Skip to content
Closed
Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
184 commits
Select commit Hold shift + click to select a range
402af22
bump llama.cpp to 5514ebea: a measured zero host figure refuses, and …
bong-water-water-bong Nov 10, 2026
cc012c4
hw-gate: the hrx_build red was a window, not a standing state - close…
bong-water-water-bong Nov 10, 2026
51b6b54
npu-pack: measure the 1.9 Token fconst rule for hidden 3072 / 24 head…
bong-water-water-bong Nov 10, 2026
a8efe87
npu-pack: evict stale canonical .fconst scratch copies; and repair th…
bong-water-water-bong Nov 10, 2026
482a7f7
docs/hw-gate: the hrx_build red was the pin, and the pins show it (co…
bong-water-water-bong Nov 10, 2026
e65f2cd
npu-pack: read a THREE-input MatMulNBits, whose implicit zero point i…
bong-water-water-bong Nov 10, 2026
4f60316
tools: place the tail grid before the padding, so a padded pack is no…
bong-water-water-bong Nov 10, 2026
d805b90
npu: read the token embedding table in place, one row per token, inst…
bong-water-water-bong Nov 10, 2026
8504001
tools+docs: the OGA-pack route's drafter arm, probed to the gate that…
bong-water-water-bong Nov 10, 2026
00de25b
tests: FLM 1.0.12's two device-free server faults — client hangup and…
bong-water-water-bong Nov 10, 2026
fa13b7c
npu-pack: measure the 1.9 Token fconst rule for hidden 2048 / 16 head…
bong-water-water-bong Nov 10, 2026
5ceed6b
hw-gate: the hrx_build window is 20:01–21:52Z and #1736 closed it, no…
bong-water-water-bong Nov 10, 2026
34ebc30
npu-pack: the external-data import writes q4nx_weight_layout, so --we…
bong-water-water-bong Nov 10, 2026
98c40eb
test(head tile): assert the blocked packer matches the pinned referen…
bong-water-water-bong Nov 10, 2026
1bcfc5d
npu-pack: the scratch sweep reaches the pre-version names and the pac…
bong-water-water-bong Nov 10, 2026
b37397b
npu-pack: locate the 1.9 head in the .fconst tail instead of requirin…
bong-water-water-bong Nov 10, 2026
6643349
npu-pack: put the 1.9 canonical .fconst copy on disk, not in the temp…
bong-water-water-bong Nov 10, 2026
ef00e33
npu: take the vocabulary from config.json, and accept a tied table sh…
bong-water-water-bong Nov 10, 2026
485425d
npu ep: the probe path never declared the `_b8` layout, so every b8 d…
bong-water-water-bong Nov 10, 2026
54a259f
npu: read the tile-packed embedding table one row per token too, so t…
bong-water-water-bong Nov 10, 2026
61ef37a
npu: the packed lm-head image as a verifiable file (engine#1716, firs…
bong-water-water-bong Nov 10, 2026
880d646
npu: serve carries the prefix cache's knobs (engine #1785, F5 flash s…
bong-water-water-bong Nov 10, 2026
90cb374
tests: register base_modes_test, and say why anchor_rerun is not a ct…
bong-water-water-bong Nov 10, 2026
0163ac8
ci: retry the apt update+install together when the archive 404s (#1815)
bong-water-water-bong Nov 10, 2026
fafabbd
release-check: the hrx-device row repeats the runtime's own init fail…
bong-water-water-bong Nov 10, 2026
d71a527
tools: expose the int8 reference contract so a family gate can be bit…
bong-water-water-bong Nov 10, 2026
3bf68c8
npu: apply the dynamic-NTK rope base rule instead of refusing it (eng…
bong-water-water-bong Nov 10, 2026
dae560b
npu-pack: take a 1.9 pack's stride and second-header offset from its …
bong-water-water-bong Nov 10, 2026
e622448
tools: whisper encoder baseline harness, with the load average in the…
bong-water-water-bong Nov 10, 2026
862b276
tools: the host reference for the per-op kernel gate, with its inputs…
bong-water-water-bong Nov 10, 2026
