Search results

[Perf] #48137 costs ~10.6% spec-decode acceptance and #48660 shifts output distributions on DeepSeek-V4-Flash — isolated via #48660-only arm on a production 2-node deployment

cpuRelated to CPU backendsRelated to CPU backends

Status: Open.

#49927 In vllm-project/vllm;

· Reederey87 opened on Jul 27, 2026

[Bug]: EngineDeadError NVFP4 marlin

bugSomething isn't workingSomething isn't working

quantization

Status: Open.

#49926 In vllm-project/vllm;

· bvowell1975 opened on Jul 27, 2026

[Bug][XPU]: GDN attention silently corrupts memory under load — fix merged in vllm-xpu-kernels but requirements/xpu.txt pins a release that predates it

bugSomething isn't workingSomething isn't working

intel-gpuRelated to Intel GPURelated to Intel GPU

Status: Open.

#49924 In vllm-project/vllm;

· frobnitzem opened on Jul 27, 2026

[Bug]: [Regression] Assertion res == CUresult::CUDA_SUCCESS failed in FlashMLA (phase1.cuh) for DeepSeek-V4 on v0.26.0 (Works in v0.25.0)

bugSomething isn't workingSomething isn't working

Status: Open.

#49922 In vllm-project/vllm;

· chendongshuo opened on Jul 27, 2026

[Perf] BF16x3 router GEMM gated off family-120 Blackwell (GB10 / DGX Spark, sm_121) — the only barrier for DeepSeek-V4-Flash's fp32 router

Status: Open.

#49921 In vllm-project/vllm;

· Reederey87 opened on Jul 27, 2026

[Bug]: DiffusionGemma - Unconditional minimax_m3 warmup import in kernel_warmup() crashes engine startup for unrelated models (Triton JIT fails to parse index_topk kernel)

bugSomething isn't workingSomething isn't working

Status: Open.

#49920 In vllm-project/vllm;

· steveh250 opened on Jul 27, 2026

[Bug]: Prefill with prompt length == 1 + num_speculative_tokens misclassified as uniform decode → FULL spec-verify cudagraph skips GDN/hybrid recurrent-state write → deterministic garbage (any spec method incl. ngram; v0.25.1 & v0.26.0)

quantization

Status: Open.

#49918 In vllm-project/vllm;

· avesed opened on Jul 26, 2026

[Feature]: DeepGEMM kernels are never warmed — only 2 of 24 entry points, so fp8_einsum JIT-loads during serving

Status: Open.

#49905 In vllm-project/vllm;

· robotnursenyc opened on Jul 26, 2026

[Bug]: Tiered KV offload promotes every waiting request (which fills primary DRAM pool)

bugSomething isn't workingSomething isn't working

Status: Open.

#49902 In vllm-project/vllm;

· t348575 opened on Jul 26, 2026

[Bug] DeepSeek-V4 on SM12x: NaN MQA logits drive top_k_per_row_prefill to emit uninitialized smem as indices -> illegal memory access

Status: Open.

#49896 In vllm-project/vllm;

· linjiapro opened on Jul 26, 2026

[Bug]: SpeculativeConfig method="draft_model" cannot load mixed-precision compressed-tensors checkpoints (config_groups)

bugSomething isn't workingSomething isn't working

quantization

Status: Open.

#49893 In vllm-project/vllm;

· Resh19S opened on Jul 26, 2026

[Bug]: GLM-5.2-NVFP4 produces garbled/incorrect output and hits NotImplementedError in forward_mha on GB10 (SM121a) with FLASHINFER_MLA_SPARSE_SM120

bugSomething isn't workingSomething isn't working

Status: Open.

#49886 In vllm-project/vllm;

· AlpaslanErdag opened on Jul 26, 2026