Can't run on AMD GPU, while llama.cpp does

[aasseman]

I'm on Linux, AMD HIP SDK v6.0.2, Radeon RX 7900 XT.

./TinyLlama-1.1B-Chat-v1.0.F16.llamafile --gpu AMD
import_cuda_impl: initializing gpu module...
get_rocm_bin_path: note: amdclang++ not found on $PATH
get_rocm_bin_path: note: $HIP_PATH/bin/amdclang++ does not exist
get_rocm_bin_path: note: /opt/rocm/bin/amdclang++ does not exist
link_cuda_dso: note: dynamically linking /home/alexis/.llamafile/ggml-rocm.so
TinyLlama-1.1B-Chat-v1.0.F16.llamafile: /usr/src/debug/hip-runtime-amd/clr-rocm-6.0.2/rocclr/os/os_posix.cpp:321: static void amd::Os::currentStackInfo(unsigned char**, size_t*): Assertion `Os::currentStackPtr() >= *base - *size && Os::currentStackPtr() < *base && "just checking"' failed.

error: Uncaught SIGABRT (SI_TKILL) at 0x3e8000051e5 on alexis-desktop pid 20965 tid 20965
  ./TinyLlama-1.1B-Chat-v1.0.F16.llamafile
  File exists
  Linux Cosmopolitan 3.3.6 MODE=x86_64; #1 SMP PREEMPT_DYNAMIC Tue, 07 May 2024 21:35:54 +0000 alexis-desktop 6.8.9-arch1-2

RAX 0000000000000000 RBX 00000000000051e5 RDI 00000000000051e5
RCX 00007ed5987e6e44 RDX 0000000000000006 RSI 00000000000051e5
RBP 00007ffc7d4eea20 RSP 00007ffc7d4ee9e0 RIP 00007ed5987e6e44
 R8 000000007f575010  R9 0000000000000007 R10 0000000000000008
R11 0000000000000246 R12 00007ed59874f740 R13 00007ed543d72478
R14 0000000000000006 R15 00007ed543d726a0
TLS 000000000084fe40

XMM0  2a202d20657361622a203d3e20292872 XMM8  2f20200a0a636376202c307830202c33
XMM1  63656863207473756a22202626206573 XMM9  00000000000000000000000000000000
XMM2  64656c696166202722676e696b636568 XMM10 00000000000000000000000000000000
XMM3  00000000000001010000000000000000 XMM11 00000000000000000000000000000000
XMM4  20202020287261502a203d3e20292865 XMM12 00000000000000000000000000000000
XMM5  ffffffffffff0000ffffffffffffffff XMM13 00000000000000000000000000000000
XMM6  00000000000000000000000000000000 XMM14 00000000000000000000000000000000
XMM7  00000000000000000000000000000000 XMM15 00000000000000000000000000000000

cosmoaddr2line /home/alexis/Downloads/TinyLlama-1.1B-Chat-v1.0.F16.llamafile 7ed5987e6e44 7ed59878ea30 7ed5987764c3 7ed5987763df 7ed598786c67 7ed543cea345

7ffc7d4eb920 7ed5987e6e44 NULL+0
7ffc7d4eea20 7ed59878ea30 NULL+0
7ffc7d4eea40 7ed5987764c3 NULL+0
7ffc7d4eeb00 7ed5987763df NULL+0
7ffc7d4eeb60 7ed598786c67 NULL+0
7ffc7d4eeb90 7ed543cea345 NULL+0
<dangerous frame>

10008004-10008009 rw-pa-      6x automap 384kB
1000800a-1000800a rw-Sa-      1x automap 64kB w/ 320kB hole
10008010-1000801e rw-pa-     15x automap 960kB w/ 2053mB hole
10010074-100184f0 r--s-- 33'917x automap 2120mB w/ 96tB hole
6fd00004-6fd0000f rw-paF     12x zipos 768kB w/ 64gB hole
6fe00004-6fe00004 rw-paF      1x g_fds 64kB
# 2122mB total mapped memory
./TinyLlama-1.1B-Chat-v1.0.F16.llamafile -m TinyLlama-1.1B-Chat-v1.0.F16.gguf --gpu AMD 
Aborted (core dumped)

Whereas llama.cpp built with

make LLAMA_HIPBLAS=1

Works perfectly well and lets me run every model I tried on the GPU.