685b6a5
tests: pin the .fconst scratch order and that a missing cache directo…
bong-water-water-bong Nov 10, 2026
b6db79e
npu: the integer host arm a kernel gate needs, because f32 cannot be …
bong-water-water-bong Nov 10, 2026
f7f8e0a
ci_targets_check: name the one route out that needs no `workflow` sco…
bong-water-water-bong Nov 10, 2026
34d976d
tests: a sweep for tests/*.sh that nothing references (#1820), and it…
bong-water-water-bong Nov 10, 2026
31f4fd4
ci: retry apt-get update+install together, so a mirror 404 stops turn…
bong-water-water-bong Nov 10, 2026
318b6ef
amdxdna: carry the two actionable drm-misc-next-2026-10-16 fixes (001…
bong-water-water-bong Nov 10, 2026
aa673e3
npu-pack: read rope_theta from its nested rope_parameters block (#150…
bong-water-water-bong Nov 10, 2026
ef865f9
npu: the batch-check oracle compares every row, bitwise, instead of o…
bong-water-water-bong Nov 10, 2026
f1e3ac2
npu: bound tensor reads by the tensor, not by the end of the file (#1…
bong-water-water-bong Nov 10, 2026
2d1f860
npu: the integer host arm's weight cache carries the layer, and never…
bong-water-water-bong Nov 10, 2026
d65b935
npu-run: refuse a sealed kernel directory by name instead of "cannot …
bong-water-water-bong Nov 10, 2026
2ac68e4
tools: choose the tail placement when the grid is padded to whole til…
bong-water-water-bong Nov 10, 2026
f4f8218
npu: refuse --npu-opt when the build links no private route, instead …
bong-water-water-bong Nov 10, 2026
486c8cb
npu-pack: a genre-A projection the decoder cannot read is refused by …
bong-water-water-bong Nov 10, 2026
64ac77f
tests: a crashed server is no longer a refusal in serve_context_limit…
bong-water-water-bong Nov 10, 2026
6bdeeb7
hrx_build_scope: pin the reason string where it names a file (#1753 r…
bong-water-water-bong Nov 10, 2026
d962438
docs/hrx: the #856 placement table names each row's load mode, since …
bong-water-water-bong Nov 10, 2026
93d0fab
fix: two reference texts that claimed more than the code does (#1805,…
bong-water-water-bong Nov 10, 2026
1174a75
npu: config.json's vocab_size is held against the embedding table it …
bong-water-water-bong Nov 10, 2026
d31a1cc
npu: the per-row embedding path is chosen with the row reader's own r…
bong-water-water-bong Nov 10, 2026
3f2f6af
npu-pack test: pin the fconst19 rows as literals, so a wrong digit or…
bong-water-water-bong Nov 10, 2026
06fccd3
tests: register #1813's batch-check oracle, which merged with no add_…
bong-water-water-bong Nov 10, 2026
4376fad
pack_import: one text copier for both import genres, refusing a faile…
bong-water-water-bong Nov 10, 2026
b1fa5dc
q4nx_tail_head: report SKIPPED, not a pass, when the onnx arms cannot…
bong-water-water-bong Nov 10, 2026
a2ddd99
hrx-build: a cmake/tests-*.cmake family file alone triggers the build…
bong-water-water-bong Nov 10, 2026
e9a7bf1
npu-pack: the header-derived layout keeps every in-block header, and …
bong-water-water-bong Nov 10, 2026
e80767c
npu-pack: a pack that DECLARES per-head q/k gains is refused when the…
bong-water-water-bong Nov 10, 2026
fe4e488
npu-run: drive the sealed kernel set, so an id run works on the packs…
bong-water-water-bong Nov 10, 2026
ae6dfd8
npu: the int4 codes and their scale are one pair, so the activation-q…
bong-water-water-bong Nov 10, 2026
466d5a1
ci: the unreferenced-tests allow-list actually enforces both directio…
bong-water-water-bong Nov 10, 2026
5a36f66
ci_targets_check: name the add_dependencies route too, because it is …
bong-water-water-bong Nov 10, 2026
39eb84b
