Datasets:
id stringlengths 7 7 | messages listlengths 1 1 | max_tokens int64 512 512 |
|---|---|---|
chat-01 | [
{
"role": "user",
"content": "Explain how a heat pump can move more heat energy than the electricity it consumes, for someone with no physics background."
}
] | 512 |
chat-02 | [
{
"role": "user",
"content": "I have a 45-minute train commute every day. Suggest a realistic plan to learn conversational Spanish in six months using that time."
}
] | 512 |
chat-03 | [
{
"role": "user",
"content": "Write a polite but firm email to a landlord asking for a leaking bathroom tap to be repaired, which I first reported three weeks ago."
}
] | 512 |
chat-04 | [
{
"role": "user",
"content": "What are the main differences between a Roth IRA and a traditional IRA, and how would someone decide between them?"
}
] | 512 |
chat-05 | [
{
"role": "user",
"content": "Summarise the causes of the 2008 financial crisis in about 300 words."
}
] | 512 |
chat-06 | [
{
"role": "user",
"content": "My sourdough comes out dense and gummy in the middle. What are the likely causes and how do I fix each one?"
}
] | 512 |
chat-07 | [
{
"role": "user",
"content": "Compare living in Amsterdam and Berlin for a software engineer: cost of living, housing, commuting and social life."
}
] | 512 |
chat-08 | [
{
"role": "user",
"content": "Explain what a transformer language model is doing when it generates text, in plain language, without equations."
}
] | 512 |
chat-09 | [
{
"role": "user",
"content": "Draft a short speech (about two minutes) for a friend's 40th birthday that is warm and a little funny."
}
] | 512 |
chat-10 | [
{
"role": "user",
"content": "What should I check before buying a used electric car? Give me a practical checklist."
}
] | 512 |
chat-11 | [
{
"role": "user",
"content": "Explain the difference between weather and climate, and why a cold winter does not disprove global warming."
}
] | 512 |
chat-12 | [
{
"role": "user",
"content": "I manage a team of five and two of them keep clashing in meetings. How should I handle it?"
}
] | 512 |
chat-13 | [
{
"role": "user",
"content": "Write a 400-word short story about a lighthouse keeper who receives a letter addressed to someone who died a century ago."
}
] | 512 |
chat-14 | [
{
"role": "user",
"content": "What are good strategies for getting back to sleep when I wake up at 4am and my mind starts racing?"
}
] | 512 |
chat-15 | [
{
"role": "user",
"content": "Explain how vaccines train the immune system, and why booster doses are sometimes needed."
}
] | 512 |
chat-16 | [
{
"role": "user",
"content": "Plan a three-day itinerary for Lisbon for someone who likes food, history and walking, on a moderate budget."
}
] | 512 |
Strix Halo llama.cpp benchmark workloads
Fixed prompt sets for measuring real-world llama-server performance of strix-llama.cpp on AMD Strix Halo (Ryzen AI Max+, gfx1151): single-stream decode and prompt throughput, time to first token, and speculative-decoding (MTP) acceptance.
Results measured with these workloads are only comparable when the workload
name and its sha256 in manifest.json match. Workloads are never edited in
place: a change gets a new version (chat-v2, ...).
Workloads
| name | requests | content | max_tokens |
|---|---|---|---|
chat-v1 |
16 | explanations, advice, emails, short fiction, planning | 512 |
code-v1 |
16 | write, refactor and review code (Python, C/C++, Rust, Go, TS, SQL, CUDA, Bash) | 768 |
long-v1 |
6 | summarise or review real llama.cpp source/docs, about 3k, 8k and 16k prompt tokens | 384 |
Each line is one request: {"id", "messages": [{"role": "user", "content"}], "max_tokens"}.
long-v1 embeds text from strix-llama.cpp commit 52d7e100b (MIT, © The ggml
authors); build_workloads.py --repo <strix-llama.cpp checkout> regenerates all
three files byte-for-byte.
Running
Start llama-server with -np 1, then replay a workload one request at a time
(one warm-up request is sent first and not recorded):
llama-server -m model.gguf -ngl 99 -fa on -np 1 --spec-type draft-mtp 2> server.log &
python3 replay.py --workload code-v1 --temperature 0 --server-log server.log --out results/code-v1.json
replay.py checks the file against manifest.json, then prints, from the
server's own timings:
| field | meaning |
|---|---|
decode_tps / decode_tps_stddev |
mean / sd of predicted_per_second over requests |
pp_tps |
mean prompt_per_second |
ttft_ms |
mean server-side prompt_ms (excludes HTTP overhead) |
prompt_tokens_mean |
mean prompt length in tokens |
accept_rate |
sum draft_n_accepted / sum draft_n (speculative runs only) |
mean_accepted_len |
accepted draft tokens per target forward pass |
mtp_source, draft_model |
with --server-log: embedded (MTP layers in the model GGUF) or sidecar (separate MTP-head GGUF) |
Compare --spec-type none and --spec-type draft-mtp on the same build, model,
server flags and temperature, alternating the two in mirrored blocks.
- Downloads last month
- 49