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

ggerganov · arthw · commit 050ba94e3821 · 2024-07-27T21:37:36.000+08:00
diff --git a/src/llama.cpp b/src/llama.cpp
@@ -15134,6 +15134,10 @@ 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
`@@ -15134,6 +15134,10 @@ static void llama_kv_cache_update_internal(struct llama_context & lctx) {`
`15134`	`15134`
`15135`	`15135`	`// apply K-shift if needed`
`15136`	`15136`	`if (lctx.model.hparams.rope_type != LLAMA_ROPE_TYPE_NONE && lctx.kv_self.has_shift) {`
	`15137`	`+ if (lctx.model.arch == LLM_ARCH_DEEPSEEK2) { // not supported due to MLA`
	`15138`	`+ GGML_ASSERT(false && "Deepseek2 does not support K-shift");`
	`15139`	`+ }`
	`15140`	`+`
`15137`	`15141`	`{`
`15138`	`15142`	`ggml_backend_sched_reset(lctx.sched);`
`15139`	`15143`