Issues

• Boot crash for fp8 block-quant MoE models under tensor parallelism when the per-rank expert intermediate is not 128-aligned (MiniMax-M2, TP8): the repack path consumes TP-refined [64,64] scales at the checkpoint's [128,128] block size

  vllm-project/vllm-gaudi#1929 · Filter by author biondogs opened on Oct 27, 2026

• EngineCore dies at the first 4K-token prefill for GPT-OSS: Dynamo fake-tensor reshape failure in HPURMSNorm.forward_oot (vllm_gaudi/ops/hpu_layernorm.py:19)

  vllm-project/vllm-gaudi#1928 · Filter by author biondogs opened on Oct 27, 2026

• PR #1895 (GDN initial-state masking): deterministic live-serving output deltas on Gaudi 3 — filed for separate triage

  vllm-project/vllm-gaudi#1902 · Filter by author biondogs opened on Oct 20, 2026
  1

• [Question] Greedy (temp=0) output is not byte-identical between a request and its byte-identical re-serve — 3 sightings at 64K/262K, both outputs retrieval-correct. Is this a known numerics consequence of bucket/padding geometry?

  vllm-project/vllm-gaudi#1860 · Filter by author biondogs opened on Oct 7, 2026
  1

• [Bug] PT_HPU_EAGER_PIPELINE_ENABLE=0: 32K streamed request dies mid-generation after ~300 s with no server-side error; engine reports healthy afterward

  vllm-project/vllm-gaudi#1855 · Filter by author biondogs opened on Oct 7, 2026

• Padding-aware bucketing removes the multi-minute first-token tail at ~1M-token context (measured exp vs pad at 262K n=4 + ~999K)

  vllm-project/vllm-gaudi#1854 · Filter by author biondogs opened on Oct 7, 2026

• [Sampling][PT_EAGER] Deterministic PT_DEVMEM OOM at the FIRST sampling pass for long-context requests: the first eager launch after a near-cap prefill triggers a 16 KiB x blockpad128(prompt+output-64) device allocation (device_memory.cpp:1129); author not identifiable from Python

  vllm-project/vllm-gaudi#1846 · Filter by author biondogs opened on Oct 5, 2026

• Prefix caching never hits for Qwen3.5/3.8 on releases/v0.30.0 at the default budget

  vllm-project/vllm-gaudi#1835 · Filter by author jbrink42 opened on Sep 30, 2026
  3

• Request: fused decode kernel for the gated delta rule (GDN) recurrence on Gaudi 2

  vllm-project/vllm-gaudi#1834 · Filter by author jbrink42 opened on Sep 30, 2026

• Request: a paged-attention decode kernel that reads KV blocks in place on Gaudi 2

  vllm-project/vllm-gaudi#1833 · Filter by author jbrink42 opened on Sep 30, 2026

• vllm-gaudi: Gemma 4 sliding-window layers get full-length KV cache

  vllm-project/vllm-gaudi#1832 · Filter by author jbrink42 opened on Sep 30, 2026
  1

• ✅ Release Build Succeeded for pre_releases/v0.30.0

  automatedrelease-report
  vllm-project/vllm-gaudi#1807 · Filter by author github-actions opened on Sep 22, 2026

• [Bug] Warmup memory miscalculation causes OOM with hybrid GDN models

  vllm-project/vllm-gaudi#1736 · Filter by author biondogs opened on Aug 13, 2026
  1

• MoE decode is slow on Gaudi 2 at low batch size

  vllm-project/vllm-gaudi#1733 · Filter by author NatTuck opened on Aug 13, 2026
  3

• Support for DeepSeek-V4-Flash-0731 on Gaudi 2

  vllm-project/vllm-gaudi#1722 · Filter by author liuende502 opened on Aug 11, 2026

• DeepSeek-V4-Flash-0731

  vllm-project/vllm-gaudi#1716 · Filter by author abilliontokens opened on Aug 10, 2026
  3

• Support for GLM-5 (or 5.1/5.2) on Gaudi 2?

  vllm-project/vllm-gaudi#1549 · Filter by author tylerebowers opened on Jun 17, 2026
  114

• 🚦 Team Review Dashboard

  vllm-project/vllm-gaudi#701 · Filter by author xuechendi opened on Dec 8, 2025
  1