Skip to content

Pull requests: gittensor-ai-lab/sparkinfer

New pull request New
15 Open 215 Closed
15 Open 215 Closed
Author
Filter by author
Loading
Label
Filter by label
Loading
Use alt + click/return to exclude labels
or ⇧ + click/return for logical OR
Projects
Filter by project
Loading
Milestones
Filter by milestone
Loading
Reviews
Filter by reviews
No reviews Review required Approved review Changes requested
Assignee
Filter by who’s assigned
Assigned to nobody Loading
Sort
Sort by
Newest Oldest Most commented Least commented Recently updated Least recently updated Best match
Most reactions
👍 👎 😄 🎉 😕 ❤️ 🚀 👀

Pull requests list

perf(moe): Q5_K fp split-K Qwen specialization + gate_up F=512 dispatch (+4.3% Qwen3.6 decode) 128-context UI-only: strongest measured context in sparkinfer eval area:kernels subsystem (emission weight 0.42) area:runtime subsystem (emission weight 0.26) eval:M sparkinfer auto-eval verdict: M needs-rebase Verified speedup but not the round winner — rebase after merge-first lands test-on-5090 Maintainer-approved to evaluate on RTX 5090 (greenlight)
#267 opened Jul 6, 2026 by inference2026 Contributor • Draft
1 task done
4
perf(qwen3.6): requantize Q8_0 attention/GDN projections to Q6_K at load (+19.6% decode) area:kernels subsystem (emission weight 0.42) area:runtime subsystem (emission weight 0.26) penalty Under a 5-day eval freeze for copycatting — not evaluated
#245 opened Jul 5, 2026 by cleanjunc Loading…
1 task done
2
perf(qwen3.6): int8 Q8_0 dp4a MMVQ for attention/GDN projections (+24% decode) area:kernels subsystem (emission weight 0.42) area:runtime subsystem (emission weight 0.26) not-tested Awaiting maintainer approval to run on RTX 5090; not evaluated
#242 opened Jul 5, 2026 by nickmopen Loading…
4
perf(qwen3.6): fast Gated-DeltaNet AR kernel + conv/L2-norm and shared-expert GEMV fusions 128-context UI-only: strongest measured context in sparkinfer eval area:kernels subsystem (emission weight 0.42) area:runtime subsystem (emission weight 0.26) eval:none sparkinfer auto-eval verdict: none re-evaluate Winner merged — rebase onto main; bot re-evaluates on push test-on-5090 Maintainer-approved to evaluate on RTX 5090 (greenlight)
#240 opened Jul 5, 2026 by fansilas Contributor • Draft
1 task done
9
perf(qwen3.6): fuse decode kernel launches — shared-expert + GDN conv/L2-norm 128-context UI-only: strongest measured context in sparkinfer eval area:kernels subsystem (emission weight 0.42) area:runtime subsystem (emission weight 0.26) eval:M sparkinfer auto-eval verdict: M needs-rebase Verified speedup but not the round winner — rebase after merge-first lands test-on-5090 Maintainer-approved to evaluate on RTX 5090 (greenlight)
#239 opened Jul 5, 2026 by minion1227 Loading…
1 task done
10
perf(qwen36): retune GEMV/router/MoE decode scheduling 128-context UI-only: strongest measured context in sparkinfer eval area:kernels subsystem (emission weight 0.42) eval:none sparkinfer auto-eval verdict: none needs-benchmark Box ticked but decode before/after not filled with a real improvement — not evaluated
#237 opened Jul 5, 2026 by DragunovX16 Contributor Loading…
3 tasks done
3
perf(qwen3.6): default the merged gdn_ar_fast_kernel ON area:kernels subsystem (emission weight 0.42) needs-benchmark Box ticked but decode before/after not filled with a real improvement — not evaluated
#236 opened Jul 5, 2026 by jaso0n0818 Contributor Loading…
1 task done
1
fix(runtime): handle tied embeddings in convert_qwen35.py (fixes #219) area:runtime subsystem (emission weight 0.26) not-tested Awaiting maintainer approval to run on RTX 5090; not evaluated
#220 opened Jul 4, 2026 by minion1227 Loading…
1 task
1
fix(runtime): guarantee forward progress in Scheduler::schedule() (fixes #217) area:runtime subsystem (emission weight 0.26) not-tested Awaiting maintainer approval to run on RTX 5090; not evaluated
#218 opened Jul 4, 2026 by minion1227 Loading…
1 task
1
perf(decode): lower adaptive 256-split threshold for long-context flash decode area:kernels subsystem (emission weight 0.42) area:runtime subsystem (emission weight 0.26) needs-benchmark Box ticked but decode before/after not filled with a real improvement — not evaluated
#194 opened Jul 3, 2026 by thomasalvaedison7777-lgtm Loading…
7 of 8 tasks
1
perf(decode): default gate/up to 1-row mmvq2_qwen + rmsnorm load elimination (sm_120) area:kernels subsystem (emission weight 0.42) not-tested Awaiting maintainer approval to run on RTX 5090; not evaluated
#188 opened Jul 3, 2026 by thomasalvaedison7777-lgtm Loading…
5 tasks done
perf(attn): cp.async double-buffered KV staging in GQA flash-decode split area:kernels subsystem (emission weight 0.42) not-tested Awaiting maintainer approval to run on RTX 5090; not evaluated
#187 opened Jul 3, 2026 by real-venus Loading…
1 task
fix(kv-cache): grow sequences by block delta, not append — prevents block leak + device block-table overrun (RTX 5090 verified) area:runtime subsystem (emission weight 0.26) needs-benchmark Box ticked but decode before/after not filled with a real improvement — not evaluated
#135 opened Jul 2, 2026 by milosde111 Loading…
1 task done
2
perf(decode): occupancy-first adaptive n_splits — 64-split mid-context plateau (skip 128) area:runtime subsystem (emission weight 0.26) needs-benchmark Box ticked but decode before/after not filled with a real improvement — not evaluated
#134 opened Jul 2, 2026 by milosde111 Loading…
1 task done
1
runtime: add Gemma 4 26B-A4B end-to-end decode path area:runtime subsystem (emission weight 0.26) not-tested Awaiting maintainer approval to run on RTX 5090; not evaluated
#85 opened Jun 28, 2026 by andriypolanski Loading…
5 tasks done
1
4
ProTip! Updated in the last three days: updated:>2026-07-03.