Support benchmarking script by using real application trace #737
Conversation
nwangfw
force-pushed
the
ning/gpu-optimizer-benchmark-with-real-prompts
branch
4 times, most recently
from
42214fb to
4761e31
Compare
Signed-off-by: Ning Wang <[email protected]>
Signed-off-by: Ning Wang <[email protected]>
Signed-off-by: Ning Wang <[email protected]>
Signed-off-by: Ning Wang <[email protected]>
Signed-off-by: Ning Wang <[email protected]>
Signed-off-by: Ning Wang <[email protected]>
Signed-off-by: Ning Wang <[email protected]>
nwangfw
force-pushed
the
ning/gpu-optimizer-benchmark-with-real-prompts
branch
from
4761e31 to
b02dabf
Compare
nwangfw
changed the title
|
@zhangjyr Can you help to review this PR? Thanks! |
| # Convert rate_start to integer (multiply by 100 and remove decimals) | ||
| req_rate=$(echo "$rate_start * 100" | bc | cut -d. -f1) | ||
| rate_limit_scaled=$(echo "$rate_limit * 100" | bc | cut -d. -f1) | ||
| while [[ $req_rate -le $rate_limit_scaled ]]; do |
There was a problem hiding this comment.
Is it because -le can compare integers only, so we need to *100? And I suppose 0.125 is also somehow supported (inaccurate is ok) by removing decimals? Just in case, I would suggest setting req_rate at least 0.01 to avoid 0.
There was a problem hiding this comment.
Yes, -le doesn't work for decimal numbers. I also modified the number to *1000 to support 0.125 case.
There was a problem hiding this comment.
Let's add a comment here to tell the maintainer why 1000 is necessary.
| # synthetic_prompt = "hi " * config_input_len | ||
| # # assign timestamp to -1 for all requests | ||
| # requests.append((synthetic_prompt, config_input_len, config_output_len, -1)) | ||
| return [] |
There was a problem hiding this comment.
Why comment on this part? If no workload is provided, the benchmark will not work without a synthetic prompt.
There was a problem hiding this comment.
Thanks for your suggestions. I have updated the benchmarking logics after our discussion today. Now the benchmarking support using real prompts from dataset and synthetic prompt.
nwangfw
force-pushed
the
ning/gpu-optimizer-benchmark-with-real-prompts
branch
2 times, most recently
from
6ee6873 to
cb0255c
Compare
|
@zhangjyr Your comments have been addressed. Please review again when you have a chance. Thanks! |
nwangfw
force-pushed
the
ning/gpu-optimizer-benchmark-with-real-prompts
branch
from
cb0255c to
5c821ff
Compare
…workload file Signed-off-by: Ning Wang <[email protected]>
nwangfw
force-pushed
the
ning/gpu-optimizer-benchmark-with-real-prompts
branch
from
5c821ff to
3db2882
Compare
| if [[ -z "$OUTPUT_FILE" ]]; then | ||
| OUTPUT_FILE="${PATH_PREFIX}/result/${MODEL}.jsonl" | ||
| fi | ||
| OUTPUT_FILE="${PATH_PREFIX}/result/${FILE_NAME}.jsonl" |
There was a problem hiding this comment.
In fact, I don't think hardcoding all files being in ${PATH_PREFIX}/result/ is a good idea. In particular, python runtime is INSTALLED in a special folder if running as a docker, and it will become difficult to know where the output is. The same problem is true for PROMPT_DIR.
| # Convert rate_start to integer (multiply by 100 and remove decimals) | ||
| req_rate=$(echo "$rate_start * 100" | bc | cut -d. -f1) | ||
| rate_limit_scaled=$(echo "$rate_limit * 100" | bc | cut -d. -f1) | ||
| while [[ $req_rate -le $rate_limit_scaled ]]; do |
There was a problem hiding this comment.
Let's add a comment here to tell the maintainer why 1000 is necessary.
Pull Request Description
This PR adding benchmarking support by using real user application prompt traces in the heterogenous gpu story. This is a general benchmarking methods working for all models. It addresses the issue where the current fix benchmarking process may not be working for certain models or certain configuration settings (e.g., temperature)
Related Issues
Resolves: #722
Important: Before submitting, please complete the description above and review the checklist below.
Contribution Guidelines (Expand for Details)
We appreciate your contribution to aibrix! To ensure a smooth review process and maintain high code quality, please adhere to the following guidelines:
Pull Request Title Format
Your PR title should start with one of these prefixes to indicate the nature of the change:
[Bug]: Corrections to existing functionality[CI]: Changes to build process or CI pipeline[Docs]: Updates or additions to documentation[API]: Modifications to aibrix's API or interface[CLI]: Changes or additions to the Command Line Interface[Misc]: For changes not covered above (use sparingly)Note: For changes spanning multiple categories, use multiple prefixes in order of importance.
Submission Checklist
By submitting this PR, you confirm that you've read these guidelines and your changes align with the project's contribution standards.