npu: the head image's cache key, because the SAME container packs dif…
bong-water-water-bong Nov 10, 2026
2625af6
npu-pack: the tied block-32 head synthesis and its untied refusal get…
bong-water-water-bong Nov 10, 2026
ee3fdcf
cmake: find the Python3 interpreter before the first add_test uses it…
bong-water-water-bong Nov 10, 2026
0df85cf
ci: a tools/*.py that declares a self-test and no runner names is now…
bong-water-water-bong Nov 10, 2026
8b8f86a
npu: the batch-check oracle compares from the same state it claims to…
bong-water-water-bong Nov 10, 2026
771727a
CONTRIBUTING: how a test reaches ctest, with the three gates that dec…
bong-water-water-bong Nov 10, 2026
03790a9
npu: the batch width falls back to the plan's prefill M, so a sealed …
bong-water-water-bong Nov 10, 2026
e2030cc
npu: the integer arm's operands come from the engine's packer, not a …
bong-water-water-bong Nov 10, 2026
da1a27d
npu-pack: the external-data codes are int4 two's complement -- readin…
bong-water-water-bong Nov 10, 2026
6890eec
amdxdna 0010: the README row stops claiming a clean failure it does n…
bong-water-water-bong Nov 10, 2026
88994cd
tests: the external-data writer imports a crafted pack to completion,…
bong-water-water-bong Nov 10, 2026
4d3d3e4
npu-pack: claim the per-head QK gains out of the attention gap, so Hu…
bong-water-water-bong Nov 10, 2026
ee5b436
tests: pin #1909's batch-check fix, which merged with no test (engine…
bong-water-water-bong Nov 10, 2026
f46ffcc
tests: register the two #1779 serve tests where CI can see them (skip…
bong-water-water-bong Nov 10, 2026
d9d80c2
bump llama.cpp to 2d62cc7af: the default host-weight mode is charged …
bong-water-water-bong Nov 10, 2026
20ff3e7
census-watch: a repo's own modeling_*.py is custom code, so MicroLlam…
bong-water-water-bong Nov 10, 2026
2411cee
npu: a pinned slot for the prefix single-turn requests share (engine#…
bong-water-water-bong Nov 10, 2026
683b16d
amdxdna: 0013 - a context that failed to restart fails its jobs, and …
bong-water-water-bong Nov 10, 2026
4e1fc55
npu: the batch width is the minimum loaded width, on top of #1904's p…
bong-water-water-bong Nov 10, 2026
70a90cc
F10: #1810 without the workflow edit that blocks it; a Gemma 4 call t…
bong-water-water-bong Nov 10, 2026
41ba36e
tools+docs: a real drafter on an OGA pack, measured: 2.96-3.4 tokens/…
bong-water-water-bong Nov 10, 2026
83b020b
npu-pack: stream the per-layer table, round it to nearest even, read …
bong-water-water-bong Nov 10, 2026
6f38b8f
moe_placement: Kernighan-Lin refinement, and a held-out share beside …
bong-water-water-bong Nov 10, 2026
9ffb5d2
Census 2026-11-10 (#1929)
bong-water-water-bong Nov 10, 2026
cedead9
npu: a serve loads the packed lm-head from the image cache instead of…
bong-water-water-bong Nov 10, 2026
8f10bfb
npu: a GDN layer's alpha/beta runs inside the region it already opens…
bong-water-water-bong Nov 10, 2026
5b1f0fe
npu-pack: Hunyuan's packed QK gains are q first, then k (Fixes #1934)…
bong-water-water-bong Nov 10, 2026
74ff67a
tests: pin the tail-head zero-run clamp, the case engine#1773's arms …
bong-water-water-bong Nov 10, 2026
578d2fa
npu: the Q4NX container cache evicts, so it stops filling the disk (f…
bong-water-water-bong Nov 10, 2026
4416920
census: DeepseekV4FlashSuperExpertForCausalLM is reviewed and report-…
bong-water-water-bong Nov 10, 2026
7fc3a27
npu: the integer host arm, and the bit-exact gate it unlocks (device …
bong-water-water-bong Nov 10, 2026
dce8b2b
docs/hrx: the IQ3_XXS table decode is the default from fork 69ee6f50e…
bong-water-water-bong Nov 10, 2026
