All issues

Search results

n > 1 in a single completions request returns identical choices (sampling-backend triton, ROCm/gfx950)

Status: Open.

#1613 In lightseekorg/tokenspeed;

· vivekkhandelwal1 opened on Sep 17, 2026

/update_weights_from_tensor and /update_weights_from_disk have no scheduler dispatch branch: the request hits the NotImplementedError fallthrough and takes the scheduler down (disk path measured: 500, then 503 on every later request)

Status: Open.

#1479 In lightseekorg/tokenspeed;

· vivekkhandelwal1 opened on Sep 10, 2026

nccl weight transfer reports success without transferring: /update_weights returns 200 and loads uninitialised device memory

Status: Open.

#1373 In lightseekorg/tokenspeed;

· vivekkhandelwal1 opened on Sep 2, 2026

[Bug] Qwen 3.8 27B FP8 inference bug

Status: Open.

#1368 In lightseekorg/tokenspeed;

· Firdavs222 opened on Sep 2, 2026

Serving throughput at concurrency 64 is bimodal on gfx950: ~8% of bench steps stall a fixed ~0.92 s before admitting the tail of the batch

Status: Open.

#1355 In lightseekorg/tokenspeed;

· vivekkhandelwal1 opened on Sep 1, 2026

add k3 gb300 tp 8 performance ci

high priority

inactive

Status: Open.

#1335 In lightseekorg/tokenspeed;

· lightseek-bot opened on Aug 31, 2026

BF16 Gluon MoE kernel syncs to host inside CUDA graph capture, blocking startup for BF16 MoE models on gfx950

Status: Open.

#1329 In lightseekorg/tokenspeed;

· vivekkhandelwal1 opened on Aug 31, 2026

Capturable grammar hangs decode CUDA-graph capture when capture batch x draft tokens exceeds 128 logits rows (xgrammar, TP=8, B300)

bugSomething isn't workingSomething isn't working

high priority

inactive

Status: Open.

#1314 In lightseekorg/tokenspeed;

· abatilo opened on Aug 30, 2026

get_available_device() masks the real error behind "Triton is not supported on the current platform"

Status: Open.

#1243 In lightseekorg/tokenspeed;

· vivekkhandelwal1 opened on Aug 25, 2026

TokenSpeed Model Support: Top Open-Source Models & Day 0 Collaboration

high priority

Status: Open.

#965 In lightseekorg/tokenspeed;

· lightseek-bot opened on Aug 6, 2026

TokenSpeed Project Ethos and External Contributions

high priority

Status: Open.

#149 In lightseekorg/tokenspeed;

· lightseek-bot opened on May 14, 2026