Issues

• [TPU/CI] test_collective_compile_cache intermittently hangs in stage-a-test-tpu-unit

  bugtpu
  sgl-project/sglang-torchtpu#1993 · Filter by author jcc-google opened on Nov 6, 2026
  

• [Test] TPU plugin unit tests that never run or can't fail

  good first issuehelp wantedtpu
  sgl-project/sglang-torchtpu#1984 · Filter by author pyc96 opened on Nov 6, 2026
  3

• [TPU/Raiden] With the default SGLANG_TPU_RAIDEN_STAGING_LAYERS=8, an asymmetric-attention-TP decode raises NotImplementedError in send_metadata, which kills the decode scheduler (SIGQUIT)

  sgl-project/sglang-torchtpu#1954 · Filter by author ap9272 opened on Nov 4, 2026

• [TPU/PD/Raiden] Raiden reshard ignores fp8 KV packing → decode KV scrambled under DP attention (Wrong output returns HTTP 200)

  sgl-project/sglang-torchtpu#1903 · Filter by author kaihmaeda opened on Nov 2, 2026
  

• [RFC] TPU serving roofline: analytical step floor, layout search, and SGLang launch flags

  sgl-project/sglang-torchtpu#1892 · Filter by author brianseung-google opened on Nov 2, 2026
  

• [TPU/PD/Raiden] Unsettled Decode receive sessions are not cancelled on request abort or transfer failure, holding Decode KV pages for 60s (#1113 follow-up)

  sgl-project/sglang-torchtpu#1829 · Filter by author Annie215 opened on Oct 29, 2026
  1

  

• [Tracking] Diffusion (multimodal_gen) on TPU: Wan2.2, LTX-2, FLUX.2-klein, Z-Image-Turbo

  sgl-project/sglang-torchtpu#1824 · Filter by author chandrasekhard2 opened on Oct 28, 2026

• [TPU] Kimi-K3 InferenceX AgentX: plan, PR series and first measurements

  enhancementtpu
  sgl-project/sglang-torchtpu#1823 · Filter by author chandrasekhard2 opened on Oct 28, 2026
  1

• [TPU] Kimi-K3: prefill/decode disaggregation and follow-up work

  sgl-project/sglang-torchtpu#1822 · Filter by author chandrasekhard2 opened on Oct 28, 2026
  

• [P2] [TPU/Memory] Cold start keeps ~1.25 GiB more host RSS per TP rank than a warm start (glibc arena fragmentation)

  sgl-project/sglang-torchtpu#1817 · Filter by author joeywan-google opened on Oct 28, 2026
  

• [Docs][TPU] Fix conflicting model support information

  sgl-project/sglang-torchtpu#1778 · Filter by author MiaoJiawei97 opened on Oct 27, 2026
  1

• [TPU/fp8] Online fp8 quantization of an unquantized draft uses a per-rank per-tensor scale, so replicated kv heads diverge across TP ranks

  sgl-project/sglang-torchtpu#1767 · Filter by author ap9272 opened on Oct 26, 2026
  

• [Model]DeepSeek-V4.1 support tracking

  deepseektpu
  sgl-project/sglang-torchtpu#1748 · Filter by author JamesBrianD opened on Oct 26, 2026
  2

  

• [Bug] MiniLB streams decode without checking prefill's body; a failed streaming prefill (200 + in-band error/abort) hangs until decode timeout

  bug
  sgl-project/sglang-torchtpu#1706 · Filter by author vsaatg opened on Oct 24, 2026
  

• [Bug] PD router treats a failed streaming prefill (200 + in-band error/abort) as success; decode waits for KV until timeout

  bug
  sgl-project/sglang-torchtpu#1704 · Filter by author vsaatg opened on Oct 24, 2026
  

• [TPU/PD/Raiden] A decode-side abort does not release the prefill's registered reads: the room is held until SGLANG_RAIDEN_CHUNK_TRANSFER_TIMEOUT_S (#1114 follow-up)

  sgl-project/sglang-torchtpu#1703 · Filter by author alisachen-google opened on Oct 24, 2026
  

• [Bug][TPU] Prefill chunks padded into the 8192-token EXTEND program return a wrong first token (Qwen3-32B TP8, v0.1.0-tpu)

  sgl-project/sglang-torchtpu#1700 · Filter by author tyuchn opened on Oct 23, 2026
  1

  

• [TPU/Raiden] Hybrid GDN under push-reshard bypasses the init-time refusal: send_metadata raises NotImplementedError and takes down the decode scheduler

  sgl-project/sglang-torchtpu#1671 · Filter by author ap9272 opened on Oct 22, 2026
  

• [TPU/Raiden] Decision criteria for turning SGLANG_RAIDEN_RESHARD_LAND_ARENA on by default

  sgl-project/sglang-torchtpu#1665 · Filter by author ap9272 opened on Oct 22, 2026
  1

  

• [TPU/Raiden] Group failure gives rank 0 a placeholder reason ("failed on another attention TP/CP rank"), so the HTTP body can drop the real refusal

  sgl-project/sglang-torchtpu#1664 · Filter by author ap9272 opened on Oct 22, 2026
  1

  

• [TPU/Raiden] LAND merge fence: measure the TPOT cost of the two blocking syncs per merging round, then keep, poll, or drop the pre-merge sync

  sgl-project/sglang-torchtpu#1662 · Filter by author ap9272 opened on Oct 22, 2026
  1

  

• [TPU/Raiden] Push-reshard receive plans bypass StagingAdmission; LAND scratch doubles each plan's host staging

  sgl-project/sglang-torchtpu#1661 · Filter by author ap9272 opened on Oct 22, 2026
  

• [TPU/Raiden] fp8 push-reshard: LAND is still refused for the sliding-window group and for 3-head sources

  sgl-project/sglang-torchtpu#1660 · Filter by author ap9272 opened on Oct 22, 2026

• [TPU/Raiden] Push-reshard into a decode with larger attention TP (split direction, src_tp < dst_tp)

  sgl-project/sglang-torchtpu#1659 · Filter by author ap9272 opened on Oct 22, 2026

• [TPU][Raiden] Hybrid-SWA models split SGLANG_RAIDEN_HOST_POOL_GB 50/50 between full and SWA engines

  sgl-project/sglang-torchtpu#1566 · Filter by author Annie215 opened on Oct 19, 2026
  1