Eval bug: Generation speed loss after b5920

[Nepherpitou]

Name and Version

.\llama-cli.exe --version
ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 3 CUDA devices:
Device 0: NVIDIA GeForce RTX 4090, compute capability 8.9, VMM: yes
Device 1: NVIDIA GeForce RTX 5090, compute capability 12.0, VMM: yes
Device 2: NVIDIA GeForce RTX 3090, compute capability 8.6, VMM: yes
load_backend: loaded CUDA backend from C:\Users\unat\llm\llamacpp\cuda12\ggml-cuda.dll
load_backend: loaded RPC backend from C:\Users\unat\llm\llamacpp\cuda12\ggml-rpc.dll
load_backend: loaded CPU backend from C:\Users\unat\llm\llamacpp\cuda12\ggml-cpu-icelake.dll
version: 5920 (d9b6910)
built with clang version 19.1.5 for x86_64-pc-windows-msvc

Operating systems

Windows

GGML backends

CUDA

Hardware

Ryzen 7900X + 64Gb DDR5@6200

Models

Qwen3-32b (Unsloth's Qwen3-32B-128K-UD-Q6_K_XL)

Problem description & steps to reproduce

When updating to any newer version than b5920 I got speed loss from 36tps to 30tps and power draw reduction on RTX 5090 from 226W to 200W.

First Bad Commit

I think it's https://github.com/ggml-org/llama.cpp/releases/tag/b5922

Relevant log output

**b5920**:


slot launch_slot_: id  0 | task 0 | processing task
slot update_slots: id  0 | task 0 | new prompt, n_ctx_slot = 65536, n_keep = 0, n_prompt_tokens = 42
slot update_slots: id  0 | task 0 | kv cache rm [0, end)
slot update_slots: id  0 | task 0 | prompt processing progress, n_past = 42, n_tokens = 42, progress = 1.000000
slot update_slots: id  0 | task 0 | prompt done, n_past = 42, n_tokens = 42
slot      release: id  0 | task 0 | stop processing: n_past = 2567, truncated = 0
slot print_timing: id  0 | task 0 |
prompt eval time =     116.05 ms /    42 tokens (    2.76 ms per token,   361.92 tokens per second)
       eval time =   69321.83 ms /  2526 tokens (   27.44 ms per token,    36.44 tokens per second)
      total time =   69437.88 ms /  2568 tokens
srv  update_slots: all slots are idle


**b5989**:


slot launch_slot_: id  0 | task 2107 | processing task
slot update_slots: id  0 | task 2107 | new prompt, n_ctx_slot = 65536, n_keep = 0, n_prompt_tokens = 42
slot update_slots: id  0 | task 2107 | need to evaluate at least 1 token for each active slot, n_past = 42, n_prompt_tokens = 42
slot update_slots: id  0 | task 2107 | kv cache rm [41, end)
slot update_slots: id  0 | task 2107 | prompt processing progress, n_past = 42, n_tokens = 1, progress = 0.023810
slot update_slots: id  0 | task 2107 | prompt done, n_past = 42, n_tokens = 1
slot      release: id  0 | task 2107 | stop processing: n_past = 2716, truncated = 0
slot print_timing: id  0 | task 2107 |
prompt eval time =     402.88 ms /     1 tokens (  402.88 ms per token,     2.48 tokens per second)
       eval time =   86797.09 ms /  2675 tokens (   32.45 ms per token,    30.82 tokens per second)
      total time =   87199.96 ms /  2676 tokens
srv  update_slots: all slots are idle

[ggerganov]

Likely already fixed with #14855

[Nepherpitou]

Likely already fixed with #14855

Checked on https://github.com/ggml-org/llama.cpp/releases/tag/b5990 - got 36tps. Yep, it was fixed. Thanks!