·
101 commits
to main
since this release
What's Changed
- Deepseek V2 FP8 support by @Concurrensee in #352
- Multi-lingual P3L by @Alexei-V-Ivanov-AMD in #356
- Upstream merge 25 01 13 by @gshtras in #358
- Enable user marker for vllm profiling by @Lzy17 in #357
- Deepseek V3 support by @gshtras in #364
- Upstream merge 25 01 20 by @gshtras in #368
- Using ROCm6.3.1 base docker and building hipblas-common by @gshtras in #366
- Update pre-commit.yml by @gshtras in #374
- Skip tokenize/detokenize when it is disabled by arg --skip-tokenizer-init by @maleksan85 in #367
- FP8 FA fixes by @ilia-cher in #381
- Returning the use of the proper stream in allreduce by @gshtras in #382
- Pytorch rowwise scaled_mm by @gshtras in #384
- Applying scales rename to fp8 config by @gshtras in #387
- Dev-docker Documentation Updates by @JArnoldAMD in #378
- Support FP8 FA from Quark format by @BowenBao in #388
- Upstream merge 25 01 27 by @gshtras in #391
New Contributors
Full Changelog: v0.6.6+rocm...v0.7.0+rocm
Contributors
BowenBao, ilia-cher, and 6 other contributors