[Startup] Parallelize torch/transformers import + weight prefetch + forkserver prewarm (#40331)

Simon Mo committed 7 months ago

[Doc] Update ViT CUDA graph doc for mixed (image+video) inputs (#40355)

Shanshan Shen committed 7 months ago

[Bugfix] Gemma4: fix multimodal embedder norm order to match HF reference (#40411)

Luciano Martins committed 7 months ago

[Misc] Reduce attention logging levels (#40086)

Chauncey committed 7 months ago

[Bugfix] Fix `_CONFIG_REGISTRY` types getting wrong config class when on-disk model_type differs (#39554)

Misa committed 7 months ago

[vLLM IR] Add IR op testing and benchmarking infrastructure (#40167)

Yanan Cao committed 7 months ago

[Bugfix] Replace code that disabled shared expert overlap (#39222)

bnellnm committed 7 months ago

[Misc][UX] Suppress confusing `num_gpu_blocks` log lines (#40402)

Matthew Bonanni committed 7 months ago

[ROCm] Hotfix: guard MLA dual RMS norm fusion against older AITer versions (#40386)

Rita Brugarolas committed 7 months ago

[EPLB] Consolidate is_unchanged/is_received_locally into TransferMetadata (#37341)

Sage Moore committed 7 months ago

[LMCache MP Connector] Add num_lmcache_extra_cached_token in KVTransferParams (#39843)

Cao Qian committed 7 months ago

Enable building MoRI with AMD AINIC stack (#38371)

Theresa Shan committed 7 months ago

[Core] Cache InductorPass.hash_source with functools.cache (#39328)

Frederik Gossen committed 7 months ago

[Bugfix][CI] Fix `tests/distributed/test_torchrun_example_moe.py` (#40349)

Nicolò Lucchesi committed 7 months ago

[WideEP] Remove naive all2all. Use allgather_reducescatter instead (#33728)

Tyler Michael Smith committed 7 months ago

[Core] Pass donate_graph_module=True to standalone_compile (#39733)

Frederik Gossen committed 7 months ago

Update flashinfer to 0.6.8 (#39959)

bai committed 7 months ago

[Core] Label torch trace logging overhead with dynamo_timed (#39329)

Frederik Gossen committed 7 months ago

[EPLB] Refactor Async EPLB synchronization logic (#37601)

Sage Moore committed 7 months ago

[MoE Refactor] Move the shared/fused expert output sum into MoERunnerBase (#35949)

bnellnm committed 7 months ago