Bug#1095201: librocblas0: missing ldc argument in bench logging
Package: librocblas0
Version: 5.5.1+dfsg-7
Severity: minor
Dear Maintainer,
When profiling the rocblas calls from llama.cpp, I was using the
following comamnd:
ROCBLAS_LAYER=2 ROCBLAS_LOG_BENCH_PATH=$HOME/bench.log \
./llama-cli -ngl 99 --color -c 2048 --temp 0.7 \
--repeat_penalty 1.1 -n -1 -m dolphin-2.2.1-mistral-7b.Q5_K_M.gguf \
-no-cnv --prompt "Once upon a time"
and got output like this in bench.log:
./rocblas-bench -f gemm_batched_ex --transposeA T --transposeB N -m 32 -n 2 -k 128 --alpha 1 --a_type f16_r --lda 1024 --b_type f16_r --ldb 4096 --beta 0 --c_type f16_r --ldc --d_type f16_r --ldd 32 --batch_count 32 --compute_type f16_r --algo 0 --solution_index 0 --flags 1
However, these arguments are incomplete, as shown when passing them to librocblas0-bench:
$ cd /usr/libexec/rocm/librocblas0-bench
$ ./rocblas-bench -f gemm_batched_ex --transposeA T --transposeB N -m 32 -n 2 -k 128 --alpha 1 --a_type f16_r --lda 1024 --b_type f16_r --ldb 4096 --beta 0 --c_type f16_r --ldc --d_type f16_r --ldd 32 --batch_count 32 --compute_type f16_r --algo 0 --solution_index 0 --flags 1
Invalid value for --ldc
This can be worked around by dropping the --ldc flag, but this is
definitely a bug. The command line emitted by the bench logging should
be suitable for pasing directly to rocblas-bench.
Sincerely,
Cory Bloor
-- System Information:
Debian Release: trixie/sid
APT prefers testing
APT policy: (500, 'testing')
Architecture: amd64 (x86_64)
Kernel: Linux 6.12.9-amd64 (SMP w/32 CPU threads; PREEMPT)
Locale: LANG=en_CA.UTF-8, LC_CTYPE=en_CA.UTF-8 (charmap=UTF-8), LANGUAGE=en_CA:en
Shell: /bin/sh linked to /usr/bin/dash
Init: systemd (via /run/systemd/system)
LSM: AppArmor: enabled
Versions of packages librocblas0 depends on:
ii libamdhip64-5 5.7.1-5
ii libc6 2.40-5
ii libgcc-s1 14.2.0-12
ii libstdc++6 14.2.0-12
librocblas0 recommends no packages.
librocblas0 suggests no packages.
-- no debconf information
Reply to: