Repository navigation

Chore(github): Create a .github folder - #401

Closed

Joe-Huber wants to merge 407 commits into

Niko1221:mainfrom

Joe-Huber:create-github-folder

Closed

Chore(github): Create a .github folder#401
Joe-Huber wants to merge 407 commits into
Niko1221:mainfrom
Joe-Huber:create-github-folder

This pull request is big! We're only showing the most recent 250 commits

Commits on Sep 29, 2026
Commits on Sep 30, 2026
- fix(engine): clear stale error/cancellation state between requests and before prompt prefill
  
  praveshkhatana
  authored and
  Niko1221
  committed
  6cad1fe
- verify: cudaFuncGetName only with CUDA 12.3+ (an sm_75 build with CUDA 12.0 failed)
  
  authored and
  Niko1221
  committed
  a897541
- Expert cache on WDDM: 1 GiB steps before quarter steps (~4 GiB more cache on a 32 GB card); Ctrl+C stops the server on Windows; no stack trace when a client hangs up (#153)
  
  authored and
  Niko1221
  committed
  6a45e59
- engine: the expert cache reserves the draft head's VRAM (0.1.27's CJK subset came out of the 700 MiB reserve, #199); a clear message when no VRAM is left for the cache (#174); the prompt's verify-w…
  
  Niko1221
  and
  claude
  committed
  b981f63
- server: /status needs the API key and forgets the answer's end (#212); constant-time key check, an explicitly empty key refuses to start (#213); an engine exit quotes its last log line (#215); tool…
  
  Niko1221
  and
  claude
  committed
  4d25c61
- setup: --host/--api-key/--draft-vocab on a start are kept for the model (#179); --draft-vocab en keeps the English/code subset (~110 MiB less VRAM); CUDA 13.0 for RTX 50 builds (#220); llama.cpp's …
  
  Niko1221
  and
  claude
  committed
  0fc1a1e
- docs: the draft layer's token subset and --draft-vocab
  Niko1221
  committed
  349f8f9
- Engine 0.1.28: the expert cache reserves the draft head's VRAM (0.1.27 left 16 GB cards below the stall line), a cancelled request no longer fails the next one, the API key protects /status; setup …
  
  Niko1221
  and
  claude
  committed
  bbaaabb
- s_gemv: the q8 split kernel reaches its barrier before a rowless warp returns
  
  gputier
  authored and
  Niko1221
  committed
  7ed0e90
- bf16: a NaN stays a quiet NaN in f32 -> bf16, as in ggml
  
  gputier
  authored and
  Niko1221
  committed
  6dc9fe8
- qsa: mask cells whose page did not resolve; refuse a clock sweep over fewer slots than the resolve block
  
  gputier
  authored and
  Niko1221
  committed
  f3925cf
- expert_source, s2_expert_grouped: guard the verify window's fixed tables
  
  gputier
  authored and
  Niko1221
  committed
  64cfe3c
- gguf_reader.hpp: define NOMINMAX before windows.h
  
  gputier
  authored and
  Niko1221
  committed
  31ed0e2
- build: guard the last two bench/micro targets on their file, as upstream does for the others
  
  gputier
  authored and
  Niko1221
  committed
  71d74cb
- E-2 on the AVX-2 path too: prefetch the i-quant expert rows (same switch, CPUs without AVX-512)
  
  pipeob0
  authored and
  Niko1221
  committed
  fa6310b
- prefill: software-pipelined loads in the column-split GDN recurrence (bit-identical; STRATA_GDN_PIPELINE=0 = old kernel)
  
  q8atnight
  authored and
  Niko1221
  committed
  ac6aad4
- decode: QSA block scores read each key block once for all of a window's queries (bit-identical; STRATA_SCORES_MULTI=0 = old grid)
  
  q8atnight
  authored and
  Niko1221
  committed
  a20f3b5
- sampler: top_k by lexicographic threshold, split over the whole GPU
  
  gputier
  authored and
  Niko1221
  committed
  47d6894
- serve: on Windows the engine, vision encoder and MCP servers end with the server
  
  authored and
  Niko1221
  committed
  e5e748b
- decode: QSA scores multi - comment names every caller that takes it (decode captured or not, prefill's pooled16); no box-specific note (review)
  
  q8atnight
  authored and
  Niko1221
  committed
  b4d788e
- prefill: GDN recurrence pipelining - the neutral 16-column variant (STRATA_GDN_CB) dropped, one kernel on upstream's split (review)
  
