Repository navigation

All pull requests

Labels (24)Milestones (0)New pull request

Open

(513)

Closed

(295)

Enable GLM 5.3 tensor parallelism on the ROCm backend
#1257davidcanar opened Oct 23, 2026
Add OdinLink (odl_tb5) TP transport: --transport odl
#1256davidcanar opened Oct 23, 2026
Link the ROCm MXFP4 and GLM tests against ROCm compat and stub
#1254fairfieldt opened Oct 23, 2026
Compile ds4_cuda.cu as C++17 like the other CUDA objects
#1253gian21391 opened Oct 22, 2026
Makefile: link ds4_rocm_unavailable.o into test_mxfp4_rocm
#1252scale03 opened Oct 22, 2026
server: stream Qwen tool-call arguments over OpenAI SSE
#1251softpudding opened Oct 22, 2026
fix typo
#1250phansel opened Oct 22, 2026
Handle quantized embeddings in non-Apple V4.1 prefill
#1248audreyt opened Oct 21, 2026Contributor
Fix Engram unmapping on larger host pages
#1247audreyt opened Oct 21, 2026Contributor
Add experimental Muse Spark Open Metal support
#1246jjakemaness opened Oct 21, 2026Contributor
Allow up to 256 images per server request
#1245ivanfioravanti opened Oct 21, 2026Contributor
metal: fix session memory leak on /ctx
#1244IngeniousIdiocy opened Oct 21, 2026Contributor
Fix DeepSeek V4.1 Q2 download checksum
#1242pminervini opened Oct 21, 2026
Batch the aligned head of ragged prefill tail chunks through the compressor replay
#1241emilianbold opened Oct 21, 2026Contributor
Qwen3.8 Flash Next: faster Metal decode (bit-exact)
#1238paradise-runner opened Oct 20, 2026
1
Run DeepSeek V4.1 Flash 1017 Q1 (Unsloth UD-IQ1_S) resident on a 128 GB Mac
#1237audreyt opened Oct 20, 2026Contributor
metal: process medium V4.1 SSD appends in one layer sweep
#1236nomishbhardwaj opened Oct 20, 2026
metal: use exact divisions in the V4.1 FP4 activation quantizer
#1234nomishbhardwaj opened Oct 19, 2026
CUDA: add resident V4.1 multi-GPU layer placement, validated on Turing
#1233agorevski opened Oct 19, 2026
download_model: accept the 1017 ds41f-q2 file
#1232nomishbhardwaj opened Oct 19, 2026
Avoid nearest_int overflow after a collapsed weighted quant fit
#1230audreyt opened Oct 18, 2026Contributor
docs: add Qwen API client setup and diagnosis guide
#1229altdotboy opened Oct 18, 2026
server: retain bounded cold cache prefixes for long prompts
#1228altdotboy opened Oct 18, 2026
speed-bench: M3 Ultra 60-core numbers for DeepSeek Flash Q2 and Qwen3.8 Q4
#1227ferni7 opened Oct 18, 2026
Replace invalid UTF-8 with U+FFFD in server JSON output
#1226Chida82 opened Oct 18, 2026

Pull requests