From a29505606ecdd0db353d7d2c1c0a79f24e0ac9d7 Mon Sep 17 00:00:00 2001 From: Jonas Knobloch Date: Mon, 24 Nov 2025 23:41:41 +0100 Subject: Use seen tokens to warm up cache --- llm/perplexity.go | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) (limited to 'llm/perplexity.go') diff --git a/llm/perplexity.go b/llm/perplexity.go index dec3945..fbec496 100644 --- a/llm/perplexity.go +++ b/llm/perplexity.go @@ -153,7 +153,7 @@ func (e *Evaluator) schedule(tokens []int64, contextSize, stride, batchSize int) j.positions = append(j.positions, n) j.tokens = append(j.tokens, tokens[i:min(i+contextSize, len(tokens))]) // TODO verify - j.seen = append(j.seen, seen) + j.seen = append(j.seen, seen-i) seen = i + contextSize n++ @@ -185,7 +185,7 @@ func (e *Evaluator) execute(j *job, device int) { panic(err) // TODO handle } - nll := negLogLikelihood(logits[:len(logits)-1], toInt(j.tokens[0][1:])) + nll := negLogLikelihood(logits[j.seen[0]:len(logits)-1], toInt(j.tokens[0][1+j.seen[0]:])) j.results = append(j.results, nll) -- cgit v1.3.1