Commit graph

809 commits

Author SHA1 Message Date
Admin
08d631a77c ai chat working 2026-05-04 08:48:38 +02:00
Admin
4a106111b8 ai manager otw 2026-04-28 12:24:40 +02:00
Admin
432ae0f8d7 aimgr 2026-04-27 14:51:24 +02:00
Admin
b04d47592b Batch Qwen CUDA prefill MoE 2026-04-20 13:48:29 +02:00
Admin
bbcfe50745 Optimize Qwen CUDA routing and exact decode 2026-04-20 13:48:29 +02:00
Admin
ca56cf3d23 Optimize Qwen CUDA MoE decode path 2026-04-20 13:48:29 +02:00
Admin
d750a764c7 Checkpoint Qwen CUDA exact progress 2026-04-20 13:48:29 +02:00
Kevin Boos
d4b278f72f Html/Markdown fixes: sub/superscript, table outlines/alignment, etc (#1051)
* draw_text: adopt outer many_instances batch on linux/windows

CodeEditor opens an outer raster batch via DrawText::begin_many_instances
before its glyph loop. The linux/windows branch of DrawText::draw_text
ignored self.many_instances and opened its own nested batch, which
resolved (via find_appendable_drawcall) to the same draw_item whose
`instances` Vec was already swapped out by the outer open, panicking on
unwrap in Cx2d::begin_many_instances.

Mirror what the other platform branch (and draw_rasterized_glyphs_abs)
already do: take self.many_instances on entry, track whether the active
raster batch is the outer one, and hand it back on exit so the caller's
end_many_instances finalizes it. Skip the !drew_raster_this_frame area
clear when the outer batch is still live.

* Html/Markdown fixes: sub/superscript, table outlines/alignment, etc

- **Sub/sup in HTML**: added a `y_shift_scales` stack on `TextFlow`, composed onto `temp_y_shift` in `draw_text`. `<sub>` pushes `+0.55`, `<sup>` pushes `-0.2` in html.rs
- **Sub/sup in Markdown**: now handles `MdEvent::InlineHtml` for `<sub>`/`<sup>` (case-insensitive), using the same stacks.
- **Space after `&amp;` (and other entities)**: the HTML lib's entity decoder now resets `last_non_whitespace` after truncate+push, so the whitespace-collapse check no longer drops the next real space.
- **Table column alignment (Markdown)**: `begin_table_cell` takes `align_x: f64`; tracks `Tag::Table` alignments and a per-row column index, passing each cell's `Alignment` through.
- **Table column alignment (HTML)**: `<td>`/`<th>` now honor `align="…"` and inline `style="text-align: …"` via new `cell_align_x` / `align_keyword_to_x` helpers in html.rs.
- **Per-row text alignment plumbing**: new `layout_align` field on `DrawText` is passed to the layouter, which already supports per-row alignment. `TextFlow` propagates a `cell_text_align_x` into it. This is the actual fix that makes cell alignment visible.
- **Wrap-flow alignment scaffolding**: implemented the previously-stubbed `Flow::Right { wrap: true }` branch in the turtle logic. Useful for non-text wrapping walks; text goes through the layouter path above.
- Add examples to uizoo: three new tables in both markdown and html tab -- a plain one, a left/center/right aligned one, and a numeric all-right-aligned one. They cover bold/italic/code/links/sub-sup/emoji/entities/strikethrough inside cells.

