[NVFP4] NVFP4 MOE emulation fallback for H100/MI300/MI350, standardize `TritonExperts` usage for OCP MX emulation (#35737)

fxmarty-amd committed 7 months ago

[MoE] Convert CT W8A8 To Oracle Structure (#39187)

Robert Shaw committed 7 months ago

[MoE Refactor] Combine MoERunnerBase + DefaultMoERunner (#40560)

bnellnm committed 7 months ago

[Bugfix][CI] Fix `v1/kv_connector/unit/test_nixl_connector_hma.py::test_fewer_blocks_with_hma` (#40597)

Nicolò Lucchesi committed 7 months ago

[xpu][rocm] Update `current_platform.supports_fp8()` for TritonExperts (#40132)

Hank_ committed 7 months ago

[Doc] Clarify supported keys for --speculative-config (#40455)

xiao committed 7 months ago

[Bugfix] LoRA for DeepSeek V3.2 (#35077)

ℍ𝕠𝕝𝕝𝕠𝕨 𝕄𝕒𝕟 committed 7 months ago

[Bugfix][CPU][RISC-V] Clamp exp() input to prevent NaN (#40428)

lyd1992 committed 7 months ago

support hotwords for FunASR model (#39674)

AllenDou committed 7 months ago

upgrade tpu-inference to v0.18.0 (#40395)

Johnny Yang committed 7 months ago

[Multimodal] Support custom video metadata for pre-extracted frame sequences (#40133)

storyicon committed 7 months ago

[Bugfix] [Reasoning] Add reasoning_start_str/reasoning_end_str properties to reasoning parsers (#40566)

Chauncey committed 7 months ago

[Model] Add block-local attention and YaRN for local layers to Gemma3 (#39823)

philip-essential committed 7 months ago

[fix] flaky test_mla_attn_quant_fusion.py (#40530)

Carl Y committed 7 months ago

[kv_offload+HMA][8/N]: Support multi-group worker transfer (#38453)

Or Ozeri committed 7 months ago

[Bugfix] Pass effective chat template kwargs to reasoning parsers (#40460)

Bugen Zhao committed 7 months ago

[Audio] Bundle `get_generation_prompt()` params into `SpeechToTextParams` (#36268)

Ekagra Ranjan committed 7 months ago

[ROCm] Support non-causal attention in ROCM_ATTN (#40176)

Micah Williamson committed 7 months ago

[compile] mla + group fp8 fusion (#38877)

Carl Y committed 7 months ago

[Bugfix] Fix quantized model initialization failure with prefetch offloading (#40432)

Rishapveer Singh committed 7 months ago