CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-31 08:51:55 +00:00
Code Issues Packages Projects Releases Wiki Activity

mpt : implement backwards compatiblity with duped output tensor (#6139)

Browse Source
This commit is contained in:
Jared Van Bortel
2024-03-18 12:49:02 -04:00
committed by GitHub
parent 104f5e0fc1
commit d199ca79f2
1 changed files with 10 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
llama.cpp
View File

@@ -540,6 +540,7 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
{ {
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" }, { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
{ LLM_TENSOR_OUTPUT_NORM, "output_norm" }, { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
{ LLM_TENSOR_OUTPUT, "output"},
{ LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
{ LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
{ LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" },
@@ -4300,9 +4301,9 @@ static bool llm_load_tensors(
{ {
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_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}); model.output_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd});
if (gguf_find_tensor(ml.ctx_gguf, tn(LLM_TENSOR_OUTPUT, "weight").c_str()) >= 0) {
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);
} else { if (!model.output) {
model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); // needs to be on GPU model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); // needs to be on GPU
ml.n_created--; // artificial tensor ml.n_created--; // artificial tensor
ml.size_data += ggml_nbytes(model.output); ml.size_data += ggml_nbytes(model.output);
@@ -4507,11 +4508,13 @@ static bool llm_load_tensors(
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_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, false); model.output_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, false);
// same as tok_embd, duplicated to allow offloading model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, false);
model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); if (!model.output) {
model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); // needs to be on GPU
ml.n_created--; // artificial tensor ml.n_created--; // artificial tensor
ml.size_data += ggml_nbytes(model.output); ml.size_data += ggml_nbytes(model.output);
} }
}
for (int i = 0; i < n_layer; ++i) { for (int i = 0; i < n_layer; ++i) {
ggml_context * ctx_layer = ctx_for_layer(i); ggml_context * ctx_layer = ctx_for_layer(i);