Bug: `GGML_SCHED_MAX_COPIES=2` produces repeated tokens when graph reuse is enabled

[KeinNiemand]

What happened?

With a CUDA build using GGML_SCHED_MAX_COPIES=2, llama-server produces bad output when graph reuse is enabled. The output repeats almost every token:

The capital capital of of France France is is Paris Paris.. Paris Paris is is not not only only the the capital capital but but also also the the largest largest city city in in France France..

Adding --no-graph-reuse fixes the output. I expected graph reuse not to change the generated text.

Name and Version

llama-server.exe --version
version: 4633 (b3dfb785)
built with MSVC 19.51.36248.0

What operating system are you seeing the problem on?

Windows

Relevant log output

llama_init_from_model: graph_reuse   = 1
llama_init_from_model: pipeline parallelism enabled (n_copies=2)
llama_init_from_model: graph splits = 3

[ikawrakow]

Adding --no-graph-reuse fixes the output. I expected graph reuse not to change the generated text.

That's I expected too. But apparently the logic that determines if a graph can be reused fails when GGML_SCHED_MAX_COPIES>1, which is something I never use.