【RUN fail】nvidia dgx spark GB10 run fail

[slffjf]

adapters/vllm/run_timing.sh
--model "$HF_MODEL"
--profile configs/profiles/nominal.json
--report-dir /dev/shm/hbfsim-vllm-exact
--num-prompts 1 --input-len 32 --output-len 8
--max-model-len 64 --max-num-batched-tokens 64
--hbf-parameter-regex '^model.layers.0.mlp.experts.w13_weight$'
--hbf-range-bytes 16384 --seed 0
[2026-10-15 16:56:21.314] [info] [syscall_context.cpp:152] Init bpftime syscall mocking..
[2026-10-15 16:56:21.314] [info] [syscall_context.cpp:153] The log will be written to: ~/.bpftime/runtime.log
[2026-10-15 16:56:21.314] [info] [syscall_server_utils.cpp:45] Starting syscall server..
[2026-10-15 16:56:21.458] [warning] [bpftime_vm_compat.hpp:251] VM factory for name: llvm already registered, overwriting
[2026-10-15 16:56:21.760] [info] [bpftime_shm_internal.cpp:1074] Registered shared memory with CUDA: addr=0xebb73fd00000 size=52428800
[2026-10-15 16:56:21.760] [info] [bpftime_shm_internal.cpp:859] Global shm constructed. shm_open_type 1 for bpftime_maps_shm
[2026-10-15 16:56:21.760] [info] [bpftime_shm_internal.cpp:48] Global shm initialized
[2026-10-15 16:56:22.856] [error] [ptx_compiler.cpp:76] Unable to compile: 3
terminate called after throwing an instance of 'std::runtime_error'
what(): Unable to compile
/home/acs/slf/HHM/HBFSim/scripts/run_with_bpftime.sh: 第 136 行： 747727 已中止 （核心已转储） "$@"
环境信息：
vllm 0.15.1 torch 2.9.1+cu128 gb10 sm_121a，Driver Version: 595.99.02 ptxas: NVIDIA (R) Ptx optimizing assembler
Copyright (c) 2005-2025 NVIDIA Corporation
Built on Wed_Aug_20_01:53:56_PM_PDT_2025
Cuda compilation tools, release 13.0, V13.0.88

[zzyuanyi]

Thanks for the detailed report. We found a PTX target/version mapping bug that can affect GB10: the old mapping selected PTX 8.7 for sm_121, which requires at least PTX 8.8. It could also lower a newer PTX version. Your current log does not establish whether this is the only cause of the failure.

We published a [fix](fccfd23) that preserves newer PTX versions and target features, and reports the actual linked compiler, options and complete compilation error. Please update and rebuild in a new directory using a toolkit that supports your target. The ptxas version on PATH alone does not identify the compiler linked into an existing build.

Experimental SASS recovery is now disabled by default in the current launcher because recovered instructions may not preserve the original kernel’s behavior. Supported native PTX instrumentation remains available; the research path can be enabled explicitly.

Our isolated checks passed on 5090, but we have not verified the complete GB10/aarch64 model path. If rebuilding does not resolve the failure, please use the [Bug report form](https://github.com/SlugLab/HBFSim/issues/new?template=bug_report.yml) to share the complete first compiler diagnostic, commit, build/run commands and selected CUDA toolkit paths.

[slffjf]

a new bug report #8 , PTX 8.8 problem I user a" ln -s" map to PTX 8.8 version , please see my new bug report, and I fix a “ARM compatibility issues”