  q8atnight
  authored and
  Niko1221
  committed
  04f4a53
- ple: --ple-io ram locks the n-gram table in RAM (no SSD read on the prompt/token path)
  
  dsh
  authored and
  Niko1221
  committed
  12fd16c
- ple: --ple-io ram: refused on Windows (no mlock), 'locked' only when mlock succeeded else 'loaded (not locked)', neutral comments + --help (review)
  
  q8atnight
  authored and
  Niko1221
  committed
  9ec069b
- perf(prefill): allow explicit routed-only staging for large chunks
  
  rluisr
  authored and
  Niko1221
  committed
  cf0d12c
- WSL2: survive the driver's ~1 GiB pinned host budget with 3 GPUs
  
  authored and
  Niko1221
  committed
  85049da
- Show prefill speed alongside decode speed in dashboard
  
  authored and
  Niko1221
  committed
  2152c8f
- setup: the 3-bit models' 262K context is counted against RAM, not a fixed 90 GB
  
  architectds
  authored and
  Niko1221
  committed
  f2b5b1c
- generate: zero session state for a native pack before the prompt path
  
  samuelishida
  authored and
  Niko1221
  committed
  853aa46
- tests: bitwise check of native_mmvq's multi_exact contract, with a negative control (#148; CMake on the backend variables, CUDA-only)
  
  authored and
  Niko1221
  committed
  9086c36
- generate: the session state is zeroed for every pack in CLI generate mode (#167 follow-up); no .pi/ ignore line
  
  Niko1221
  and
  claude
  committed
  ec64796
- A fatal CUDA fault in the prompt path exits at once instead of hanging until the watchdog (#224); CMake warns when sm_120 is built with CUDA < 13.0 (#220, #224); iq_pack writes native_experts.txt u…
  
  Niko1221
  and
  claude
  committed
  89d423f
- Engine 0.1.29: sampled answers up to 40% faster with the same tokens (#197), faster prompt kernels (#187, #188, #207), correctness fixes (#154); setup requires it
  
  Niko1221
  and
  claude
  committed
  d6708a4
- kernels: rope scaling config and the scaled cos/sin table (none/linear/YaRN)
  
  authored and
  Niko1221
  committed
  249e27a
- rope: thread scaling through the native, indexer and prefill rotation sites
  
  authored and
  Niko1221
  committed
  05d96df
- engine: --rope-scaling/--rope-scale/--yarn-* flags; rope keys as GGUF defaults
  
  authored and
  Niko1221
  committed
  1018f29
- setup: --rope-scaling/--rope-scale; contexts past 262144 require scaling
  
  authored and
  Niko1221
  committed
  53b5426
- docs: context extension (rope scaling)
  
  authored and
  Niko1221
  committed
  eaeb0c0
- generate: an explicit --rope-scaling none or --rope-scale 1 overrides the GGUF rope keys
  
  authored and
  Niko1221
  committed
  3bb8abd
- bench: rope-scaling needles past the trained 262K (linear 2, yarn 2)
  
  authored and
  Niko1221
  committed
  0c1efc1
- indexer: the batched append takes the one rope config, with a parity test
  
  authored and
  Niko1221
  committed
  a564868
- setup: the documented automatic rope configuration actually happens now
  
  authored and
  Niko1221
  committed
  55ec15b
- setup: a free-form --context past the menu got a raw StopIteration
  
  authored and
  Niko1221
  committed
  52b2e81
- setup: the automatic rope factor derives from the context actually served
  
  authored and
  Niko1221
  committed
  04949fe
- test: pin the short-context rope table (§4, no factor-2 fallback in the window)
  
  authored and
  Niko1221
  committed
  d5aab1a
- test: the indexer batch-vs-sequential parity covers linear and compares bits
  
  authored and
  Niko1221
  committed
  36edb03
- prefill: the draft layer's batched RoPE takes the one rope config
  
  authored and
  Niko1221
  committed
  c264737
- generate: the in-window scaling note stops claiming YaRN's magnitude correction (TODO 22)
  
  authored and
  Niko1221
  committed
  0a6a180
- Integrate validated conversation snapshot core with upstream 0.1.25
  
  jeremiahritchey
  authored and
  Niko1221
  committed
  ccc660b
- test: require a real physical-memory sample on Windows
  
  midhatn
  authored and
  Niko1221
  committed
  bca029f
- Add reproducible Pi cache comparison and document current integration
  
  jeremiahritchey
  authored and
  Niko1221
  committed
  75948df
- Enforce Pi benchmark request limits and engine memory evidence
  
