Repository navigation

All pull requests

Labels (11)Milestones (0)New pull request

Open

(17)

Closed

(39)

build(deps): bump the minor-update group across 1 directory with 205 updates
#129dependabot[bot] opened Oct 20, 2026Bot
qwen4_exp: DFlash/DFlash2 speculative decoding for Qwen3.8-Flash-Next
#128JC1DA opened Oct 19, 2026Contributor
build(deps): bump fsspec from 2024.12.0 to 2026.9.0
#125dependabot[bot] opened Oct 14, 2026Bot
build(deps): bump open-clip-torch from 2.32.0 to 3.3.0
#124dependabot[bot] opened Oct 14, 2026Bot
build(deps): bump psutil from 6.1.0 to 7.2.2
#123dependabot[bot] opened Oct 14, 2026Bot
build(deps): bump tokenspeed-mla from 0.1.8 to 2.0.0
#121dependabot[bot] opened Oct 14, 2026Bot
[GLM5][Spec Decode] Group sliding-window draft layers separately in the GLM-5.3 KV builder (DFlash2 on A100)
#120ogidog opened Oct 13, 2026
1
docs: add pinned_max_round_threshold_mb to the Qwen3.8-Flash-Next recipe
#115EchterAgo opened Oct 1, 2026
attention: route small-q spec-decode verify batches onto the 3D split-KV kernel (7-20x per-call at long context)
#110kaka86mm opened Sep 28, 2026
Fix MiMo-V2.6 TP>4 serving: fused-fp8-QKV sharding, MTP loader, sm86 attention path
#106avtc opened Sep 23, 2026
3
Warn when sliding-window KV capacity is pipeline-bound
#104kaka86mm opened Sep 23, 2026
Accept IEC suffixes (kiB/MiB/GiB/TiB) in human_readable_int
#103kaka86mm opened Sep 23, 2026
Fixes GLM-5.3 (glm_moe_dsa, FP8 checkpoint) startup on Ampere. Refs #38.
#100noah80 opened Sep 22, 2026
feat(chat): support OpenAI "file" content parts by inlining documents as text
#95ming616 opened Sep 20, 2026
feat(qwen4exp): support fp8_e4m3 KV cache for QSA
#90me040526 opened Sep 17, 2026
4
Speculative-decoding verify attention: split-KV kernel for sm8x (+12% decode on the FA2 path)
#82Karl0007 opened Sep 14, 2026
2
[Qwen4Exp][Bugfix] Slice the FlashInfer prefill query by the real (unpadded) token count
#81Karl0007 opened Sep 14, 2026
2

Navigation Menu

Repository navigation

All pull requests

build(deps): bump the minor-update group across 1 directory with 205 updates

qwen4_exp: DFlash/DFlash2 speculative decoding for Qwen3.8-Flash-Next

build(deps): bump fsspec from 2024.12.0 to 2026.9.0

build(deps): bump open-clip-torch from 2.32.0 to 3.3.0

build(deps): bump psutil from 6.1.0 to 7.2.2

build(deps): bump tokenspeed-mla from 0.1.8 to 2.0.0

[GLM5][Spec Decode] Group sliding-window draft layers separately in the GLM-5.3 KV builder (DFlash2 on A100)

docs: add pinned_max_round_threshold_mb to the Qwen3.8-Flash-Next recipe

attention: route small-q spec-decode verify batches onto the 3D split-KV kernel (7-20x per-call at long context)

Fix MiMo-V2.6 TP>4 serving: fused-fp8-QKV sharding, MTP loader, sm86 attention path

Warn when sliding-window KV capacity is pipeline-bound

Accept IEC suffixes (kiB/MiB/GiB/TiB) in human_readable_int

Fixes GLM-5.3 (`glm_moe_dsa`, FP8 checkpoint) startup on Ampere. Refs #38.

feat(chat): support OpenAI "file" content parts by inlining documents as text

feat(qwen4exp): support fp8_e4m3 KV cache for QSA

Speculative-decoding verify attention: split-KV kernel for sm8x (+12% decode on the FA2 path)

[Qwen4Exp][Bugfix] Slice the FlashInfer prefill query by the real (unpadded) token count

Pull requests