040a0c7
npu-pack: the tile writer refuses a 4-bit code outside 0..15 instead …
bong-water-water-bong Nov 10, 2026
df3ff03
npu: two cold serves of one pack both finish writing the head image (…
bong-water-water-bong Nov 10, 2026
0ae65c8
bump-hrx: say when AMD rebased its line, instead of calling AMD's com…
bong-water-water-bong Nov 10, 2026
abd2cb1
bump llama.cpp to a5dc597b8: one-token decode at any head size, and a…
bong-water-water-bong Nov 10, 2026
a8e7745
npu/mtp: the distribution-following acceptance rule, as a tested head…
bong-water-water-bong Nov 10, 2026
a0f17ac
tools: dflash_draft_info, the gate before a DFlash draft loader: the …
bong-water-water-bong Nov 10, 2026
74f6db6
npu-pack: read the 1.9 block-32 head from the .fconst tail (Fixes #17…
bong-water-water-bong Nov 10, 2026
0c852e8
npu: the integer host arm packs an unaligned geometry row-major, so i…
bong-water-water-bong Nov 10, 2026
bc67179
npu-pack: refuse a one-header vote on a pack that declares per-head q…
bong-water-water-bong Nov 10, 2026
f7ecac4
serve: refuse --npu-opt on a model-generic container, where the per-o…
bong-water-water-bong Nov 10, 2026
f1ad35e
bump llama.cpp to fa533bd21: a large expert activation no longer ends…
bong-water-water-bong Nov 10, 2026
e6f93ac
build: every cmake/tests-*.cmake is included by a glob, so a new test…
bong-water-water-bong Nov 10, 2026
2f8dc3b
npu: a Q4NX cache hit counts as a use, and the cache bound cannot wra…
bong-water-water-bong Nov 10, 2026
701c6dd
npu: NPU_INFER_DUMP_OPS prints one magnitude line per projection and …
bong-water-water-bong Nov 10, 2026
c6ef5b1
npu-pack: read a FastFlowLM NPU2 model source's container, report its…
bong-water-water-bong Nov 10, 2026
2d6c039
npu-pack: sweep stale fconst scratch in every candidate directory, ne…
bong-water-water-bong Nov 10, 2026
0f305ed
bump llama.cpp to 50d28600a: Qwen3-VL's rotation runs on the device (…
bong-water-water-bong Nov 10, 2026
6925067
npu-pack: say what a rule row's `prefix` actually is (a twin-diff byt…
bong-water-water-bong Nov 10, 2026
b0f550d
npu: a launch's result is dequantised with one multiply per element, …
bong-water-water-bong Nov 10, 2026
003aad1
tools: hf_tensor.py reads one safetensors tensor with the offset rule…
bong-water-water-bong Nov 10, 2026
7c94d5a
npu-pack: import a 1.9 pack whose every MatMulNBits is block 32 (engi…
bong-water-water-bong Nov 10, 2026
b91119b
npu-pack: Qwen3-8B 1.9 pack imports end to end (hidden 4096 / 32 head…
bong-water-water-bong Nov 10, 2026
7e4f7a9
docs: a 1.9 pack's ONNX components are vendor-EP graphs, and ort_run …
bong-water-water-bong Nov 10, 2026
89a7ca2
npu: the attention rows of a prefill block share one thread team (Ref…
bong-water-water-bong Nov 10, 2026
5cb49f4
forward-serve: --prefix-pin, the pinned shared prefix on the device p…
bong-water-water-bong Nov 10, 2026
223d8e2
npu-pack: the container writer refuses a 4-bit code outside 0..15 too…
bong-water-water-bong Nov 10, 2026
8e9e871
docs(npu): the head-image cache is swept with the Q4NX cache, as one …
bong-water-water-bong Nov 10, 2026
a2ab46b
Merge main into main-audioops (engine#1931)
Nov 10, 2026
233fcad
npu-pack: a qwen35 GGUF is packed by an in-tree converter, and the GG…
bong-water-water-bong Nov 10, 2026
49b27b4
npu: the audio fork-join is tested for bit-identity and survives a th…
Nov 10, 2026
7a6d168
serve: --flash, the single-turn NPU profile; --ctx-size reaches the f…
bong-water-water-bong Nov 10, 2026
468d029
npu-pack test: the q4nx-container arm removes its fixture directory (…
