feat: DFlash support for tensor-parallel Qwen3.6 targets - #533

Open

jkyamog wants to merge 12 commits into

Luce-Org:mainfrom

jkyamog:feat-qwen35-dflash-tensor-parallel

Commits on Jul 30, 2026

vendor(ggml): add selected-device tensor parallel support

jkyamog
authored and
davide221
committed
feat(qwen35): add target tensor parallelism

jkyamog
authored and
davide221
committed
fix(meta): preserve repeated split layouts

jkyamog
authored and
davide221
committed
fix(qwen35): validate tensor parallel placement

jkyamog
authored and
davide221
committed
fix(cuda): validate communicator devices and preserve API

jkyamog
authored and
davide221
committed
fix(tp): preserve validation and add NCCL smoke coverage
davide221
committed
docs(vendor): record tensor parallel GGML source
davide221
committed
feat(dflash): support tensor-parallel targets

jkyamog
authored and
davide221
committed
perf(dflash): keep TP speculative state on device

jkyamog
authored and
davide221
committed
perf(dflash): batch TP feature mirror transfers

jkyamog
authored and
davide221
committed
fix(dflash): address tensor parallel review feedback

jkyamog
authored and
davide221
committed
fix(tp): bound feature sync memory and propagate failures
davide221
committed