Merge pull request #2941 from JuliaGPU/tb/backwards_compatibility

Relax checks as per backwards compatibility.

Author: maleadt

Project.toml

@@ -55,7 +55,7 @@ AbstractFFTs = "0.5, 1.0"
 Adapt = "4.4"
 BFloat16s = "0.5, 0.6"
 CEnum = "0.2, 0.3, 0.4, 0.5"
-CUDA_Compiler_jll = "0.2"
+CUDA_Compiler_jll = "0.3"
 CUDA_Driver_jll = "13"
 CUDA_Runtime_Discovery = "1"
 CUDA_Runtime_jll = "0.19"
@@ -65,7 +65,7 @@ DataFrames = "1.5"
 EnzymeCore = "0.8.2"
 ExprTools = "0.1"
 GPUArrays = "11.2.4"
-GPUCompiler = "1.1"
+GPUCompiler = "1.4"
 GPUToolbox = "0.3, 1"
 KernelAbstractions = "0.9.38"
 LLVM = "9.3.1"

src/compatibility.jl

@@ -296,15 +296,19 @@ function llvm_compat(version=LLVM.version())
     return (cap=cap_support, ptx=ptx_support)
 end
 
-function cuda_compat(driver=driver_version(), compiler=compiler_version())
-    # devices have to be supported by both the compiler and the driver
-    driver_cap_support = cuda_cap_support(driver)
-    compiler_cap_support = cuda_cap_support(compiler)
-    cap_support = sort(collect(driver_cap_support ∩ compiler_cap_support))
-
-    # PTX code only has to be supported by the compiler
-    compiler_ptx_support = cuda_ptx_support(compiler)
-    ptx_support = cuda_ptx_support(compiler)
+function cuda_compat(version=runtime_version())
+    # we don't have to check the driver version, because it offers backwards compatbility
+    # beyond the CUDA toolkit version (e.g. R580 for CUDA 13 still supports Volta as
+    # deprecated in CUDA 13), and we don't have a reliable way to query the actual version
+    # as NVML isn't available on all platforms. let's instead simply assume that unsupported
+    # devices will not be exposed to the CUDA runtime and thus won't be visible to us.
+
+    # we also don't have to check the compiler version, because CUDA_Compiler_jll is
+    # guaranteed to have the same major version as CUDA_Runtime_jll, meaning that the
+    # compiler will always support at least the same devices as the runtime.
+
+    cap_support = sort(collect(cuda_cap_support(version)))
+    ptx_support = sort(collect(cuda_ptx_support(version)))
 
     return (cap=cap_support, ptx=ptx_support)
 end

src/compiler/compilation.jl

@@ -204,7 +204,7 @@ end
             error("CUDA.jl requires PTX $requested_ptx, which is not supported by LLVM $(LLVM.version())")
         llvm_ptx = maximum(llvm_ptxs)
         isempty(cuda_ptxs) &&
-            error("CUDA.jl requires PTX $requested_ptx, which is not supported by CUDA driver $(driver_version()) / runtime $(runtime_version())")
+            error("CUDA.jl requires PTX $requested_ptx, which is not supported by CUDA $(runtime_version())")
         cuda_ptx = maximum(cuda_ptxs)
     end
 
@@ -229,7 +229,7 @@ end
         ## use the highest capability supported by CUDA
         cuda_caps = filter(<=(capability(dev)), cuda_support.cap)
         isempty(cuda_caps) &&
-            error("Compute capability $(requested_cap) is not supported by CUDA driver $(driver_version()) / runtime $(runtime_version())")
+            error("Compute capability $(requested_cap) is not supported by CUDA $(runtime_version())")
         cuda_cap = maximum(cuda_caps)
     end
 

src/initialization.jl

@@ -66,9 +66,9 @@ function __init__()
         return
     end
 
-    if !(v"12" <= driver < v"14-")
-        @error "This version of CUDA.jl only supports NVIDIA drivers for CUDA 12.x or 13.x (yours is for CUDA $driver)"
-        _initialization_error[] = "CUDA driver unsupported"
+    if driver < v"12"
+        @error "This version of CUDA.jl requires an NVIDIA driver for CUDA 12.x or higher (yours only supports up to CUDA $driver)"
+        _initialization_error[] = "NVIDIA driver too old"
         return
     end
 
@@ -133,10 +133,6 @@ function __init__()
     if runtime < v"12"
         @error "This version of CUDA.jl only supports CUDA 12 or higher (your toolkit provides CUDA $runtime)"
     end
-    if runtime.major != driver.major
-        @warn """You are using CUDA $runtime with a driver for CUDA $(driver.major).x.
-                 It is recommended to upgrade your driver, or switch to automatic installation of CUDA."""
-    end
 
     # ensure the loaded runtime matches what we precompiled for.
     if toolkit_version == nothing