bong-water-water-bong Nov 10, 2026
8f1eddd
npu: NPU_INFER_DUMP_OPS says so once at model load, so a stale binary…
bong-water-water-bong Nov 10, 2026
0a46707
npu: execute on an imported dma-buf -- the engine's arm, and the clai…
bong-water-water-bong Nov 10, 2026
5a693b2
tests: scratch goes to ~/.cache/scratch-tests, not /tmp (tmpfs, and s…
bong-water-water-bong Nov 10, 2026
153b7b6
bump llama.cpp to 2293f95ea: prefill attention keeps its rate at long…
bong-water-water-bong Nov 10, 2026
391ebd1
tools/xhci_boot_capture: the fallback outdir is the cache, never /tmp…
bong-water-water-bong Nov 10, 2026
d8a7c4c
npu: read a pack's vision component (vision_weight.q4nx): every tenso…
bong-water-water-bong Nov 10, 2026
a4fca85
tests: batch_check_state runs in CI, on the runtime stub and a fabric…
bong-water-water-bong Nov 10, 2026
e35ee7f
npu: bound every token pick by the tokenizer, not only the grammar's …
bong-water-water-bong Nov 10, 2026
fcc50c1
amdxdna: a kernel-log report of what TDR recovery did, and a test-onl…
bong-water-water-bong Nov 10, 2026
69f917e
tools: hrx_gtt_split.py -- re-measure #856's duplicate from a clean c…
bong-water-water-bong Nov 10, 2026
cfa2312
Merge main into cl-flm/1783-unblock (engine#2035 review)
Nov 10, 2026
e6d42ee
forward-serve: a streamed reply is sent as it is generated, not after…
bong-water-water-bong Nov 10, 2026
3151a82
npu: the decode-family block gets its own method, body unchanged (eng…
bong-water-water-bong Nov 10, 2026
de0bf51
gitignore: npu_pack_import_test's q4nx-container fixtures are run out…
bong-water-water-bong Nov 10, 2026
dd585f2
npu/plan: a fifth OpId cannot arrive without a registry entry (engine…
bong-water-water-bong Nov 10, 2026
a2f180d
npu-pack: the container genre's geometry, as a pure function with its…
bong-water-water-bong Nov 10, 2026
aca7437
docs/axb35: the rear-USB-A and Type-C halves of engine#823 need diffe…
bong-water-water-bong Nov 10, 2026
0073fcd
xhci capture: record WHAT is on each connected port, so a plugged dri…
bong-water-water-bong Nov 10, 2026
0db2358
npu: the MiniCPM-V 4.7 vision tower on the host, float32, matching th…
bong-water-water-bong Nov 11, 2026
1af5725
npu: the DFlash draft block's acceptance arithmetic, with the measure…
bong-water-water-bong Nov 11, 2026
72ba575
npu: load a DFlash draft package as a model source: config + containe…
bong-water-water-bong Nov 11, 2026
d4dc5e7
bump-hrx: the failure report carries the failing step and the conflic…
bong-water-water-bong Nov 11, 2026
f9dcbc4
npu-pack: the 1.9 canonical copy is written under a temporary name an…
bong-water-water-bong Nov 11, 2026
b2f17e2
npu: a sealed set can carry a wide prefill family, named by the pack …
bong-water-water-bong Nov 11, 2026
e731df9
npu-pack: the shape rule for a 1.8 external-data initializer, pinned …
bong-water-water-bong Nov 11, 2026
fa86a09
forward-serve: --prefix-pin pins the end of the system turn on the fi…
bong-water-water-bong Nov 11, 2026
8079720
ci: the runner health check reads the restart policy, not just whethe…
bong-water-water-bong Nov 11, 2026
0c64350
serve: let the operator choose the load mode, so #1668's prefill leve…
bong-water-water-bong Nov 11, 2026
0dbb593
tests/census_watch: derive "reviewed classes are documented" from sig…
bong-water-water-bong Nov 11, 2026
59d342d
ci: a comment cancels the PR's pr-agent review, because both enter on…
bong-water-water-bong Nov 11, 2026
f308fa9
npu-pack: plan() fills a container's geometry and names the source th…
bong-water-water-bong Nov 11, 2026
e27e7da
