13 Commits
Author SHA1 Message Date
Kurt ShusterandKe Bao 6460082c05 [Mamba] Fix checkpoint depth for prefixes that end off the radix page (#39115)
Co-authored-by: Ke Bao <ispobaoke@gmail.com>
2026-09-17 14:30:57 +08:00
Kurt Shuster f81fbc749a [Fix] Merge adjacent KV-row frees so a mid-page split under DCP cannot double-free (#38941) 2026-09-14 21:44:32 +08:00
Kurt ShusterandBaizhou Zhang f2111715cd [fa] Make the FlashAttention backend extensible by subclasses (#33426)
Signed-off-by: Kurt Shuster <kurt@thinkingmachines.ai>
Co-authored-by: Baizhou Zhang <sobereddiezhang@gmail.com>
2026-09-13 21:21:25 -07:00
ed39568e79 [MoE] Add extension points for custom runner backends (#32665)
Co-authored-by: Alex Nails <alex.nails@radixark.ai>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-08-29 19:03:41 -07:00
Kurt Shuster b7686e17d6 [k3] declare packed_modules_mapping on KimiK3ForConditionalGeneration (#36211) 2026-08-28 10:22:36 +08:00
Kurt ShusterandClaude Fable 5 3fe65e0654 Deterministic gumbel sampling: clamp u=1 so masked tokens can't be sampled (#33423)
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
2026-08-09 21:43:27 +08:00
Kurt ShusterandEthan Su f34c20af86 [VLM] Fix Kimi-K2.5 CPU path: rename grid_thws -> image_grid_thw (#23501)
Co-authored-by: Ethan (Yusheng) Su <yushengsu.thu@gmail.com>
2026-04-27 20:34:28 +00:00
Kurt Shuster 32d9fe5a32 [lora] Speedup triton backend sgemm calls with better grid (#22386) 2026-04-15 13:47:07 -07:00
Kurt ShusterandYusheng Su ff13dfee45 [lora][moe] Virtual experts for LoRA MoE (#22122)
Co-authored-by: Yusheng Su <yushengsu.thu@gmail.com>
2026-04-13 21:19:30 +00:00
Kurt ShusterandYusheng Su f81b6df3a3 [lora] Fix partial MoE rank loading, VL lm_head, strict loading, deepseek on-demand (#21864)
Co-authored-by: Yusheng Su <yushengsu.thu@gmail.com>
2026-04-12 16:25:02 -07:00
Kurt Shuster 0e0091c6c8 [server] Add --quantization unquant to explicitly opt out of quantization (#21863) 2026-04-12 02:17:22 -07:00
Kurt Shuster 8da1cfb30d [lora][moe] Decoupled LoRA MoE backend with Marlin support (#21858) 2026-04-11 14:59:27 -07:00
Kurt Shuster db30a63a13 [sgl-kernel] support > 1024 experts in moe_align_block_size kernel (#21610) 2026-04-08 11:45:13 -07:00