Commit ad17016
Support dense Qwen3 generation in EMLXAxon (#119)
* feat: add dense qwen3 generation to EMLXAxon
* Update emlx_axon/lib/emlx_axon/qwen3/generate.ex
Co-authored-by: Paulo Valente <16843419+polvalente@users.noreply.github.com>
* Update emlx_axon/lib/emlx_axon/qwen3/generate.ex
Co-authored-by: Paulo Valente <16843419+polvalente@users.noreply.github.com>
* Update emlx_axon/lib/emlx_axon/qwen3/generate.ex
Co-authored-by: Paulo Valente <16843419+polvalente@users.noreply.github.com>
* Update emlx_axon/lib/emlx_axon/qwen3/generate.ex
Co-authored-by: Paulo Valente <16843419+polvalente@users.noreply.github.com>
* Address text generation style feedback
* Use microsecond timing for Qwen3 profiling
* Update emlx_axon/lib/emlx_axon/qwen3/dense_loader.ex
Co-authored-by: Paulo Valente <16843419+polvalente@users.noreply.github.com>
* Simplify Qwen3 dense parameter lookup
* Rename Qwen3 quantized attention path
* Leave dense Qwen3 state evaluation lazy
* Keep Qwen3 greedy sampler generic
* Isolate Qwen3 native accelerators
* Inline Qwen3 native error returns
* Move Qwen3 native code under emlx_fast
* Move memory include to Qwen3 source
---------
Co-authored-by: Paulo Valente <16843419+polvalente@users.noreply.github.com>1 parent 25f5340 commit ad17016
20 files changed
Lines changed: 7079 additions & 200 deletions
File tree
- emlx_axon
- lib/emlx_axon
- qwen3
- test/emlx
- emlx
- c_src
- emlx_fast
- lib
- test/emlx
| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
47 | 47 | | |
48 | 48 | | |
49 | 49 | | |
50 | | - | |
51 | | - | |
| 50 | + | |
| 51 | + | |
52 | 52 | | |
53 | 53 | | |
54 | 54 | | |
| |||
59 | 59 | | |
60 | 60 | | |
61 | 61 | | |
62 | | - | |
| 62 | + | |
63 | 63 | | |
64 | 64 | | |
65 | 65 | | |
| |||
| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
640 | 640 | | |
641 | 641 | | |
642 | 642 | | |
| 643 | + | |
0 commit comments