pali112
/
ollama
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

tests: move csv output to benstat format

This commit is contained in:
Patrick Devine 2025-10-26 18:24:35 -07:00
parent b97eb2b858
commit 61349a8ec6
1 changed files with 17 additions and 21 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

38
integration/model_perf_test.go
View File

@ -161,11 +161,12 @@ func doModelPerfTest(t *testing.T, chatModels []string) {
} }
testCases := []struct { testCases := []struct {
name string
prompt string prompt string
anyResp []string anyResp []string
}{ }{
{blueSkyPrompt, blueSkyExpected}, {"blue_sky", blueSkyPrompt, blueSkyExpected},
{maxPrompt, []string{"shakespeare", "oppression", "sorrows", "gutenberg", "child", "license", "sonnet", "melancholy", "love", "sorrow", "beauty"}}, {"max", maxPrompt, []string{"shakespeare", "oppression", "sorrows", "gutenberg", "child", "license", "sonnet", "melancholy", "love", "sorrow", "beauty"}},
} }
var gpuPercent int var gpuPercent int
for _, tc := range testCases { for _, tc := range testCases {
@ -259,25 +260,20 @@ func doModelPerfTest(t *testing.T, chatModels []string) {
} }
} }
} }
// Round the logged prompt count for comparisons across versions/configurations which can vary slightly prefillTimePerToken := float64(resp.PromptEvalDuration.Nanoseconds()) / float64(resp.PromptEvalCount)
fmt.Fprintf(os.Stderr, "MODEL_PERF_HEADER:%s,%s,%s,%s,%s,%s,%s\n", prefillTokensPerSec := float64(resp.PromptEvalCount) / (float64(resp.PromptEvalDuration.Nanoseconds()) + 1e-12) * 1e9
"MODEL", fmt.Fprintf(os.Stderr, "BenchmarkModel/name=%s-%s/%d/step=%s %d %.2f ns/token %.2f token/sec\n",
"CONTEXT", model, tc.name, numCtx, "prefill", resp.PromptEvalCount, prefillTimePerToken, prefillTokensPerSec)
"GPU PERCENT",
"APPROX PROMPT COUNT", evalTimePerToken := float64(resp.EvalDuration.Nanoseconds()) / float64(resp.EvalCount)
"LOAD TIME", evalTokensPerSec := float64(resp.EvalCount) / (float64(resp.EvalDuration.Nanoseconds()) + 1e-12) * 1e9
"PROMPT EVAL TPS", fmt.Fprintf(os.Stderr, "BenchmarkModel/name=%s-%s/%d/step=%s %d %.2f ns/token %.2f token/sec\n",
"EVAL TPS", model, tc.name, numCtx, "generate", resp.EvalCount, evalTimePerToken, evalTokensPerSec)
)
fmt.Fprintf(os.Stderr, "MODEL_PERF_DATA:%s,%d,%d,%d,%0.2f,%0.2f,%0.2f\n", fmt.Fprintf(os.Stderr, "BenchmarkMode/name=%s-%s/%d 1 %d ns/request\n",
model, model, tc.name, numCtx, resp.TotalDuration.Nanoseconds())
numCtx, fmt.Fprintf(os.Stderr, "BenchmarkMode/name=%s-%s/%d/step=%s 1 %d ns/request\n",
gpuPercent, model, tc.name, numCtx, "load", resp.LoadDuration.Nanoseconds())
(resp.PromptEvalCount/10)*10,
float64(resp.LoadDuration)/1000000000.0,
float64(resp.PromptEvalCount)/(float64(resp.PromptEvalDuration)/1000000000.0),
float64(resp.EvalCount)/(float64(resp.EvalDuration)/1000000000.0),
)
} }
} }
}) })