llama : assume tied weights if lm_head/output weights is missing (ggml-org#5824)

dmahurin · hazelnutcloud · commit e30f26b988f0 · 2024-03-10T15:38:20.000+08:00
This is to support model configurations with "tie_word_embeddings" set to true.

Co-authored-by: Don Mahurin &lt;2797413+dmahurin@users.noreply.github.com&gt;
diff --git a/llama.cpp b/llama.cpp
@@ -3888,7 +3888,13 @@ static bool llm_load_tensors(
                     {
                         model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
                         if (model.arch != LLM_ARCH_MINICPM){
-                            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);
+                            }
                         }
                     }
 

Original file line number	Diff line number	Diff line change
`@@ -3888,7 +3888,13 @@ static bool llm_load_tensors(`
`3888`	`3888`	`{`
`3889`	`3889`	`model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});`
`3890`	`3890`	`if (model.arch != LLM_ARCH_MINICPM){`
`3891`		`- model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab});`
	`3891`	`+ model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, false);`
	`3892`	`+ // if output is NULL, init from the input tok embed`
	`3893`	`+ if (model.output == NULL) {`
	`3894`	`+ model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});`
	`3895`	`+ ml.n_created--; // artificial tensor`
	`3896`	`+ ml.size_data += ggml_nbytes(model.output);`
	`3897`	`+ }`
`3892`	`3898`	`}`
`3893`	`3899`	`}`
`3894`	`3900`