[Model] Pass intermediate_tensors to the model when capturing CUDA gr… (#57745)

Mi Jiazhi committed about 1 month ago

[XPU] fix incorrect gdn kernel log for XPU path (#54871)

Yan Ma committed about 1 month ago

[Perf] Give a prefill-only batch the model state's number of logit rows (#57416)

Matt Mastracci committed about 1 month ago

[Bugfix] DiffusionGemma: cast the self-conditioning soft embed to the buffer dtype (#57462)

Matt Mastracci committed about 1 month ago

[CPU][Perf] refactor paged attention for Arm CPUs (#56045)

Fadi Arafeh committed about 1 month ago

[Feature][EPD] Support dynamic EPD (EC-connector) proxy (#54176)

Tianyu Guo committed about 1 month ago

[ROCm][DSv4.1][Perf] Reuse the decode topk ragged metadata across layers (#57434)

Fangzhou Ai committed about 1 month ago

Add one-step recipe serving (#57306)

Louie Tsai committed about 1 month ago

[XPU][UT]Fix test_mnnvl_alltoall device and distributed backend (#57591)

YiSheng5 committed about 1 month ago

[XPU][UT]Using the device_control_env_var to restrict visible device on different platform (#57581)

YiSheng5 committed about 1 month ago

[Test][Core] Compute the expected hybrid prefix-cache hit instead of hard-coding it (#57467)

Fanli Lin committed about 1 month ago

[Bugfix] Fix SM100 fp8_ds_mla cache scales (#49435)

Ariel Bereslavsky committed about 1 month ago

[CI][Bugfix] Fix MoE reprocess test mock after #57405's kernel refactor (#57641)

vllm-agent committed about 1 month ago

[GLM5.3 Perf] Size the GLM-5 sparse indexer decode workspace, 3072 MiB GPU memory saved (#57701)

Wentao Ye committed about 1 month ago

[Bugfix] Take padded path for ragged decode batches in sparse_attn_indexer (#52500)

Pavel Zakharov committed about 1 month ago

[Model Runner V2] Support custom logits processors (#56497)

David Cheung committed about 1 month ago

[Refactor] Use dynamic MM cache in processor (#57674)

Cyrus Leung committed about 1 month ago

[Kernel][Perf] Add sm_120 (RTX PRO 6000 / RTX 50) tuned configs for batch-invariant persistent matmul (#57456)

Lio Einaudi committed about 1 month ago

[ROCm][Perf] Enable the AITER GDN decode fast path for flat qkvz layouts (#53623)

Mikko Tukiainen committed about 1 month ago

[Bugfix][Model] Fix MiMo-V2.5 fused fp8 qkv_proj sharding (pre-shard count is num_key_value_heads; MTP path too) (#57508)

vllmellm committed about 1 month ago