Issues

[bug] Flash-Next (mHC highway): a decode step between another request's prefill chunks corrupts output, and the process stays bad for concurrent requests until restart
Atlas-Inf/atlas#147 · opened on Oct 1, 2026
[job] AMD launchers: one thinking switch and a serve-flag cleanup
Atlas-Inf/atlas#146 · opened on Oct 1, 2026
2
[job] Strix Halo: multi-session serving (parallel or sequential) and long context, Flash-Next and the 27B
Atlas-Inf/atlas#145 · opened on Sep 30, 2026
[job] Qwen3.8-Flash-Next on Strix Halo: prefill and TTFT
Atlas-Inf/atlas#144 · opened on Sep 30, 2026
1
[job] Qwen3.8-Flash-Next on Strix Halo: decode speed (serial and MTP)
Atlas-Inf/atlas#143 · opened on Sep 30, 2026
[job] Qwen3.8-Flash-Next: prompt-lookup + MTP hybrid drafting for agentic decode
Atlas-Inf/atlas#130 · opened on Sep 28, 2026
1
[bug] Flash-Next NVFP4: shutdown exits before teardown finishes, and teardown hits ~147k invalid frees
Atlas-Inf/atlas#122 · opened on Sep 27, 2026
6
[job] Repo diet: env-var registry + prune 497 ATLAS_* vars, large files, merged branches
Atlas-Inf/atlas#117 · opened on Sep 27, 2026
3
[job] Qwen3.8-Flash-Next: long-context cold prefill to SGLang parity (stage 1 = #68)
Atlas-Inf/atlas#74 · opened on Sep 25, 2026
7
[job] Qwen3.8-27B on Strix Halo: prefill 2.4x gap + DP4A verify tiers for K=5..9
Atlas-Inf/atlas#72 · opened on Sep 25, 2026
10
[job] GB10: raise the Atlas tensor-core GEMM ceiling (29-39 vs cuBLAS 75-206 TFLOP/s)
Atlas-Inf/atlas#71 · opened on Sep 25, 2026
6
[job] Qwen3.8-Flash-Next: serve EXL3 checkpoints natively on GB10
Atlas-Inf/atlas#65 · opened on Sep 25, 2026
28
[job] GLM-5.3-Flash: new model support on GB10
Atlas-Inf/atlas#59 · opened on Sep 24, 2026
5
[job] Qwen3.8-Flash-Next: land the Strix Halo Windows port
Atlas-Inf/atlas#57 · opened on Sep 23, 2026
8