ds4-winhip / docs /serving-capabilities.md
staccs's picture
ds4-winhip: Windows/HIP 7.2 build tree, launcher, supervisor, bench + notes (part 6)
5396c20 verified
|
Raw History Blame Contribute Delete
7.57 kB

Serving capabilities

This table is generated from ds4_core::serving_caps and the resolved plan's controls, which also supply ds4-perf serving-controls. qualified describes the recorded artifact and workload scope, not every quantization, context, concurrency, media input or feature combination. unverified means implemented (Support::Present); none means unsupported. Consult the release ledger and family contracts for evidence and limits.

Execution

Cells show the available lane or mode and its qualification. An opt-in capability still requires explicit selection and successful native fitting. The request's effective and qualified plan remains authoritative.

Model Banks Prefix reuse Disk KV Snapshot MTP kind / lane / support
DeepSeek V4 Flash persistent / qualified exact / qualified qualified qualified DeepSeek / bank / qualified
DeepSeek V4 Pro persistent / qualified exact / qualified qualified qualified DeepSeek / bank / qualified
DeepSeek V4.1 Flash serial / none exact / unverified unverified none Embedded / serial / qualified
Solar Open2 250B persistent / qualified partial / qualified qualified qualified None / none / none
Motif-3 persistent / qualified partial / qualified qualified qualified None / none / none
K-EXAONE 236B A23B persistent / qualified partial / unverified qualified qualified None / none / none
dots3-note-prev opt_in / unverified partial / unverified unverified qualified Embedded / serial / unverified
Qwen3.8-Flash-Next persistent / qualified partial / qualified qualified qualified Embedded / bank / qualified
GLM 5.3 Flash persistent / unverified partial / unverified unverified unverified Embedded / serial + bank / unverified
K2-Horizon 375B A23B persistent / qualified exact / qualified unverified unverified None / none / none
Inkling Small serial / none exact / qualified unverified unverified Sidecar / serial / qualified
Step-3.7-Flash opt_in / qualified partial / qualified qualified qualified Sidecar / serial + bank / qualified
Ling-3.0-flash-VL persistent / qualified partial / qualified qualified qualified None / none / none
MiMo-V2.6-Flash-RL opt_in / unverified partial / unverified unverified unverified Embedded / serial / qualified
Naive-N0.5-Flash opt_in / qualified partial / qualified qualified qualified External / serial + bank / unverified
IQuest-Q1 persistent / unverified partial / unverified unverified unverified Embedded / serial + bank / unverified
Ling-3.1-flash-VQ persistent / unverified partial / unverified unverified unverified Embedded / serial + bank / unverified
Beam-1.0-501B-A23B-VQ persistent / unverified partial / unverified unverified unverified Sidecar / serial / unverified

Bounds and allocator controls

The runtime context cap is an admission bound, not a qualification claim. β€” means the capability table declares no bound or recorded value; it does not mean unlimited capacity. Qualified columns retain their workload scope.

Model Runtime context cap Qualified context Qualified banks Qualified prompt Native prefill environment
DeepSeek V4 Flash β€” β€” β€” β€” DS4_METAL_PREFILL_CHUNK
DeepSeek V4 Pro β€” β€” β€” β€” DS4_METAL_PREFILL_CHUNK
DeepSeek V4.1 Flash β€” β€” 1 β€” β€”
Solar Open2 250B β€” β€” β€” β€” DS4_METAL_PREFILL_CHUNK
Motif-3 262144 β€” β€” β€” DS4_MOTIF3_PREFILL_CHUNK
K-EXAONE 236B A23B β€” β€” β€” β€” DS4_EXAONE_PREFILL_CHUNK
dots3-note-prev 524288 β€” 1 β€” DS4_DOTS3_PREFILL_CHUNK
Qwen3.8-Flash-Next 1048576 262144 2 β€” DS4_QWEN_PREFILL_CHUNK
GLM 5.3 Flash 1048576 2048 1 β€” DS4_GLM53_PREFILL_ROWS
K2-Horizon 375B A23B β€” 32768 1 β€” DS4_EXAONE_PREFILL_CHUNK
Inkling Small 1048576 1024 1 β€” DS4_INKLING_PREFILL_CHUNK
Step-3.7-Flash 262144 65536 2 6300 DS4_STEP37_PREFILL_CHUNK
Ling-3.0-flash-VL 262144 65536 2 β€” DS4_LING3VL_PREFILL_CHUNK
MiMo-V2.6-Flash-RL 1048576 524288 1 β€” DS4_MIMO2_PREFILL_CHUNK
Naive-N0.5-Flash 1048576 8192 2 702 DS4_NAIVE_PREFILL_CHUNK
IQuest-Q1 524288 β€” β€” β€” DS4_IQUEST_PREFILL_CHUNK
Ling-3.1-flash-VQ 262144 β€” β€” β€” DS4_LING31_PREFILL_CHUNK
Beam-1.0-501B-A23B-VQ 1048576 β€” β€” β€” DS4_BEAM_PREFILL_CHUNK

Scheduler chunk candidates: 256, 512, 1024, 2048, 4096, 8192. The resolved plan only offers values at or below its known native capacity. Unknown capacity offers no candidate. These are scheduler controls, not measurements or speed claims.

Qualification notes

  • DeepSeek V4.1 Flash: recorded Spark gates cover greedy single-request embedded speculation with decode graphs (56-token and 1,100-token byte-identical runs and matching engine traces); the sampling profile is not qualified; the artifact fixes the draft block and verification width is scheduled per round (--mtp-draft is unsupported; --dspark-verify pins verification); native prompt-end exact prefix reuse and its KVC disk checkpoint are present, unqualified; --prefix-reuse off|exact|auto controls live and disk reuse (--no-kv-reuse aliases off); partial reuse, banks and session snapshots are unavailable; context comes from model metadata, --ctx is rejected.
  • K-EXAONE 236B A23B: exact reuse qualified; LLLG partial checkpoints await live qualification.
  • dots3-note-prev: text banks, local-window partial reuse and serial MTP are present but unqualified.
  • Qwen3.8-Flash-Next: common UX baseline; configured values and verified combinations differ.
  • GLM 5.3 Flash: compact banks, partial reuse, snapshots and embedded MTP are present; 1M structural capacity and historical 2K qualification are separate.
  • K2-Horizon 375B A23B: 32K one-bank serving is qualified; disk KV and external owner import are not.
  • Inkling Small: serial text snapshots present; media snapshots unsupported.
  • Step-3.7-Flash: text banks are opt-in; Chat restart hits need history-stable identity; images serial.
  • MiMo-V2.6-Flash-RL: 512K serial text and 256K serial media/DFlash are prior gates. With MTP off, 256K two-bank text plus serial media passed bounded checks at chunk 2048 with Q8 repack off, including live partial reuse and restart disk continuation. 1M one-bank text passed a bounded 1,040,506-token prompt; two banks did not fit. 512K two-bank media exceeds Spark memory.
  • Naive-N0.5-Flash: main-only chunk-2048 buffered retrieval and disk continuation: 256K/two banks, 512K/one bank; draft-loaded and other shapes retain the bounded 8K gate; DSpark acceleration unqualified.
  • IQuest-Q1: 8K/two-bank thinking HTTP passed at chunk 128 with short prompts and MTP off/on (draft 3, margin 0); plan bounds stay unqualified because reasoning and margin are not represented; no-thinking output, other shapes and 512K remain unqualified.

Regenerate without loading a model:

cargo run -q -p ds4-core --example serving_docs > docs/serving-capabilities.md
cargo test -p ds4-core --test serving_docs