Merge main into cl-flm/1783-unblock again (engine#2035)
Nov 11, 2026
1788487
tools: hrx_dispatch_profile.py, time per HRX dispatch from the backen…
bong-water-water-bong Nov 11, 2026
540e1fd
npu: a MiniCPM-V 4.7 image from bytes to the tower's input: plan, tor…
bong-water-water-bong Nov 11, 2026
3725f12
forward-serve: the tokenizer's eos_token and the ChatML turn end are …
bong-water-water-bong Nov 11, 2026
c4826e6
tools: hrx_dispatch_profile.py names unfinished dispatches, and its s…
bong-water-water-bong Nov 11, 2026
0b88efa
weekly: the runner preflight names a restart-policy failure as one (r…
bong-water-water-bong Nov 11, 2026
75a8924
npu: the text side of a MiniCPM-V 4.7 image: its tokens and the canva…
bong-water-water-bong Nov 11, 2026
d113064
tests: the sweep covers tests/*_test.py too, as a set baseline (#2018…
bong-water-water-bong Nov 11, 2026
6e28f10
bump llama.cpp to 7a1f2b81c: the prefill kernels stay finite above th…
bong-water-water-bong Nov 11, 2026
fcd8412
npu: a sealed set can carry a narrow decode family: 3.26-3.33 -> 3.69…
bong-water-water-bong Nov 11, 2026
61ed559
tests: register moe_expert_chunks_accum_test and shrink the unreferen…
bong-water-water-bong Nov 11, 2026
24479ae
npu-pack test: the pack leg's head tile count follows the stored rows…
bong-water-water-bong Nov 11, 2026
50550b2
forward-serve: an answer that starts inside a reasoning block the pro…
bong-water-water-bong Nov 11, 2026
5acb6e7
bump-hrx: a mirror hold, so a rewrite of AMD's line that needs a hand…
bong-water-water-bong Nov 11, 2026
0f68f1e
weekly: the runner preflight warns on a short restart policy and bloc…
bong-water-water-bong Nov 11, 2026
697db7d
npu-pack: say what main does with a graph head (stored as is; an unal…
bong-water-water-bong Nov 11, 2026
aa11d51
ci: a skipped test is a test that did not run, and now something says…
bong-water-water-bong Nov 11, 2026
cad81fd
tools: fit the routing head over --dump-encoder output, so before/aft…
bong-water-water-bong Nov 11, 2026
875272b
npu-pack: a tied head must correlate AND not be mis-scaled, or it imp…
bong-water-water-bong Nov 11, 2026
afb4640
forward-serve: a text request to a decision-only pack is a 400 that n…
bong-water-water-bong Nov 11, 2026
a97e398
tools: npu_serve_smoke, the release gate's check that npu serve is on…
bong-water-water-bong Nov 11, 2026
b96a8bc
forward-serve: whitespace around a reasoning block is layout, not tex…
bong-water-water-bong Nov 11, 2026
7a62de2
tools: logits_kld, the comparator #803's KLD gate needs before anyone…
bong-water-water-bong Nov 11, 2026
30123f7
npu-plan: plan.json can name the wide and narrow sealed families (pre…
bong-water-water-bong Nov 11, 2026
c9b38d4
npu: the dense-tail cache is on by default from a 128 MiB tail, the s…
bong-water-water-bong Nov 11, 2026
077f38d
laya: tools/laya_finetune, the routing head's CPU fine-tune over --du…
bong-water-water-bong Nov 11, 2026
a5e8415
npu serve: say it at ready when no 16-row decode kernel was found (re…
bong-water-water-bong Nov 11, 2026
7e10fcc
Merge main into cl-flm/1783-unblock (fourth time)
Nov 11, 2026
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
24 changes: 20 additions & 4 deletions .github/workflows/ci.yml
Original file line number Diff line number Diff line change
Expand Up @@ -35,10 +35,26 @@ jobs:
steps:
- name: Tools for checkout and the build
run: |
apt-get update
DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends ca-certificates git \
python3 python3-venv curl jq file procps make gcc-16 g++-16 cmake ninja-build libcurl4-openssl-dev libzstd-dev \
