Default NvvmArch::Compute75 silently produces InvalidPtx on pre-Turing GPUs (Pascal/Maxwell/Volta)

Inizia da crates/nvvm/src/lib.rs e dalla guida introduttiva per confermare il valore predefinito Compute75 documentato e il modo in cui l’esempio configura la propria architettura. Poi esamina cust::Module::from_ptx e i punti di ingresso del loader PTX per valutare l’opzione diagnostica. Il lavoro è completo quando un miglioramento concordato della guida o della diagnostica affronta le GPU pre-Turing e include una copertura o una documentazione per il comportamento risultante.

What happened

Following the getting-started guide on a GTX 1070 (Pascal, sm_61) with CUDA 12.9, cargo run builds successfully but fails at runtime with:

Error: InvalidPtx

No indication of why the PTX is invalid.

Root cause

cuda_builder::CudaBuilder defaults to NvvmArch::Compute75 (crates/nvvm/src/lib.rs):

/// This default value of 7.5 corresponds to Turing and later devices. We default to this
/// because it is the minimum supported by CUDA 13.0 while being in the middle of the range
/// supported by CUDA 12.x.
#[default]
Compute75,

This is a deliberate, documented choice — but it means the getting-started example silently generates PTX targeting sm_75, which the CUDA driver refuses to load on any GPU below Turing (Pascal, Maxwell, Volta — all still fully supported for building/running on CUDA 12.x, just not compiling for them by default). The failure surfaces as a generic cust::error::CudaError::InvalidPtx with nothing pointing at the arch mismatch, which is a confusing first-run experience for anyone following the guide on older hardware.

Suggestions (either would help; happy to send a PR for whichever is preferred)

1. Guide fix: add a line to the getting-started example / Windows setup section noting that .arch(NvvmArch::ComputeXX) should be set explicitly to match the target GPU's actual compute capability, with a pointer to Compute Capability Gating and to nvidia-smi --query-gpu=compute_cap --format=csv for finding it.
2. Better diagnostics: have cust::Module::from_ptx (or the PTX loader) detect a CUDA_ERROR_INVALID_PTX / CUDA_ERROR_NO_BINARY_FOR_GPU outcome and, when possible, compare the PTX's .target smXX line against the current device's actual compute capability, surfacing a clearer error like "PTX built for sm_75 but device is sm_61 — rebuild with .arch(NvvmArch::Compute61)".

Either fix would have saved a fair amount of debugging time here since the error gives no hint that architecture — not toolchain or driver version — is the issue.

Environment

• GPU: GTX 1070 (Pascal, sm_61)
• Driver: 582.28
• CUDA Toolkit: 12.9.2
• Rust nightly: pinned per rust-toolchain.toml at the commit in question
• OS: Windows 11