llama : disable context-shift for DeepSeek v2 (#8501)

ggerganov · web-flow · commit d65a8361fed8 · 2024-07-17T10:32:59.000+03:00
diff --git a/src/llama.cpp b/src/llama.cpp
@@ -15133,6 +15133,10 @@ static void llama_kv_cache_update_internal(struct llama_context & lctx) {
 
     // apply K-shift if needed
     if (lctx.model.hparams.rope_type != LLAMA_ROPE_TYPE_NONE && lctx.kv_self.has_shift) {
+        if (lctx.model.arch == LLM_ARCH_DEEPSEEK2) { // not supported due to MLA
+            GGML_ASSERT(false && "Deepseek2 does not support K-shift");
+        }
+
         {
             ggml_backend_sched_reset(lctx.sched);
 

Original file line number	Diff line number	Diff line change
`@@ -15133,6 +15133,10 @@ static void llama_kv_cache_update_internal(struct llama_context & lctx) {`
`15133`	`15133`
`15134`	`15134`	`// apply K-shift if needed`
`15135`	`15135`	`if (lctx.model.hparams.rope_type != LLAMA_ROPE_TYPE_NONE && lctx.kv_self.has_shift) {`
	`15136`	`+ if (lctx.model.arch == LLM_ARCH_DEEPSEEK2) { // not supported due to MLA`
	`15137`	`+ GGML_ASSERT(false && "Deepseek2 does not support K-shift");`
	`15138`	`+ }`
	`15139`	`+`
`15136`	`15140`	`{`
`15137`	`15141`	`ggml_backend_sched_reset(lctx.sched);`
`15138`	`15142`