libssl-dev libdrm-dev libpcre2-dev
# engine#1814: `apt-get update` and the install must be retried TOGETHER, with a back-off. Twice on 2026-11-10 the
# install 404'd on .debs the fresh index had just listed (perl-base/perl-modules-5.40/libperl5.40, mirror IPs
# 91.189.92.22 and 91.189.91.82) and turned main red for every PR; `gh run rerun --failed` passed with no code
# change, so it is the mirror, not the tree. Retrying only the install would reuse the same stale index, so each
# attempt re-reads the index -- which is also what lets it reach a different mirror.
ok=0
for attempt in 1 2 3 4; do
if apt-get update &&
DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends ca-certificates git \
python3 python3-venv curl jq file procps make gcc-16 g++-16 cmake ninja-build libcurl4-openssl-dev \
libzstd-dev libssl-dev libdrm-dev libpcre2-dev; then
ok=1; break
fi
echo "apt attempt $attempt failed (mirror 404 is a known transient, engine#1814); retrying in $((attempt * 15))s"
sleep $((attempt * 15))
done
if [ "$ok" != 1 ]; then
echo "::error::apt-get update+install failed 4 times -- if the 404s name .debs the index lists, it is the mirror (engine#1814), not this tree; re-run the job"
exit 1
fi
git config --global --add safe.directory "$GITHUB_WORKSPACE"
python3 -m venv /opt/ci-venv && echo /opt/ci-venv/bin >> "$GITHUB_PATH"
- uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
Expand Down
1 change: 1 addition & 0 deletions .github/workflows/hrx-build.yml
Original file line number Diff line number Diff line change
Expand Up @@ -54,6 +54,7 @@ on:
- third_party/llama.cpp
- third_party/hrx-system
- CMakeLists.txt
- cmake/tests-*.cmake
- cmake/hrx.cmake
- config/hrx-toolchain.config
- config/hrx-system-hold.config
Expand Down
7 changes: 6 additions & 1 deletion .github/workflows/pr-agent.yml
Original file line number Diff line number Diff line change
Expand Up @@ -38,7 +38,12 @@ jobs:
runs-on: [self-hosted, pr-agent]
# on the job, not the workflow: a comment that skips the job (a bot's, say) must not cancel a review
concurrency:
group: pr-agent-${{ github.event.pull_request.number || github.event.issue.number }}
# A comment and a push are different jobs on one serialised model runner: an `/ask` must not cancel the review of
# the PR it is asked about, and a `synchronize` must cancel the review of the head it replaced. Grouping by event
# name keeps push-cancels-push (right: a newer head makes the older review moot) without letting a busy thread
# starve it. Measured 2026-11-10: PR #2007's review ran 19:21:26 -> cancelled 19:30:31, and `issue_comment:
# [created]` puts every comment on the PR into this same group with `cancel-in-progress: true`.
group: pr-agent-${{ github.event.pull_request.number || github.event.issue.number }}-${{ github.event_name }}
cancel-in-progress: true
timeout-minutes: 30
# same-repo PRs, and comments by the owner, members or collaborators on a pull request
Expand Down
5 changes: 5 additions & 0 deletions .gitignore
Original file line number Diff line number Diff line change
Expand Up @@ -41,3 +41,8 @@ pin-report.json
# probe - a red that is the tree's, not the repository's (engine#1564). Ignoring them also stops run output being
# committed by accident.
.hrx123/logs/

# npu_pack_import_test's container-genre fixtures (test_q4nx_container): written to <golden dir>/q4nx-container, and to
# ./q4nx-container when the test runs with no golden dir -- so a run from the source root left them where `git add -A`
# picks them up, and #1986 committed five of them (820 KB) and reddened the copyright gate. Run output, not source.
q4nx-container/
507 changes: 452 additions & 55 deletions CMakeLists.txt

Large diffs are not rendered by default.

20 changes: 20 additions & 0 deletions CONTRIBUTING.md
Original file line number Diff line number Diff line change
Expand Up @@ -78,6 +78,26 @@ an RFC or issue.
project: one label, a request for the bug-report fields when a bug skipped the form, a registry
lookup for linked Hugging Face models, and a pointer to Q&A for questions. The model only picks
the label; every comment is a template. Maintainers can run it on any issue from the Actions tab.
