Repository navigation

All pull requests

[draft] feat(rdna2): LeapQuant config flag, validator, numerics tests and opt-in page packing (follow-up to #168)
#170BlivionIaG opened Nov 14, 20269 tasks
Add native GGUF load path for Qwen4Exp (Flash-Next) mixed-quant shard sets
#169crispyging opened Nov 14, 2026
docs+tools(rdna): contribution contract, one-command PR receipt, archdiff, async/stall checks [skip ci]
#166BlivionIaG opened Nov 11, 2026
feat(qwen4-prep): Qwen4 config probe, monokernel stage planner, opt-in dense MLP + 4 upstream Qwen4Exp backports
#164BlivionIaG opened Nov 11, 2026
feat(rdna): HeteroMoEMethod, H1 recipes, decode dispatch knob (stacked on hetero-moe/hx/expert-stats)
#163BlivionIaG opened Nov 11, 2026
feat(rdna): hx hub/expert-server exchange kernels with digest guard
#162BlivionIaG opened Nov 11, 2026
docs(rdna2): hetero MoE spec (W7800 hub + V620 expert servers) and perf model
#161BlivionIaG opened Nov 11, 2026
feat(rdna3): gfx1100 Flash-Next W4A16 port (fat gfx1030;gfx1100 _rocm_C)
#160BlivionIaG opened Nov 11, 2026
spike(disagg): P/D resume exact vs the cold run (opt-in prefill_monolithic_split + unpadded_resume_step)
#159BlivionIaG opened Nov 11, 2026
feat(rdna2): tiered expert budgets and native hybrid KV paging
#148BlivionIaG opened Oct 29, 2026
Glm53 flash rdna correctness v2
#133JoseMiralles opened Oct 20, 20263 of 4 tasks
feat(rdna): dynamic speculative drafting for Flash-Next (builds on #119)
#125BlivionIaG opened Oct 18, 20262 of 8 tasks
[backlog] perf(rdna2): cut the EXL3 fused-MoE overhead (Flash-Next EXL3 decode/prefill)
#124BlivionIaG opened Oct 17, 2026
[backlog] feat(qwen4_exp): vision tower for EXL3 Qwen3.8-Flash-Next
#123BlivionIaG opened Oct 17, 2026
[backlog] feat(qwen4_exp): MTP2 for EXL3 Qwen3.8-Flash-Next
#122BlivionIaG opened Oct 17, 2026
[backlog] perf(rdna2): serve the v1 worker/sampler/spec-decode Triton kernels from HIP (VLLM_RDNA_ALL_HIP)
#121BlivionIaG opened Oct 16, 2026
docs(rdna2): how to reproduce the MonoKernel (#94) numbers
#120BlivionIaG opened Oct 16, 20263 of 4 tasks
[backlog] feat(monokernel): pipeline-parallel (PP3) support
#116BlivionIaG opened Oct 16, 2026
[backlog] feat(monokernel): asymmetric W4A16 experts in the k2 routed phases
#115BlivionIaG opened Oct 16, 2026
[backlog] perf(monokernel): GDN projections at widths 12-24 (skinny GEMM vs Tensile)
#113BlivionIaG opened Oct 16, 2026
[backlog] perf(monokernel): fold the QSA glue kernels and fix the HIP sparse-decode floor
#112BlivionIaG opened Oct 16, 2026
[backlog] perf(monokernel): k2 FFN block above 8 rows (MTP2 c=8, width 24)
#110BlivionIaG opened Oct 16, 2026
[backlog] perf(monokernel): run the MTP drafter layer on the MonoKernel
#109BlivionIaG opened Oct 16, 2026
feat(rdna2): opt-in W8A8 int8 sdot4 dense and MoE kernels for gfx1030
#93BlivionIaG opened Oct 13, 20263 of 4 tasks
fix(rdna2): gate W4A8 MoE strictly above live cudagraph capture size
#90BlivionIaG opened Oct 12, 2026

Pull requests