correct mistake in conditionality for attn.k

Nexesenex · web-flow · commit 1118c046df26 · 2024-08-08T18:56:20.000+02:00
diff --git a/src/llama.cpp b/src/llama.cpp
@@ -15414,10 +15414,10 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
             // TODO: explore better strategies
             new_type = GGML_TYPE_Q8_0;
         }
-        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS && (qs.model.hparams.n_gqa() < 2 || qs.model.hparams.n_expert < 2)) {
+        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS && qs.model.hparams.n_gqa() < 2 && qs.model.hparams.n_expert < 2) {
             new_type = GGML_TYPE_IQ3_XXS;
         }
-        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS && (qs.model.hparams.n_gqa() < 2 || qs.model.hparams.n_expert < 2)) {
+        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS && qs.model.hparams.n_gqa() < 2 && qs.model.hparams.n_expert < 2) {
             new_type = GGML_TYPE_IQ2_S;
         }
     } else if (name.find("attn_q.weight") != std::string::npos) {

Original file line number	Diff line number	Diff line change
`@@ -15414,10 +15414,10 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n`
`15414`	`15414`	`// TODO: explore better strategies`
`15415`	`15415`	`new_type = GGML_TYPE_Q8_0;`
`15416`	`15416`	`}`
`15417`		`- else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS && (qs.model.hparams.n_gqa() < 2 \|\| qs.model.hparams.n_expert < 2)) {`
	`15417`	`+ else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS && qs.model.hparams.n_gqa() < 2 && qs.model.hparams.n_expert < 2) {`
`15418`	`15418`	`new_type = GGML_TYPE_IQ3_XXS;`
`15419`	`15419`	`}`
`15420`		`- else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS && (qs.model.hparams.n_gqa() < 2 \|\| qs.model.hparams.n_expert < 2)) {`
	`15420`	`+ else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS && qs.model.hparams.n_gqa() < 2 && qs.model.hparams.n_expert < 2) {`
`15421`	`15421`	`new_type = GGML_TYPE_IQ2_S;`
`15422`	`15422`	`}`
`15423`	`15423`	`} else if (name.find("attn_q.weight") != std::string::npos) {`