vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Merge branch 'ggerganov:master' into k-shift2

Browse source
This commit is contained in:
MaggotHATE 2024-10-28 14:39:10 +05:00 • committed by GitHub
commit 968b4bac5b
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
3 changed files with 11 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

1
examples/server/server.cpp
View file

@ -1882,6 +1882,7 @@ struct server_context {
if (slot.state == SLOT_STATE_STARTED) {
slot.t_start_process_prompt = ggml_time_us();
slot.t_start_generation = 0;
slot.n_past = 0;
slot.n_prompt_tokens = prompt_tokens.size();
slot.state = SLOT_STATE_PROCESSING_PROMPT;

6
examples/server/utils.hpp
View file

@ -266,8 +266,10 @@ static llama_tokens format_infill(
}
// for now pick FIM context to fit in a batch (ratio prefix:suffix = 3:1, TODO: configurable?)
const int n_suffix_take = std::min<int>(tokens_suffix.size(), (n_batch/4));
const int n_prefix_take = std::min<int>(tokens_prefix.size(), 3*(n_batch/4) - 3);
const int n_prefix_take = std::min<int>(tokens_prefix.size(), 3*(n_batch/4));
const int n_suffix_take = std::min<int>(tokens_suffix.size(), std::max<int>(0, (n_batch/4) - (2 + tokens_prompt.size())));
SRV_DBG("n_prefix_take = %d, n_suffix_take = %d, total = %d\n", n_prefix_take, n_suffix_take, (n_prefix_take + n_suffix_take));
// fill the rest of the context with extra chunks
const int n_extra_take = std::min<int>(std::max<int>(0, n_ctx - (n_batch) - 2*n_predict), extra_tokens.size());

7
ggml/src/ggml-cuda.cu
View file

@ -1484,14 +1484,19 @@ static void ggml_cuda_op_mul_mat(
const size_t nbytes_data = ggml_nbytes(src0);
const size_t nbytes_padding = ggml_row_size(src0->type, MATRIX_ROW_PADDING - ne00 % MATRIX_ROW_PADDING);
dev[id].src0_dd = dev[id].src0_dd_alloc.alloc(ctx.pool(id), nbytes_data + nbytes_padding);
// TODO: remove this for MUSA once the Guilty Lockup issue is resolved
#ifndef GGML_USE_MUSA
CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd, 0, nbytes_data + nbytes_padding, stream));
#else // GGML_USE_MUSA
CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd + nbytes_data, 0, nbytes_padding, stream));
#endif // !GGML_USE_MUSA
}
// If src0 is on a temporary compute buffer (partial offloading) there may be some padding that needs to be cleared:
if (ne00 % MATRIX_ROW_PADDING != 0 && ggml_is_quantized(src0->type) && ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE && src0->view_src == nullptr) {
const size_t nbytes_data = ggml_row_size(src0->type, (dev[id].row_high - dev[id].row_low)*ne00);
const size_t nbytes_padding = ggml_row_size(src0->type, MATRIX_ROW_PADDING - ne00 % MATRIX_ROW_PADDING);
CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd + nbytes_data , 0, nbytes_padding, stream));
CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd + nbytes_data, 0, nbytes_padding, stream));
}
if (src1_on_device && src1_is_contiguous) {