cuda: allow ggml_cuda_cpy to copy contiguous F32 and F16 tensors greater than INT_MAX

[jukofyork]

This is a partial fix for #15049 and #14325, so the specific case where we are using cudaMemcpyAsync doesn't fail.

See #15298 and the discussion #15049 (comment) for details on fixing this for the non- cudaMemcpyAsync case.

I'm unsure about mudnnMemcpyAsync and the only reference it I can find is here:

https://github.com/SJTU-IPADS/PowerInfer/blob/d3ebd7c5666348cf43c22f0d62dfbc9a763cffb8/smallthinker/ggml/src/ggml-musa/mudnn.cu#L88

// Asynchronous memory copy using mudnn::Unary::IDENTITY
musaError_t mudnnMemcpyAsync(ggml_backend_cuda_context& ctx, const ggml_tensor* dst, const ggml_tensor* src) {
    mudnn::Tensor tensor_dst, tensor_src;

    MUDNN_CHECK(tensor_dst.SetType(ggml_type_to_mudnn_type(dst->type)));
    MUDNN_CHECK(tensor_src.SetType(ggml_type_to_mudnn_type(src->type)));

    std::vector<int64_t> dims, strides;
    const int ndims = get_ggml_dims_and_strides(src, dims, strides);

    MUDNN_CHECK(tensor_dst.SetNdInfo(ndims, dims.data(), strides.data()));
    MUDNN_CHECK(tensor_src.SetNdInfo(ndims, dims.data(), strides.data()));
    MUDNN_CHECK(tensor_dst.SetAddr(dst->data));
    MUDNN_CHECK(tensor_src.SetAddr(src->data));

    mudnn::Unary op;
    MUDNN_CHECK(op.SetMode(mudnn::Unary::Mode::IDENTITY));
    MUDNN_CHECK(op.SetAlpha(0.0f));
    MUDNN_CHECK(op.SetBeta(0.0f));

    mudnn::Handle* handle = get_cached_handle(ctx.device);
    MUDNN_CHECK(handle->SetStream(ctx.stream()));
    MUDNN_CHECK(op.Run(*handle, tensor_dst, tensor_src));

    return musaSuccess;
}

it doesn't look to need the <= INT_MAX assertion, but not 100% sure without any proper API references.

[jukofyork]

Just testing this now and will un-draft after.

[jukofyork]

It doesn't fix my crash sadly:

slot get_availabl: id  0 | task -1 | selected slot by LRU, t_last = -1
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, task.n_tokens = 42410
slot update_slots: id  0 | task 0 | n_tokens = 0, memory_seq_rm [0, end)
slot update_slots: id  0 | task 0 | prompt processing progress, n_tokens = 18432, batch.n_tokens = 18432, progress = 0.434614
slot update_slots: id  0 | task 0 | n_tokens = 18432, memory_seq_rm [18432, end)
slot update_slots: id  0 | task 0 | prompt processing progress, n_tokens = 36864, batch.n_tokens = 18432, progress = 0.869229
/home/juk/llama.cpp/ggml/src/ggml-cuda/cpy.cu:326: GGML_ASSERT(ggml_nbytes(src0) <= INT_MAX) failed
/home/juk/llama.cpp/build/bin/libggml-base.so(+0x16298)[0x7f4c5c95d298]
/home/juk/llama.cpp/build/bin/libggml-base.so(ggml_print_backtrace+0x1e4)[0x7f4c5c95d664]
/home/juk/llama.cpp/build/bin/libggml-base.so(ggml_abort+0x11e)[0x7f4c5c95d7ee]
/home/juk/llama.cpp/build/bin/libggml-cuda.so(_Z13ggml_cuda_cpyR25ggml_backend_cuda_contextPK11ggml_tensorPS1_+0xef7)[0x7f4c552eb337]
/home/juk/llama.cpp/build/bin/libggml-cuda.so(+0x133b28)[0x7f4c55333b28]
/home/juk/llama.cpp/build/bin/libggml-cuda.so(+0x13401f)[0x7f4c5533401f]
/home/juk/llama.cpp/build/bin/libggml-base.so(ggml_backend_sched_graph_compute_async+0x807)[0x7f4c5c977a67]
/home/juk/llama.cpp/build/bin/libllama.so(_ZN13llama_context13graph_computeEP11ggml_cgraphb+0xa1)[0x7f4c5c69c1c1]
/home/juk/llama.cpp/build/bin/libllama.so(_ZN13llama_context14process_ubatchERK12llama_ubatch14llm_graph_typeP22llama_memory_context_iR11ggml_status+0xe1)[0x7f4c5c69d831]
/home/juk/llama.cpp/build/bin/libllama.so(_ZN13llama_context6decodeERK11llama_batch+0x291)[0x7f4c5c6a3021]
/home/juk/llama.cpp/build/bin/libllama.so(llama_decode+0xb)[0x7f4c5c6a3f1b]
/home/juk/llama.cpp/build/bin/llama-server(+0xdd166)[0x555ade363166]
/home/juk/llama.cpp/build/bin/llama-server(+0xa29d9)[0x555ade3289d9]
/home/juk/llama.cpp/build/bin/llama-server(+0x62a52)[0x555ade2e8a52]
/lib/x86_64-linux-gnu/libc.so.6(+0x2724a)[0x7f4c5c44624a]
/lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0x85)[0x7f4c5c446305]
/home/juk/llama.cpp/build/bin/llama-server(+0x64791)[0x555ade2ea791]

Closing for now and will take a look at that other PR that tries to do this for all types.