Search results

server: inconsistent created timestamp in /v1/models endpoint breaks client caching

Status: Open.

#26169 In ggml-org/llama.cpp;

· XBold opened on Jul 27, 2026

Vulkan: AMD flash-attention tuning gated on maxComputeSharedMemorySize == 65536 is skipped when driver reports 32768 (Vega/gfx90c, Adrenalin 26.5.2) - ~17% slowdown

Status: Open.

#26163 In ggml-org/llama.cpp;

· writingdeveloper opened on Jul 27, 2026

Eval bug: crash with MiniMax-M3 MSA in and KV cache-type q8_0 q4_x q5_x

bug-unconfirmed

Status: Open.

#26155 In ggml-org/llama.cpp;

· smalinin opened on Jul 27, 2026

[MoE] Scheduler crash during partial offload (GGML_SCHED_MAX_SPLIT_INPUTS)

bug-unconfirmed

Status: Open.

#26152 In ggml-org/llama.cpp;

· CodeToFreedom opened on Jul 26, 2026

[RPC] Crash with --override-tensor (-ot) due to missing memset_tensor

bug-unconfirmed

Status: Open.

#26151 In ggml-org/llama.cpp;

· CodeToFreedom opened on Jul 26, 2026

Eval bug: ROCm gfx1151, GGML_CUDA_ENABLE_UNIFIED_MEMORY=1 causes silent output corruption with Qwen3-Coder-Next

bug-unconfirmed

Status: Open.

#26148 In ggml-org/llama.cpp;

· Vinz2008 opened on Jul 26, 2026

Eval bug: RPC Node Cache Grows without Bound?

bug-unconfirmed

Status: Open.

#26143 In ggml-org/llama.cpp;

· d-shehu opened on Jul 26, 2026

llama_get_embeddings_ith returns NaN/zeros for output slots 2+ with embd-batch decode on Qwen3-VL

Status: Open.

#26142 In ggml-org/llama.cpp;

· smellslikeml opened on Jul 26, 2026

Feature Request: Gigatoken Tokenizer

enhancementNew feature or requestNew feature or request

Status: Open.

#26139 In ggml-org/llama.cpp;

· Nocturna22 opened on Jul 26, 2026

Feature Request: server: expose per-device memory usage (weights / context / compute)

Status: Open.

#26129 In ggml-org/llama.cpp;

· tobocop2 opened on Jul 26, 2026

Misc. bug: server prompt cache is incompatible with the RPC backend — GGML_ASSERT(tensor->data != NULL) in prompt_save() → state_seq_get_data(), triggered as soon as -np > 1

Status: Open.

#26128 In ggml-org/llama.cpp;

· doornail opened on Jul 26, 2026

Windows HIP release omits gfx1152; Radeon 860M fails with invalid device function

Status: Open.

#26127 In ggml-org/llama.cpp;

· AIwork4me opened on Jul 26, 2026