Fix Ctrl-C not working in run_chat, run_llama (#11)

* Fix Ctrl-C not working in run_chat, run_llama

Fixes #7.

Also only set GGML_METAL_PATH_RESOURCES on Apple machines.

* Fix run_llama

* Silence stdout output when running tests, try to fix MacOS random test fails

src/run-programs.jl

@@ -22,9 +22,14 @@ See also: `run_chat`, `run_server`
 """
 function run_llama(; model::AbstractString, prompt::AbstractString="", nthreads::Int=Threads.nthreads(), n_gpu_layers::Int=99, ctx_size::Int=2048, args=``)
     cmd = `$(llama_cpp_jll.main()) --model $model --prompt $prompt --threads $nthreads --n-gpu-layers $n_gpu_layers --ctx-size $ctx_size $args`
-    # Provides the path to locate ggml-metal.metal file (must be provided separately)
-    cmd = addenv(cmd, "GGML_METAL_PATH_RESOURCES" => joinpath(llama_cpp_jll.artifact_dir, "bin"))
-    return read(cmd, String)
+    if Sys.isapple()
+        # Provides the path to locate ggml-metal.metal file (must be provided separately)
+        cmd = addenv(cmd, "GGML_METAL_PATH_RESOURCES" => joinpath(llama_cpp_jll.artifact_dir, "bin"))
+    end
+    s = disable_sigint() do
+        read(cmd, String)
+    end
+    return s
 end
 
 """
@@ -54,9 +59,15 @@ See also: `run_llama`, `run_server`
 """
 function run_chat(; model::AbstractString, prompt::AbstractString="", nthreads::Int=Threads.nthreads(), n_gpu_layers::Int=99, ctx_size::Int=2048, args=``)
     cmd = `$(llama_cpp_jll.main()) --model $model --prompt $prompt --threads $nthreads --n-gpu-layers $n_gpu_layers --ctx-size $ctx_size $args -ins`
-    # Provides the path to locate ggml-metal.metal file (must be provided separately)
-    cmd = addenv(cmd, "GGML_METAL_PATH_RESOURCES" => joinpath(llama_cpp_jll.artifact_dir, "bin"))
-    run(cmd)
+    if Sys.isapple()
+        # Provides the path to locate ggml-metal.metal file (must be provided separately)
+        cmd = addenv(cmd, "GGML_METAL_PATH_RESOURCES" => joinpath(llama_cpp_jll.artifact_dir, "bin"))
+    end
+    disable_sigint() do
+        # disallow julia's SIGINT (Ctrl-C) handler, and allow Ctrl-C
+        # to be caught by llama.cpp
+        run(cmd)
+    end
 end
 
 """

test/run-programs.jl

@@ -1,15 +1,20 @@
 @testset verbose=true "Llama, no model needed" begin
     showtestset()
 
+    default_run_kwargs = Dict(:n_gpu_layers => 1, :ctx_size => 8)
     @testset "run_llama" begin
         showtestset()
-        @test run_llama(; model="", args=`-h`) isa String
-        @test run_llama(; model="", prompt="", args=`-h`) isa String
+        redirect_stdio(stdout=devnull) do
+            @test run_llama(; model="", args=`-h`, default_run_kwargs...) isa String
+            @test run_llama(; model="", prompt="", args=`-h`, default_run_kwargs...) isa String
+        end
     end
 
     @testset "run_chat" begin
         showtestset()
-        @test run_chat(; model="", args=`-h`) isa Base.Process
+        redirect_stdio(stdout=devnull) do
+            @test run_chat(; model="", args=`-h`, default_run_kwargs...) isa Base.Process
+        end
     end
 
     # @testset "LlamaContext" begin