  jeremiahritchey
  authored and
  Niko1221
  committed
  77e4830
- Match upstream K8V4 flags when sizing conversation snapshots
  
  jeremiahritchey
  authored and
  Niko1221
  committed
  cea72be
- Allow bounded Pi task completion and record benchmark provenance
  
  jeremiahritchey
  authored and
  Niko1221
  committed
  db65252
- Record upstream integration, Linux acceptance and controlled Pi results
  
  jeremiahritchey
  authored and
  Niko1221
  committed
  7fa893c
- Focus shared-core patch on runtime, regression tests and current documentation
  
  jeremiahritchey
  authored and
  Niko1221
  committed
  cb9b3b1
- Clear recoverable snapshot errors before batched draft prefill
  
  jeremiahritchey
  authored and
  Niko1221
  committed
  33fe713
- Verify restored draft KV and fail incomplete state fingerprints
  
  jeremiahritchey
  authored and
  Niko1221
  committed
  50c7d40
- Refresh core validation record for upstream 0.1.27
  
  jeremiahritchey
  authored and
  Niko1221
  committed
  6622371
- Report actual draft residency in snapshot verification
  
  jeremiahritchey
  authored and
  Niko1221
  committed
  2cc4e76
- Record linked 0.1.27 core build and host validation
  
  jeremiahritchey
  authored and
  Niko1221
  committed
  980b354
- Record final Linux snapshot and RAM cache acceptance
  
  jeremiahritchey
  authored and
  Niko1221
  committed
  f9d99af
- Ignore llama.cpp dependency directories and symlinks
  
  jeremiahritchey
  authored and
  Niko1221
  committed
  8d5f80b
- Remove planning documents from the core PR
  
  jeremiahritchey
  authored and
  Niko1221
  committed
  6bf9c68
- Compare cache-disabled engines against untouched upstream
  
  jeremiahritchey
  authored and
  Niko1221
  committed
  d9a02e6
- Retain unchanged KV storage across repeated conversation parking
  
  jeremiahritchey
  authored and
  Niko1221
  committed
  cecf697
- Bound incremental snapshot fragmentation and preserve cache admission
  
  jeremiahritchey
  authored and
  Niko1221
  committed
  8dddc9e
- Require fixed binaries and budgets in the incremental capture gate
  
  jeremiahritchey
  authored and
  Niko1221
  committed
  c0d575c
- Document retained snapshot ownership and admission contract
  
  jeremiahritchey
  authored and
  Niko1221
  committed
  6865d4a
- prefill: every expert streams from 1024-token chunks (was 2048): +17-28% on 1-4K-token prompts, the same output; STRATA_PREFILL_STREAM_MIN overrides
  
  Niko1221
  and
  claude
  committed
  8acd17c
- server: optional idle unload, POST /unload and /load, a free-VRAM guard and a before_load hook
  
  authored and
  Niko1221
  committed
  84c19a4
- server: the image encoder unloads and loads with the engine; docs for sharing the GPU
  
  authored and
  Niko1221
  committed
  915773a
- server (#208 follow-up): a request stamps last_request_at in load() (the idle thread cannot unload between load and run); unload() sends QUIT before terminate; the VRAM guard checks the first GPU (…
  
  Niko1221
  and
  claude
  committed
  9bbee42
- Docker build (#96): the engine compiled in docker build, STRATA_EXECV so the server is PID 1, SIGTERM stops it cleanly, KV/GPUS/LOW_RAM as env vars
  
  authored and
  Niko1221
  committed
  a94fe10
- Pascal (sm_60): bring the engine up on compute capability 6.0
  
  ruibeikaa
  authored and
  Niko1221
  committed
  01bca8a
- core: add guarded shared expert arena backing
  
  rhgo1749
  authored and
  Niko1221
  committed
  dfa2bda
- docs: limit shared arena backing to dev shm
  
  rhgo1749
  authored and
  Niko1221
  committed
  f9dad01
- Multi-GPU: the session carve and the per-stage prompt loans
  
  authored and
  Niko1221
  committed
  77709ef
- wip: version 0.1.30
  Niko1221
  committed
  1f5db5f
- STATE_HASH: pooled= keeps its 0.1.29 meaning (the completed indexer rows); the extent with the spare row is pooled_full=
  
  Niko1221
  and
  claude
  committed
  01df436
- qsa: the table indexer's spare rotates at position 0 like the native one (TODO 7)
  
  Niko1221
  and
  claude
  committed
  93d81fb
- rope: one validity rule for the resolved config; none passes the identity constants to every analytic site
  
