common : ignore all EOG tokens

ggerganov · slaren · web-flow · commit f8754ba8fa81 · 2024-12-12T22:50:34.000+02:00
Co-authored-by: Diego Devesa &lt;slarengh@gmail.com&gt;
diff --git a/common/common.cpp b/common/common.cpp
@@ -941,8 +941,12 @@ struct common_init_result common_init_from_params(common_params & params) {
     }
 
     if (params.sampling.ignore_eos) {
-        LOG_INF("%s: added EOS logit bias = %f\n", __func__, -INFINITY);
-        params.sampling.logit_bias.push_back({llama_token_eos(model), -INFINITY});
+        for (llama_token i = 0; i < llama_n_vocab(model); i++) {
+            if (llama_token_is_eog(model, i)) {
+                LOG_INF("%s: added %s logit bias = %f\n", __func__, common_token_to_piece(lctx, i).c_str(), -INFINITY);
+                params.sampling.logit_bias.push_back({i, -INFINITY});
+            }
+        }
     }
 
     if (params.sampling.penalty_last_n == -1) {

Original file line number	Diff line number	Diff line change
`@@ -941,8 +941,12 @@ struct common_init_result common_init_from_params(common_params & params) {`
`941`	`941`	`}`
`942`	`942`
`943`	`943`	`if (params.sampling.ignore_eos) {`
`944`		`- LOG_INF("%s: added EOS logit bias = %f\n", __func__, -INFINITY);`
`945`		`- params.sampling.logit_bias.push_back({llama_token_eos(model), -INFINITY});`
	`944`	`+ for (llama_token i = 0; i < llama_n_vocab(model); i++) {`
	`945`	`+ if (llama_token_is_eog(model, i)) {`
	`946`	`+ LOG_INF("%s: added %s logit bias = %f\n", __func__, common_token_to_piece(lctx, i).c_str(), -INFINITY);`
	`947`	`+ params.sampling.logit_bias.push_back({i, -INFINITY});`
	`948`	`+ }`
	`949`	`+ }`
`946`	`950`	`}`
`947`	`951`
`948`	`952`	`if (params.sampling.penalty_last_n == -1) {`