Commit Graph

Select branches

Hide Pull Requests

dependabot/pip/pip-63a259946c

main

#1

#1/merge

Mono Color

• b46ca8e025 Bump the pip group across 1 directory with 2 updates dependabot/pip/pip-63a259946c dependabot[bot] 2026-11-01 02:12:35 +00:00
• c62d5e3a15 DeepSeek plan: the NVIDIA cards it counts are the ones setup runs (a card of 8 GB or less starts Off beside a bigger one), and the plan says which card is not used main Niko1221 2026-10-31 03:23:01 +01:00
• 7aa9a584aa Setup: say which card is not used when a card under 8 GB sits beside a bigger NVIDIA card (a line only, no change to the cards used) Niko1221 2026-10-31 03:10:00 +01:00
• 987ebd442f Launch: list the Arc Pro B65 (0xe222) and the other discrete Battlemage/Alchemist ids, so two cards are both named for a launch that names none Niko1221 2026-10-31 02:57:34 +01:00
• e4031fc4d0 AMD cards fix, review round: setup's pinned one-card pick is still written (with gpus_by), only DeepSeek on AMD is judged when a copy adopts an earlier card, the launcher uses the plan's card rule and remembers a named card, Windows uses setup's amd_rank and no Linux arch list, message names strata launch --gpu Niko1221 2026-10-31 03:05:41 +01:00
• dec8f1366c Tests for the AMD card rules (4 cards, APU, mixed architectures, Windows, --gpu, --gpus, old configs, Intel, NVIDIA unchanged); stale-stamp test stubs setup.gpus Niko1221 2026-10-31 02:56:19 +01:00
• 6f7f763ce5 DeepSeek on several AMD cards uses every card when none was named: setup's own pick is no longer 'asked'; old configs of that pick move to every card at the start; the launcher and the app stop keeping it Niko1221 2026-10-31 02:56:19 +01:00
• f93ec8a924 DeepSeek card rule shared by the plan, the options and setup: every card that counts, one AMD card on Windows, only cards the engine package has code for, an APU beside a discrete card left out Niko1221 2026-10-31 02:56:19 +01:00
• 5322f1c2bd NOTICE for libstrix (asf0, MIT): the gfx1151 fused-expert barrier follows its barrier() Niko1221 2026-10-31 01:21:21 +01:00
• 4cf0d8cd38 Merge f-039-mac-dsu-rel (33bbbca44) Niko1221 2026-10-30 22:59:51 +01:00
• 33bbbca442 Offer UD-IQ1_M on a Mac too: not run there, said plainly in its plan (the Mac gate stays as an empty mechanism) Niko1221 2026-10-30 22:59:35 +01:00
• 38f56619cc Both Unsloth DeepSeek sets are offered on a DGX Spark: both passed a real Spark run (AARCH64_NOT_YET is empty); tests keep the mechanism Niko1221 2026-10-30 22:50:34 +01:00
• 1a26492064 Merge f-039-mac-dsu-rel (6a7d75057) Niko1221 2026-10-30 22:46:09 +01:00
• bee469c462 serve: the qualified GDN speedups are on from the server (Linux CUDA: GDN_TC + GDN_CONVL2=2; Linux HIP gfx12 only: GDN_CHUNKED), Qwen3.8 only, explicit env wins Niko1221 2026-10-30 22:40:20 +01:00
• 6a7d750578 UD-IQ1_S on a Mac: the measured line from the M5 Max run (alone: 15-18 tok/s, 3K prompt 230-310, 14K 520) Niko1221 2026-10-30 22:44:19 +01:00
• 417765b8d8 tests: the Spark gate still holds both Unsloth sets Niko1221 2026-10-30 22:36:49 +01:00
• 5c25767fcd Offer Unsloth UD-IQ1_S on a Mac (run on an M5 Max); UD-IQ1_M stays off until it has run Niko1221 2026-10-30 22:14:50 +01:00
• ea5dbffa7f Metal: Q6_K multi-column dense matvec (ds4_q6k_mvn) Niko1221 2026-10-30 20:09:02 +01:00
• a3a1103c13 dsu_parity: --bench and the ds4_mm pipelines Niko1221 2026-10-30 20:08:18 +01:00
• ba3d28b7e3 Q6_K matvec: packed loads via uchar4 Niko1221 2026-10-30 20:03:40 +01:00
• 0dba618d34 Metal Q6_K dense matvec: unaligned block loads (210-byte blocks), found by dsu_parity; test fix Niko1221 2026-10-30 20:03:18 +01:00
• 5ff7da4a70 Metal tests for the Unsloth set kernels: IQ1_S matvec and tile loader, Q6_K dequant and matvec, reference checked against ggml Niko1221 2026-10-30 20:00:05 +01:00
• 1b6c67122a Metal DeepSeek engine: IQ1_S expert kernels (decode GEMV, window, prompt tile) and a Q6_K dense dequant for the Unsloth sets Niko1221 2026-10-30 19:58:32 +01:00
• 7851a8dab6 Merge f-038-ds-unsloth (final 4148c5125) Niko1221 2026-10-30 20:26:56 +01:00
• 4148c51257 launch: ask dsv41_fit.not_here directly (the launch tests stub the setup object) Niko1221 2026-10-30 20:25:25 +01:00
• bd3c5cd4cb Keep Unsloth's 1-bit DeepSeek sets off a DGX Spark until a Spark run passes (AARCH64_NOT_YET, one rule beside the Mac one) Niko1221 2026-10-30 20:23:59 +01:00
• 99866a47a2 Merge f-029-urgent (bc694233b) Niko1221 2026-10-30 20:06:15 +01:00
• 73e0a3b6ed Merge f-038-ds-unsloth (final 9144da7b1) Niko1221 2026-10-30 20:00:26 +01:00
• 9144da7b1a tests: the launch list's download sizes include the two Unsloth sets Niko1221 2026-10-30 19:59:31 +01:00
• a833fb46f3 Merge f-038-ds-unsloth (macOS gate 32806f82a) Niko1221 2026-10-30 19:58:24 +01:00
• 32806f82aa Keep Unsloth's 1-bit DeepSeek sets off the Mac until the Metal engine has an IQ1_S kernel (lists, --model, --model-path, app check, setup) Niko1221 2026-10-30 19:57:24 +01:00
• bc694233bf Hardware list: AMD cards are counted from the kernel's KFD files when there is no setup module and no rocm-smi (ROCm from a Python venv), instead of not at all Niko1221 2026-10-30 19:45:23 +01:00
• dd3f9ec82f tests: the catalog download sizes cover the two Unsloth DeepSeek sets (241.0 and 258.4 GB) Niko1221 2026-10-30 19:35:11 +01:00
• 47cbb76690 Merge f-038-ds-unsloth (current head) Niko1221 2026-10-30 19:28:13 +01:00
• ab083a693f Chat: streamed text is applied about ten times a second instead of once per token, so a long answer no longer saturates the page (main-thread time 15.7 s to 9.0 s, peak heap 114 MB to 62-75 MB for a 40,000-token answer; Mio's Out of Memory report) Niko1221 2026-10-30 19:24:02 +01:00
• ad4755c249 Say plainly that the 1-bit UD-IQ1 sets are weaker on long or careful work Niko1221 2026-10-30 19:23:07 +01:00
• b60fbb54f7 serve: the parallel stack default is Linux CUDA only (HIP and Windows get nothing: no measurement outside the replica config) Niko1221 2026-10-30 19:20:29 +01:00
• 8f13b8c4ad Server: one plain line at start when a graphics card in use is of an architecture the compiled engine was not built for (#4662) Niko1221 2026-10-30 19:18:45 +01:00
• 0b1808f072 Update: a failed test start of the new engine names the kept log file to report (#4692) Niko1221 2026-10-30 19:18:01 +01:00
• cfa7f737cc Setup: a compiled engine gets code for every usable NVIDIA card in the PC, not only the card setup picked (a TITAN RTX + RTX 3080 build for sm_75 only failed every long prompt with unspecified launch failure) (#4662) Niko1221 2026-10-30 19:18:01 +01:00
• b3de587e03 Merge f-036-strix-u (7f6bc3308): the Windows gfx1151 default set holds both STRATA_SELECT_WMMA and STRATA_PF_F16 (seven switches); test counts updated Niko1221 2026-10-30 19:17:27 +01:00
• 7f6bc33081 gfx1151 on Windows: the prompt experts run on the F16 matrix-core kernels by default (STRATA_PF_F16) Niko1221 2026-10-30 19:15:54 +01:00
• 1ead709790 Control: the time-skip check and tune engine reads also tolerate non-UTF-8 output (#4675) Niko1221 2026-10-30 19:12:42 +01:00
• dd63f88fdd Qwen3.6 engine: the warm-up reads only cells that exist and starts its small probes in a fresh sequence, so a prompt chunk equal to the context no longer writes past the KV cache (#4708) Niko1221 2026-10-30 19:12:21 +01:00
• 2a9ccfe3ea serve: the parallel stack default only where it was run (Linux CUDA nine switches, Linux HIP seven, nothing on Windows or macOS) Niko1221 2026-10-30 19:07:47 +01:00
• dd471d3140 serve: the exact parallel stack is the default with batch slots (explicit env still wins) Niko1221 2026-10-30 18:43:45 +01:00
• 4a75fa2b62 API gateway: a client that hangs up right after data: [DONE] no longer turns a finished request into "Request cancelled." (#4690) Niko1221 2026-10-30 19:07:42 +01:00
• 3282c61343 Server: a tool the conversation already called counts as declared, so a call to it comes back as a tool call, not raw text (#4663) Niko1221 2026-10-30 19:06:50 +01:00
• 209dcadb3f Parser: a fence is closed only by one as long as it opened with, and an indented code block is a quote, so calls shown in them are not run (#4702 #4703) Niko1221 2026-10-30 19:06:11 +01:00
• e66a445892 Server: a picture reader written into an older engine folder follows the engine in use after an update (#4658) Niko1221 2026-10-30 19:03:21 +01:00
• 68718c33fa Monitor: the drive-temperature reader (every physical drive, from the first second) runs only while the Monitor is open, not at app start (#4661) Niko1221 2026-10-30 19:02:16 +01:00
• ea52e55cd7 Chat: only the streaming message renders again on each token, so a long chat no longer redoes its whole history per token (Out of Memory report from Mio; heap and main-thread time measured) Niko1221 2026-10-30 18:58:30 +01:00
• 08e8770c6e Prefill (layer-major): an expert with no mapped blob is copied by its source instead of ending the request (#4695) Niko1221 2026-10-30 18:46:16 +01:00
• 7146d6983f Setup, Control and Terminal: a child's non-UTF-8 output (GBK adapter names) no longer kills Step 1 or launch (#4675) Niko1221 2026-10-30 18:44:41 +01:00
• 01ab6b51bb Merge f-036-strix-ds (957038a1c) Niko1221 2026-10-30 18:43:54 +01:00
• 957038a1cd dsv41 (Windows HIP, integrated GPU): the shards' views are dropped after the load by default (DS4_NO_VIEW=0 turns it off, =1 forces it elsewhere) Niko1221 2026-10-30 18:43:29 +01:00
• d340dc8dd1 apps/host test: planning a whole-drive models folder no longer deletes <drive>\Strata (#4689) Niko1221 2026-10-30 18:43:11 +01:00
• ffaa6ad78c tests: the app preview and launch list know the two Unsloth sets Niko1221 2026-10-30 18:41:24 +01:00
• 4006305476 tests: the lockfile versions are 0.2.9 Niko1221 2026-10-30 18:40:48 +01:00
• 97e6f1fe19 Version 0.2.9: Terminal and app 0.2.9, engine and Control 0.1.55 (every Qwen and DeepSeek package, Metal and SYCL included) Niko1221 2026-10-30 18:40:46 +01:00
• 33bd74182b Merge f-038-ds-unsloth Niko1221 2026-10-30 18:40:13 +01:00
• 4d12701181 Merge f-0281-ds Niko1221 2026-10-30 18:40:13 +01:00
• aed382885e Merge f-036-strix-ds Niko1221 2026-10-30 18:40:12 +01:00
• 274257630b Merge f-036-strix-u Niko1221 2026-10-30 18:40:12 +01:00
• d96873fc56 Merge f-036-strix-q Niko1221 2026-10-30 18:40:11 +01:00
• 86bed1c92f Merge f-037-arc (5eaee81a4) Niko1221 2026-10-30 18:40:11 +01:00
• 203df526f1 serve/test_admit_order: the engine stand-in sets the arrivals deque the coalescing code reads Niko1221 2026-10-30 07:27:44 +01:00
• b6fc410ee5 Metal window.cpp: qualify NG_HC_DIM / NG_HIST in finish_rows (the --batch rows code did not compile in the host build) Niko1221 2026-10-30 07:04:13 +01:00
• 0a76d62e68 pinned.cu: the shared-arena tmpfs check is Linux-only (sys/vfs.h and statfs f_type), so the Metal host build compiles again Niko1221 2026-10-30 06:53:55 +01:00
• c0e245c8be Merge f-029-par-metal (766b54afb) Niko1221 2026-10-30 18:39:27 +01:00
• 16e08b50cd Merge f-029-par-amd (4772f0cd9) Niko1221 2026-10-30 18:39:27 +01:00
• 1fb3e9a326 Merge f-029-par-harness Niko1221 2026-10-30 18:37:54 +01:00
• 1f8f2fa8ad Merge f-029-par-big Niko1221 2026-10-30 18:37:53 +01:00
• 3a76d5eaf6 Merge f-029-gdn-mx Niko1221 2026-10-30 18:37:53 +01:00
• 01ef876562 Merge f-029-gdn-amd Niko1221 2026-10-30 18:37:53 +01:00
• 4746bfb165 Merge f-029-gdn-cuda Niko1221 2026-10-30 18:37:52 +01:00
• 0b1fbf4012 gfx1151 on Windows: STRATA_SELECT_WMMA joins the default prompt set (x1.008 at 8K, x1.037 at 32K, x1.077 at 64K; about 10 of 2,051 selected cells flip at the cutoff for queries past ~5K tokens) Niko1221 2026-10-30 18:22:35 +01:00
• 6923421a61 Accept Unsloth's DeepSeek UD-IQ1_S and UD-IQ1_M sets (pins, models, app file check, fit numbers; opt-in names) Niko1221 2026-10-30 18:05:31 +01:00
• 09dc5c52b3 gfx11 gr_upmix: 256 tokens a block (16 waves) halves the w_up staging per token; prefill timing splits the QSA projection stage Niko1221 2026-10-30 17:55:46 +01:00
• 5eaee81a43 [SYCL] kprof: profiling-tag sections time stretches without events of their own (the oneMKL GEMMs) Niko1221 2026-10-30 16:33:01 +01:00
• d936db6547 gfx1151: a Q8_0 down expert layer takes the F16 matrix-core prompt kernel by default Niko1221 2026-10-30 16:15:54 +01:00
• ac3e0e17b5 GDN gfx12: the chunked path from 64 tokens, as on gfx11 (owner: every gain counts; final gate at 64: +1.7% to +6.1%, KL 0.0025/0.038, 55/61 tasks, soak clean) Niko1221 2026-10-30 16:05:14 +01:00
• b3bc477512 diagnostics: STRATA_QSA_STAT (overlap of consecutive queries' selections), hip_pa_bench target and PA_BENCH_REF (compare two variants of the prompt attention on a recorded selection) Niko1221 2026-10-30 15:41:53 +01:00
• db729163ce Revert "setup: a Strix Halo's BIOS carve-out holds expert slots of the Unsloth RAM-budget plans" Niko1221 2026-10-30 15:41:41 +01:00
• 19b5f873f2 dsv41 (Windows): DS4_NO_VIEW=1 drops every view of the shards and the load's buffered handles after the load, reads go unbuffered Niko1221 2026-10-30 14:56:43 +01:00
• 0122eceb4c [SYCL] --batch-groups: 'auto' is understood, and two Arc B-series cards with 8 or more slots default to 2 pipelined groups (INFO batch_groups); STRATA_BATCH_GROUPS_AUTO=0 / --batch-groups 1 turn it off Niko1221 2026-10-30 13:56:52 +01:00
• 8401c21f9d [SYCL] prompt attention on the int8 DPAS (ESIMD, Battlemage): hi/lo int8 Q and P, 2D block loads for the four-cell selection blocks; default on for Battlemage, STRATA_PROMPT_ATTN_ESIMD=0 off Niko1221 2026-10-30 13:22:33 +01:00
• 9c2121cf5c dsv41: the CPU lane's service thread spins through a decode token (DS4_LANE_SPIN_MS, 250) and naps on a high-resolution timer; DS4_PROF_SYNC decode stage profile; DS4_LANE_SKIP timing oracle Niko1221 2026-10-30 12:26:18 +01:00
• 1cfffa2ede gfx1151 fused experts: one barrier per stage (explicit LDS wait, after libstrix's barrier(), asf0/libstrix MIT) and 3 blocks per WGP for IQ2_S/IQ2_XS gate/up on gfx11 Niko1221 2026-10-30 12:14:41 +01:00
• a83252a7b7 gfx1151: the VGPR cap of the occupancy expert kernels applies on clang 24 as on clang 23 (IQ3_S/IQ4_NL layer 79.7 -> 69.9 ms at 16K tokens, parity unchanged) Niko1221 2026-10-30 11:40:16 +01:00
• d8056fbe00 [SYCL] STRATA_KPROF=1: per-kernel GPU timer (event-profiled queues, kp() launch wrapper, decode/prompt tables); tools/arc A/B harness Niko1221 2026-10-30 11:27:59 +01:00
• 6fb110e557 Opt-in research switches: STRATA_PAR_COUNTERS, STRATA_PIPED_CONCURRENT_MIN, multi-prompt layer-major read (Prefill::run_multi, BCO co-reads, server STRATA_COALESCE_PREFILL), STRATA_CRASH_BT Niko1221 2026-10-30 11:08:53 +01:00
• 28c4543e48 tools: parallelism load harness (par_load, par_analyze, par_run, sweep, recon), E00 capability probe, batch_test --co/--seq/--pad and the phase-1/2 variant lists Niko1221 2026-10-30 11:08:53 +01:00
• e1e7e32680 prefill timing: STRATA_PREFILL_TIMING also splits the hyper-connection read into down, up, inject and mix Niko1221 2026-10-30 11:03:01 +01:00
• 94fc01c6fd setup: a Strix Halo's BIOS carve-out holds expert slots of the Unsloth RAM-budget plans Niko1221 2026-10-30 11:00:42 +01:00
• b0ba459514 Merge f-029-par-sycl (SYCL batch windows, GDN/attn rows kernels) Niko1221 2026-10-30 10:56:29 +01:00
• 2fef430202 Merge f-034-cand (0.2.9 candidate) into f-036-strix Niko1221 2026-10-30 10:52:31 +01:00
• 7a9245b6c0 [SYCL] batch window: port the one-launch-for-all-rows kernels (GDN groups, block scores rows, decode attention rows; opt-in STRATA_BATCH_GDN_GROUPS / STRATA_BATCH_ATTN_ROWS) and the batch-window stage profile; no speed gain on 2x Arc B65 Niko1221 2026-10-30 08:25:37 +01:00
• 368bcb7a26 serve/test_admit_order: the engine stand-in sets the arrivals deque the coalescing code reads Niko1221 2026-10-30 07:27:44 +01:00
• 5ad0135c00 Metal window.cpp: qualify NG_HC_DIM / NG_HIST in finish_rows (the --batch rows code did not compile in the host build) Niko1221 2026-10-30 07:04:13 +01:00