[KDA] Update flashKDA to support bf16 checkpoint state (#56485)

Wei Zhao committed about 1 month ago

[Rust Frontend][gRPC] Surface engine generation errors (#56405)

Alec committed about 1 month ago

[ROCm][CI] Extend timeout for `Basic Models (other)` (#56522)

Micah Williamson committed about 1 month ago

[ROCm][Kernel][DSV4] Remove tl.constexpr to avoid cold-compile churn in indexer gather kernel (#56153)

amd-xavierwang committed about 1 month ago

[BugFix] Fix DP token padding in dflash attention metadata (#56181)

Giancarlo Delfin committed about 1 month ago

[Bugfix][KV Connector] Only enforce disk block alignment for O_DIRECT (#55424)

mevince committed about 1 month ago

[MRV1] Scope breakable cudagraphs to the piecewise path only (#56312)

Nick Hill committed about 1 month ago

Upgrade tpu-inference to v0.29.0 (#56388)

Lyle Lai committed about 1 month ago

[Bugfix][Kimi-K3] Fix KDA projection overlap on Hopper (#55426)

chengchengpei committed about 1 month ago

[Agents] Link Triton skill to JIT kernel warmup guide (#56499)

Misha Goin committed about 1 month ago

[Kernel][MoE] Optimize batched_moe_align_block_size with cooperative writes (#53280)

Misha Goin committed about 1 month ago

[Bugfix][MLA] Read sparse model settings from text config (#56160)

Andy Lo committed about 1 month ago

[Spec][V2] Support MTP speculative decoding under pipeline parallelism (#46994)

Chris Eastwood committed about 1 month ago

[Docs] Correct API key authentication scope (/inference does NOT bypass the API key) (#56269)

Sören Dréano committed about 1 month ago

[Model][ROCm] Enable DeepSeek V4 Vision (#55107)

Andreas Karatzas committed about 1 month ago

Revert "[Rocm][Kimi-k3] Fix pipeline_parallel support for the kimik3 DCP mode (#53664)" (#56429)

Shanshan Shen committed about 1 month ago

[httpx migration] Import httpx from huggingface_hub (#56460)

Lucain committed about 1 month ago

[Deprecation] Deprecate items scheduled for 0.29 (#55353)

Wentao Ye committed about 1 month ago

[Kimi Perf] Group fp8 mla cahche insertion, 4~6x kernel level performance improvement for small batch (#55356)

Wentao Ye committed about 1 month ago

[Multimodal] Use GPU NVDEC for EPD encoder-only instance video media IO (#53675)

Isotr0py committed about 1 month ago