Skip to content
XMPPwockyPublic

About

DeepSeek v4.1 Flash inference server, specialized for Strix Halo + dGPU configurations

Resources

Stars

2 stars

Watchers

0 watching

Forks

Latest commit

 

History

1,238 Commits

Folders and files

Repository files navigation

deepstrix docs — start here

Index written 2026-10-04 with the docs-vs-code audit. Most files in docs/ are dated journals; this page says which ones describe the system as it is.

What the system is (2026-10-04)

  • Model: DeepSeek-V4.1-Flash (40 layers, 384 routed experts, top-6, CSA2 compressed attention with the sparse indexer, Engram, mHC, the DSpark drafter). Spec: v41/ARCH_SPEC.md. Vocabulary: GLOSSARY.md (canonical).
  • Two boxes. Box 1 = the hub (deepstrix-server, 9070 XT dGPU + Strix Halo iGPU). Box 2 = deepstrix-expertd, serving routed experts over USB4 / thunderbolt0. Routed experts are native MXFP4, paged through LRU pools on both boxes from NVMe; box 1 owns each layer's hottest experts (T2 partition + hot set), box 2 the rest.
  • Serving: the multistream arena (8 slots, prefill/decode bursts, one or two lanes per step), arena DSpark (speculative decoding while at most V41_MS_DSPARK_STREAMS streams are live; two since 2026-10-06), arena stage graphs keyed by (stage, rows), snapshot / prefix restore, vision (v41/VISION_PORT.md), OpenAI-compatible API, and /v1/embeddings through the embed phase (Qwen3-Embedding-4B on the dGPU, --embed-gguf).
  • Dropped: V4-Flash / GGUF (2026-09-24). Its code is still in the tree until the removal branch merges; its docs are history.
  • Laguna support lives beside it (laguna/), CLI only.

Living references — keep these in sync with the code

doc what
v41/ARCH_SPEC.md model constants, tensor names, engine presentation
GLOSSARY.md canonical names (owner sign-off 2026-10-04); scripts/ontology_census.py ratchets it
v41/TUNING.md knobs and host settings; §0 = production vs code default
v41/KNOWN_BUGS.md correctness ledger; its status index lists what is open
v41/REMOTE_EXPERTS.md box-2 daemon and wire protocol (cited from the code)
crates/v4flash-kernels/src/knobs.rs (header) how a knob is resolved: default < env < legacy file < V41_KNOBS_FILE; live vs static
v41/KNOB_AUDIT_2026-10-04.md dated inventory of env reads still outside knobs.rs

Production configuration is not in the repo. The hub is launched by ~/run_v41_server.sh plus env files under ~/scratch-ms/, and live knobs come from /dev/shm/deepstrix-knobs.txt (box 2: ~/expertd-knobs.txt). Where a doc says "default", it should mean the code default; v41/TUNING.md §0 has the production values as of 2026-10-04.

Designs and plans in flight

doc state (2026-10-04)
v41/MS_DSPARK_STREAMS_DESIGN.md built; two speculating streams live since 2026-10-06, default 2 (b5def35)
v41/GRAPH_KEYS_DESIGN.md built; V41_MS_GRAPH_KEYS=stage_b live since 2026-10-05, default since b5def35
v41/EMBED_PHASE_DESIGN.md built; embed phase deployed 2026-10-05
v41/DSPARK_ARENA_PLAN.md built: the live DSpark path
v41/MULTISTREAM_DECODE_PLAN.md M1 built; later milestones open or replaced
v41/KV_PREFIX_STORE_DESIGN.md approved design; M1 code on branch worktree-kv-prefix-store
v41/EVTRACE_REBUILD_PLAN.md P0, P1, P3 built
ONTOLOGY_REVIEW_2026-10-04.md glossary adopted; renames pending
v41/HOT_SPLIT_SIM.md simulator + its recommendations (scripts/split_sim)

History

Everything else is a dated record. Docs overtaken by later work carry a Status (docs audit 2026-10-04) banner under the title saying what changed and where the truth now lives. Rough groups:

  • V4.1 engine bring-up (09-11..09-18): v41/PLAN.md, v41/ENGINE_PORT.md, v41/STATE_2026-09-13.md, v41/REVIEW_2026-09-13.md, v41/M0_AUDIT.md, v41/M7_EXPERT_TIER.md, v41/INDEXER_PORT_PLAN.md, v41/INDEXER_DENSE_VS_SPARSE.md, v41/ROADMAP_2026-09-18.md, v41/SECOND_BOX.md (+ v41/second_box_flake.patch, superseded by the lumi flake).
  • Decode, paging and placement investigations: v41/DECODE_*, v41/decode_designs/, v41/MULTISTREAM_M1A_*, v41/COLD_EXPERT_CACHING.md, v41/EXPERT_ALLOCATION_REWORK.md, v41/ALLOCATION_STRATEGY.md, v41/GLOBAL_POOL_FEASIBILITY.md, v41/BOX2_MISS_SUBSTITUTION.md, v41/LINK_IDLE_LATENCY.md, v41/PREFETCH_STUDY_2026-09-18.md, v41/ROUTER_LOOKAHEAD_AND_PREFETCH.md, v41/STATIC_PLACEMENT_DOES_NOT_GENERALIZE.md, v41/WHY_THE_BIG_POOL_REGRESSED.md, v41/SLOT_RESPLIT_MEASURED.md, v41/MIXED_PRECISION_BY_RESIDENCY.md, v41/CLAIM_CAP_ZERO_RETRACTED.md, v41/MATCHING_THE_SINGLE_BOX_DECODE.md, v41/EXPERT_FORMAT_IS_THE_REMAINING_LEVER.md, v41/ROUTED_MOE_IS_NOT_A_KERNEL.md.
  • DSpark before the arena (legacy prefill-shaped verify, 09-13..09-16): v41/DSPARK_DESIGN.md, v41/DSPARK_BUILD_PLAN.md, v41/DSPARK_STATE_2026-09-15.md, v41/DSPARK_VERIFY_*, v41/DSPARK_WHAT_IS_LEFT.md, v41/DSPARK_DRAFTER_GAP.md, v41/DSPARK_ACCEPTANCE_INVESTIGATION.md, v41/DSPARK_SINGLE_STREAM_PERF.md, v41/VERIFY_*, v41/WHY_THE_VERIFY_SPLIT_CHANGES_FAILED.md.
  • Audits and reviews: ARCHITECTURE_REVIEW_2026-09-24.md, v41/AUDIT_2026-09-22.md, v41/PROFILING_AUDIT_2026-09-21.md, v41/KERNEL_PERF_REVIEW.md, v41/OVERNIGHT_2026-09-22.md, v41/PREFILL_100K_PROFILE.md, v41/PREFILL_BALANCE_2026-09-12.md, TOOL_PROMPT_FIDELITY.md.
  • V4-Flash / GGUF era (05..09-10): DESIGN.md (the original design), PHASE0.md, PHASE1_*, PHASE2_KERNEL_VALIDATION.md, M40_*, M5x_*, FP8_KV_IMPL_2026-09.md, E2M1_INDEXER_KEYS_2026-09.md, IQ2_*, IQ3_S_*, IGPU_VRAM_POOL_PLAN.md, VRAM_FREE_PLAN_2026-09.md, EJPIR_DS4HIP_PREFILL_ANALYSIS.md, UNSLOTH_UD_IQ2XXS.md, VISION_STATUS.md.
  • Laguna: laguna/ (laguna/PLAN.md §0 has the outcome).
  • CREDITS.md.

Conventions

  • Dates in docs are UTC (YYYY-MM-DD, HH:MM UTC).
  • "Default" = the code default. A launch-script value is "production sets X".
  • When a measurement or figure is retracted, put a banner on every doc that stated it, not only on the new doc that retracts it.
  • A plan's Status: line says built / partly built (which parts) / abandoned / superseded by X. Update it when the code lands, not later.
  • Link Claude memory notes sparingly: they are outside the repo and a reader without them cannot follow the reference.

About

DeepSeek v4.1 Flash inference server, specialized for Strix Halo + dGPU configurations

Resources

Stars

2 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages