Niko1221
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
2026-10-31 03:23:01 +01:00
Niko1221
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)
2026-10-31 03:10:24 +01:00
Niko1221
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
2026-10-31 03:10:24 +01:00
Niko1221
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
2026-10-31 03:05:41 +01:00
Niko1221
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
2026-10-31 02:56:19 +01:00
Niko1221
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
2026-10-31 02:56:19 +01:00
Niko1221
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
2026-10-31 02:56:19 +01:00
Niko1221
5322f1c2bd
NOTICE for libstrix (asf0, MIT): the gfx1151 fused-expert barrier follows its barrier()
2026-10-31 01:21:21 +01:00
Niko1221
4cf0d8cd38
Merge f-039-mac-dsu-rel ( 33bbbca44)
2026-10-30 22:59:51 +01:00
Niko1221
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)
2026-10-30 22:59:35 +01:00
Niko1221
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
2026-10-30 22:50:34 +01:00
Niko1221
1a26492064
Merge f-039-mac-dsu-rel ( 6a7d75057)
2026-10-30 22:46:09 +01:00
Niko1221
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
2026-10-30 22:46:09 +01:00
Niko1221
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)
2026-10-30 22:44:19 +01:00
Niko1221
417765b8d8
tests: the Spark gate still holds both Unsloth sets
2026-10-30 22:36:49 +01:00
Niko1221
5c25767fcd
Offer Unsloth UD-IQ1_S on a Mac (run on an M5 Max); UD-IQ1_M stays off until it has run
2026-10-30 22:36:26 +01:00
Niko1221
ea5dbffa7f
Metal: Q6_K multi-column dense matvec (ds4_q6k_mvn)
2026-10-30 22:36:13 +01:00
Niko1221
a3a1103c13
dsu_parity: --bench and the ds4_mm pipelines
2026-10-30 22:36:13 +01:00
Niko1221
ba3d28b7e3
Q6_K matvec: packed loads via uchar4
2026-10-30 22:36:12 +01:00
Niko1221
0dba618d34
Metal Q6_K dense matvec: unaligned block loads (210-byte blocks), found by dsu_parity; test fix
2026-10-30 22:36:12 +01:00
Niko1221
5ff7da4a70
Metal tests for the Unsloth set kernels: IQ1_S matvec and tile loader, Q6_K dequant and matvec, reference checked against ggml
2026-10-30 22:36:12 +01:00
Niko1221
1b6c67122a
Metal DeepSeek engine: IQ1_S expert kernels (decode GEMV, window, prompt tile) and a Q6_K dense dequant for the Unsloth sets
2026-10-30 22:36:12 +01:00
Niko1221
7851a8dab6
Merge f-038-ds-unsloth (final 4148c5125)
2026-10-30 20:26:56 +01:00
Niko1221
4148c51257
launch: ask dsv41_fit.not_here directly (the launch tests stub the setup object)
2026-10-30 20:25:25 +01:00
Niko1221
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)
2026-10-30 20:23:59 +01:00
Niko1221
99866a47a2
Merge f-029-urgent ( bc694233b)
2026-10-30 20:06:15 +01:00
Niko1221
73e0a3b6ed
Merge f-038-ds-unsloth (final 9144da7b1)
2026-10-30 20:00:26 +01:00
Niko1221
9144da7b1a
tests: the launch list's download sizes include the two Unsloth sets
2026-10-30 19:59:31 +01:00
Niko1221
a833fb46f3
Merge f-038-ds-unsloth (macOS gate 32806f82a)
2026-10-30 19:58:24 +01:00
Niko1221
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)
2026-10-30 19:57:24 +01:00
Niko1221
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
2026-10-30 19:45:23 +01:00
Niko1221
dd3f9ec82f
tests: the catalog download sizes cover the two Unsloth DeepSeek sets (241.0 and 258.4 GB)
2026-10-30 19:35:11 +01:00
Niko1221
47cbb76690
Merge f-038-ds-unsloth (current head)
2026-10-30 19:28:13 +01:00
Niko1221
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)
2026-10-30 19:24:02 +01:00
Niko1221
ad4755c249
Say plainly that the 1-bit UD-IQ1 sets are weaker on long or careful work
2026-10-30 19:23:07 +01:00
Niko1221
b60fbb54f7
serve: the parallel stack default is Linux CUDA only (HIP and Windows get nothing: no measurement outside the replica config)
2026-10-30 19:22:21 +01:00
Niko1221
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 )
2026-10-30 19:18:45 +01:00
Niko1221
0b1808f072
Update: a failed test start of the new engine names the kept log file to report ( #4692 )
2026-10-30 19:18:01 +01:00
Niko1221
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 )
2026-10-30 19:18:01 +01:00
Niko1221
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
2026-10-30 19:17:27 +01:00
Niko1221
7f6bc33081
gfx1151 on Windows: the prompt experts run on the F16 matrix-core kernels by default (STRATA_PF_F16)
...
Owner rule: a consistent Strix Halo speed gain goes on by default. Windows gfx1151, engine 0.1.54 builds, interleaved, idle box
(ab_f16def, 2 rounds, no environment on either side): IQ3_S prompts 4K 1,180 -> 1,279, 8K 1,289 -> 1,398, 32K 1,299 -> 1,404 tok/s
(x1.08); UD-IQ4_XS 8K 1,092 -> 1,223, 32K 1,135 -> 1,248 (x1.10-1.12). Decode over 200 tokens unchanged (58.6 -> 58.2, 48.8 -> 48.5),
first greedy ids identical, 38 of 38 runs exit 0. The rounding differs from the int8 kernels; KL is information only.
STRATA_PF_F16=0 turns it off. arch_defaults_test expects six switches on Windows gfx1151.
2026-10-30 19:15:54 +01:00
Niko1221
1ead709790
Control: the time-skip check and tune engine reads also tolerate non-UTF-8 output ( #4675 )
2026-10-30 19:12:42 +01:00
Niko1221
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 )
2026-10-30 19:12:21 +01:00
Niko1221
2a9ccfe3ea
serve: the parallel stack default only where it was run (Linux CUDA nine switches, Linux HIP seven, nothing on Windows or macOS)
2026-10-30 19:08:30 +01:00
Niko1221
dd471d3140
serve: the exact parallel stack is the default with batch slots (explicit env still wins)
...
With --parallel/--batch >= 2 on a CUDA or HIP Qwen engine the server sets CKPT_POOL=1, CKPT_POOL_READY=4, SLOT_CKPT_MOVE, SLOT_COPY_DIRECT, SLOT_COPY_BATCHSYNC, BATCH_GDN_GROUPS, BATCH_ATTN_ROWS, BATCH_SELF_COMMIT and QSA_SLOTS unless the environment or the config's env already names them (=0 turns one off). DeepSeek, SYCL, Metal and Vulkan engines are untouched.
2026-10-30 19:08:30 +01:00
Niko1221
4a75fa2b62
API gateway: a client that hangs up right after data: [DONE] no longer turns a finished request into "Request cancelled." ( #4690 )
2026-10-30 19:07:42 +01:00
Niko1221
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 )
2026-10-30 19:06:50 +01:00
Niko1221
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 )
2026-10-30 19:06:11 +01:00
Niko1221
e66a445892
Server: a picture reader written into an older engine folder follows the engine in use after an update ( #4658 )
2026-10-30 19:03:21 +01:00
Niko1221
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 )
2026-10-30 19:02:16 +01:00
Niko1221