CUDA: extend K-type validation to V-types for flash attention

[sanmai]

Overview

Extends the K-type validation to V-types in ggml_cuda_get_best_fattn_kernel.

Additional information

As with GGML_CUDA_FA_ALL_QUANTS=ON all combinations are possible and unsupported combinations lead to crash instead of falling back to the CPU backend, say -ctk iq4_nl -ctv q8_0 falls back to the CPU, but -ctk q8_0 -ctv iq4_nl leads to crash.

[Thread debugging using libthread_db enabled]
Using host libthread_db library "/usr/lib/x86_64-linux-gnu/libthread_db.so.1".
0x00007ff360aa5ffe in ?? () from /usr/lib/x86_64-linux-gnu/libc.so.6
#0  0x00007ff360aa5ffe in ?? () from /usr/lib/x86_64-linux-gnu/libc.so.6
#1  0x00007ff360a9a7a4 in ?? () from /usr/lib/x86_64-linux-gnu/libc.so.6
#2  0x00007ff360a9a7ed in ?? () from /usr/lib/x86_64-linux-gnu/libc.so.6
#3  0x00007ff360b06c07 in wait4 () from /usr/lib/x86_64-linux-gnu/libc.so.6
#4  0x00007ff360f1d973 in ggml_print_backtrace () from bin/libggml-base.so.0
#5  0x00007ff360f1dabf in ggml_abort () from bin/libggml-base.so.0
#6  0x00007ff35c609b33 in ggml_cuda_flash_attn_ext(ggml_backend_cuda_context&, ggml_tensor*) () from bin/libggml-cuda.so.0
#7  0x00007ff35c65ba26 in ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context*, ggml_cgraph*, bool, bool, void const*) () from bin/libggml-cuda.so.0
#8  0x00007ff35c65bfd2 in ggml_backend_cuda_graph_compute(ggml_backend*, ggml_cgraph*) () from bin/libggml-cuda.so.0
#9  0x00007ff360f39953 in ggml_backend_sched_graph_compute_async () from bin/libggml-base.so.0
#10 0x00007ff3600cb3b0 in llama_context::graph_compute(ggml_cgraph*, bool) () from bin/libllama.so.0
#11 0x00007ff3600cdfd5 in llama_context::process_ubatch(llama_ubatch const&, llm_graph_type, llama_memory_context_i*, ggml_status&) () from bin/libllama.so.0
#12 0x00007ff3600d4ac5 in llama_context::decode(llama_batch const&) () from bin/libllama.so.0
#13 0x00007ff3600d66be in llama_decode () from bin/libllama.so.0
#14 0x00007ff3605ec61c in common_init_from_params(common_params&, bool) () from bin/libllama-common.so.0
#15 0x00007ff361174540 in server_context_impl::load_model(common_params&) () from bin/libllama-server-impl.so
#16 0x00007ff3610c5542 in llama_server(int, char**) () from bin/libllama-server-impl.so
#17 0x00007ff360a31f77 in ?? () from /usr/lib/x86_64-linux-gnu/libc.so.6
#18 0x00007ff360a32027 in __libc_start_main () from /usr/lib/x86_64-linux-gnu/libc.so.6

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES, to debug the issue