* minor cleanup; prefer `style` over `align` HTML tag
2026-04-18 10:29:50 +02:00
Admin
8f611181bd Add Qwen runtime and Windows CUDA build support 2026-04-17 11:44:43 +02:00
Admin
1ba10212dc Clean up MLX backend boundary config 2026-04-16 19:07:04 +02:00
Admin
4b27bbaac4 Add CUDA multimodal exact path 2026-04-15 21:33:39 +02:00
Admin
cf711852ca Fix CUDA chat repeated prompt loop 2026-04-15 21:33:39 +02:00
Admin
0c1aa42156 Add CUDA exact chat windowing and kernels 2026-04-15 21:33:39 +02:00
Admin
6cffd711c7 cuda otw 2026-04-13 11:09:08 +02:00
Admin
9809128567 cuda exact chat reuse and decode tuning 2026-04-13 11:09:08 +02:00
Admin
b5e31a352a wip cuda long-context profiling 2026-04-13 11:05:49 +02:00
Admin
77904dec0d cuda prefill metrics and chunked kernels 2026-04-13 11:05:49 +02:00
Admin
90ae117d84 cuda exact prefill to ~87 tok/s 2026-04-13 11:05:49 +02:00
Admin
d2d8e0459b Add FLUX warm pipeline and reference benchmarks 2026-04-13 11:01:13 +02:00
Admin
33261dcfc7 flux 1 works 2026-04-13 11:01:13 +02:00
Admin
43ac724e91 rotor quant for metal 2026-04-13 11:01:13 +02:00
Admin
8026215d2f Document Rotor divergence against BF16 baseline 2026-04-13 11:01:13 +02:00
Admin
f79724af4f Add optional Rotor-style K-cache compression for Gemma 4 2026-04-13 11:01:13 +02:00
alanpoon
5af1c812fc audio_input_panic_fix for macos (#1038) 2026-04-11 13:52:25 +02:00
Admin
b50c05dfe7 Add CUDA NVFP4 backend and GPU decode path 2026-04-11 11:44:04 +02:00
Admin
4d0a4c44f0 mlx image opt 2026-04-11 09:51:37 +02:00
Admin
bff75eb694 fix 2026-04-10 21:14:32 +02:00
Admin
0a8376e39f mlx multimodal 2026-04-10 15:10:25 +02:00
Admin
9ca672088f mlx optimisations 2026-04-10 14:48:17 +02:00
Admin
8f542c0d8a mlx working 2026-04-10 14:48:17 +02:00
Admin
1d1e91dae7 cleanup 2026-04-10 14:48:17 +02:00
Admin
e19a79d5a4 mlx otw 2026-04-10 14:48:17 +02:00
Admin
69aa615947 mlx 5x->2x 2026-04-10 14:46:22 +02:00
Admin
3d0171c8d6 mlx 2026-04-10 14:46:22 +02:00
Admin
f30a62422f mlx otw 2026-04-10 14:46:22 +02:00
Admin
00eed2e268 mlx otw 2026-04-10 14:46:22 +02:00
Ruben Daniels
c1056ea4f8 fix(base64): handle single '=' padding in base64_decode (#1027)
The decoder only checked for double padding ('==') at input[len-2],
subtracting 1 output byte. Single padding ('=') at input[len-1] was
not handled, leaving 1 extra garbage byte in the decoded output.

This affected 2 out of 3 input lengths (any input where len % 3 == 2),
producing decoded output 1 byte longer than expected.

Fix: check input[len-1] for '=' first (subtract 1 byte), then check
input[len-2] for '=' (subtract another byte for double padding).

Added 7 roundtrip tests covering: no padding (3n bytes), single
padding (3n+2 bytes), double padding (3n+1 bytes), empty input,
lengths 1-20, all 256 byte values, and URL-safe alphabet.

Co-authored-by: prime intellect <prime@prime-intellects-Mac-Studio.local>
2026-04-09 20:05:45 +02:00
Admin
f6c32eae92 stdin 2026-04-05 11:30:53 +02:00
Admin
d9d92a557e llama works 2026-04-05 11:30:53 +02:00
Admin
d4cbbaea87 llama works atleast 2026-04-05 11:30:53 +02:00
Admin
1e8f8cebbb churn 2026-04-05 11:28:47 +02:00
Admin
ddec6fc602 churn 2026-04-05 11:28:47 +02:00
Admin
080794781f driveable 2026-04-05 11:28:47 +02:00
Admin
a8ee2b194b drivable 2026-04-05 11:28:47 +02:00
Admin
357a0bd51b drivable 2026-04-05 11:28:47 +02:00
Admin
6cf21b67bd iterating 2026-04-05 11:28:47 +02:00
Admin
0a0652cfb0 otw 2026-04-05 11:28:47 +02:00
Admin
d134df9f06 xr llama and slug 2026-04-01 12:19:03 +02:00
Admin
1b9906f144 llama + xr 2026-04-01 12:19:03 +02:00
Admin
6836d3ac07 fixup 2026-04-01 12:19:03 +02:00