- **A test runs only if a job builds it.** A ctest row is not enough. `ci.yml` builds a fixed `--target` list, so a
target that nothing builds never reaches ctest: hook it with `add_dependencies(onebit <target>)`, the way
`onebit_npu_forward_syntax` is, and `tools/ci/ci_targets_check.py --check` keeps the count from growing. A
`tests/*.sh` script has to be named by a row — or listed with a reason in `tools/ci/unreferenced_tests.txt` if it is
not a row (`anchor_rerun.sh`, an operator measurement that always exits 0) — and `tools/ci/tests_referenced_check.py
--check` fails on one that is referenced by nothing. A row that runs Python goes through `add_python_test()`, which
hands the script to the interpreter instead of relying on the file's exec bit; `tests/python_test_order_check.py
--check` enforces that it follows `find_package(Python3)`. Two rows may not share a name: CMake stops at configure with
`add_test given test NAME "…" which already exists in this directory`.
- **A new test's registration goes in its own file.** Put the `add_test(...)` (or `add_python_test(...)`) row, its
comment, and its `set_tests_properties` in a new `cmake/tests-<name>.cmake` with the Apache notice, and do not touch
`CMakeLists.txt` for it: every `cmake/tests-*.cmake` is included by a glob near the end of `CMakeLists.txt`, after all
targets exist. A new file cannot conflict with another pull request, and a row appended to `CMakeLists.txt` does —
that file changed 37 times in one day and three open pull requests had a registration hunk that no longer applied
(engine [#1979](https://github.com/1bit-MONSTER/engine/issues/1979)). A row that belongs to an existing family goes in
that family's file (`cmake/tests-gpu-hold.cmake`). An `add_executable` for a test binary still goes in
`CMakeLists.txt` next to the sources it builds; only the registration moves. Existing rows stay where they are.
A registration file that needs a fixed position may still be `include()`d by name from `CMakeLists.txt`; the glob
then leaves it alone. `tests/registration_glob_test.sh` (itself registered from `cmake/tests-registration-glob.cmake`)
runs the glob block and holds both rules.
- **Pins only move forward.** CI fails a pull request whose submodule pin is behind, or has
diverged from, the pin on `main` (`tools/check_pins.py`), because that drops commits the engine
already shipped. A rollback on purpose carries the `pin rollback` label.
Expand Down
11 changes: 10 additions & 1 deletion app/embed_serve.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -184,7 +184,11 @@ int serve_only(const std::string& model_dir, const std::string& host, int port,
};
srv.Post("/v1/embeddings", handler);
srv.Post("/embeddings", handler);
if (!srv.bind_to_port(host, port)) {
// bind_to_port() answers bool; the port comes back from bind_to_any_port(), which is the same bind_internal(host, 0)
// with the chosen port returned. Same idiom as app/serve.cpp:2023 and :2694 - a caller that asked for port 0 has no
// other way to learn which one it got.
const int bound_port = port == 0 ? srv.bind_to_any_port(host) : (srv.bind_to_port(host, port) ? port : -1);
if (bound_port < 0) {
std::fprintf(stderr, "1bit serve: cannot listen on %s:%d\n", host.c_str(), port);
return 2;
}
Expand All @@ -195,6 +199,11 @@ int serve_only(const std::string& model_dir, const std::string& host, int port,
std::thread listener([&] { srv.listen_after_bind(); });
std::fprintf(stderr, "1bit serve: %s on %s embedding (%s), in-process: /v1/embeddings, dim %d, %d-token cap\n",
id.c_str(), backend.device().c_str(), model_dir.c_str(), backend.dim(), backend.max_tokens());
// engine#1779: the readiness line every other serve route prints, and the one `serve_port_from_log()`
// (tests/serve_port.sh) reads back. The in-process embedding route returned before either of the other two
// "ready on" sites (serve.cpp:2050, :2765), so a caller that asked for --port 0 could not discover the port it was
// listening on - which is what engine #1485 fixed for the other routes and left true only for them.
std::fprintf(stderr, "1bit serve: ready on http://%s:%d\n", host.c_str(), bound_port);
listener.join();
return 0;
}
Expand Down
Loading
Loading