  Niko1221
  and
  claude
  committed
  71f031f
- docs: rope scaling is experimental and opt-in; the measured perplexity at 293K
  
  Niko1221
  and
  claude
  committed
  7699309
- Coupled draft sampling for the MTP drafter (STRATA_SPEC_COUPLED=1, off by default)
  
  Niko1221
  and
  claude
  committed
  1643965
- Low-RAM mode, resident variant: the experts the GPU does not hold stay in RAM (--resident-experts), chosen by setup
  
  Niko1221
  and
  claude
  committed
  872af82
- Merge f-84 into rel-0.1.30
  Niko1221
  committed
  483953f
- setup: the resident low-RAM variant keeps its RAM copy pageable on ROCm (STRATA_RESIDENT_PIN=0 in the config's env)
  
  Niko1221
  and
  claude
  committed
  7e4ee54
- Merge f-coupled into rel-0.1.30
  Niko1221
  committed
  1afdd47
- HIP: enable RDNA4 gfx1201 (RX 9070 / AI PRO R9700)
  
  doplxyz
  authored and
  Niko1221
  committed
  0704e7d
- generate: the option parser's else-if chain in two statements (MSVC C1061: more than 128 nested blocks after the new options); the same order and effects
  
  Niko1221
  and
  claude
  committed
  8616c60
- Merge f-resram into rel-0.1.30
  Niko1221
  committed
  538a0df
- Merge f-189 (PR #189, the conversation snapshot core and RAM cache) into rel-0.1.30
  
  Niko1221
  and
  claude
  committed
  703ef30
- Conversation snapshots and checkpoints follow the session carve (#216): a session saves and restores only the layers it owns
  
  Niko1221
  and
  claude
  committed
  e46ccca
- HIP: RDNA4 gfx1201 beside gfx1100; the startup check compares the card with the compiled archs; setup picks ROCm and the hipBLASLt table per arch
  
  Niko1221
  and
  claude
  committed
  9907379
- CMake: the D-1 prompt-attention parity tool is CUDA-only; the HIP build with tests configures again
  
  Niko1221
  and
  claude
  committed
  65e3df2
- doorbell_ring: a volatile store, so the host sees the ring on RDNA4 without a stream sync
  
  Niko1221
  and
  claude
  committed
  53f9e7a
- docs: RDNA4 (gfx1201) in AMD_HIP.md with the validation on the RX 9070 XT and R9700; how to calibrate a hipBLASLt table; README and setup's texts name the RDNA4 cards
  
  Niko1221
  and
  claude
  committed
  d8848f3
- Merge f-rdna4 (AMD RDNA4 gfx1201, with doplxyz's patch) into rel-0.1.30
  Niko1221
  committed
  3d99089
- CPU experts: the multi-token kernels from one token on (#152): a token's expert rows round the same whatever its verify window held, so output no longer depends on the drafts; STRATA_IQ_MT_MIN=2 re…
  
  Niko1221
  and
  claude
  committed
  a74b1cd
- tools: the cache-disabled gate compares every resolved setting except the engine version (release vs previous release)
  
  Niko1221
  and
  claude
  committed
  bcac869
- #152 as an opt-in (the owner's choice): STRATA_IQ_MT_MIN=1 makes CPU expert rounding independent of the drafting; the default stays 0.1.29's rule (IQ3_S decoded 1-3% slower with it on); docs; the w…
  
  Niko1221
  and
  claude
  committed
  b8c4bd1
- tests/hip: prefill_native_batch passes the indexer a RopeScaling (the #84 signature; plain rope over base 1e4 as before) - a HIP-only test the CUDA build never compiled
  
  Niko1221
  and
  claude
  committed
  86fd27f
- docs: the layer split on two AMD RDNA4 cards (identity with one card, checkpoints, speed; setup offers one AMD card)
  
  Niko1221
  and
  claude
  committed
  fb73292
- Engine 0.1.30: short prompts up to 28% faster (every expert streams from 1024-token chunks), the resident low-RAM variant, the multi-GPU session carve (#216), AMD RDNA4, idle unload (#208); opt-ins…
  
  Niko1221
  and
  claude
  committed
  30ec18e
- verify: STRATA_LOGPOS hook for per-position teacher-forced log-probabilities
  
