Skip to content

Pull requests: gittensor-ai-lab/sparkinfer

New pull request New
12 Open 358 Closed
12 Open 358 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(qwen35): int8 tensor-core prefill GEMM for Qwythos [DRAFT — awaits #398]
#422 opened Jul 15, 2026 by inference2026 Contributor • Draft
1 task
perf(qwen35): weight-amortized Q4_K tile-GEMM prefill for Qwythos above 64k (5.5x pp @128k) area:kernels subsystem (emission weight 0.42) area:runtime subsystem (emission weight 0.26) test-on-5090 Maintainer-approved to evaluate on RTX 5090 (greenlight)
#409 opened Jul 14, 2026 by blinkeye-lcm Contributor Loading…
1 task done
perf(qwen36): specialize fixed-shape decode kernels area:kernels subsystem (emission weight 0.42) area:runtime subsystem (emission weight 0.26) test-on-5090 Maintainer-approved to evaluate on RTX 5090 (greenlight)
#403 opened Jul 14, 2026 by 9876543210-tc-0123456789 Contributor Loading…
1 task done
perf(qwen35): batched prompt prefill for Qwythos (one weight-amortized GEMM pass) area:kernels subsystem (emission weight 0.42) area:runtime subsystem (emission weight 0.26) test-on-5090 Maintainer-approved to evaluate on RTX 5090 (greenlight)
#398 opened Jul 14, 2026 by fansilas Contributor Loading…
1 task done
2
fix(attn): sparse-KV combine drops the attention output gate (Qwythos >= 8k) 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
#393 opened Jul 14, 2026 by reyanthony062001-ops Contributor Loading…
1 task done
2
fix(kernels): re-land the hd256 int8-MMA correctness fixes reverted by #294 area:kernels subsystem (emission weight 0.42)
#389 opened Jul 13, 2026 by reyanthony062001-ops Contributor • Draft
1 task done
1
2
perf(qwen35): skip LM head on prefill + dedicated prefill CUDA graph area:runtime subsystem (emission weight 0.26) needs-rebase Verified speedup but not the round winner — rebase after merge-first lands
#387 opened Jul 13, 2026 by James-CUDA Contributor Loading…
1 task done
1
14
perf(qwen36): tune hd256 combine for 160-split long-context decode 128k-context area:kernels subsystem (emission weight 0.42) copycat-warn eval:S sparkinfer auto-eval verdict: S eval-qwen35:S eval-qwen36:none 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)
#363 opened Jul 13, 2026 by claytonlin1110 Contributor Loading…
4 of 5 tasks
1
22
perf(qwen36): Q4_K shared-expert FFN kernels + Q8_0/Q5_K→Q4_K requant of the MoE FFN stage (+3.1–3.8%) 32k-context UI-only: strongest measured context in sparkinfer eval area:kernels subsystem (emission weight 0.42) area:runtime subsystem (emission weight 0.26) copycat-warn eval:REJECT sparkinfer auto-eval verdict: REJECT eval-qwen35:REJECT eval-qwen36:REJECT test-on-5090 Maintainer-approved to evaluate on RTX 5090 (greenlight)
#360 opened Jul 13, 2026 by petermilord Loading…
5 tasks done
3
perf(qwen36): Q4_K shared-expert FFN kernels + Q8_0/Q5_K→Q4_K requant of the MoE FFN stage (+2.8–3.2%) 32k-context UI-only: strongest measured context in sparkinfer eval area:kernels subsystem (emission weight 0.42) area:runtime subsystem (emission weight 0.26) eval:REJECT sparkinfer auto-eval verdict: REJECT eval-qwen35:REJECT eval-qwen36:REJECT test-on-5090 Maintainer-approved to evaluate on RTX 5090 (greenlight)
#359 opened Jul 13, 2026 by milosde111 Loading…
6 tasks done
2
perf(qwen36): Q8_0→Q4_K requant of GDN ssm_out projections (+2.8% @128) 128-context UI-only: strongest measured context in sparkinfer eval area:runtime subsystem (emission weight 0.26) eval:REJECT sparkinfer auto-eval verdict: REJECT eval-qwen35:REJECT eval-qwen36:REJECT merge-first Round's biggest verified speedup — merge this first test-on-5090 Maintainer-approved to evaluate on RTX 5090 (greenlight)
#355 opened Jul 12, 2026 by blinkeye-lcm Contributor Loading…
1 task done
22
perf(attn): grid-target KV-split cap for int8-MMA long-context decode 128-context UI-only: strongest measured context in sparkinfer eval area:runtime subsystem (emission weight 0.26) eval:none sparkinfer auto-eval verdict: none eval-qwen35:none eval-qwen36:none re-evaluate Winner merged — rebase onto main; bot re-evaluates on push test-on-5090 Maintainer-approved to evaluate on RTX 5090 (greenlight)
#345 opened Jul 11, 2026 by fansilas Contributor • Draft
1 task done
8
ProTip! no:milestone will show everything without a milestone.