-
Notifications
You must be signed in to change notification settings - Fork 187
Pull requests: lightseekorg/tokenspeed
Author
Label
Projects
Milestones
Reviews
Assignee
Sort
Pull requests list
perf(deepseek-v4): sanitize SWA slot mapping once per step
#614
opened Jul 8, 2026 by
Xiangyi1996
Collaborator
•
Draft
[DO NOT MERGE] fix(scheduler): keep hybrid mamba match on device path
#612
opened Jul 8, 2026 by
lightseek-bot
Contributor
Loading…
feat: enable the prefill CUDA graph by default
#611
opened Jul 8, 2026 by
nperrin-fr
Collaborator
Loading…
perf(multimodal): optimize SHM input transport
#610
opened Jul 8, 2026 by
yechank-nvidia
Collaborator
•
Draft
perf(multimodal): optimize Qwen vision runtime
#609
opened Jul 8, 2026 by
yechank-nvidia
Collaborator
•
Draft
perf(multimodal): optimize Qwen3.5 M-RoPE preparation
#608
opened Jul 8, 2026 by
yechank-nvidia
Collaborator
•
Draft
refactor(glm): reuse DeepseekV3 base implementation via template methods
#604
opened Jul 7, 2026 by
wtsung
Contributor
Loading…
Add Gluon fused MoE support for dynamic MXFP4 activations
#603
opened Jul 7, 2026 by
qedawkins
Contributor
Loading…
fix(runtime): restore EnvField overrides after exceptions
#601
opened Jul 7, 2026 by
Xiangyi1996
Collaborator
•
Draft
fix(deepseek-v4): gate experimental TRT-LLM mHC backend
#600
opened Jul 7, 2026 by
Xiangyi1996
Collaborator
•
Draft
[AMD] Initial support for qwen3.5-397b-a17b-mxfp4 on AMD hardware
#592
opened Jul 4, 2026 by
raikonenfnu
Contributor
•
Draft
fix(output): prevent prompt-tail tokens leaking into streamed output_ids
#584
opened Jul 3, 2026 by
minedec
Contributor
Loading…
feat(distributed): wire TensorRT-LLM all-reduce backend
#580
opened Jul 2, 2026 by
Xiangyi1996
Collaborator
•
Draft
fix(evalscope): preserve plain completion prefixes
#579
opened Jul 2, 2026 by
dongjiyingdjy
Contributor
•
Draft
perf(deepseek-v4): accelerate indexer Q and packed FP8 quantization
#563
opened Jun 30, 2026 by
dongjiyingdjy
Contributor
Loading…
perf(multimodal): Prewarm Kimi vision encoder before readiness
#556
opened Jun 30, 2026 by
qimcis
Contributor
Loading…
perf(scheduler): Split mixed prefill/decode forwards to preserve decode graphs
#555
opened Jun 30, 2026 by
qimcis
Contributor
Loading…
feat(serve): export Prometheus metrics via /metrics endpoint
#553
opened Jun 29, 2026 by
ilyaters
Loading…
feat(platform): model the Blackwell family — datacenter / Thor / consumer
#551
opened Jun 28, 2026 by
jasl
Contributor
Loading…
ci(disaggregation): add Qwen3.5 EPD encode coverage
#549
opened Jun 28, 2026 by
chenht2022
Contributor
•
Draft
feat(disaggregation): add EPD encode pipeline
#548
opened Jun 28, 2026 by
chenht2022
Contributor
Loading…
perf(deepseek-v4): integrate TRT-LLM mHC kernels
#547
opened Jun 28, 2026 by
dongjiyingdjy
Contributor
•
Draft
feat(rl): weight-transfer control plane for RL online weight sync
#546
opened Jun 28, 2026 by
qywu
Collaborator
Loading…
Previous Next
ProTip!
Follow long discussions with comments:>50.