LLDB Crashes When Following the Steps in Puzzle 9

[isuckatcs]

uv/pixi CLI Version

pixi 0.52.0

Mojo Version

Mojo 25.6.0.dev2025081505 (b13372f2)

GPU Information

Tesla T4 (Google Colab)

Operating System

Ubuntu 22.04.4 LTS

Related Puzzle

Puzzle 9: GPU Debugging Workflow

What happened?

The compiler/JIT compiler in LLDB crashes with the following error:

/content/mojo-gpu-puzzles/solutions/p01/p01.mojo:1:1: error: ptxas application ptx input, line 78; error   : Feature 'Defining labels in .section' requires PTX ISA .version 7.0 or later
ptxas fatal   : Ptx assembly aborted due to errors

from memory import UnsafePointer
^
/content/mojo-gpu-puzzles/.pixi/envs/default/bin/mojo: error: failed to run the pass manager

IIUC for PTX ISA 7.0 we need at least CUDA 11.0, but running pixi run nvcc --version says CUDA 12 is installed.

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2025 NVIDIA Corporation
Built on Tue_May_27_02:21:03_PDT_2025
Cuda compilation tools, release 12.9, V12.9.86
Build cuda_12.9.r12.9/compiler.36037853_0

Apparently running the compiler without using the -O0 and the -g flags simultaneously works, so both of the following commands execute fine:

• pixi run mojo build -O0 solutions/p01/p01.mojo -o solutions/p01/p01_debug
• pixi run mojo build -g solutions/p01/p01.mojo -o solutions/p01/p01_debug

The latter can be used to proceed with the Binary + LLDB and the Binary + CUDA-GDB approaches with some limitations and for some files.

pixi run mojo build -g problems/p09/p09.mojo -o problems/p09/p09_debug crashes as well, but after deleting the elif argv()[1] == "--second-case": and elif argv()[1] == "--third-case": branches from the source code makes it compile. For the other 2 branches I couldn't find a workaround that makes them compile.

Steps to Reproduce

(Option 1)

1. Complet the prerequisits in the README
2. Go to https://puzzles.modular.com/puzzle_09/essentials.html#your-first-debugging-commands
3. Proceed all the way to step 4 in the guide.

(Option 2)

1. Run pixi run mojo build -O0 -g solutions/p01/p01.mojo -o solutions/p01/p01_debug

Relevant Log Output

Additional Context

No response

Code of Conduct

• I agree to follow this project's Code of Conduct

[ehsanmok]

Interesting! thanks for filing the issue. T4 and older GPUs won't be supported and there're definitely more cases like this that the book assumes new architectures. I'll document this along side #63

[al626]

Not sure if this is a separate issue, but lldb debugging also does not work on Apple Silicon:

➜ mojo-gpu-puzzles git:(main) ✗ pixi run mojo debug solutions/p01/p01.mojo
Current executable set to '.../src/mojo/mojo-gpu-puzzles/.pixi/envs/default/bin/mojo' (arm64).
(lldb) br set -n main
Breakpoint 1: 16 locations.
(lldb) run
Process 80035 launched: '.../src/mojo/mojo-gpu-puzzles/.pixi/envs/default/bin/mojo' (arm64)
Process 80035 stopped
* thread #1, queue = 'com.apple.main-thread', stop reason = breakpoint 1.1
    frame #0: 0x000000010002e934 mojo`main
mojo`main:
->  0x10002e934 <+0>:  stp    x26, x25, [sp, #-0x50]!
    0x10002e938 <+4>:  stp    x24, x23, [sp, #0x10]
    0x10002e93c <+8>:  stp    x22, x21, [sp, #0x20]
    0x10002e940 <+12>: stp    x20, x19, [sp, #0x30]
(lldb) c
Process 80035 resuming
function declaration may only have a unique !dbg attachment
ptr @air.thread_position_in_threadgroup.x
.../src/mojo/mojo-gpu-puzzles/solutions/p01/p01.mojo:1:1: error: failed to lower module to LLVM IR for archive compilation, translate module to LLVMIR failed
from memory import UnsafePointer
^
.../src/mojo/mojo-gpu-puzzles/.pixi/envs/default/bin/mojo: error: failed to run the pass manager
Process 80035 exited with status = 1 (0x00000001)

This post seems to indicate puzzles 1-15 should work in apple silicon. If this is a known issue, it would be great if there was e.g. a shell script a user could run that would indicate what puzzles will work with their hardware and what puzzles are expected to fail.

Thank you by the way for the recent addition of Apple silicon GPUs!

GPU: Apple M3
Mojo: Mojo 0.25.7.0.dev2025092205 (22942900)
Pixi: pixi 0.55.0
OS: Mac OS Sequoia 15.5
Xcode: Xcode 16.4 Build version 16F6

[al626]

Also, interestingly, the first case and third case in the p09 puzzles do not have issues in Apple silicon:

➜ mojo-gpu-puzzles git:(main) ✗ pixi run p09 --first-case
✨ Pixi task (p09 in default): mojo problems/p09/p09.mojo --first-case
First Case: Try to identify what's wrong without looking at the code!

result: HostBuffer([10.0, 10.0, 10.0, 10.0])
➜ mojo-gpu-puzzles git:(main) ✗ pixi run p09 --third-case
✨ Pixi task (p09 in default): mojo problems/p09/p09.mojo --third-case
Third Case: Advanced collaborative filtering with shared memory...
WARNING: This may hang - use Ctrl+C to stop if needed

Input array: [1, 2, 3, 4]
Applying collaborative filter using shared memory...
Each thread cooperates with neighbors for smoothing...
Waiting for GPU computation to complete...
Result: HostBuffer([1.0, 2.5, 4.0, 4.0])
[SUCCESS] Collaborative filtering completed successfully!

[ehsanmok]

Yes, puzzles 9 and 10 are nvidia only. Our tooling is ramping up quickly so please stay tuned.