Skip to content

PXA v4.0.3

Choose a tag to compare

View all tags
@poisonxa16 poisonxa16 released this 04 Nov 06:36
· 25 commits to main since this release
v4.0.3
1808b9d

PXA 4.0.3

4.0.3 in one line: faster long prompts on Tesla P100, LFM support, and misc speed improvements and fixes.

4.0.3 includes everything from 4.0.2, which went to Valued Supporters as a beta and was not released on its own. Both sets of changes are below.

Install on Linux: download the Linux tarball below, then install it with the pxa-update in your current PXA folder's bin/ (it checks the file, installs 4.0.3 next to your current version and keeps the old one, so you can go back):

bin/pxa-update apply-file pxa-v4.0.3-linux-x86_64-cuda12.8-sm60_61_70_75_80_86_89.tar.gz --sha256 c49b6897138e9f9c09d78f019689973dc9405b7baa4a694e3f4973f6668610c0

New to PXA? Unpack the tarball into a new folder and follow START-HERE.md inside it.

Windows: download PXA Control for Windows, the engine zip and the PXQN library (ggml-pxqn.dll, published beside the zip, not inside it) below. RELEASE-NOTES-windows.md says where each file goes.

Docker: docker pull ghcr.io/poisonxa16/pxa:v4.0.3, or download pxa-v4.0.3-docker.tar.gz and docker load -i it; both give the same image. Start it the same way as the 4.0.1 image.

Check a downloaded file with sha256sum -c SHA256SUMS (or against the table below) before you unpack or load it.

Downloads

Platform File sha256
Linux x86_64 (NVIDIA) pxa-v4.0.3-linux-x86_64-cuda12.8-sm60_61_70_75_80_86_89.tar.gz c49b6897138e9f9c09d78f019689973dc9405b7baa4a694e3f4973f6668610c0
Windows x64 (NVIDIA) pxa-engine-win64-cuda12.zip 210a472884cc63429bf446bb4e7093bf6ca746dc5e5e845349972bd87950b13b
PXA Control for Windows pxa-control-win64.exe 02a2f2db4898db668fb931304518200ca44a0226c7e71a2c381a1a577318414c
PXQN library for Windows (copy into engine\bin) ggml-pxqn.dll f4d66603f0c05b17abe766a02eb5b512fd42978c66e430dbf8b92d0bf4c90caa
Terms for the PXQN library LICENSE-ggml-pxqn.txt 6192118880ed8098a8dcb9a61e90afd269b1250ea3ba86a53faa35bba6b73a82
Windows install notes RELEASE-NOTES-windows.md 186f2264d33680a91228d10baaae883598ce846f85771510b23d1185d041a903
Docker (Linux, NVIDIA) pxa-v4.0.3-docker.tar.gz f4dd14a326317a76c6e3a8a09012311714fff22a14cba89ce745b33948cbeed3
Docker image (same image) ghcr.io/poisonxa16/pxa:v4.0.3 image id ce2e9d3f219b9f8f45c67f1e8cecbfc49a4fc580ec3c4d29ca585bc43f034fba

What is new in 4.0.3

Faster long prompts on Tesla P100

  • Two Tesla P100 cards with Qwen 27B (our 6-bit supporter file) and MTP on: the first 128k tokens of a long prompt fill about 36% faster than in 4.0.2, about 309 tokens per second, up from about 227.
  • Decoding speed is the same as in 4.0.2.

What is new in 4.0.2 (first public release of these changes)

LFM support

  • LFM2 models load and run.
  • LFM2-MoE models (8B and 24B) load and run.
  • LFM3 models load and run.

Misc speed improvements

  • GTX 10xx (1080 Ti class): prompt processing is faster, about 40% on long prompts measured on a GTX 1080 Ti.
  • Two Tesla P100 cards with a 27B model and MTP: decoding is faster than in 4.0.1.
  • Tesla P100: k-quant layers decode a little faster, now on by default, with the same output.

Fixes

  • Speculative decoding on Tesla P100 gives the same text as plain decoding on code, prose and chat prompts.
  • PXA Control no longer keeps its sign-in token in a browser cookie.
  • Release builds check every file they ship more strictly before they go out.
  • Misc fixes.

Good to know

  • Every build carries native code for GTX 10xx, Tesla P100/P40, V100, Turing (RTX 20xx, GTX 16xx, T4), RTX 30xx and RTX 40xx. Turing cards are not run-tested by us yet; tell us how it goes.
  • Multi-GPU GTX 10xx and P40 setups are untested by us: please report how they run.
  • The Docker tag latest moves with the public release; v4.0.3 is fixed.
  • A PXA Control for macOS build of this release is planned.

Known issues

Everything under "Known issues" in the 4.0 and 4.0.1 notes still applies.

  • Linux in-app update: PXA Control's "Install update" and pxa-update apply stop with "download failed" in 4.0.3 and earlier. Use the steps under "Install on Linux" above; this is fixed in the next update.
  • GLM-5.3-Flash support is still experimental and was not re-tested on this release.
  • Tesla P100: with Qwen 27B on one card, some replies can differ slightly from 4.0.1 after a few hundred tokens, where two next words were almost equally likely. We measured the difference: it is smaller than the engine's own run-to-run variation from batch size, and quality (perplexity) is unchanged. To get 4.0.1's exact output back, start the engine with PXA_KQMMV_VERIFY_ID=4.
  • On one V100 (sm_70), dense Qwen3.8-27B decode without speculation is 6 to 9% slower than in 4.0.1 and earlier, in our tests (4.0.2 and 4.0.4 have the same slowdown). Output is identical. A fix is in testing for 4.0.5.

How to report problems

Use Control's "Report a problem", or write in the PXA Discord #support channel (link below). Say the version, 4.0.3, plus the card, the model file and the command.

Community

Assets 19
Loading