  Enky
  authored and
  Niko1221
  committed
  06b82fe
- gemm: name the cuBLAS/CUDA call that failed at init, and flag resource shortages
  
  lukmanfauzie
  authored and
  Niko1221
  committed
  70204ca
- tests: make iq_parity reproducible from a fresh checkout (TODO 24)
  
  authored and
  Niko1221
  committed
  18827d8
- fused_gr: sm_75 down kernel with TILE=1280 staging, other archs unchanged
  
  authored and
  Niko1221
  committed
  68f1362
- q2_avx2: two-block 256-bit unpack for rows of 1-2 tokens (bit-exact)
  
  Hozák Martin
  authored and
  Niko1221
  committed
  9e9dce4
- serve: a request settles its status, history and totals while it still holds the fifo (#266)
  
  Niko1221
  and
  claude
  committed
  daaaf9d
- gemm (#240 follow-up): setup calls that failed before without a word are logged and survived as before (not fatal - hipBLAS can return not-supported), init_external names its failures too, the MiB …
  
  Niko1221
  and
  claude
  committed
  ba2cf61
- tests (#264 follow-up): the fixture step skips (exit 3) instead of failing when the vendored gguf-py is absent; no unused SCALE_F16
  
  Niko1221
  and
  claude
  committed
  e224a16
- tokenizer: a word longer than 64 symbols merges with a heap and a linked list, not a rescan per merge (#268)
  
  Niko1221
  and
  claude
  committed
  8aa8f51
- fix(serve): report a tool call cut off by the end of the output as unfinished
  
  alphastorm
  authored and
  Niko1221
  committed
  925c354
- serve: a whole OpenAI answer leaves out a tool call whose arguments don't parse unless it ends in tool_calls (#211)
  
  Niko1221
  and
  claude
  committed
  9a1fc19
- setup: the low-RAM mode's one-GPU warning says why, with the RAM math; --low-ram off --yes installs (#250)
  
  Niko1221
  and
  claude
  committed
  b1d808d
- kernels: faster grouped and per-hit Q2_0 expert kernels, bitwise identical
  
  gputier
  authored and
  Niko1221
  committed
  1d0c6e1
- s2_expert_grouped (#241 follow-up): the new chunk kernels' dp4a through STRATA_DP4A, so the sm_60 experimental build keeps its exact fallback
  
  Niko1221
  and
  claude
  committed
  a9868ca
- iq_kernels: decode each IQ weight part once for every column and entry
  
  gputier
  authored and
  Niko1221
  committed
  8ec94aa
- qsa_prompt_attn: the f16 tensor-core path runs on Turing (sm_75) as two m16n8k8 steps
  
  authored and
  Niko1221
  committed
  b6bdc70
- qsa_prompt_attn (#270 follow-up): the m16n8k8 pair only for sm_75+ CUDA device code (AMD and pre-Turing builds keep the trap, which the host never reaches); Turing chosen per call from the current …
  
  Niko1221
  and
  claude
  committed
  d9cbe3c
- setup: reproducible installs - the checkout's own engine release, pinned Hugging Face revisions, pinned Python packages (#214)
  
  Niko1221
  and
  claude
  committed
  ba5c387
- serve: an opt-in hard thinking budget, reasoning_budget_tokens (#123)
  
  Niko1221
  and
  claude
  committed
  48614ae
- README: where things are stored (#226)
  
  Niko1221
  and
  claude
  committed
  c8357b5
- load_experts_gguf reads with fread on a FILE* (#230)
  
  Niko1221
  and
  claude
  committed
  74ccff3
- Windows: cap the sliced arena pin at the GPU's shared-memory budget (#243)
  
  Niko1221
  and
  claude
  committed
  75c40ef
- Verify window: bounded final wait; release the GPU's flag waits before the engine ends (#267)
  
  Niko1221
  and
  claude
  committed
  7c24aed
- Stall report: name the prompt chunk and layer; mark a finished verify window (#251)
  
  Niko1221
  and
  claude
  committed
  9631c96
- Refuse a CPU without AVX2 at start, by name
  
  Niko1221
  and
  claude
  committed
  d4279d5
- Experimental Pascal/Volta build: admit sm_70, and the device check honours it (#236)
  
  Niko1221
  and
  claude
  committed
  3206593
- Merge branch 'f-131-fix' into r131-fix
  Niko1221
  committed
  95ac751
- mcp: a server that stops quotes its last log line reliably (the stdout reader waited no time for the stderr reader, so the reason was sometimes missing - the flaky test_a_server_that_fails_is_left_…
  
  Niko1221
  and
  claude
  committed
  3b1cbf5
- Split GGUF artifacts: every shard found (a missing one is an error), tensors from their owning shard, per-role expert files (native_experts.txt v4)
  
