Default Branch

main 990ee89877 · [Fix] distributed: Prevent unavailable PyNccl enablement (#45503) · Updated 2026-10-28 15:44:24 +08:00

ZIP  TAR.GZ

Branches

xpu/qwen3-fp8-w8a16-dense 7c98507040 · Remove unnecessary comments in fp8_utils.py · Updated 2026-10-28 16:00:43 +08:00	0 9		ZIP  TAR.GZ
xpu/qwen3-moe-tuning-xpu 168154d20c · [Fix] [XPU] benchmark: Run the fused MoE Triton tuner on XPU · Updated 2026-10-28 16:00:43 +08:00	0 6		ZIP  TAR.GZ
xpu/qwen3-ccl-arc-allreduce 520b855d09 · [Perf] [XPU] distributed: Default oneCCL to its Arc all-reduce on Arc GPUs · Updated 2026-10-28 16:00:42 +08:00	0 4		ZIP  TAR.GZ
xpu/qwen3-qwen3-fused-qknorm-rope 49657e2b5a · [Perf] [XPU] qwen3: Fuse q/k RMSNorm and RoPE on XPU · Updated 2026-10-28 16:00:42 +08:00	0 5		ZIP  TAR.GZ
xpu/qwen3-eagle-draft-attn-fallback 6e61eca8ee · [Fix] [XPU] spec: Give EAGLE drafts triton attention when intel_xpu lacks their head size · Updated 2026-10-28 16:00:41 +08:00	0 2		ZIP  TAR.GZ
xpu/qwen3-eagle-target-only-sampling c536c44202 · [Fix] [XPU] spec: Sample EAGLE verification on XPU instead of forcing argmax · Updated 2026-10-28 16:00:41 +08:00	0 3		ZIP  TAR.GZ
xpu/qwen3-rope-native-fallback e7fbf03f74 · [Fix] [XPU] rope: Fall back to native RoPE for head sizes the XPU kernel lacks · Updated 2026-10-28 16:00:40 +08:00	0 1		ZIP  TAR.GZ
cheng/refactor/unified-kv-tree-components 6aeefd6986 · test: supply state tree slots in remaining MLX component fixtures · Updated 2026-10-28 15:57:53 +08:00	7 109		ZIP  TAR.GZ
cheng/refactor/unified-kv-typed-views 1086d4ad14 · fix: initialize FlashInfer test routing through its pool · Updated 2026-10-28 15:56:48 +08:00	7 96		ZIP  TAR.GZ
cheng/refactor/unified-kv-arena-and-ids c51c0b8c00 · unified-kv-id-spaces(unit-ids,non_mechanical_provable): Keep a sub-pool's unit names in its own record · Updated 2026-10-28 15:56:21 +08:00	7 61		ZIP  TAR.GZ
cheng/refactor/unified-kv-requests a5d88ad44c · unified-kv-lifecycle(lifecycle-through-storage,non_mechanical_provable): Take a request's rows and KV ids through the KV storage · Updated 2026-10-28 15:56:21 +08:00	7 63		ZIP  TAR.GZ
cheng/refactor/unified-kv-storage-consumers a558f73489 · style: format static KV storage transfer boundary · Updated 2026-10-28 15:56:21 +08:00	7 68		ZIP  TAR.GZ
cheng/refactor/unified-kv-draft-views 266e581a62 · fix: use resolved decode graph configuration for speculative capabilities · Updated 2026-10-28 15:56:02 +08:00	7 56		ZIP  TAR.GZ
cheng/refactor/unified-kv-pins 568669aebe · fix: retain one current-role transfer pin across PD switches · Updated 2026-10-28 15:51:30 +08:00	7 53		ZIP  TAR.GZ
cheng/refactor/unified-kv-drop-facades af9e5981a7 · unified-kv-drop-facades(combination-ratchet,non_mechanical_provable): Fail a new unified pool class named for a combination of KV kinds · Updated 2026-10-28 15:51:15 +08:00	7 50		ZIP  TAR.GZ
cheng/hot-fix/unified-kv-hip-fused-kv-write 13042505d3 · unified-kv-pins(hip-fused-window-write,non_mechanical_provable): Leave a unified window's KV writes to the pool on ROCm · Updated 2026-10-28 15:51:14 +08:00	7 47		ZIP  TAR.GZ
cheng/hot-fix/unified-kv-tri-host-stack c512039b81 · unified-kv-transfer(tri-host-stack,non_mechanical_provable): Build the full+SWA+state HiCache stack like the other hybrid stacks · Updated 2026-10-28 15:51:14 +08:00	7 42		ZIP  TAR.GZ
cheng/refactor/unified-kv-state-access 2faf40c84a · fix: bind EAGLE buffers through KV layer routes · Updated 2026-10-28 15:51:14 +08:00	7 41		ZIP  TAR.GZ
cheng/refactor/unified-kv-transfer e1644d157b · unified-kv-transfer(transfer-compare,non_mechanical_provable): Add a script comparing what PD peers and HiCache see of a unified pool · Updated 2026-10-28 15:51:14 +08:00	7 46		ZIP  TAR.GZ
cheng/refactor/unified-kv-addressing 9f0af30395 · unified-kv-addressing(drop-target-allowlists,non_mechanical_provable): Stop gating the unified target's attention backend by speculative algorithm · Updated 2026-10-28 15:51:13 +08:00	7 25		ZIP  TAR.GZ