model card: https://huggingface.co/0xdfi/GLM-5.2-1M-context-NVFP4-4x-DGX-Spark

Legacy exp1

Historical evidence only — not selectable profiles. The July 2026 exp1 work used DCP1/DCP2/DCP4 labels before the O14 profile names were adopted. Those settings and limits are not current selector values. DCP4 remains Legacy exp1 DCP4; it is not “O14 High.”

Legacy exp1 topology observations

Legacy topology	Historical max context with full graphs	Historical peak decode	Historical fastest prefill
Legacy exp1 DCP1	375K	~42 tok/s	~820 tok/s
Legacy exp1 DCP2	625K	~37 tok/s	746 tok/s
Legacy exp1 DCP4	1M	~38 tok/s short / ~30 tok/s deep	614 tok/s

These are preserved as historical observations from one rig, not deployment promises or selectable capacities. The retained DCP1 fixture values used in the current matched-speed table are 819 tok/s prefill, 42.3 tok/s peak decode, and 29–33 tok/s at ~53K actual prompt tokens.

Legacy exp1 engineering findings

• Sparse-indexer scratch scaled with context × batch and was not sharded by DCP.
• One versus two active RoCE links changed measured throughput by only about 1.5–3%; this repository imposes no fixed link-count or link-rate requirement.
• CUDA graphs improved decode but consumed memory otherwise available to context.
• Decode varied with MTP acceptance at short context and converged toward the sparse-indexer ceiling at depth.
• Unified-memory load pressure required operational care around page cache and swap.

The archived matrix is in FINDINGS.md. The parameterized historical launcher and benchmark scripts remain for evidence reproduction only; they are not current profile recipes.

Validation

python3 scripts/render_profile_chart.py --check
python3 scripts/validate_profiles.py
python3 -m unittest discover -s tests -v

The validator checks exact selector values/statuses, null artifact claims, chart determinism, private defaults, and stale current-profile capacities.

License and attribution

Apache-2.0. Community attribution is preserved in NOTICE.