profiling radeon 7900xt (20gb vram)

[cryptopsy0]

/s/bee-rocm/build/bin/llama-server --version

ggml_cuda_init: found 1 ROCm devices (Total VRAM: 20464 MiB):
Device 0: AMD Radeon RX 7900 XT, gfx1100 (0x1100), VMM: no, Wave Size: 32, VRAM: 20464 MiB
version: 9459 (07ac3ce)
built with GNU 16.1.0 for Linux x86_64

1. Baseline – no DFlash

remove draft model / --spec-type dflash / other --spec-* DFlash args

export GGML_DFLASH_PROFILE=1
/s/bee-rocm/build/bin/llama-server
-m /root/.models/dflash/Qwen3.6-27B-Q4_K_M.gguf \
-md /root/.models/dflash/dflash-draft-3.6-q4_k_m.gguf
--host 0.0.0.0 --port 8080 \
--jinja --metrics -ngl all -ngld all
--reasoning on -fa on --mlock --no-mmap -np 1
-ctk turbo4 -ctv turbo4
-lv 3 --log-timestamps --log-prefix --log-colors off
-c 128000
##result
tokens: 481
wall tokens/s: 37.03
decode tokens/s: 41.01

2. DFlash default

same command, with DFlash enabled, but no forced --spec-draft-n-max 3

added
--spec-type dflash
--spec-dflash-cross-ctx 512
--spec-draft-ngl 999 \
##result
tokens: 479
wall tokens/s: 38.07
decode tokens/s: 42.30

3. DFlash with n-max 3

same as #2, but add --spec-draft-n-max 3

##result
tokens: 481
wall tokens/s: 37.22
decode tokens/s: 41.14

Test: Write a complete Python 3 module implementing a doubly-linked list with the following methods: append, prepend, insert_at, remove_at, find, reverse, to_list, length, is_empty, iter. Include comprehensive docstrings, type hints, and pytest unit tests for every method. Return only the code, no commentary.

[Anbeeld]

Please provide the logs generated by these runs.

[cryptopsy0]

reran with -ctk q4_0 -ctv q4_0 .. speeds are consistent but overall lower than original turbo4

test3.log
test2.log
test1.log

error log with turbo4/turbo4: turbo.log
rocminfo | head -100 &> /rocminfo.log rocminfo.log

hipcc --version
HIP version: 7.2.53210-9999
clang version 22.1.6
Target: x86_64-pc-linux-gnu
Thread model: posix
InstalledDir: /usr/lib/llvm/22/bin
Configuration file: /etc/clang/22/x86_64-pc-linux-gnu-clang++.cfg

/s/bee-rocm/build/bin/llama-server --version
ggml_cuda_init: found 1 ROCm devices (Total VRAM: 20464 MiB):
Device 0: AMD Radeon RX 7900 XT, gfx1100 (0x1100), VMM: no, Wave Size: 32, VRAM: 20464 MiB
version: 9459 (07ac3ce)
built with GNU 16.1.0 for Linux x86_64

[cryptopsy0]

#disable kv

export GGML_DFLASH_DISABLE_KV_CACHE=1 && export GGML_DFLASH_KV_CACHE_MODE=off

disable-kv.log

before:
wall tokens/s: 30.41
decode tokens/s: 31.71

after:
wall tokens/s: 23.41
decode tokens/s: 24.16

#sanity check