Qwen2 : assume tied weights if lm_head/output weights is missing (#6738)

jklj077 · web-flow · commit e11b2e6e1e18 · 2024-04-18T14:38:04.000+03:00
diff --git a/llama.cpp b/llama.cpp
@@ -5184,7 +5184,13 @@ static bool llm_load_tensors(
                     // output
                     {
                         model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
-                        model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab});
+                        model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, false);
+                        // if output is NULL, init from the input tok embed
+                        if (model.output == NULL) {
+                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+                            ml.n_created--; // artificial tensor
+                            ml.size_data += ggml_nbytes(model.output);
+                        }
                     }
 
                     for (int i = 0; i < n_layer; ++i) {

Original file line number	Diff line number	Diff line change
`@@ -5184,7 +5184,13 @@ static bool llm_load_tensors(`
`5184`	`5184`	`// output`
`5185`	`5185`	`{`
`5186`	`5186`	`model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});`
`5187`		`- model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab});`
	`5187`	`+ model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, false);`
	`5188`	`+ // if output is NULL, init from the input tok embed`
	`5189`	`+ if (model.output == NULL) {`
	`5190`	`+ model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});`
	`5191`	`+ ml.n_created--; // artificial tensor`
	`5192`	`+ ml.size_data += ggml_nbytes(model.output);`
	`5193`	`+ }`
`5188`	`5194`	`}`
`5189`	`5195`
`5190`	`5196`	`for (int i = 0; i < n_layer; ++i) {`