Offload Bitnet token embeddings to the GPU - the right way (#2)

OK, I should have checked how it was done for Gemma and do
the same for Bitnet. But better late than never.

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>

1 files changed, 4 insertions, 18 deletions

diff --git a/llama.cpp b/llama.cpp
index dba3b1ce..169f7d68 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -5355,22 +5355,7 @@ static bool llm_load_tensors(
     bool use_mmap_buffer = true;
 
     // there is very little benefit to offloading the input layer, so always keep it on the CPU
-    //model.buft_input = llama_default_buffer_type_cpu(true);
-    //
-    // Well, this is not really true when the model uses the same tensor for token embeddings and for output
-    // (e.g., Bitnet, Gemma). If we use the above, then the matrix multiplication with the output tensor runs
-    // on the CPU, which can have quite a significant impact on performance. For instance, for 3B-Bitnet, I get
-    // TG-128 = ~240 t/s on an RTX-4080 with the above, and TG-128 = 320 t/s with the version below.
-    // The issue with just generically putting token embeddings on the GPU is that CUDA supports the GET_ROWS
-    // operation only for F16 and legacy quants, and this leads to a massive drop in performance when token embeddings
-    // are quantized with a k- or i-quant (which is almost always true). The back-end related stuff and offloading
-    // to the GPU has become quite opaque and hard to understand, so for now we fix this just for Bitnet
-    // (where token_embeddings is quantized with Q8_0).
-    if (model.arch == LLM_ARCH_BITNET) {
-        model.buft_input = llama_default_buffer_type_offload(model, main_gpu);
-    } else {
-        model.buft_input = llama_default_buffer_type_cpu(true);
-    }
+    model.buft_input = llama_default_buffer_type_cpu(true);
 
     model.buft_layer.resize(n_layer);
 
@@ -6729,7 +6714,8 @@ static bool llm_load_tensors(
 
                     // output
                     {
-                        model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output      = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); // same as tok_embd, duplicated to allow offloading
                     }
 
                     const uint32_t n_ff = hparams.n_ff;
@@ -12055,7 +12041,7 @@ struct llm_build_context {
         cb(cur, "result_norm", -1);
 
         // lm_head
-        cur = ggml_mul_mat(ctx0, model.tok_embd, cur);
+        cur = ggml_mul_mat(ctx0, model.output, cur);
         cb(cur, "result_output", -1);
 
         ggml_build_forward_expand(gf, cur);