Repository navigation

All pull requests

Labels (9)Milestones (0)New pull request

Open

(24)

Closed

(30)

Choose the decode context bucket from active requests only
#86Quentin-Anthony opened Oct 27, 2026
Create per-request random generators on CPU
#85Quentin-Anthony opened Oct 27, 2026
Turn off dynamo automatic dynamic shapes so warmup works with more than one bucket
#84Quentin-Anthony opened Oct 27, 2026
Use an exact top-k without sort in the MoE router fallback
#83Quentin-Anthony opened Oct 27, 2026
Add MTP decode
#78qingzwang opened Oct 23, 20264 tasks done
Llama 3: use the per-head block table for GQA decode at TP < num_kv_heads
#77niklas-palm opened Oct 23, 2026
contrib/benchmark: one serving benchmark for every contrib model
#75whn09 opened Oct 23, 2026
contrib: add Hy3-FP8 (HYV3ForCausalLM) static-FP8 model
#73qingzwang opened Oct 22, 20269 tasks done
[feat] Add DeepSeek V4.1 Flash with DSpark
#71OswaldHe opened Oct 21, 202611 of 13 tasks
4
[feat] Add generic speculative decoding adapter for custom drafter so that models in contrib/ can be self-contained
#70OswaldHe opened Oct 21, 2026
7
Add Voxtral audio model support with DFlash speculative decoding
#69jimburtoft opened Oct 21, 2026
1
contrib: Qwen3.5 (hybrid DeltaNet, dense + MoE) on Trn2, on the recurrent-state core
#68whn09 opened Oct 21, 202611 of 13 tasks
4
contrib: MiMo-V2.5, V2.5-Pro and V2.6 Flash-RL / Pro-RL (MiMoV2ForCausalLM) on Trn2
#67whn09 opened Oct 21, 202611 of 13 tasks
contrib: GLM-5.3 (GlmMoeDsaForCausalLM) on Trn2
#66whn09 opened Oct 21, 202611 of 13 tasks
contrib: Tencent Hy4-preview (HYV4ForCausalLM) on Trn2
#65whn09 opened Oct 21, 202611 of 13 tasks
Fix text-only vision buckets and warmup block-table alignment for non-power-of-2 blocks
#64whn09 opened Oct 21, 2026
Add Hy3-FP8 (HYV3ForCausalLM) static-FP8 model with DCP support
#59qingzwang opened Oct 17, 20263 of 5 tasks
Add Qwen3-MoE (Qwen3-30B-A3B) support with optional YaRN long context
#56jlonge4 opened Sep 11, 2026
Fix auto-label workflow failing on labels that don't exist in the repo
#52MichaelWalker-git opened Aug 25, 20261 task
Add Qwen3 DFlash support and fix tied embeddings
#50Shekswess opened Aug 25, 20264 of 6 tasks
Add gpt-oss 20B DFlash recipe
#49Shekswess opened Aug 25, 20264 of 5 tasks
Add Llama 3.1 8B DFlash recipe
#48Shekswess opened Aug 25, 20264 of 5 tasks
Add DFlash speculative decoding
#47Shekswess opened Aug 25, 20263 of 5 tasks
feat(model): add native BF16 Qwen3 and Qwen3MoE support with FP8 optimizations
#39wutongabc opened Jul 23, 20263 of 5 tasks

Pull requests