  Niko1221
  and
  claude
  committed
  02cfe36
- iq_pack: per-role shards and data_start (v4 only for a split layer), split keys checked, experts.bin reused only with its source sidecar, atomic publication
  
  Niko1221
  and
  claude
  committed
  53f967a
- iq_pack: each tensor in the form the engine reads (FORM), exact F32->BF16 by default, PLE conv F32 -> F16 by the loader, conversions.json; the engine refuses a wide conv1d
  
  Niko1221
  and
  claude
  committed
  bede5ca
- tests: --real modes for gguf_split_test and expert_layout_test (a real model's shards / a real pack's spans, headers only)
  
  Niko1221
  and
  claude
  committed
  2786095
- iq_kernels: Q4_K / Q5_K / Q5_1 / Q8_0 native experts (Unsloth UD-Q4_K_XL), per-role format lists, native_expert_supported at startup, Q8_0 embedding and PLE key
  
  Niko1221
  and
  claude
  committed
  efeffd8
- CS-T: native packs read their experts from the GGUF shards in place (--mmap-experts without experts.bin), --resident-budget-gib N, per-tier counters
  
  Niko1221
  and
  claude
  committed
  3889344
- verify (#267 follow-up): the window's final wait stays a blocking sync - the cudaStreamQuery poll cost IQ3_S ~3% decode; the stall watchdog's release of every verifier's GPU waits is what keeps a s…
  
  Niko1221
  and
  claude
  committed
  0e23f57
- HIP: __byte_perm as v_perm_b32 and SWAR __vsub4/__vsubss4/__vcmpne4 (from #262 by ttio2tech)
  
  committed
  87fbc00
- AMD: gfx1101 (RX 7800 XT, #254 by jhohertz) and gfx1200 (RX 9060 XT, #256 by Efeisot) as community-validated cards
  
  committed
  b0dda4b
- setup: AMD layer splits with --gpus (every chosen card a supported AMD arch; the engine compiled for each)
  
  Niko1221
  and
  claude
  committed
  cc97fcd
- pinned (#243 follow-up): the shared-memory-budget cap on the sliced pin is opt-in (STRATA_ARENA_PIN_GIB=auto) - by default the arena pins as in 0.1.30; on this 64 GB PC the cap pinned 26 instead of…
  
  Niko1221
  and
  claude
  committed
  e9e2ba3
- CS-T speed: the file tier fetched in parallel, PCIe planning without file reads, a partial pin of the RAM tier, bit-exact AVX2 multi-token kernels for Q4_K / Q5_1 / Q8_0 experts, decode-only tier s…
  
  Niko1221
  and
  claude
  committed
  cacffa9
- Merge branch 'f-131-srv' into r131-q4
  Niko1221
  committed
  a348c7a
- Merge branch 'f-q4-on242' into r131-q4
  Niko1221
  committed
  b93a00c
- CS-T: the partial pin at most 24 GiB (STRATA_PARTIAL_PIN_GIB), the Q4_K/Q5_1/Q8_0 multi-token kernels from two tokens on
  
  Niko1221
  and
  claude
  committed
  8f5148c
- CS-T: the GGUF-in-place prefetch asks Windows for all of a layer's missed slices at once (PrefetchVirtualMemory)
  
  Niko1221
  and
  claude
  committed
  5ffda19
- #253: the 8 GiB arena pin cap only under WDDM; STRATA_ARENA_PIN_GIB; the split's prompt chunk steps down
  
  Niko1221
  and
  claude
  committed
  ed6c124
- --split-skip-if-fits (opt-in): an auto split runs on the main card alone when it holds every profiled expert + KV
  
  Niko1221
  and
  claude
  committed
  450149d
- Merge branch 'f-131-amd' into r131-amd
  Niko1221
  committed
  6926e0b
- Revert #257 (AVX2 two-block Q2_0 unpack): bit-exact but not faster on an AVX2-only CPU (Ryzen 9 3900X: gate/up -2.6..-4.7% at 1 token, down -3..-5% at 2 tokens)
  
  Niko1221
  and
  claude
  committed
  c7aa3e5
- CS-T: the partial pin locks only what it does not register and ends at an expert boundary; the Q4_K/Q5_1/Q8_0 multi-token kernels opt-in (STRATA_KQ256=1)
  
  Niko1221
  and
  claude
  committed
  2736bd0
Commits on Oct 1, 2026

Commits

Commits on Sep 29, 2026

Commits on Sep 30, 2026